1. Ecology
  2. Epidemiology and Global Health
Download icon

Physiology and ecology combine to determine host and vector importance for Ross River virus

  1. Morgan P Kain  Is a corresponding author
  2. Eloise B Skinner  Is a corresponding author
  3. Andrew F van den Hurk
  4. Hamish McCallum
  5. Erin A Mordecai
  1. Department of Biology, Stanford University, United States
  2. Natural Capital Project, Woods Institute for the Environment, Stanford University, United States
  3. Centre for Planetary Health and Food Security, Griffith University, Australia
  4. Public Health Virology, Forensic and Scientific Services, Department of Health, Australia
Research Article
  • Cited 0
  • Views 561
  • Annotations
Cite this article as: eLife 2021;10:e67018 doi: 10.7554/eLife.67018

Abstract

Identifying the key vector and host species that drive the transmission of zoonotic pathogens is notoriously difficult but critical for disease control. We present a nested approach for quantifying the importance of host and vectors that integrates species’ physiological competence with their ecological traits. We apply this framework to a medically important arbovirus, Ross River virus (RRV), in Brisbane, Australia. We find that vertebrate hosts with high physiological competence are not the most important for community transmission; interactions between hosts and vectors largely underpin the importance of host species. For vectors, physiological competence is highly important. Our results identify primary and secondary vectors of RRV and suggest two potential transmission cycles in Brisbane: an enzootic cycle involving birds and an urban cycle involving humans. The framework accounts for uncertainty from each fitted statistical model in estimates of species’ contributions to transmission and has has direct application to other zoonotic pathogens.

Introduction

More than 60% of existing infectious diseases of humans are multi-host pathogens (i.e. moving between non-human and human populations) and approximately 75% of emerging and re-emerging infectious diseases affecting humans have a non-human origin (Taylor et al., 2001; van Doorn, 2014). It it therefore critical to identify the role that different vertebrate host and vector species play in maintaining transmission and facilitating spillover into humans. However, identifying which species enable pathogen persistence and quantifying the relative contribution that each species makes to transmission is notoriously difficult, particularly because definitions for vectors and hosts vary greatly within the literature (Appendix 1—table 1). The dynamics of multi-host pathogen systems can range in complexity from spillover between a single source population to a single target population (e.g. from bats to humans as has been postulated for SARS-CoV-1 and SARS-CoV-2: Boni et al., 2020) to large interconnected networks of species that maintain a pathogen in a given environment and facilitate spillover into a target population (e.g. zoonotic arboviruses, such as West Nile (WNV) and Rift Valley fever (RVFV) viruses: Viana et al., 2014).

Developing appropriate mitigation strategies for zoonotic pathogens hinges on quantifying which processes have the largest influence over each species’ importance in transmission cycles. Studies characterising zoonotic arbovirus transmission often focus on pairwise transmission between non-human hosts and vectors, or vectors and humans (for example work in WNV: Marm Kilpatrick et al., 2006, Ross River virus: Koolhof and Carver, 2017, Stephenson et al., 2018, leishmaniasis: Stephens et al., 2016, Chagas disease: Gürtler and Cardinal, 2015, Jansen et al., 2018). However, these and other proposed approaches (Appendix 1—table 1) that capture only a portion of a pathogen’s transmission cycle cannot completely quantify a species’ contribution to transmission within a community. Understanding the ecological importance of host and vector species for transmission requires modeling the complete transmission cycle (host-vector-host or vector-host-vector transmission), ‘closing the loop’ by estimating the number of new infections in the next generation. This is needed to quantify each species’ contribution to 0, defined as the number of new infections arising from a single case in an otherwise susceptible population. While this is well understood (e.g. see Turner et al., 2013, Fenton et al., 2015, Webster et al., 2017), this approach is used less frequently for multi-vector, multi-host pathogens because of the need for data across multiple phases of transmission for multiple host and vector species.

Here, we present a general framework (Box 1) that: (1) quantifies host and vector species’ relative importance across a complete transmission cycle of zoonotic arboviruses (Figure 1), using Ross River virus (RRV) as the model virus—a system for which we have data for many host and vector species for nearly all components of the transmission process; (2) identifies which of the many interacting physiological and ecological processes have the largest control over the importance of each species; and (3) helps to reveal where the largest sources of uncertainty occur in order to identify which datasets require additional collection for more robust predictions (Restif et al., 2012). The approach uses three nested metrics of increasing biological complexity: physiological competence; transmission over one half of the pathogen’s life cycle (half-cycle transmission; that is, host-to-vector or vector-to-host transmission); and transmission over the pathogen’s complete life cycle (complete-cycle transmission) (Box 1). This strategy has application to other zoonotic pathogens for which some physiological and ecological data exist across vectors and hosts. Even for systems with limited data, a framework that integrates the entire transmission cycle can be useful for hypothesis testing and for guiding data collection by identifying the processes that most contribute to uncertainty in competence (i.e. model-guided fieldwork, sensu Restif et al., 2012).

The transmission cycle of Ross River virus, a multi-host, multi-vector arbovirus, and the components our framework uses to model this transmission cycle.

The first requirements for transmission are physiologically competent hosts that become infected (A: ‘proportion viremic’) and are able to replicate the virus to suitable levels to infect vectors (A: ‘titer profiles’) and vector species that can become infected (B: ‘Mosquito infection probability’) and eventually are able to transmit virus (B: ‘Mosquito transmission probability’). Physiologically competent hosts and vectors contribute to the transmission of the virus through a continuous cycle of transmission, which can be viewed from two perspectives, either starting with an infected host or starting with an infected vector; regardless of perspective, a single complete cycle contains a single set of physiological and ecological components. Each of these components are used in our framework in one of three ways: statistical models fit to empirical data, from which uncertainty is propagated into the final calculations of transmission (boxes outlined in black); raw empirical data (boxes outlined in blue); and point estimates (boxes outlined in red). Italic bold numbers and text next to the boxes outlined in black describe, in brief, the type of statistical model used to estimate each component (GLMM stands for generalized linear mixed model). Details on all components are provided in the Materials and Methods, Supplementary files, and Appendix Figures that are listed next to framework components; associated raw Source Data files are also listed.

Box 1.

Nested approach for characterising the complete transmission cycle of zoonotic arboviruses.

Stage 1: Physiological competence.

Characterizing the physiological response a species has to infection is fundamental to estimating its potential as a host or vector within a community. We define the physiological competence of a host species as its viremic response to infection multiplied by the proportion of individuals of that species that develop a viremic response when exposed to infection. We model each host species’ viremic response as a continuous function over time (Appendix 1—figure 1); to compare hosts’ physiological competences, we summarize their titer profiles using the area under the curve (AUC), which simultaneously captures the magnitude and duration of titer (Appendix 1—figure 2). For vectors, we quantify physiological competence using the product of the proportion of individuals that get infected following exposure to a given dose (Appendix 1—figure 3) and the proportion that go on to transmit the virus (Appendix 1—figure 4). Specifically, we quantify physiological vector competence using the multiplication of the AUC of these two curves (Appendix 1—figure 5, Appendix 1—figure 6). For a visualization of these components within an arbovirus life cycle see Figure 1.

Stage 2: Transmission over one half of the pathogen’s life cycle (host-to-vector or vector-to-host transmission).

To begin to understand the role species play in community transmission, we quantify how many vectors an infected host will generate or how many new host infections an infected vector will create. To do this, we combine host and vector physiological competence (Stage 1) with host and vector abundances and contact rates. Specifically, to quantify host-to-vector transmission we combine estimates (while propagating uncertainty) from host titer profiles over time, mosquito infection probabilities given titer (infectious dose), mosquito feeding behavior (which combines vector preference and host abundance), and mosquito abundance (Figure 1). For vector-to-host transmission, we combine estimates from mosquito transmission probabilities, survival, mosquito feeding behavior, and host abundance.

Stage 3: Transmission over the pathogen’s complete life cycle (host-vector-host or vector-host-vector transmission).

A complete transmission cycle can be achieved by multiplying the two half-transmission calculations from Stage 2 (host-to-vector and vector-to-host) in either order; the 0 calculated from either order will be identical. However, each of the two multiplication orders reveals something different. Multiplying host-to-vector by vector-to-host transmission gives host-vector-host transmission (a complete transmission cycle from the perspective of a host), which can be used to reveal all host-to-host pairwise transmission pathways. In other words, beginning with an infected host, how many (and which) other hosts become infected? Conversely, multiplying vector-to-host transmission by host-to-vector will reveal all vector-to-vector transmission pathways starting with an infected vector.

As a case study, we focus on RRV, an alphavirus that causes a disease syndrome characterized by polyarthritis, which is responsible for the greatest number of mosquito-borne human disease notifications in Australia, with approximately 5000 cases notified annually (Australian Govt. Dept. of Health, 2020). It has also caused major epidemics in Pacific Islands involving tens of thousands of cases (Aaskov et al., 1981; Tesh et al., 1981; Harley et al., 2001), and may have the potential to emerge and cause explosive epidemics out of its current geographical range (Flies et al., 2018; Shanks, 2019). Understanding the drivers of epidemic and endemic transmission of RRV in Australia and Pacific Island countries has remained challenging because of the number of hosts and mosquitoes that potentially become infected and the large uncertainty around which of these vectors and hosts contribute most to transmission. Under controlled laboratory conditions, more than 15 species of mosquitoes from at least five genera have demonstrated the physiological ability to transmit RRV. The disease has long been considered to exist in a zoonotic transmission cycle, primarily because the number of human cases during winter months was considered to be too low to sustain community transmission (Harley et al., 2001). However, the most important vertebrate hosts of RRV are highly ambiguous because more than 50 species have demonstrated serological evidence of natural exposure to RRV (reviewed in Stephenson et al., 2018). Much uncertainty remains as to which vertebrate species contribute to RRV community transmission and how the importance of these species in transmission varies by locations (such as urban vs. rural settings, or in Australia vs. the Pacific Islands, where there are different vertebrate communities). Although insights have previously been gained through modeling approaches (Carver et al., 2009; Denholm et al., 2017; Koolhof and Carver, 2017), these studies note that future progress in RRV modeling requires consideration of the dynamics of multiple mosquito species and multiple hosts, accounting for their differing availability and physiological capability to transmit RRV.

We parameterize our framework for RRV to quantify the relative importance of hosts and vectors for disease transmission and to illustrate how the relative importance of these species changes depending on what metric is used. Specifically, we ask the following questions for RRV transmission in Brisbane, Australia, a community in which RRV is endemic:

  1. Which host and vector species are most physiologically competent for transmitting RRV?

  2. How does integrating species ecology change the most important hosts and vectors when considering a half (host-to-vector or vector-to-host) or complete (host-vector-host or vector-host-vector) transmission cycle?

  3. How do viruses circulate through different species in the community, that is, which hosts and vectors contribute to intra- and inter-species transmission?

Results

Physiological competence

Host competence

To quantify a host species’ physiological competence we multiplied the proportion of individuals of that species that developed a viremic response by the area under that species’ estimated titer profile over time, which we fit to the individuals that mounted a viremic response. This AUC metric captures both the absolute magnitude and duration of a host species’ viremic response, weighted by how common this response is. Of the vertebrate species available for the analysis in Brisbane, we estimated that rats and macropods had the strongest viremic response to RRV infection (Figure 2A). Sheep, rabbits, humans, and possums formed a distinct cluster of hosts with the next strongest responses; uncertainty in host titer profiles obscures our ability to differentiate among the responses of these species. Of the remaining species, we estimated that ‘birds’ (an average of Gallus gallus domesticus [Chicken], Cacatua sanguinea [Little corella], and Anas superciliosa [Pacific black duck]) had a stronger viremic response than flying foxes, horses, and cattle. No dogs or cats developed detectable viremia when exposed to RRV experimentally (N = 10 for each species), resulting in the lowest physiological competence. Fitted titer profiles for all hosts for which data were available are presented in Appendix 1—figure 1 (AUC for these profiles are presented in Appendix 1—figure 2), while the proportion of the cohort of each host species that developed a viremic response when exposed to RRV is listed in Supplementary file 1.

The most competent host species for Ross River virus (RRV) transmission in Brisbane change when considering physiological traits alone (A) or also considering ecological traits (B, C).

(A) Estimated physiological response of hosts to experimental infection with RRV, summarized using the area under their estimated titer profiles over time (AUC). In all panels, points show median estimates; error bars are 95% confidence intervals (CIs) that combine the uncertainty from all statistical sub-models used to obtain the estimates presented in that panel (see Figure 1 and Box 1 for these components). Titer profile AUC is used only to quantify host physiological competence, while time-dependent titer profiles (pictured in Appendix 1—figure 1) are used in half-cycle and complete-cycle transmission. The ordering of hosts based on highest (top) to lowest (bottom) physiological competence in A is conserved in B and C to aid visualization of host order changes among panels. (B) Host-to-vector transmission; matrices show the median estimated number of vectors infected by each host species, while the points show infection totals (sums across matrix rows), with error bars. (C) Host-vector-host transmission. As in B, the matrices show estimated median numbers of next-generation host infections for all host species pairs, while the points show sums across rows of the matrices (left plot) and the proportion of infections in the second generation that are in the same species as the original infected individual (center plot).

Vector competence

To quantify mosquito physiological competence, we used the area under the infection probability versus dose curve multiplied by the area under the transmission probability over time since infection curve. We estimated that the mosquito species with the highest physiological potential for RRV transmission (susceptibility of mosquitoes to infection, and of those that become infected, their potential to transmit RRV) was Coquillettidia linealis, although the 95% CI for this species overlaps with four species with the next highest median estimates (Aedes procax, Verrallina funerea, Aedes vigilax, and Mansonia uniformis) (Figure 3A). In contrast, Culex annulirostris, Culex quinquefasciatus, Aedes notoscriptus, and Culex sitiens were estimated to all have low physiological potential. Infection probability curves for all mosquito species for which we gathered data, including those in the Brisbane community and from elsewhere in Australia, are shown in Appendix 1—figure 3 and Appendix 1—figure 5.

Ross River virus (RRV) transmission capability of Brisbane mosquitoes remained consistent when considering physiological traits alone (A) or also considering ecological traits (B, C).

(A) Physiological response of mosquitoes to experimental infection with RRV, summarized using the area under (AUC) of their estimated infection probability versus dose curves multiplied by the area under their transmission probability versus time curves. Points show median estimates; the error bars in each panel are 95% confidence intervals (CIs) that combine the uncertainty from all statistical sub-models used to obtain the estimates presented in that panel (see Figure 1 and Box 1 for these components). AUC is used only to quantify mosquito physiological competence; raw infection and transmission profiles (pictured in Appendix 1—figure 3 and Appendix 1—figure 4, respectively) are used in calculations of half-cycle and complete-cycle transmission. The ordering of vector species based on highest (top) to lowest (bottom) physiological competence in A is conserved in B and C to aid visualization of vector order changes among panels. (B) Vector-to-host transmission; matrices show the median numbers of hosts infected by each vector species, while the points show infection totals (sums across matrix rows), with error bars. (C) Vector-host-vector transmission. As in B, the matrices show median numbers of next-generation vector infections for all vector species pairs, while the points show sums across rows of the matrices (left plot) and the proportion of infections in the second generation that are in the same species as the original infected individual (center plot).

Half-transmission cycle

Host-to-vector transmission

Integrating host physiological competence with ecological factors governing host-vector contacts (see Figure 1 and Box 1) can dramatically change estimated host importance (Figure 2B). Despite large uncertainty in estimates for the number of mosquitoes that a single infected host can infect while infectious, humans have both the largest estimated median and highest estimated potential (upper 95% CI bound) for infecting mosquitoes in Brisbane. We predict that an infected human would predominantly infect Ae. vigilax, followed by Ae. procax and Cx. annulirostris. Both rats and macropods, which had the highest physiological potential for transmission (Figure 2A), dropped beneath possums, birds, and horses according to median estimates, though overlapping 95% CIs obscure our ability to determine which host is able to infect more mosquitoes while infectious. Similarly, sheep dropped from being in the cluster of the most important species when using physiological response alone (Figure 2A) to one of the lowest potential hosts for RRV transmission to mosquitoes in Brisbane (Figure 2B). Conversely, horses, which had one of the lowest estimated viremic responses, increased in importance when considering the contribution of ecological traits to community transmission. Cats and dogs were estimated to be unable to transmit RRV to any mosquitoes given that neither mount a viremic response.

Vector-to-host transmission

While host relative importance markedly changed between physiological competence and transmission over half a transmission cycle, mosquito estimates did not. Cq. linealis, Ae. procax, Ae. vigilax, and Ve. funerea were estimated to infect the largest number of hosts (using median estimates) after embedding mosquito physiological competence into vector-to-host transmission (Figure 3B), although wide overlapping 95% CI make it impossible to differentiate among these species. We estimated that an infected Cq. linealis would mostly infect birds, while an infected Ae. procax and Ae. vigilax would infect a larger diversity of host species including birds, humans, and dogs. Of the remaining species, Cx. annulirostris, Cx. quinquefasciatus, and Cx. sitiens remained poor vectors, infecting only a small number of hosts.

Complete-transmission cycle

We calculated the number of second generation hosts an infected host would infect (or the number of second generation mosquitoes an infected mosquito would infect) in a Brisbane host community using a next generation matrix (NGM). Our estimates across a complete-transmission cycle combine all the components listed in Figure 1 and described in Box 1; uncertainty is propagated from fitted statistical sub-models (see Table 1).

Table 1
Model components, the transmission metrics in which they are used, and the data and statistical modeling choices used to estimate each.

The column 'Parameter' lists the parameters as they appear in Equation 1 and Equation 2. Abbreviations for the transmission metrics are: HC = host competence; H-to-V = host-to-vector transmission; V-to-H = vector-to-host; H-to-H = host-vector-host; V-to-V = vector-host-vector. The ‘Data’ column lists the name of the Source data file containing the raw data; all citations are listed in the online supplement (Supplementary file 3). Data sources are described in the Supplemental Methods: Data. The ‘Methodological Details’ column lists where in the manuscript methods are described.

Model ComponentParameterTransmission MetricsDataStatistical ModelUncertaintyMethodological Details
Proportion of individuals of host species i exposed to infection that produce viremiaωiHC H-to-V H-to-H V-to-Vhost_response.csv human_titer.csvRaw DataNone (Raw Data)Methods: Vertebrate hosts: titer profiles; Supplemental Methods: Host physiological competence; Supplementary file 1
Host titer (in species i on day j)θidiHC H-to-V H-to-H V-to-Vhost_response.csv human_titer.csvLinear model with a quadratic term for days post infection1000 simulated titer curves for each speciesMethods: Vertebrate hosts: titer profiles; Supplemental Methods: Host physiological competence; Appendix 1—figure 1; Supplementary file 1
Proportion of host species i that are seronegativeηjV-to-H H-to-H V-to-Vhost_ seroprevalence.csvRaw DataNone (Raw Data)Supplementary file 1
Infection probability of mosquito species j as a function of dosepjVC H-to-V V-to-H H-to-H V-to-Vmosquito_ infection.csvGeneralized linear model (logistic regression)1000 samples from a multivariate Normal distribution using the estimated means and vcov matrixMosquito vectors: infection and transmission probability; Supplemental Methods: Vector physiological competence; Appendix 1—figure 3; Supplementary file 2
Transmission probability of mosquito species j r days post infectionpirjVC V-to-H H-to-H V-to-Vmosquito_ transmission.csvGeneralized linear model (logistic regression)1000 samples from a multivariate Normal distribution using the estimated means and vcov matrixMosquito vectors: infection and transmission probability; Supplemental Methods: Vector physiological competence; Appendix 1—figure 4; Supplementary file 2
Survival probability of mosquito species j up to r days post infectionλjrjV-to-H H-to-H V-to-VExponential decay using point estimate for daily mortality probabilityNoneMethods: Mosquito survival; Appendix 1—figure 7
Proportion of mosquito species j’s blood meals that are obtained from host species iβijαii=1IβijαiV-to-H H-to-H V-to-Vmosquito_ feeding.csv host_ abundance.csvCustom Bayesian regression modelBayesian posteriorMethods: Mosquito feeding preference; Supplemental Methods: Mosquito feeding preference; Supplementary file 2; Supplementary file 3
Number of susceptible mosquitoes of species i per host species jϕijH-to-V H-to-H V-to-Vmosquito_ abundance.csvRaw Data + AssumptionNone (Raw Data + Point Estimate)
Daily biting rate of mosquito species jσjH-to-V V-to-H H-to-H V-to-VAssumptionNone (Point Estimate)Assumed value of 0.5 Day−1

Host-vector-host transmission

Estimated host importance changed little between host-to-vector and host-vector-host transmission: humans, birds, possums, horses, and macropods remained in the top cluster of hosts (Figure 2C). Despite wide 95% CI of humans that overlapped with birds, possums, horses, and macropods, much of the density distribution of host-vector-host transmission estimates (obtained by propagating uncertainty from all statistical sub-models) for humans falls above that of other species (Appendix 2—figure 1). For example, 32% of the distribution of total host-to-host infections for humans is at higher estimates than the upper bound of the 95% CI for birds, the next highest species by median estimate. We estimated that the mosquitoes that would acquire RRV from humans mostly go on to infect humans (‘self-infections’), followed by birds, dogs, and to a lesser extent possums. Even when weighting second generation infections by the proportion of hosts that mount a viremic response (i.e., ignoring all sink infections in dogs and thus counting second generation infectious hosts only), humans still produce the most second-generation infectious hosts by median estimate, though CI once again overlap with birds, macropods, horses, and possums (Appendix 2—figure 2). We predicted that an infected bird (the species with the second highest estimated median) would primarily infect other birds, followed by dogs and humans, respectively (Figure 2C).

Because humans are the only species without data from experimental infection studies (titer was measured when infected humans began showing symptoms), we checked the robustness of our results by re-running analyses assuming a host titer duration for humans reflecting only the observed human viremic period. Even when human titer duration was reduced, humans remained in the top cluster of hosts (with birds, possums, horses, and macropods) for RRV transmission potential despite an overall lower total number of second-generation infections (Appendix 2—figure 3, Appendix 2—figure 4). This highlights the robust result that humans likely contribute to the RRV transmission cycle in Brisbane due to their physiological competence, abundance, and attractiveness to competent mosquitoes like Ae. vigilax and Ae. procax.

Vector-host-vector transmission

Across a complete vector-host-vector transmission cycle, confidence intervals remained wide for the estimated number of mosquitoes an infected mosquito of each species would infect over its lifetime (Figure 3C left panel). Nonetheless, the results suggest that Cq. linealis, Ae. procax, Ve. funerea, Ae. vigilax, and Ma. uniformis have a much higher maximum transmission potential than Cx. annulirostris, Cx. quinquefasciatus, Cx. sitiens, and Ae. notoscriptus.

Importantly, the results pictured in Figure 3C calculate second generation mosquito infections conditional on starting with a mosquito exposed to 6.4 log10 infectious units of RRV per mL (the median dose used in experimental infection studies); if it is a rare event that a given mosquito species becomes exposed in the first place, basing mosquito importance on this metric could be misleading. For example, regardless of the species of the originally infected mosquito (rows of the Figure 3C matrix), we predict that most second generation infections will be in Ae. vigilax, followed by Ae. procax and Cq. linealis (columns of the Figure 3C matrix), because of their abundance and feeding preferences. Similarly, while an individual Ve. funerea or Ma. uniformis mosquito could potentially have the highest ability for producing second-generation infections in mosquitoes (Figure 3C), their rarity (0.27% and 0.14% of the Brisbane mosquito community, respectively; Supplementary file 2) means that few second generation infections from any source mosquito occur in Ve. funerea or Ma. uniformis. Thus, unlike Ae. vigilax, Ae. procax, and Cq. linealis, the rare mosquitoes Ve. funerea or Ma. uniformis are very unlikely to play an important role in RRV transmission over multiple generations in this ecological context.

Multiple generations of transmission

To estimate which host and mosquito species drive RRV spread as it invades a naive host population, we approximated transmission over five complete RRV life cycles using the next-generation matrix (NGM) approach to calculate transmission in discrete time steps where each time step represents a complete cycle of transmission. Simulating the spread of infection over multiple generations, starting with one initially infected human in an otherwise susceptible vertebrate population in Brisbane, shows that infections tend to propagate through humans, birds, dogs, and horses (median estimates: Figure 4; estimates with uncertainty: Appendix 2—figure 5). Overall, while infection does circulate largely in the broader vertebrate community (as opposed to continuously cycling between a small subset of vectors and hosts), we estimated that at the beginning of an epidemic in Brisbane, many infections would occur in humans and birds, a moderate number in horses, and many sink infections in dogs. These new infected individuals (apart from dogs and cats) continue to spread infection in the community, and already by the third generation of infection, the most dominant pathways of transmission have converged to birds infecting other birds, humans infecting other humans, humans infecting birds, horses infecting humans, and ‘wasted’ transmissions from both humans and birds to dogs, a dead-end host (Figure 4 Generation 3).

RRV epidemic dynamics propagate through initially naïve host and vector communities.

Epidemics are simulated in two ways: transmission in the host community resulting from an initial infection in a human (top row), or transmission in the mosquito community arising from a source infection in a Ma. uniformis mosquito (bottom row). Each matrix cell contains the median estimated number of new infections in a given species (columns) arising from all infected individuals of a given species in the previous generation (rows). The red arrow shows the direction of infection. We show generations 1–3 here to illustrate how quickly infections propagate through the community and converge on dominant transmission pathways, by generation 3. Uncertainty in the number of new infections in each host and mosquito species over five generations is shown in Appendix 2—figure 5 and Appendix 2—figure 6, respectively.

Starting with an initial infection in a Ma. uniformis mosquito (to illustrate the effect of beginning with an infection in a rare species), the multi-generation approximation shows that after only a single generation the framework predicts that the majority of infected mosquitoes will be Ae. vigilax and Ae. procax, and to a lesser extent Cq. linealis and Cx. annulirostris (median estimates: Figure 4; estimates with uncertainty: Appendix 2—figure 6), which mirrors the results in Figure 3C. Despite the potentially high competence of Ma. uniformis, their rarity in the Brisbane mosquito community causes them to participate little in sustained community transmission. After only three generations, we predicted that most transmission of RRV in Brisbane was occurring from Ae. vigilax, Ae. procax, and Cq. linealis; the dominance of these three species can be seen in Figure 4 by the large number of pairwise transmission events between them.

Discussion

Motivated by a practical need to identify the relative importance of hosts and vectors for zoonotic arboviral transmission, we developed a nested approach that incorporates existing data, uncertainty, and the complex, dynamic interactions that underpin the transmission of multi-host, multi-vector pathogens. We applied this approach to RRV transmission in Brisbane, which is thought to have multiple transmission cycles (Stephenson et al., 2018; Claflin and Webb, 2015), and contributes a significant public health burden (Jansen et al., 2019). Our approach highlights how species importance changes across physiological and ecological drivers of transmission across half, complete, and multiple generations of transmission cycles, thus isolating the factors that contribute most to vector or host importance.

Physiology meets ecology: changes in species importance

The first aim of this study was to characterise which hosts and vectors had high physiological competence for RRV. Species must be able to acquire and propagate the virus to be an important host or vector. Our results corroborate some of what has been previously reported (Stephenson et al., 2018; Harley et al., 2001), but also generated some surprising results. The strong physiological competence of macropods has long been acknowledged, while cats and dogs have never been considered to play a role as hosts; our research supported both of these ideas. By contrast, horses, which occasionally develop high viremia in response to RRV infection and have been previously considered a moderately competent host (described in Stephenson et al., 2018), have low physiological competence on average because less than 15% of exposed horses develop a viremic response when infected. Conversely, humans, which have not been considered important for local transmission, had a moderate-to-high physiological competence following infection with RRV (Figure 2A). For vectors, RRV has long been considered a generalist virus, capable of persisting across climates and habitats within Australia; our result that no single species was dominant in its physiological competence supports this view.

Physiological competence alone, without ecological data, provides an incomplete picture of transmission and can be misleading. For example, a host’s physiological competence is of little importance if that host is rare or adopts behaviors that prevents exposure (Downs et al., 2019). Further, mosquito feeding preferences can drive pathogen transmission more strongly than host competence (Simpson et al., 2012). There are many documented circumstances in which species that are highly competent for transmission under controlled conditions play a minor role in community transmission (Levin et al., 2002; Marm Kilpatrick et al., 2006), or conversely, where species with apparently low competence in laboratory studies are highly important for transmission in nature (Brady et al., 2014; Brook and Dobson, 2015). We found the former to be the case for RRV hosts across half and complete transmission cycles. For example, we estimated that humans contributed more mosquito infections (Figure 2B) and second generation host infections (Figure 2C) than the most physiologically competent species (rats, sheep, and macropods; although human 95% CI overlapped that of macropods). There are longstanding debates within disease ecology surrounding how ecological interactions moderate disease dynamics, for example, through dilution effects (Johnson and Thieltges, 2010) and zooprophylaxis (Donnelly et al., 2015). The nested approach is useful for identifying specific mechanisms because it analyzes transmission as a step-wise process with increasing ecological complexity by integrating different forms of trait data. Specifically, the results from a half transmission cycle represent the pairwise interactions between host and vector species. For example, a physiologically competent host with low community competence based on host-to-vector cycles (for RRV this includes rats, sheep, and rabbits) occurs due to low rates of contact between this host and vectors with a high infection probability. By contrast, a host with low competence across a complete transmission cycle, but high host-to-vector transmission competence, would reflect more on the transmission ability of the vectors that host infects. By separating transmission in this way, we can examine the contribution each trait makes to species importance and test hypotheses such as whether it is more important for a host to infect a greater number and diversity of vectors, or fewer, more competent vectors.

In our study, different ecological drivers likely underpin the importance of humans and birds, the two species with the highest median estimates for complete-cycle transmission (Figure 2, Appendix 2—figure 1). For example, when compared to all other hosts, humans had the highest susceptible population (contributing 66% of the total community abundance, with less than 14% seropositivity). This, in combination with their moderately-high physiological competence (Figure 2B) contributes to their overall importance. These factors are more important than other ecological drivers. For example, although humans infect a large number of moderately competent vectors (Ae. vigilax and Ae. procax; Supplementary file 3), the mosquito feeding patterns potentially limit human importance because many of the mosquitoes reported to feed on humans have lower competence for RRV (such as Cx. annulirostris and Ae. notoscriptus). That being said, the number of Ae. vigilax that humans infect (Figure 2B) suggests that a potentially fruitful path for reducing human infections is vector control of Ae. vigilax populations, which is already one of the primary targets of mosquito control operations in Brisbane (Brisbane City Council,, 2019). In contrast, birds were estimated to be only approximately 5% of the host community composition and almost a third were seropositive, further reducing the total number of susceptible individuals. Despite this relative scarcity, birds were highly important in the half and complete transmission cycles. This high importance is likely driven by the strong feeding association with the highly physiologically competent mosquito Cq. linealis rather than birds’ physiological competence or abundance.

Transmission pathways of RRV in Brisbane

Moving beyond single transmission cycles, when we approximate transmission through the Brisbane community over five generations (approximately the transmission season: Australian Govt. Dept. of Health, 2020), we estimate that infection spreads widely through the community, with the largest number in humans, birds, dogs, and horses. The physiologically competent, abundant, and generalist feeder Ae. vigilax plays an important role in this propagation. Despite large uncertainty, our findings for RRV transmission cycles in Brisbane point to two overlapping transmission cycles: an enzootic cycle, characterized primarily by transmission between birds and Cq. linealis, and a domestic cycle characterized by human-to-human infections facilitated by Ae. vigilax and Ae. procax. These two cycles are linked by these feeding generalists, which transfer infection between birds and humans. Within each of these overlapping cycles, dogs play a diluting role by absorbing infectious bites as they are not able to transmit RRV.

Multiple transmission cycles for RRV have long been hypothesized (Harley et al., 2001), yet no previous studies have implicated the species involved in these cycles or quantified their contribution to transmission. Humans and birds have been greatly understudied as potential hosts of RRV, yet unlike marsupials, they persist across the geographic distribution of RRV. Despite frequent detection of RRV in major metropolitan centers (Claflin and Webb, 2015), the potential for humans to contribute to endemic transmission (as opposed to epidemic transmission: Rosen et al., 1981; Aaskov et al., 1981) has empirically been understudied. Although our predictions provide some support for the importance of these understudied pathways, because we were unable to model seasonal changes in vector abundance or the correlated seasonal changes in human RRV cases in Brisbane (which generally peak in late summer through early autumn: Australian Govt. Dept. of Health, 2020), more modeling and empirical work is needed. Hopefully our identification of multiple transmission pathways will allow for future research to formulate hypotheses for RRV seasonality. For such work data would need to be collected across seasons to distinguish the role of seasonality and the timing/drivers of spillover that shift transmission from an enzootic to domestic cycle.

The vectors identified in Brisbane transmission cycles, Ae. vigilax, Ae. procax and Cq. linealis, are recognised as important vectors for RRV and are regularly targeted in vector control programs. However, we predicted that Cx. annulirostris and Ae. notoscriptus are less competent vectors, although they are often cited as key RRV vectors in Brisbane (Kay and Aaskov, 1989; Russell, 1995; Watson and Kay, 1998). The evidence in favor of Cx. annulirostris as a vector is that RRV is frequently detected in wild-caught individuals, and that abundance has been high during previous outbreaks of RRV (Jansen et al., 2019). RRV has also been isolated from Ae. notoscriptus during outbreaks in Brisbane (Ritchie et al., 1997); however, the species had relatively low abundance in this study, and low transmission ability (Appendix 1—figure 4) in comparison to other potential vectors. This suggests a new hypothesis that Cx. annulirostris and Ae. notoscriptus are secondary RRV vectors (capable of playing a supplemental role in transmission but unable to maintain an epidemic) to other species such as Ae. vigilax which are primary RRV vectors (capable of starting and maintaining epidemics). Although novel for RRV, the distinction between primary and secondary vectors has been made for other arboviruses (Turell et al., 2005). Finally, the isolation of RRV from wild caught mosquitoes demonstrates that a particular species is infected with the virus, it is incomplete evidence for mosquito species’ specific role in virus transmission. Even if found infected in the field, the lower transmission capability of Cx. annulirostris or Ae. notoscriptus relative to Ae. vigilax, Ae. procax and Cq. linealis means that the former are likely to transmit infection to fewer hosts than the latter.

Caveats and uncertainty

It is important to acknowledge a number of caveats with the data and modeling assumptions we used. For physiological competence, experimental studies vary substantially in their methods. We overcame some of this variation by transforming published data into the same viral units between studies (e.g., infectious units were converted to per milliliter: IU/mL). However, not all variation in experimental approaches could be included in our regression model because of data sparsity. Thus, it is possible that some of the variation we attribute to species may in fact be explained by methodology used in different studies. For the ecological data, the methods used to collect species abundance data can also result in bias, as different traps and survey types detect different species (Brown et al., 2008; Lühken et al., 2014). For example, the species trapped using CO2-baited light traps in this study may not be a true representation of the entire mosquito community in Brisbane. Similarly, vertebrate survey methods are biased against detecting species with cryptic behavior, and thus represent a biased sample of the host community available to host-seeking mosquitoes. While the uncertainty captured in the reported data were propagated through our estimates of competence, unmeasured uncertainty arising due to experimental methods could additionally affect the results. However, compared with approaches that focus solely on a single physiological or ecological data source to infer competence, the approach presented here allows for a more detailed investigation of vector and host competence and their drivers.

There are many potential hosts that are not included in this analysis due to data limitations. As a minimum requirement, host species were only included if they were included in mosquito blood meal field observations, were experimentally exposed to the virus, and were measured for background seroprevalence and abundance in Brisbane. In some instances, to meet these minimum data requirements, species were aggregated by taxonomic group. For example, we averaged the responses of chickens, little corellas, and Pacific black ducks to ’birds’ (while a strong simplifying assumption, the clustering of these species’ physiological response does provide support for this choice: Appendix 1—figure 2). In other instances (such as the potential for koalas to be hosts of RRV), species were unable to be modeled because of an absence of viremia data. Further, we ignore seasonal matching of transmission with host reproduction, ignore duration of host life stages, and either make a snapshot measure of host transmission capability (Figure 2, Figure 3) or make a simple five-generation approximation that averages across host and vector infectious periods (Figure 4). Finally, some hosts and vectors may only be locally important for RRV transmission, as opposed to being important over the entire geographic distribution of the virus. For example, though sheep have high physiological importance, they were not locally important in Brisbane. However, sheep could play a greater role in the maintenance and spillover of RRV in rural areas where they are more abundant and/or where other species of mosquitoes with higher biting affinity for sheep may occur.

For mosquitoes, data sets with the most substantial gaps included host feeding data, physiological transmission capability, and mosquito survival. Blood meal data is difficult to collect, but is very important because feeding patterns enter into the equation twice for vector-host-vector transmission. Limited blood meal counts (Supplementary file 3) led to high uncertainty in feeding patterns for many species (e.g. Ma. uniformis), which can have a large influence over the width of the 95% CI (Figure 3C). Addressing these data gaps is critical for refining vector predictions for RRV, though these data are logistically difficult and costly to obtain. More laboratory experiments on mosquito transmission probability over time, especially for those understudied species that we predict have the potential to be important transmitters would also help to better resolve transmission patterns in the Brisbane community. For example, the 95% CIs for Ma. uniformis and Ve. funerea are particularly wide, which could place them as either highly important vectors or inefficient vectors. Finally, because we assumed identical survival for all species, with no uncertainty (i.e., survival did not contribute to the widths of the confidence intervals across species), the uncertainty we present is an underestimate. Species-specific field-based mortality rates are a crucial data source that needs to be obtained for more accurate measures of mosquito transmission capability. It is important to note, however, that even in spite of large uncertainty for vector-host-vector transmission (Figure 3C), the rarity of many of these mosquito species make them mostly irrelevant when approximating transmission over multiple generations (Figure 4, Appendix 2—figure 6).

While all of these modeling choices and data shortcomings can influence model outcomes, a clear advantage of the framework is that uncertainty from each statistical sub-model fit to independent data sets is accounted for in the overall estimates. In doing so, parameters with high uncertainty, such as mosquito feeding preferences or transmission probabilities, can be targeted in future studies to help refine the framework’s predictions.

Applications for other vector borne diseases

This framework can be applied to other vector-borne pathogens in a number of ways. A principal application would be to identify important vectors and hosts for other multi-host, multi-vector pathogens, including RVFV (Turell et al., 2008; Davies and Karstad, 1981; Gora et al., 2000; Busquets et al., 2010); WNV (Kain and Bolker, 2019), or yellow fever virus (Rosen, 1958; Jupp and Kemp, 2002), for which competence data exist for several species. For these viruses, our framework and code can be used by substituting data and modifying the underlying statistical sub-models (e.g., titer profiles) to match the dynamics of the pathogen of interest; the subsequent calculations for host and vector competence, half-cycle transmission, and complete-cycle transmission are usable without modification. The generality of this framework and its nested approach can also support (with minimal modification) additional transmission pathways such as vertical transmission (where mosquitoes emerge from immature stages already infected with a given pathogen), or direct vertebrate-to-vertebrate transmission as can occur for some vector-borne diseases such as RVFV (Wichgers Schreur et al., 2016) or Zika virus (D’Ortenzio et al., 2016).

Secondary applications for this framework could include identifying the largest gaps and uncertainties within datasets. This is advantageous because in light of finite resources, model-guided research (Restif et al., 2012) can identify the most important data needed to improve predictions for disease emergence and transmission. Another application would be to apply the framework for a single pathogen across space and time, such as across the geographic range of RRV or between seasons. This is useful to compare shifts in transmission dynamics, identify hotspots or potential for spillover. Though our framework has not been developed to predict the timing and peak of epidemic events, it can be used to disentangle the underlying transmission dynamics of vector-borne pathogens in specific locations, which allows for the development of predictive modeling.

Finally, the generality and multi-phase nature of this framework provide a common language to compare and contrast the transmission dynamics not just within a single pathogen, but also between them. Until now, the highly diverse methods, definitions and data required to characterise vectors and hosts has hindered the ability to make comparisons between pathogens. The integration of multidisciplinary data in this framework is done in a way that could be used to compare host or vector physiological competence and ecological traits for other multi-host, multi-vector pathogens.

Conclusion

Identifying important vectors and hosts of zoonotic pathogens is critical for mitigating emerging infectious diseases and understanding transmission in a changing world. However, attempts to do so have been hampered by the multidisciplinary datasets required and differing definitions that can alter the importance of a species. Here we developed a nested approach that can be applied to any multi-host, multi-vector pathogen for which some competence data exists. Applying this approach to RRV transmission in Brisbane, we were able to: (a) identify two hosts of potentially high importance that deserve further investigation (humans and birds), (b) two potential transmission cycles (an enzootic cycle and a domestic cycle), and (c) datasets that should be targeted (bloodmeal studies, vector transmission experiments, field-based mosquito survival estimates) to reduce overall uncertainty and ultimately increase the future power of the framework. Future studies that aim to identify and quantify the importance of different species in virus transmission cycles must integrate both physiological competence data and ecological assessments to more fully understand the capacity of species to transmit pathogens. The nested approach here provides a tool to integrate these different datasets while acknowledging uncertainty within each, which could be applied to any multi-host, multi-vector pathogen for which some competence data exists.

Materials and methods

The methods are presented in three sections to reflect our three focal questions. First, we describe the calculation of host and vector physiological competence. Second, we describe half-cycle (host-to-vector and vector-to-host transmission) and complete-cycle (host-vector-host or vector-host-vector) transmission. Third, we describe how we use complete-cycle transmission to approximate transmission over multiple generations. We introduce data and calculations for components that are used in multiple transmission metrics (e.g., host virus titer profiles) with the first metric in which they are used.

Host and vector physiological competence

Vertebrate hosts: virus titer profiles

Request a detailed protocol

We fit host virus titer profiles as continuous functions over time to published data on host vertebrate responses to infection. For each of 15 experimentally infected non-human vertebrate species we extracted the proportion of exposed individuals that developed detectable viremia, their duration of detectable viremia in days, their peak viremia titer, and the unit of measure of this titer (such as median lethal dose (LD50), suckling mouse intracerebral injection (SMIC50)) (from Whitehead, 1969; Spradbrow, 1973; Rosen et al., 1981; Kay et al., 1986; Ryan et al., 1997; Boyd et al., 2001; Boyd and Kay, 2002). All reported viral concentrations were converted to infectious units per millilitre (IU/mL) values, rather than 0.1 mL or 0.002 mL as reported in some studies. Titer data are summarized in Supplementary file 1 and a summary of these studies’ methodological details can be found in Stephenson et al., 2018; all data extracted from these publications are available in Source data 1.

For non-human species, only means and standard deviations for peak titer and duration of detectable titer were reported. We transformed these summary measures into continuous titer profiles (continuous functions of titer over time that are needed to quantify mosquito infection probability) by modeling titer profiles as quadratic functions of time since infection, based on observed patterns in the data. For human titer profiles, for which experimental infection studies were not available, we used data from one observational study (Rosen et al., 1981) that measured titer in humans exhibiting disease symptoms during an outbreak in the Cook Islands in 1980. Details on how we constructed continuous titer curves, with uncertainty, for all hosts are available in Appendix 1; for raw human titer data see Source data 2. In Appendix 1—figure 1 we show 95% confidence intervals (CI) for each of the hosts’ quadratic profiles generated from this procedure with the summary values of peak and duration of titer extracted from the literature overlayed. To quantify host physiological competence we summarized the titer profiles into a single metric using the area under the curve (AUC) of the time-dependent titer curves. We use AUC because it simultaneously captures both titer magnitude and the duration of detectable titer (the host’s infectious duration). AUC is used only to summarize host competence; raw time-dependent titer values are used to calculate mosquito infection. The AUC for the fitted titer profiles (Appendix 1—figure 1) are shown in Appendix 1—figure 2.

Mosquito vectors: infection and transmission probability

Request a detailed protocol

We fit mosquito infection probabilities and mosquito transmission probabilities using published data from laboratory experimental exposure of mosquitoes to RRV. From experimental infections of mosquitoes we collected information on the infectious dose they were exposed to, the number of mosquitoes receiving an infectious dose, the proportion of mosquitoes that became infected, the proportion of mosquitoes that went on to become infectious (i.e., transmitted the virus), and the time it took for mosquitoes to become infectious (the extrinsic incubation period) (from Kay et al., 1979; Kay et al., 1982; Kay, 1982; Kay et al., 1982; Ballard and Marshall, 1986; Fanning et al., 1992; Vale et al., 1992; Wells et al., 1994; Doggett and Russell, 1997; Watson and Kay, 1998; Jennings and Kay, 1999; Ryan et al., 2000; Doggett et al., 2001; Jeffery et al., 2002; Kay and Jennings, 2002; Jeffery et al., 2006; Webb et al., 2008; Ramírez et al., 2018). Mosquito infection and transmission data are summarized in Supplementary file 2; raw data files are included as Source data 3 and Source data 4, respectively.

We modeled both mosquito infection probability (the proportion of all experimentally exposed mosquitoes with virus detected in their bodies) and transmission probability (the proportion of all experimentally exposed mosquitoes with virus detected in their saliva, measured via feeding on a susceptible vertebrate species or using an in vitro method of saliva collection) using generalized linear mixed effects models (GLMM) with Binomial error distributions, fit in R using the package lme4 (Bates et al., 2015). For each model, the proportion of mosquitoes infected or transmitting was taken as the response variable and the total number exposed to infection was used as weights; species were modeled using random effects. For additional details see the Supplemental Methods. Fitted infection probability curves for all mosquito species for which we gathered data—those found in Brisbane and elsewhere in Australia—are shown in Appendix 1—figure 3; transmission probability curves are shown in Appendix 1—figure 4. To quantify mosquito physiological competence we summarized mosquito infection and transmission probabilities into a single metric using the area under the curve (AUC) of the dose-dependent infection curve multiplied by the area under the curve (AUC) of the time-dependent transmission curve. AUC is used only to summarize mosquito competence; raw probability values are used to calculate the probability a mosquito becomes infected when feeding on an infected host (given the titer in that host) and the probability they are able to transmit to a susceptible host (given the number of days post infection that the feeding occurs). The AUC for the fitted infection probability (Appendix 1—figure 3) and transmission probability (Appendix 1—figure 4) curves are shown in Appendix 1—figure 5 and Appendix 1—figure 6, respectively.

Half-cycle and complete-cycle transmission

Both half-cycle (host-to-vector and vector-to-host) and complete-cycle (host-vector-host and vector-host-vector) transmission nest host and vector physiological competence in an ecological context (Figure 1). To quantify each of these metrics we used a next-generation matrix (NGM) model (Diekmann et al., 1990; Hartemink et al., 2009), which, for a vector-borne disease, requires the construction of two matrices of transmission terms. The first matrix (denoted HV, where bold terms refer to matrices) contains species-specific host-to-vector transmission terms, which we write with hosts as rows and vectors as columns. The second matrix (VH) contains vector-to-host transmission terms and has vectors as rows and hosts as columns. Cells of HV and VH contain the expected average number of infections between pairs of species over the whole infectious period of the infector (host in HV, vector in VH); each pairwise transmission term is a function of host and vector physiological competence as well as ecological factors. Row sums of HV give the total number of vectors (of all species) infected by each host (total host-to-vector transmission); similarly row sums of VH give the total number of hosts (of all species) infected by infectious vectors.

We calculate the total number of individuals of each mosquito species j that a host species i infects over its infectious period d (which gives entry [i, j] of HV) as:

(1) Ivij=di=19(pj|θidi)ωiϕijσjβijαii=1Iβijαi,

where pj|θidi is the probability that a susceptible species of mosquito (j) would become infected when biting host i on day di when it has titer θidi. We model infection over a period of 9 days for all host species given that the estimated titer of all host species is predicted to be undetectable by 9 days, equating to a very small mosquito infection probability (Appendix 1—figure 1). The proportion of individuals of species i that manifest an infection with detectable titer θidi is given by ωi, while ϕij is the number of susceptible mosquitoes of species i per host species j, σj is the daily biting rate of mosquito species j, and βijαii=1Iβijαi is the proportion of all mosquito species j’s bites on host species i, which is jointly determined by the relative abundance of host i (αi) and the intrinsic feeding preference of mosquito j on host i (βij) (details given in Mosquito feeding behavior below). Equation 1 assumes no species specific host-by-mosquito interactions for infection probability; mosquito infection probability is uniquely determined by the level and duration of titer within a host (i.e., a dose-response function of host titer). The only direct evidence against this assumption that we are aware of is an example where more Cx. annulirostris became infected when feeding on a bird than on a horse despite there being a lower viremia in the bird (Kay et al., 1986).

The total number of individuals of each host species i that a mosquito of species j infects over its infectious period rj (which gives entry [j, i] of VH) is given by:

(2) Ihji=rj=138pirjηjλjrjσjβijαii=1Iβijαi,

where pirj is the probability an infected mosquito of species j transfers infection to a given susceptible host by bite on day rj of their infectious period, λjrj is the probability of survival of mosquito species j until day rj, σj is the daily biting rate of mosquito species j, and βijαii=1Iβijαi is the proportion of all mosquito species j’s bites on host species i. We calculate mosquito-to-host transmission over 38 days given that we assume mosquitoes do not survive longer than 38 days (see Mosquito survival below).

The key differences between the host-to-vector (HV; Ivij) and vector-to-host (VH; Ihji) transmission matrix entries are two-fold. First, HV assumes that host infectivity is titer- and time-dependent and depends on mosquito density per host; conversely, VH assumes that mosquito infectiousness is titer-independent (dose-independent) but time-dependent and depends on daily mosquito survival and host species relative abundance. Second, for HV we assume a single infected host of a given species enters into a community of susceptible mosquitoes, while for VH we assume that a single mosquito of a given species becomes exposed to a dose of 6.4 log10 infectious units per mL (the median dose used across all mosquito infection studies) and then enters a host community with empirically estimated background host immunity (from Doherty et al., 1966; Marshall et al., 1980; Vale et al., 1991; Boyd and Kay, 2002; Faddy et al., 2015; Skinner et al., 2020; see Supplementary file 1 and Source data 7 for sample sizes and the proportion of each host testing seropositive for RRV). The primary similarity between these matrices is that mosquito biting rate, host abundance, and mosquito feeding preference (σj times the fraction of α and β terms) are used in both matrix calculations as the components that control the contact rate between infected hosts and susceptible mosquitoes (VH) or infected mosquitoes and susceptible hosts (VH).

Complete-cycle transmission is calculated using the matrix product of HV and VH, which is commonly referred to as the ‘who acquires infection from whom’ matrix (Schenzle, 1984; Anderson and May, 1985; Dobson, 2004). Specifically, using HV*VH gives GHH, in which each cell describes the total number of pairwise host-vector-host transmission events, assuming a single infected host appears at the start of its infection in an otherwise susceptible host population. Likewise, using VH*HV gives GVV, in which each cell describes the total number of pairwise mosquito-to-mosquito transmission events, assuming a single infected mosquito appears at the start of its infectious period in an otherwise susceptible mosquito population. Row sums of GHH give the total number of new host infections in the second generation that originate from single source infections in each host species (total host-vector-host transmission), or the total number of mosquito-to-mosquito transmission events in the case of GVV. Column sums of GHH or GVV give the total number of newly infected individuals of each host or mosquito species arising from one infection in each host or mosquito, respectively. These properties can be used to find, for example, dead-end hosts (i.e., ‘diluters’; Schmidt and Ostfeld, 2001), which would be captured by host species with a small row sum and large column sum in GHH. Further, Diekmann et al., 1990 show that the dominant eigenvalue of either GHH or GVV describes 0, the typical number of secondary cases, resulting from pathogen transmission in the heterogeneous community whose pairwise transmission dynamics are described in HV and VH.

We estimated each of the parameters of HV and VH using either statistical sub-models fit to empirical data or directly from empirical data taken from the literature. Uncertainty from all statistical sub-models was propagated into the calculations of HV and VH in one of three ways: (1) titer: by simulating 1000 titer curves given the uncertainty in peak titer and duration of titer in the published data sources (see Supplemental Methods); (2) mosquito infection probability and mosquito transmission probability: by constructing density distributions using the means and variance-covariance matrix of the estimated coefficients assuming univariate or multivariate normality (using 1000 samples; see Kain and Bolker, 2017, Kain and Bolker, 2019 for two examples using this method of uncertainty propagation in similar frameworks); (3) mosquito feeding behavior: using the estimated Bayesian posterior. We do not consider uncertainty for those framework components that rely on raw data (the proportion of hosts that mount a viremic response, host and mosquito relative abundance, and host seroprevalence) or point estimates (mosquito to host ratio, mosquito biting rate, and mosquito survival). Thus, the 95% CIs we present contain uncertainty from fitted statistical models but do not account for the full uncertainty. All of our framework’s parameters, the data used to parameterize all sub-models within the framework, and methods of uncertainty propagation are listed in Table 1. Details on vertebrate host and mosquito abundance, mosquito survival, and mosquito feeding behavior are described below.

Vertebrate host abundance

Request a detailed protocol

Vertebrate abundance data for Brisbane was calculated from a variety of sources including published literature and technical reports (see Supplementary file 1 and Source data 5). Data on livestock species (cattle, sheep, horses) and humans arose from technical reports undertaken by agricultural and government agencies (Australian Bureau of Statistics, 2018; Meat and Livestock Australia, 2019a; Meat and Livestock Australia, 2019b; Ward et al., 1996). Cat and dog abundance was derived from a general pets per human ratio from a technical report (Animal Medicines Australia, 2019), and scaled to the human population in Brisbane. Abundance for wildlife was derived either from citizen science reports (birds, possums and macropods: Australian EPA, 2019), or published fauna surveys undertaken in Brisbane (flying foxes: Queensland Government, 2020; rats, rabbits: Skinner et al., 2021). Host abundance was calculated as a measure of density within Brisbane (hosts per km2). We used the relative densities of each of these species as reported in these sources as the species’ proportions in our community for our analysis.

Mosquito abundance

Request a detailed protocol

Mosquito relative abundances were estimated for Brisbane by combining data from mosquito surveys (requested from the Brisbane City Council mosquito surveillance program). In brief, Brisbane City Council operates weekly CO2 plus 1-octen-3-ol baited Centers for Disease Control (CDC)-style light traps across 10 sites in Brisbane. Traps are set 1.5 m off the ground before dusk, and collected just after dawn the following morning. Any trapped mosquitoes are stored in −20°C until identification to species level by a single person. This data is not publicly available, but has been analyzed and described in Skinner et al., 2021. Mosquito abundance from these surveys was calculated as an average weekly total during peak mosquito season (October to May). Mosquito species abundance data was also supplemented with the results of analyses of the vertebrate host origin of mosquito blood meals presented in previous published studies (Ryan et al., 1997; Kay et al., 2007; Jansen et al., 2009). Mosquito abundance data is summarized in Supplementary file 2; raw data is available in Source data 8.

We used the observed proportion of each mosquito species detected in these surveys as the proportion of that species in our community for our analysis, which assumes that the observed species proportions are unbiased predictors of their true proportions. Because the number of mosquitoes per host (Equation 1: ϕ) is needed to calculate the absolute number of mosquitoes an infected host would infect, we multiplied the relative abundances of mosquitoes by 40 (our assumed value for overall raw number of mosquitoes per host in the community). While this may be an over- (or under-) estimate of the true value in Brisbane, because this value is only a scalar in the NGM framework it will only affect the magnitude of estimates and not the relative estimates among species.

Mosquito survival

Request a detailed protocol

Survival data (either field or laboratory derived) for the mosquito species present in Brisbane, Australia, is not available for most species. For this reason, we modeled mosquito survival as being identical for all species. Specifically, we used an exponential decay model for mosquito survival using a daily survival probability that is half of the daily maximum survival rate of Cx. annulirostris (calculated as 1/lifespan) measured in optimal laboratory conditions (from Shocket et al., 2018 who used data from McDonald et al., 1980, which may over-estimate survival rates in nature). However, we assume that mosquito survival probability falls to zero after day 38.

Mosquito feeding behavior

Request a detailed protocol

We modeled the observed blood meals in wild-caught mosquitoes (the number of blood fed mosquitoes and the source of the blood meals) as arising jointly from the abundance of each host in the community and each mosquitoes’ intrinsic feeding preference on each host species (the latent variable that we model here). Data was extracted from published blood meal surveys specific to Brisbane (from Ryan et al., 1997; Kay et al., 2007; Jansen et al., 2009); mosquito blood meal data is summarized in Supplementary file 2 and Supplementary file 3; raw data is available in Source data 6. Specifically, we modeled the number of blood meals a mosquito of species j obtains from host species i (δij) as:

(3) δijMulti(N,βijαii=1Iβijαi),

where δij is a multinomially distributed random variable (the extension of the binomial distribution for greater than two outcomes) with probability equal to the intrinsic preference of mosquito j for host species i (βij), weighted by the abundance of host species i (αi), relative to all host species in the community (sum over all host species in the denominator). Written in this way, βij is the ratio of the proportion of bites mosquito species j takes on host species i relative to biting host species j in proportion to their abundance in the community (which would occur if a mosquito were biting randomly). We fit this multinomial model in a Bayesian context in Stan (Carpenter et al., 2017), interfaced with R using the package rstan (Stan Development Team, 2020). For details on the fitting of this Bayesian model see Appendix 1; the full Stan model is also available in the GitHub repository hosting the code: Kain, 2021a.

Tailoring the model to the Brisbane community

Request a detailed protocol

One difficulty with the integration of diverse data types is variation in the biological scale at which these data are collected. For our model, vertebrate host types are recorded at different taxonomic levels across data sets (e.g. laboratory infection experiments are conducted at the species level while mosquito blood meal surveys report identification of the blood meal host source at a taxonomic level ranging from species through to higher level classification such as class or family). In order to integrate the predictions from our individual sub-models fit to single data types (e.g. infection experiments and blood meal surveys) to parameterize HV and VH, and thus draw inference on the importance of different hosts and mosquitoes in RRV transmission in Brisbane, Australia, we made three simplifying assumptions. First, we averaged each mosquito’s infection probability when biting ‘birds’ (the taxonomic level available for blood meal data) for the three species of birds with a measured viremic response (Pacific black duck: Anas superciliosa, domestic chicken: Gallus gallus domesticus, and little corella: Cacatua sanguinea) and ‘macropods’ for the two macropod species with a measured viremic response (agile wallaby: Macropus agilis and eastern grey kangaroo: Macropus giganteus). This averaging implicitly assumes (in the absence of species-level information) that all birds and all macropods respond identically to infection. Although a strong simplifying assumption, the three bird species have very similar viremic responses, as do the two macropod species (Appendix 1—figure 2). Second, we summed all individuals of all bird species and all macropod species recorded in the Brisbane host surveys in order to calculate the relative abundance of each of these host types to match the aggregation of titer profiles (see Supplementary file 1 for the relative abundance of each host type in Brisbane). Finally, we retained only nine mosquito species for which we had both abundance data and blood meal data (Supplementary file 2), although this excludes many potentially relevant mosquito species, the nine species we retained account for 90% of the Brisbane mosquito community according to our abundance data (Supplementary file 1). Our inference on host importance in Brisbane, Australia is thus focused on the following host groupings: birds, cats, cattle, dogs, flying foxes, horses, humans, macropods, possums (namely Brushtail possums Trichosurus vulpecula), rats, rabbits, and sheep. We consider the importance of the following mosquito species: Ae. notoscriptus, Ae. procax, Ae. vigilax, Cq. linealis, Cx. annulirostris, Cx. australicus, Cx. quinquefasciatus, Cx. sitiens, Ve. funerea, and Ma. uniformis.

Multi-generation approximation

Request a detailed protocol

We approximated how RRV would spread in a naive host and mosquito community at the start of an epidemic to highlight which infection pathways drive transmission as RRV invades. To approximate epidemic transmission, we used the next-generation matrix (NGM) approach to calculate the progression of the disease in discrete time steps where each time step represents a complete cycle of transmission. Because this method relies on the total number of mosquitoes infected over a host’s entire infectious period (9 days) and the total number of hosts infected by a mosquito over its entire lifespan (38 days; weighted by their probability of surviving over this period), it approximates how epidemics would propagate if pathogen transmission occurred in discrete generations, rather than continuously in overlapping generations. It is therefore a simplification that does not fully represent time-dependent epidemic dynamics. We use this simulation simply to highlight the host and mosquito species that would experience the most infections early in an epidemic (given by the total transmission potential across both a host’s and mosquito’s infectious period).

Specifically, we first calculated the number of hosts of each species that would become infected starting with a single infected host individual of one species using GHH. To calculate which hosts would become infected in the next generation, we then used GHH starting with the individuals infected from the previous step. We repeated this process over only five generations to avoid modeling transmission over a longer period than one transmission season in Brisbane. By using the Brisbane community in which RRV is endemic, we use this analysis as an illustrative example of disease emergence and not to provide specific predictions for RRV emergence in any specific new location with no prior exposure to RRV. To estimate how infection spreads in the mosquito community, we used a similar approach, but instead started with one infected mosquito and used GVV. As with host-vector-host transmission using GHH, while this strategy provides only a coarse approximation of transmission over time by assuming discrete generations of infection, it is useful for revealing important pathways of transmission and identifying species that remain important transmitters over multiple generations without the need to parameterize a dynamic, continuous-time epidemic model.

All data used in this study are uploaded as Source Data files. All codes are hosted on GitHub (https://github.com/morgankain/RRV_HostVectorCompetence, copy archived at swh:1:rev:be7e87c3c4c8af0420a8dd42cdcff5586fdbad90): Kain, 2021a; Kain, 2021b.

Appendix 1

Statistical sub-models

Vertebrate hosts: titer profiles

We converted reported means and standard deviations for peak titer and duration of detectable titer into continuous titer profiles, which are needed to translate titer into mosquito infection probability given a feeding event. For each species, we first simulated N titer values at each of the first day, the day hosts reached their peak titer, and the last day of infection (where N is the total number of individuals of each species in the infection experiment that developed detectable viremia). We simulated the last day of infection and the log of peak titer for each species by drawing N samples from a Gaussian distribution using the reported means and standard deviations for infection duration and peak titer. We assumed titre on day 1 and the last day of infection were at a detectability threshold of 102.2 infectious units/ml blood (the detection limit of RRV in African green monkey kidney (Vero) cells: McLean et al., 2021), and that simulated peak titer occurred at the midpoint between the first and simulated last day of infection. We then fit a linear model in R to these simulated data using linear and quadratic terms for day post infection. To quantify uncertainty in quadratic titer profiles, we simulated and fit linear models to 1000 simulated sets of titer curves; in Appendix 1—figure 1 we show the 95% CI for each of the 15 hosts’ quadratic profiles generated from this procedure with the raw summary values of peak and duration of titer extracted from the literature overlayed (the area under the curve for these titer profiles are shown in Appendix 1—figure 2).

For human titer profiles, we used data obtained during an epidemic of RRV in the Cook Islands in 1980 (Rosen et al., 1981). This study measured human titer from the day of symptom onset; raw data showed that humans experienced peak titer on day 1 of symptoms. To remain consistent with how we modeled non-human titer curves, we fit quadratic curves to the human titer data, which predict a peak at the first day of symptoms and that humans have detectable titer approximately 3 days prior to symptom onset. While it is uncertain how many days prior to symptom onset humans manifest a detectable viremic response, expert opinion on RRV (Leon Hugo and John Mackenzie pers com) is that it is likely at least 1 day, and for other arboviruses such as dengue, humans produce virus titers sufficient to infect mosquitoes for multiple days prior to symptom onset (Duong et al., 2015). Because our assumption of a quadratic titer curve extends titer to 3 days that have no direct quantitative empirical support—which results in humans having a longer duration of titer than any other host—as a conservative estimate of human physiological competence, we also run our model assuming that human titer increases from an undetectable level to a peak on day 1 of symptom onset after only a single day (instead of approximately three as predicted with the quadratic model).

Mosquito vectors: infection and transmission probability

In total, we gathered data for 17 experimentally infected mosquito species (all extracted data is available as Source Data Files). In these experiments, mosquitoes were fed a given dose of RRV via an artificial blood source which contained diluted stock virus or, in limited cases, from living organisms, such as suckling mice. The proportion that went on to become infected (RRV detected in the body) and infectious (RRV detected in the saliva measured artificially or via feeding on a susceptible vertebrate) was recorded. In the generalized linear mixed effects model (GLMM) for mosquito infection probability, we used virus dose as the sole fixed effect and modeled variation among mosquito species using a random intercept and slope over dose. For transmission probability over time, we used days since infection as the sole fixed effect and modeled variation among mosquito species’ transmission over time using a random intercept and slope over time (days since feeding). While the maximum transmission probability is sometimes allowed to vary by mosquito species, we lacked the data to estimate different maxima for each species. Thus, we used simple logistic regression which models probability using an asymptote of one. Uncertainty among mosquito species (which were modeled using a random effect) were obtained from the conditional modes and conditional covariances of the random effect for species (for further details see the code available on GitHub: Kain, 2021a).

Mosquito vectors: feeding behavior

We fit our multinomial model in a Bayesian context because a Bayesian model allows us to incorporate prior probabilities in order to model feeding patterns on species that were either: (A) not detected in the host survey but appear in the blood meal data; or (B) detected in the host survey but do not show up in the blood meal data. Specifically, for case (A), priors allow us to model a mosquito’s feeding patterns on a species that would otherwise have an abundance of zero without having to make an arbitrary assumption such as, for example, that a given host species that was not observed in the community but whose blood was observed in a mosquito was exactly equal in rarity to the rarest detected species (e.g. see Hamer et al., 2009). For case (B), priors allow us to avoid the biologically implausible assumption that a mosquitoes’ preference for a host that simply was not recorded in that specific blood meal survey is exactly zero. For example, in our blood meal data, zero Culex quinquefasciatus were recorded to have taken a blood meal from humans, although it is well understood that this species does occasionally bite humans and can lead to human infection of, for example, West Nile virus (Molaei et al., 2007). We used a Dirichlet distribution for our prior on host abundance, which is the conjugate prior to the multinomial distribution (Tu, 2014). The Dirichlet distribution is parameterized with a vector of positive reals (α), with length equal to the number of categories being modeled (for us, hosts). For our Dirichlet prior we smoothed the observed host proportions in the data in an attempt to control for the low detection probability of more cryptic species to produce the following α vector (rounded for display): human = 917, dog = 187, cat 138, bird = 73, possum = 22, flying_fox = 19, cattle = 14, macropod = 7, sheep = 0.4, horse = 0.2, rabbit = 0.2, rat = 0.2.

We assume that the underlying feeding preference of each mosquito species (proportional increases or decreases in biting host species relative to biting those species in proportion to their relative abundance) across host species is Gamma distributed (a flexible two-parameter distribution on [0, inf] that can resemble an exponential distribution with mode at zero or a Gaussian-like distribution with strictly positive values). We allow the shape of this Gamma distribution to vary among mosquito species, which, in biological terms, flexibly allows the model to capture mosquitoes with specialist feeding preferences (skewed Gamma across host species—mosquitoes bite many host species rarely and a few species often) and generalist feeding tendencies (flatter Gamma—mosquitoes bite hosts in accordance with their relative abundance). To do so, we use a multi-level model in which we assume that the shape of the Gamma distributions describing each mosquito species’ preference are in turn Gamma distributed. This can be interpreted as being used to model the distribution of specialists and generalists mosquitoes in the sample. Specifically, to allow the ‘shape’ of the species-level Gamma distributions to vary, we assume that the two parameters that describe those Gamma distributions are drawn from two higher-level Gamma distributions; we used a prior of gamma(4, 4) for each of the higher-level Gamma distributions which are minimally informative priors used to constrain the model to search a realistic space of feeding preferences (e.g. not a perfectly uniform case or an extremely skewed exponential case).

Appendix 1—table 1
Reviews suggesting frameworks on how to define the terms ‘host’ and ‘vector’ vary greatly in which physiological and ecological criteria they consider (indicated with ”X’) contribute to the importance of a species as hosts or vectors.
ReferenceHost or vectorPhysiologicalEcological
Pathogen load (e.g. titre duration and magnitude)Pathogen detected (e.g. virus isolation)Immune response (e.g. detectable antibodies)Survival (i.e. survives long enough to transmit)Population susceptibilityAbundanceContact between vector and hostBreeding patternsActivity patterns
DeFoliart et al., 1987HostXXXXX
Levin et al., 2002HostXXXX
Ashford, 1997HostXXXX
Haydon et al., 2002HostXXXX
Kuno et al., 2017HostXXXX
Cleaveland and Dye, 1995HostXXX
Silva et al., 2005HostXXXX
WHO Scientific Group on Arthropod-Borne and Rodent-Borne Viral Diseases, 1985HostXXXXX
Scott, 1988HostXXXX
Wilson et al., 2017Vector
DeFoliart et al., 1987VectorXXX
Kahl et al., 2002VectorXXX
Killick-Kendrick, 1990VectorXXXXX
Beier, 2002Vector
WHO Scientific Group on Arthropod-Borne and Rodent-Borne Viral Diseases, 1985VectorXXX
Appendix 1—figure 1
Continuous virus titer profiles over hosts’ infectious periods constructed using empirical estimates of peak titer and titer duration.

For all non-human species ‘Day’ represents days since experimental exposure to Ross River virus (RRV). Solid black curves and gray envelopes show predicted medians and 95% CI calculated from all simulated titer curves. Horizontal dashed blue lines show empirically estimated peak titers (Supplementary file 1) for each species and horizontal dotted blue lines show ± 1 SD. Vertical dashed red lines show empirically estimated end dates of detectable titer and vertical dotted red lines show ± 1 SD. Horizontal solid black lines show the maximum detectable titer. For humans, points show reported means from raw data and error bars show ± 1 SD. The human titer data is shifted in time for visualization purposes (in the raw data the first observation of human titer is recorded on day 1 of symptoms not exposure). Our predictions for humans ignore the outlier data point pictured at day 10, but do simulate titer on days prior to empirically observed titer. For further details see commenting in the R code available on GitHub: Kain, 2021a.

Appendix 1—figure 2
Area under the curve (AUC) calculated from the host virus titer curves pictured in Appendix 1—figure 1.

We use AUC to collapse the continuous host titer curves (Appendix 1—figure 1) into a single metric because it simultaneously captures both the height of the curve (actual titer values) and duration of detectable titer (infectious duration). We use AUC to quantify host physiological responses (see Figure 2A); however, the complete titer curves (Appendix 1—figure 1) are used to host-to-mosquito or mosquito-to-host transmission, not AUC. Orange points and error bars (95% CI) show calculated AUC multiplied by the proportion of all of the individuals of each species that develop detectable viremia when exposed to virus (see Supplementary file 1 for the proportion of individuals of each species that developed a viremic response in infection experiments). Green points and error bars show calculated AUC ignoring ignoring the proportion of hosts that display a viremic response. Note, for example, the large difference in the physiological competence of horses using these two metrics; horses have been considered important hosts historically, although this claim has ignored the large proportion that do not produce detectable viremia (see Stephenson et al., 2018).

Appendix 1—figure 3
Probability mosquitoes become infected with Ross River virus as a function of infectious dose.

Data points show the proportion of mosquitoes with infection detected at a given infectious dose in laboratory experiments; point size reflects the total number of mosquitoes exposed to infection. Model predictions are from a binomial GLMM, with dose as a fixed effect and mosquito species as a random effect (intercept and slope over dose), which was fit in R using the package lme4 (Bates et al., 2015). Solid black lines show predicted medians, and gray envelopes are 95% CI constructed from the conditional modes and conditional covariances of the random effect (for further details see the code on GitHub: Kain, 2021a).

Appendix 1—figure 4
Probability over time that an infected mosquito transmits Ross River virus to a susceptible host given a feeding event.

Data points show the proportion of mosquitoes transmitting virus in laboratory experiments ; point size reflects the total number of mosquitoes exposed to infection and color shows the experimental dose mosquitoes were exposed to. Model predictions are from a binomial GLMM, with day as a fixed effect and random effects of mosquito species (intercept and slope over day) and reference (intercept), fit in R using the package lme4 (Bates et al., 2015). Solid black lines show predicted medians, and grey envelopes are 95% CI constructed from the conditional modes and conditional covariances of the random effect. We did not include dose as a fixed effect because of model fitting/parameter identifiability issues, but show the doses used in the laboratory experiments here (color). Dotted lines connect data points that are from the same experiment.

Appendix 1—figure 5
Area under the curve of the mosquito infection probability curves shown in Appendix 1—figure 3.

Points show medians and error bars show 95% CI.

Appendix 1—figure 6
Area under the curve of the mosquito transmission probability curves shown in Appendix 1—figure 4.

Points show medians and error bars show 95% CI. Of all mosquitoes without data just Ve lineata is pictured here as in Appendix 1—figure 4.

Appendix 1—figure 7
Culex annulirostris daily survival in laboratory conditions using the half-max of survival in optimal conditions.

In the absence of species-specific survival for most of our species we use this survival curve (from Shocket et al., 2018 who used data from McDonald et al., 1980) for all of the species in our model, but assume that survival after day 38 falls to zero.

Appendix 2

Results figures

Appendix 2—figure 1
Complete density distributions for total estimated host-to-host transmission for the the top five species by median estimates (humans, birds, possums, horses, macropods).

Distributions show the 1000 samples obtained by propagating uncertainty from all statistical sub-models see Table 1 for details. The vertical dotted lines show distribution medians.

Appendix 2—figure 2
Ross River virus transmission capability of hosts as measured by the number of second generation hosts exposed to infection vs virus transmission capability of hosts as measured by the total number of second-generation hosts that mount a viremic response.

The top panel is recreated from Figure 2C; the bottom row uses the same calculation for transmission but weights all second generation hosts by the proportion of those hosts that display a viremic response (i.e. dogs do not contribute to the sum in the bottom row). Although host ranks do not change depending on the method of quantifying host transmission importance, overall estimates of transmission decrease when removing sink infections (bottom panel).

Appendix 2—figure 3
Ross River virus transmission capability of hosts based on physiological traits alone or with consideration of ecological traits that drive transmission — assuming human titer begins only 1 day prior to symptom onset instead of assuming a full quadratic titer profile as we do in the main text.

A. Physiological response of hosts to experimental infection with RRV, summarized using the area under their estimated titer profiles over time (AUC). In all panels, points show median estimates; error bars are 95% confidence intervals (CIs) that combine the uncertainty from all statistical sub-models used to obtain the estimates presented in that panel (see Figure 1 and Box 1 for these components). Titer profile AUC is used only to quantify host physiological competence, while raw titer profiles (pictured in Appendix 1—figure 1) are used in half-cycle and complete-cycle transmission. The ordering of hosts based on highest (top) to lowest (bottom) physiological competence in A is conserved in B and C to aid visualization of host order changes among panels. B. Host-to-vector transmission; matrices show the median numbers of vectors infected by each host species, while the points show infection totals (sums across matrix rows), with error bars. C. Host-vector-host transmission. As in B, the matrices show median numbers of next-generation host infections for all host species pairs, while the points show sums across rows of the matrices (left plot) and the proportion of infections in the second generation that are in the same species as the original infected individual (center plot).

Appendix 2—figure 4
Ross River virus transmission capability of mosquitoes based on physiological traits alone or with consideration of ecological traits that drive transmission — assuming human titer begins only 1 day prior to symptom onset instead of assuming a full quadratic titer profile as we do in the main text.

A. Physiological response of mosquitoes to experimental infection with RRV, summarized using the area under (AUC) of their estimated infection probability versus dose curves multiplied by the area under their transmission probability versus time curves. Points show median estimates; the error bars in each panel are 95% confidence intervals (CIs) that combine the uncertainty from all statistical sub-models used to obtain the estimates presented in that panel (see Figure 1 and Box 1 for these components). AUC is used only to quantify mosquito physiological competence; raw infection and transmission profiles (pictured in Appendix 1—figure 3 and Appendix 1—figure 4, respectively) are used in calculations of half-cycle and complete-cycle transmission. The ordering of vector species based on highest (top) to lowest (bottom) physiological competence in A is conserved in B and C to aid visualization of vector order changes among panels. B. Vector-to-host transmission; matrices show the median numbers of hosts infected by each vector species, while the points show infection totals (sums across matrix rows), with error bars. C. Vector-host-vector transmission. As in B, the matrices show median numbers of next-generation vector infections for all vector species pairs, while the points show sums across rows of the matrices (left plot) and the proportion of infections in the second generation that are in the same species as the original infected individual (center plot).

Appendix 2—figure 5
An initial human infection propagates infection through the host community.

Starting with a single infected human in generation ‘zero’ (all hosts begin with zero infected individuals except humans), the next generation matrix approach can be used to approximate (using the time step of a generation) how an epidemic would unfold in the community. Here, we show the total number of new infections of each species as the infection spreads in the community across generations beginning with the source infection in one human. In generation one, all infections arise from the source human infection. In subsequent generations, the plotted number of infections for each species is the estimated total number of infections in that species arising from all transmission pathways. Our median 0 estimate for Ross River virus transmission in Brisbane is just above one, which results in a very slow increase in cases over generations (solid lines); however, large uncertainty for the number of infections produced by each infected host and mosquito (see Figure 2, Figure 3) results in the possibility of explosive epidemics and thousands of infected individual hosts after a few generations. The thin grey black lines are 500 epidemic realizations. Because we assume a fully susceptible host and vector population, this is an epidemic simulation, which would over-estimate the amount of RRV transmission in Brisbane because of the high host immunity in the host population that is ignored here.

Appendix 2—figure 6
An initial Ma. uniformis infection propagates through the mosquito community.

Starting with a single infected Mansonia uniformis in generation ‘zero’, the next generation matrix approach approximates the number of mosquitoes infected in subsequent generations. All generation one mosquito infections arise from the source Ma. uniformis infecting hosts and those hosts infecting mosquitoes; the plotted number of infections for each mosquito species is the estimated total number of infections in that species arising from all transmission pathways. As these results are generated from the same model that produced the results in Appendix 2—figure 5 (simply with a different perspective) median estimates (bold black line) show slightly increasing numbers of infections in mosquitoes over generations. However, large uncertainty for the number of infections produced by each infected host and mosquito (see Figure 2, Figure 3) results in the possibility of explosive epidemics and thousands of infected individual mosquitoes after a few generations. As in Appendix 2—figure 5, the thin grey black lines are 500 epidemic realizations. Because we assume a fully susceptible host and vector population, this is an epidemic simulation, which would over-estimate the amount of Ross River virus transmission in Brisbane because of the high host immunity in the host population that is ignored here.

Appendix 2—figure 7
Simulated illustrative example for how host species can change rank between host-to-mosquito (panels A-C) and host-to-host (panels D-F) definitions of competence, even without considering host abundance, mosquito abundance, mosquito biting preference, or differences in mosquito survival (each of these variables makes increases the possible routes to host rank reversal).

In this example, host species A has a more peaked titer curve than host species B (panel A). Here, when each of these host species are bit by two different mosquito species with different infection probability curves (panel B), host species B has an overall higher probability of infecting these two mosquitoes (panel C). To the right of the top panel shows the total number of mosquitoes infected over the course of 8 days of infection in these two host species, assuming five susceptible mosquitoes of each species per host and a daily biting rate of 0.4 for each mosquito species. When these mosquito species differ in their incubation rate and thus transmission probability (panel D), and the same survival probability (differential survival makes the reversal of ranks easier – if mosquito species two has lower survival the gap between host species will widen) even if they have the same survival probability (panel E), they will have different survival-weighted transmission rates per bite over time (panel F). Taking the total number of infected mosquitoes of each species in the host-to-mosquito infection step and multiplying by the total number of transmissions over the mosquitoes lifetime, considering mosquito biting rate, results in host species A producing a fraction more host-to-host infections than species B.

Data availability

All data analyzed and all code generated during this study are included in the manuscript and supporting files.

References

  1. Book
    1. Animal Medicines Australia
    (2019)
    Pets in Australia: A National Survey of Pets and People
    Canberra, Australia: Animal Medicines Australia.
    1. Ashford RW
    (1997)
    What it takes to be a reservoir host
    Belgian Journal of Zoology 127:85–90.
  2. Book
    1. Beier JC
    (2002) Vector incrimination and entomological inoculation rates
    In: Doolan D. L, editors. Malaria Methods and Protocols. New York: Springer. pp. 3–11.
    https://doi.org/10.1385/1-59259-271-6:01
  3. Conference
    1. Cleaveland S
    2. Dye C
    (1995)
    Rabies in the serengeti: the role of domestic dogs and wildlife in maintenance of disease
    Proceedings of the Third International Conference of the Southern and Eastern African Rabies Group.
    1. Doggett S
    2. Klowden M
    3. Russell R
    (2001)
    Are vector competence experiments competent vector experiments
    Arbovirus Research in Australia. 8:126–130.
    1. Doggett S
    2. Russell R
    (1997)
    Aedes notoscriptus can transmit inland and coastal isolates of ross river and barmah forest viruses from New South Wales
     Arbovirus Research in Australia 7:79–81.
  4. Book
    1. Kahl O
    2. Gern L
    3. Eisen L
    4. Lane RS
    (2002) Ecological research on Borrelia burgdorferi sensu lato: terminology and some methodological pitfalls
    In: Gray J, Kahl O, Lane R, Stanek G, editors. Lyme Borreliosis: Biology, Epidemiology and Control. New York: CABI Publishing. pp. 29–46.
    https://doi.org/10.1079/9780851996325.0029
    1. Kay BH
    (1982) Three modes of transmission of Ross River virus by Aedes vigilax (Skuse)
    The Australian Journal of Experimental Biology and Medical Science 60:339–344.
    https://doi.org/10.1038/icb.1982.37
    1. Kay B
    2. Hall R
    3. Fanning I
    4. Mottram P
    5. Young P
    6. Pollitt C
    (1986)
    Experimental infection of vertebrates with Murray Valley encephalitis and Ross River viruses
    Arbovirus Research in Australia.  4::71–75.
    1. Kay B
    2. Aaskov JG
    (1989)
    Ross river virus (epidemic polyarthritis)
    The Arboviruses: Epidemiology and Ecology 4:93–112.
    1. Marshall ID
    2. Woodroofe GM
    3. Gard GP
    (1980) Arboviruses of coastal south-eastern Australia
    Australian Journal of Experimental Biology and Medical Science 58:91–102.
    https://doi.org/10.1038/icb.1980.9
    1. Russell R
    (1995)
    Arboviruses and their vectors in Australia: an update on the ecology and epidemiology of some mosquito-borne arboviruses
    Rev Med Vet Entomol 83:141–158.
  5. Book
    1. Scott T
    (1988) Vertebrate host ecology
    In: Monath T. P, editors. The Arboviruses: Epidemiology and Ecology. Boca Raton: CRC Press. pp. 257–280.
    https://doi.org/10.1201/9780429289170
    1. Silva ES
    2. Gontijo CM
    3. Melo MN
    (2005)
    Criteria for an infected host to be considered a primary reservoir
    Trends in Parasitology 12:550–552.
  6. Software
    1. Stan Development Team
    (2020) RStan, version 2.21.2
    The {R} Interface to Stan.
    1. Taylor LH
    2. Latham SM
    3. Woolhouse MEJ
    (2001) Risk factors for human disease emergence
    Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences 356:983–989.
    https://doi.org/10.1098/rstb.2001.0888
  7. Book
    1. Tu S
    (2014)
    The Dirichlet-Multinomial and Dirichlet-Categorical Models for Bayesian Inference
    Berkeley: Computer Science Division.
    1. WHO Scientific Group on Arthropod-Borne and Rodent-Borne Viral Diseases
    (1985)
    Arthropod-borne and rodent-borne viral diseases. Report of a WHO scientific group
    World Health Organization Technical Report Series 719:1–116.
    1. Wilson AJ
    2. Morgan ER
    3. Booth M
    4. Norman R
    5. Perkins SE
    6. Hauffe HC
    7. Mideo N
    8. Antonovics J
    9. McCallum H
    10. Fenton A
    (2017) What is a vector?
    Philosophical Transactions of the Royal Society B: Biological Sciences 372:20160085.
    https://doi.org/10.1098/rstb.2016.0085

Decision letter

  1. Thomas S Churcher
    Reviewing Editor; Imperial College London, United Kingdom
  2. Miles P Davenport
    Senior Editor; University of New South Wales, Australia
  3. Gregory Albery
    Reviewer; Department of Biology, Georgetown University

Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Acceptance summary:

This paper nicely highlights the huge amount of data needed to understand the complexities of vector borne disease transmission and control. It produces an elegant framework to rigorously bring together disparate sources of data from multiple hosts and vectors and the results give clear policy relevant results for the control of Ross River Virus in Brisbane.

Decision letter after peer review:

Thank you for submitting your article "Physiology and ecology together regulate host and vector importance for Ross River virus and other vector-borne diseases" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Miles Davenport as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Gregory Albery (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

The manuscript represents a considerable body of work and brings together disparate sources of data in a single framework for understanding the relative importance of different vector and host species. This is a welcomed approach to allow the relative contributions (and importantly, the relative uncertainties) to be explicitly investigated.

Essential revisions:

1 – Title should be revised. Though I would agree that physiology and ecology regulate other vector borne diseases this is not shown in this manuscript. The authors should therefore consider removing "and other vector-borne diseases" from the title.

2 – Could the authors explain the rationale for using the ranking framework over something more quantitative such as the use of the basic reproduction number (R0). Ranking systems have the potential to hide considerable nuance, especially when parameterisation is unclear.

3 – The authors mention many of the uncertainties in the discussion, but this is largely ignored in the abstract and results where point estimates which rely on small sample datasets are presented (see comments 1.1 and 1.2 in particular). Given the inevitable lack of good quality data to parameterise the model uncertainty is always likely to be underestimated. Nevertheless, it would be great if the uncertainties could be more fully reflected, including statistical analyses if these were thought to be appropriate. Reviewer 1 highlights that this could be done in a single Bayesian framework. If this is not possible perhaps a figure (along the lines of Figure 1) highlighting the uncertainty of the result would be appropriate. Either way, the abstract and conclusions should be tempered to reflect this.

4 – The model appears to be parameterised for an endemic setting but the multigenerational model highlights invasion dynamics. Why were generations 1,3 and 5 chosen to highlight? The authors state that vector control is conducted in Brisbane to control the disease though this doesn't appear in the parameterisation. Why was this the case and would it not influence the conclusions i.e. it is more likely to target anthropophagic vectors so could reduce their importance.

5 – The paper could be shorted substantially given some of the repetition in the introduction and discussion (a ~10% reduction should be easily possible). Could the authors also considering adding a few sentences of to the results very briefly outlining the methods used in each section (see reviewers comment 2.1).

Reviewer #1 (Recommendations for the authors):

The manuscript proposed an interesting approach for an important and famously hard problem. There was much I liked about it, such as approaching the vector-host system in both directions for investigating the complete transmission cycle. I also like the attempt to integrate different sources of data and the results figures were very clear. However, I found the manuscript poorly written (too long, repetitive and not very clear) and there were important gaps in the description of the data that raised concerns about the validity of the methods and strength of the results. I expand on this and provide other suggestions below.

Major concerns (gaps in data and methods description)

1.1 – Methods. It would be important to add statistical support to the results. In particular Figure 2 and 3. Without it the rank results are based on visual observations but with the confidence intervals overlapping so much, they are only very weakly supported. To me, in most cases there is no "winning" host/vector. Humans also have such wide confidence intervals that strong conclusions about it are difficult.

1.2 – Samples sizes. There are no mention of samples sizes. These need to be added for all data. It is particularly important to determine the validity of some results, for example:

– Are the points in Figure Sm3 the data points for vector competence? If so, how can a model be fitted to a single data point and the uncertainty in the predictions is not even wider than for other species with more data? Same seem to occur in Figure Sm5.

– How many bird species and are their competences not variable?

1.3 – Origin of data. Where and how each data type was obtained is never explained. This needs to be clarified for all data. For example, the metric used for physiological competence are titers but how they are obtained?. I assume these are viral titers from titration assays and not PCRs. Are the assays equally sensitive for all species? Are they from vertebrate blood samples and vector saliva? Where do these data come from? Provide references. Perhaps a data section would be helpful.

1.4 – Parameter values. There are multiple assumptions/values that seem taken from literature included in the models that are never described/explained. E.g., infectious period of each host, host abundance, expected average number of infections etc. I appreciate that some data have been previously described, such as vertebrate abundance but a summary here would help understand some of the differences in the results e.g. weighted and non weighted AUCs. Perhaps expansion of Table 1.

1.5 – The use of the term "reservoir" for vector-borne diseases is more confusing than helpful. Who is the reservoir, the vertebrate host or the vector? Maintenance and transmission can't happen without either. Is the complex host-vector the reservoir then? I would suggest avoiding this term throughout for clarity.

1.6 – The first part of the introduction (first 5-6 paragraphs) is mostly redundant and ends up being a repetition of itself and the discussion. Could be reduced to a single paragraph and then straight to the RRV case study, which is an interesting system in itself. The wide applicability of the method is much better explained in the discussion.

1.7 – Figure 1. I didn't find this Figure helpful. On the contrary. It has too many elements that are too small to be visible and understood. Are the graphs from real data or are they make up curves and numbers? I agree that it would be useful to understand where each type of data was used for each step and what model was applied to it. But this Figure is not helping me with that. I would suggest a simplified version, for example in a diagram or schematic without images but that explains the framework, perhaps more like a workflow.

1.8 – Throughout the manuscript the authors mention "the model" to refer to the methodology or framework they develop. This is confusing because there are many models within 'the model' and the model is not actually a model per se… I suggest updating this terminology for clarity.

1.9 – The results seem to fully hang on the definition and estimation of physiological competence. The implications should be well explained.

1.10 – Why was "study" not included as random effect in the models?

1.11 – Why do a single Bayesian model for the mosquito feeding behaviour but not for the other models, and why leave out the rest of the transmission cycle? The approach is piece-meal like with multiple predictions from different GLMMs that are disconnected with one another and their summary statistics are put together at the end. Could these be put into a single Bayesian framework that allows uncertainties and data be propagated throughout the different model components?

1.12 – Bayesian model not defined, what was the likelihood and the priors? Provide reference for priors too.

1.13 – With the size of the human confidence intervals, can we make any statements about the role of human to community transmission?

1.14 – First paragraph of discussion is repetition of introduction.

1.15 – Table S1. Add disease associated with study.

1.16 – Data does not seem available in the GitHub and the model codes need to be organised. Is the Bayesian model missing?

Reviewer #2 (Recommendations for the authors):

The work presents a formidable amount of data and I particularly liked the distinction between half cycles and full cycles, and the explicit differentiation between physiological competence and exposure/behavioral processes. The finding that physiological competence may not align with epidemiological salience is an important one for vector-borne pathogens in general, and a useful lesson going forward. The paper falls short a bit on its description of the modelling process in relating them to the results, and the novelty of the finding that physiological competence does not fully explain epidemiological importance needs to be clarified, but overall this is a solid and useful contribution. The following points should be addressed

2.1 If the paper is going to be in results-first format, the results need a bit more description of the models to facilitate understanding. I found myself wondering how exactly these results were being produced, and needing to flit back and forth from the results to the methods, which isn't ideal. Including the right level of detail (i.e. without recapitulating the methods) is tricky in a results-first Results section, but I think this paper could definitely do with including more information. A few sentences at the start of each section giving a bit of background on the model formulation and how the answers were produced would do the job.

2.2 The models are doing a very heavy lift here (particularly the NGM's), and I appreciate the authors' decision to include model limitations fairly prominently in their discussion. I have no personal experience with models like these, but they appear to have been conducted well.

2.3 Although the authors' reference to host species' ecology as an important determinant of their epidemiological role is useful, the discussion particularly could do with zeroing in more precisely on what aspects of the hosts' ecology are driving the disparity between physiological competence and epidemiological importance (i.e., behaviors, vector biting preferences, and population dynamics). The findings in the paper are definitely novel, but without clarifying what might be going on to drive this disconnect there's a risk of reinventing the wheel of behavioral competence a bit. These processes are all decomposed in Cynthia Downs and colleagues' 2019 competence paper in Trends in Parasitology (for example), and I think the paper should be a bit more explicit in the introduction and discussion that there are well-appreciated reasons that we might not expect physiology to paint the whole picture of competence. For example, it is relatively unsurprising that very rare hosts are epidemiologically unimportant in the system irrespective of their physiological competence. Is it possible that humans' role in these dynamics could be reduced to "there are lots of humans, so they end up being important"? The fact that physiological competence is only one component is in the introduction (lines 54-58), but a more detailed outline of the traits (behavioral and demographic) that could override it is necessary. This will then set the stage for the findings of the models.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for submitting your article "Physiology and ecology combine to determine host and vector importance for Ross River virus" for consideration by eLife. Your article has been reviewed by 1 peer reviewer, and the evaluation has been overseen by a Reviewing Editor and Miles Davenport as the Senior Editor. The reviewers have opted to remain anonymous.

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

The revised manuscript is much clearer and nicely highlights the huge amount of data needed to understand the complexities of vector borne diseases. We feel that the article is very close to being acceptable for publication but would like to clarity on points 4 and 5 highlighted by Reviewer #1 below before we proceed. This would allow clarity on the fitting process and would improve interpretation of the titre data (no changes needed to model, just discussion). When preparing the final manuscript, the authors are encouraged to consider the other points highlighted below by Reviewer #1 which could provide further clarification to the reader.

Reviewer #1 (Recommendations for the authors):

The authors seem to have put a lot of thought in the revised manuscript and have addressed most of my concerns well. Thank you. However, this is a very dense manuscript and I still find the methods confusing in places.

1 – Figure 1 is much more informative now but perhaps could be expanded to also provide an indication of the type of model used in each parameter to show the workflow.

2 – On previous comment 1.1. I appreciate the revisions and the more honest presentation of the results. All much clearer. I also welcome the explanation of propagation of uncertainty, this is an important feature that I have missed. One of my key points here though was that regardless of how uncertainty is estimated, a formal quantitative way of comparing the densities estimated (such as z-test, Wilcoxon or Kolmogorov tests?) could strengthen the outcomes.

3 – On previous comment 1.11. I don't agree with the reasons mentioned for not using a single Bayesian model instead of the piece-meal approach – 2/3 are considered advantages of Bayesian modelling and the less-friendly approach is debatable (indeed the framework proposed in this manuscript does not seem very user-friendly for empiricists either). That said, although I would approach it differently, I cannot say it is incorrect – though I raise here the concern about how Bayesian and ML-type outcomes are combined in the NGM. The outcomes of a Bayesian model are fundamentally different from those so-called frequentist approaches. All your parameters are 'frequentist' apart from feeding behaviour. Are these being incorporated in a comparable way?

4 – Titre profiles. I have two standing concerns with these. Not sure if problems or if they just need to be explained and discussed.

4.1. Titration values tend to be highly variable within species, let alone between species. Given the small differences in the results, how confident can we be in that the results are meaningful at the ecological scale used?

4.2. The analysis are based on peak titre but considering the different profiles this could be problematic. E.g. in the example shown (Appendix 2 Figure 7: A) where one species has a marked high peak and the other has a constant titre lower than the peak of the other species. How can we link this to an infection/transmission probability? Especially the short time frame of the experiments.

5 – Appendix 1 Figure 4: Still don't understand how these curves are generated. I appreciate that a baseline curve is being generated from all the data combined (all studies and species). But then how can you differentiate among species. A species random effect is added but, on most occasions, there is 1 transmission value for 1 time point for a single species. Did the model pass validation diagnostics? There is also some confusion with the use of dose: the legend first mentions that dose was used as a fixed effect and then a few lines down says it couldn't be used as fixed effect. This also raises issue about the meaning of the transmission probability, especially post ~15 days as all is driven by a single species (no data for all others).

6 – Appendix 1 Table 1: swap word 'reservoir' for 'host' to be consistent with main text.

7 – Figure 2 add sample sizes from each species.

– L98-100. cattle, horses and flying foxes seem all pretty similar to me

https://doi.org/10.7554/eLife.67018.sa1

Author response

Essential revisions:

1 – Title should be revised. Though I would agree that physiology and ecology regulate other vector borne diseases this is not shown in this manuscript. The authors should therefore consider removing "and other vector-borne diseases" from the title.

We have removed the "and other vector-borne diseases" phrase as suggested.

2 – Could the authors explain the rationale for using the ranking framework over something more quantitative such as the use of the basic reproduction number (R0). Ranking systems have the potential to hide considerable nuance, especially when parameterisation is unclear.

We now clarify that we used a quantitative R0 framework (the Next Generation Framework [NGM]); results now focus on interpreting the continuous values estimated by this quantitative framework and their confidence intervals (which Figures 2 and 3 present). The contribution that each host and mosquito species makes to transmission as calculated in the NGM model is detailed on lines 481-492 and lines 529-544.

– We can see that our previous over-emphasis on species ranks has generated confusion regarding what our model actually calculates. We have revised this language to make it clearer that our framework produces continuous estimates of species competence, and that we discuss grouping of species because the confidence intervals on species quantitative competence values overlap substantially with some species but are distinct from others. To clarify that our framework estimates continuous, quantitative measures of transmission potential (e.g., the number of hosts infected per host) and not ranks, we have:

A) Rephrased goal #1 in the Introduction (removing the term "rank" in favor of species' relative importance)

B) Replaced explicit statements about species ranks with more descriptive statements using the outcomes that the model actually calculates (e.g., the number of mosquitoes infected per host or the number of hosts infected per mosquito). For example, see lines lines 98-100; 126; 129-131; 172-174. While we retain the use of the term "rank" in a few instances in the Discussion as a means of distilling some of the results, we have removed "rank" entirely from the Introduction, Methods, and Results.

C) Explicitly stated what our model estimates (and thus what we show in the Results) in a series of highlight/recap sentences at the start of each Results section that state the most important points of the Methods that are needed for a reader to interpret that specific Results section given the results-first format of eLife (a change suggested by Reviewer 2).

3 – The authors mention many of the uncertainties in the discussion, but this is largely ignored in the abstract and results where point estimates which rely on small sample datasets are presented (see comments 1.1 and 1.2 in particular). Given the inevitable lack of good quality data to parameterise the model uncertainty is always likely to be underestimated. Nevertheless, it would be great if the uncertainties could be more fully reflected, including statistical analyses if these were thought to be appropriate. Reviewer 1 highlights that this could be done in a single Bayesian framework. If this is not possible perhaps a figure (along the lines of Figure 1) highlighting the uncertainty of the result would be appropriate. Either way, the abstract and conclusions should be tempered to reflect this.

This is an important point. We have worked to improve our description of how uncertainty is incorporated into the model throughout the manuscript. Specifically, our framework propagates all uncertainty from all fitted statistical sub-models into the final transmission calculations. We now state more clearly in the Methods which parameters are derived from statistical sub-models (which propagate uncertainty) versus point estimates, raw data, or assumptions (which do not propagate uncertainty) (see lines 546-560). In this section we describe the three ways in which uncertainty from host titer curves, mosquito infection probabilities, mosquito transmission probabilities, and mosquito biting preferences are incorporated into all measures of competence in which these parameters appear. It is true that we used a few point estimates and in some cases raw data instead of statistical model predictions, mainly in cases where the data were too limited to formally fit a statistical model. To clarify these points we have:

A) Have added an "uncertainty" column to Table 1 that describes whether and how uncertainty was propagated.

B) Explicitly state in Results Figures 2 and 3 what uncertainty is contained within the CI.

C) Simplified Figure 1 and included the source for each component involved in transmission (e.g., model predictions or raw data) and mention which components have uncertainty propagated.

D) Added a new 'Box 1' in the Introduction (which synthesizes the three nested components of our framework; we give more information on 'Box 1' in our response to 5 below).

E) Included sample sizes in all Appendix Figures and Supplemental Tables.

F) Reduced the strength of our inference about human importance in the Discussion and Abstract.

– We note that our strategy achieves a similar effect to a full Bayesian model (more detail on why we chose to fit sub-models separately instead of a full Bayesian model in our response to Comment 1.11).

– Finally, we note that despite overlapping 95% CI between humans and the other top species (birds, horses, and macropods), much of the density of human host-to-host transmission potential resides above the upper bound of the 95% CI of all of the other species (specifically, 32% of the human density distribution lies above the upper 95% CI of birds, the next-highest species). This indicates that despite overlap in the predicted density functions of transmission potential, we are confident that humans are at least among the most competent, and potentially the single most competent, host species we studied. We point this out in the Results on line 153 and cite a new Appendix Results Figure (now Appendix 2-Figure 1) which we hope will also help to convey the uncertainty propagation (by showing the full density distribution of the outcome instead of just a CI).

4 – The model appears to be parameterised for an endemic setting but the multigenerational model highlights invasion dynamics. Why were generations 1,3 and 5 chosen to highlight? The authors state that vector control is conducted in Brisbane to control the disease though this doesn't appear in the parameterisation. Why was this the case and would it not influence the conclusions i.e. it is more likely to target anthropophagic vectors so could reduce their importance.

We can see how these decisions generated some confusion. In addition to considering the role of each host and vector species in the endemic setting (Figures 2, 3), we also wanted to consider their role in an epidemic setting (Figure 4) to highlight those hosts and mosquitoes that would play the largest role in disease emergence in a naive community. We focused on an emergence in a new community given evidence that RRV distribution is expanding. We appreciate that our focus on Brisbane, an endemic community, is not ideal, but this was the only location with sufficient data available to parameterize the model. Further, while it is true that at no time or location would all hosts be susceptible to infection in Brisbane, spatial and temporal heterogeneity in Brisbane will lead to different transmission scenarios than the specific endemic setting we consider for the main analyses. Such variation often leads endemic diseases to be analyzed with R0 to provide a baseline metric of transmission potential, which is similar to what we provide here. We have revised the language in the Methods section "Multi-generation approximation" (lines 645-670) to clarify these points. We have also added a recap sentence at the start of the Results section "Multiple generations of transmission” (lines 191-194) to remind the reader why we made this shift (or to first explain it given the Results-first structure).

– Your point about generations 1, 3, and 5 is a good one. We have revised Figure 4 to show generations 1-3 sequentially, rather than generations 1, 3, and 5, and we state in the Results (lines 200-204) and in the Figure 4 caption that pairwise transmissions have converged by generation 3 (which is why we don’t show generations 4 and 5 in the main text). We also now describe our rationale for five generations (which we show in what is now Appendix 2-Figure 5 and Appendix 2-Figure 6) on lines 661-662 of the Methods and restate it on line 280 of the Discussion.

– Though vector control does occur in Brisbane, it is often reactive and sporadic. We do not have data for these sporadic efforts and thus do not include it in the framework. We also note that we do not model seasonal transmission (with, for example, fluctuations in the abundance of hosts and mosquitoes); see lines 295-301 for an expanded discussion about seasonality. While we recognize that RRV transmission will vary by season (lines 295-301), we believe our data synthesis and model framework is novel for RRV transmission within Brisbane even without including seasonality or vector control. We have added details stating that the host and mosquito abundance surveys were conducted between October and May (lines 580-581) to help provide context for which season the presented predictions are best suited to represent.

5 – The paper could be shorted substantially given some of the repetition in the introduction and discussion (a ~10% reduction should be easily possible). Could the authors also considering adding a few sentences of to the results very briefly outlining the methods used in each section (see reviewers comment 2.1).

We have streamlined the Introduction to remove the redundancies highlighted by Reviewer 1. Also, to help streamline the Introduction (and to assist in the Results-first format) we have moved the details about the specific steps of the framework (physiology, half-cycle and complete-cycle transmission) into a new 'Box 1'. We tie this to Figure 1 to help link the life-cycle of RRV, the workflow of the model, and the details about the three ways of estimating competence. We hope this will both help the flow of the Introduction as well as assist with the Results-first format (re: Reviewer 2).

– We have reduced the length of the Discussion by removing similar restatements about physiology vs. ecology and host and vector importance.

– We have also followed the excellent suggestion to add a few sentences to the start of each Results section to assist readers given the Results-first format (re: Reviewer 2).

– The combined Introduction and Discussion are now ~4,500 words (including Box 1), down from ~5,650 words in our original submission. However, we note that despite cutting over 1,000 words from the Introduction and Discussion, the new text in the Methods and in the Results to help clarify our framework has resulted in our revision having an almost identical word count to the original submission (just about 10,200).

Reviewer #1 (Recommendations for the authors):

The manuscript proposed an interesting approach for an important and famously hard problem. There was much I liked about it, such as approaching the vector-host system in both directions for investigating the complete transmission cycle. I also like the attempt to integrate different sources of data and the results figures were very clear. However, I found the manuscript poorly written (too long, repetitive and not very clear) and there were important gaps in the description of the data that raised concerns about the validity of the methods and strength of the results. I expand on this and provide other suggestions below.

Major concerns (gaps in data and methods description)

1.1 Methods. It would be important to add statistical support to the results. In particular Figure 2 and 3. Without it the rank results are based on visual observations but with the confidence intervals overlapping so much, they are only very weakly supported. To me, in most cases there is no "winning" host/vector. Humans also have such wide confidence intervals that strong conclusions about it are difficult.

This is an important point. We can see how some of our previous language in the Results and Discussion may have inadvertently oversimplified the results and glossed over the degree of overlap in confidence intervals. We have carefully revised the text to make sure that all references to human and bird importance are stated with a caveat about large uncertainty or accompanied with words such as ("potential" or "may"). For example, see lines 248 and 408 in the Discussion or lines 96-98 and 120-122 in the Results. In the Abstract we have removed explicit reference to humans in favor of sticking to a broader statement about the species with the highest physiological competence were not the most important for community transmission (left unsaid but for rats and sheep 95% CI don’t overlap with humans for for host-to-host).

– We can see that our previous version did not make it clear enough that all of the uncertainty in all sub-models fit to data was propagated through to the estimates of half-cycle and full-cycle transmission. We now clarify how uncertainty is propagated from fitted sub-models (lines 545-560) and list the sources of data and parameters that do not have uncertainty in Table 1. In the captions of Figures 2 and 3 we also now state what uncertainty is propagated. We also briefly state in the Results that uncertainty is propagated from all statistical sub-models (lines 146, 152). Finally, in our revamped Figure 1, we list which components of the framework have uncertainty propagated. We address your other points about data and sample sizes in our responses to your other comments below.

– We also now more clearly emphasize the point that despite overlapping 95% CI between humans and the other top species (birds, horses, and macropods), much of the density of human host-to-host transmission potential resides above the upper 97.5% confidence bound of all of the other species (specifically, 32% of the human density distribution lies above the upper 95% CI of birds, the next-highest species; Results on lines 122-126). We have added Appendix 2-Figure 1 to further emphasize the point that we propagated all uncertainty from statistical sub-models (more on uncertainty propagation in our comments below).

– Finally, given that the estimates we present in the Figures are quantitative measures of transmission (the number of infected mosquitoes per infected host) and not ranks, we have removed the term "rank" from the text (except for a few intentional uses in the Discussion) to avoid confusion about what our model estimated.

1.2 Samples sizes. There are no mention of samples sizes. These need to be added for all data. It is particularly important to determine the validity of some results, for example:

– Are the points in Figure Sm3 the data points for vector competence? If so, how can a model be fitted to a single data point and the uncertainty in the predictions is not even wider than for other species with more data? Same seem to occur in Figure Sm5.

– How many bird species and are their competences not variable?

Good point. We have added sample sizes to all Supplemental Figures (now Appendix Figures) and Tables (now attached as Supplementary files: Tables S1-S3). The points in Figure Sm3 (now Appendix 1- Figure 3) are raw data from individual infection experiments (the proportion of mosquitoes found infected given a specific dose). We have expanded the caption of Appendix 1-Figure 3 to clarify this, and we have now sized the points by the number of mosquitoes exposed in each of those infection experiments.

– The confidence intervals in Appendix 1-Figure 3 and Sm5 (now Appendix 1-Figure 4) are generated directly from the logistic regression mixed model, which estimates the infection probability versus dose relationship for each mosquito species while, effectively, borrowing information on the more-studied species to inform the curves for the less-studied species. The confidence intervals are wider for species with less data (e.g., Ma. uniformis and Ma. septempunctata), which is displayed in Appendix 1-Figure 5 (they have the widest 95% CIs) as the area under the curves pictured in Appendix 1-Figure 3 (considering the uncertainty in those curves). Similarly, for the mosquitoes in Appendix 1-Figure 4, the species with fewer data points have wider confidence intervals, which can be seen in Appendix 1-Figure 6. However, it is true that they are only a little wider. Because of the information borrowing in the mixed model, the estimates for these species (that have very little data) are strongly pulled to the overall "grand" mean, which is estimated with a reasonable amount of certainty given the overall number of data points and the number of species with relatively similar responses. Also, we note that an individual species with even one data point that falls very near that mean response (e.g., Ma. septempunctata, which had been a large sample size) will produce narrower 95% CIs than will be produced for a species with one data point away from this grand mean with a smaller sample size (e.g., Ma. uniformis which has the largest 95% CIs). As a specific example: given that so few species have means that rise to, for example, an infection probability of 50% at a low dose, the CI for even mosquitoes with no data around these doses range will not include an infection probability of 50%. This uncertainty for less studied mosquitoes is propagated into the calculations for transmission, which can be seen in the very wide CI for Ma uniformis (all panels Figure 3).

– We describe our simplification of the bird's responses into "bird" on lines 626-633. We state there that this is a necessary simplification to run the model given how few birds have been infected. We now point out that the physiological response of the three birds are clustered together in Appendix 1-Figure 2 (Methods lines 626-633 and Discussion (Caveats section) lines 341-342), which provides some support for this simplification.

1.3 Origin of data. Where and how each data type was obtained is never explained. This needs to be clarified for all data. For example, the metric used for physiological competence are titers but how they are obtained?. I assume these are viral titers from titration assays and not PCRs. Are the assays equally sensitive for all species? Are they from vertebrate blood samples and vector saliva? Where do these data come from? Provide references. Perhaps a data section would be helpful.

We can see that some of our explanations of data were not clear enough. We have now:

A) Provide additional details about the data (and the data collection) used in each component of the framework in Methods sub-section that describes that framework component (the largest amount of new material being added to explain the host and mosquito abundance data; see lines 561-572 and 573-592 respectively).

B) Made sure to cite all papers from which we gathered data in these sections.

C) Expanded the details about the data and citations in what are now Tables S1-S3 (attached as Supplementary files), which should also help with this.

D) Added a Methods section titled "Mosquito abundance" (lines 573-592) which was missing in our first submission.

– All of the raw data have now been uploaded to GitHub. Our mistake of not uploading those data until now.

– We recognize that the previous submission was light on the details about the experimental methods in the publications from which we extracted data and acknowledge the influence that different experimental assays and methods can have on results. We feel, however, that reporting these details and summarizing the variation among studies (n = 22 if vector and host studies are considered) to a high level of detail would add relatively little value (and some clutter) to an already long manuscript, especially because these details can be found in one place (reviewed in detail in Stephenson et al., 2018). As such, we have now added a reference to Stephenson et al., 2018 which reviews these experimental methods for RRV hosts. We have also added the infectious units each study reports in the reworked supplemental data tables which provides at least some indication to the assay conducted.

1.4 Parameter values. There are multiple assumptions/values that seem taken from literature included in the models that are never described/explained. E.g., infectious period of each host, host abundance, expected average number of infections etc. I appreciate that some data have been previously described, such as vertebrate abundance but a summary here would help understand some of the differences in the results e.g. weighted and non weighted AUCs. Perhaps expansion of Table 1.

Thank you for pointing out these omissions. We now state the exact values that we assumed for mosquito-to-host ratio and how this relates to mosquito abundance on lines 585-592 and have added the assumed point estimate value for mosquito biting rate (0.5/day) in Table 1.

– We can also see that some of our descriptions of model parameters and assumptions were insufficient. We now specifically define host "infectious period" in the equation for host-to-mosquito transmission (Equation. 1) and lines 496-497 and line 650 of the Methods. For mosquitoes, we now state in Equation. 2 and lines 514, 600, and 651 of the Methods that we assume that no mosquitoes survive beyond day 38, and thus we model transmission of all mosquitoes only until day 38 (weighted by an exponentially decaying survival). In continuation of our overall effort to more directly describe what we modeled, we have removed most of our uses of "infectious period."

– We have also reorganized the Methods sections "Vertebrate hosts: titer profiles” (starting on line 425) and "Mosquito vectors: infection and transmission probability” (starting on line 450) so that they first state that we modeled continuous functions over time (or dose), and second describe how we summarize those continuous curves using AUC as a means of collapsing these continuous curves into a metric to compare among hosts and mosquitoes. In these sections we also point out that AUC is used only as a metric to describe these continuous curves, and that it is not used to calculate metrics of transmission (e.g., host-to-mosquito transmission) (see our comment below and in response to Comment 1.9 as well).

– We have rewritten the caption for Appendix 1-Figure 2 to more clearly state how the weighted and non-weighted AUC differ (AUC conditional on the proportion of hosts showing a viremic response or not, respectively). In the caption we also note that the weighted vs. non-weighted comparison is primarily for visualization purposes to highlight how large of a difference considering the proportion of hosts that display a viremic response makes to AUC (in the text we point out that horse titer curves have been viewed historically without conditioning on the fact that few horses actually produce detectable viremia). Finally, we also added a statement in the caption about the use of AUC to summarize the continuous curves to compare hosts (more on this in our response to Comment 1.9) and not in transmission calculations.

– The expected average number of infections is a modeling outcome (for example, the number of mosquito infections per host infection, which is pictured in Figure 2B, which is dependent on mosquito abundance, biting preferences, host titer). Hopefully our refined language around AUC, removal of the term "rank", and our response to your Comment 1.8 that we use "model" to refer to different levels of analysis (sub-models and the overall framework) will resolve this confusion.

1.5 – The use of the term "reservoir" for vector-borne diseases is more confusing than helpful. Who is the reservoir, the vertebrate host or the vector? Maintenance and transmission can't happen without either. Is the complex host-vector the reservoir then? I would suggest avoiding this term throughout for clarity.

We appreciate that infected hosts without mosquitoes would have little importance. To reduce confusion we have removed the term "reservoir" and just use "host."

1.6 – The first part of the introduction (first 5-6 paragraphs) is mostly redundant and ends up being a repetition of itself and the discussion. Could be reduced to a single paragraph and then straight to the RRV case study, which is an interesting system in itself. The wide applicability of the method is much better explained in the discussion.

Following this suggestion, we have streamlined the Introduction down from 8 to 5 paragraphs (and moved the details about the three nested steps of competence into Box 1) and left most of the discussion of the broader applicability of the method to the Discussion.

1.7 – Figure 1. I didn't find this Figure helpful. On the contrary. It has too many elements that are too small to be visible and understood. Are the graphs from real data or are they make up curves and numbers? I agree that it would be useful to understand where each type of data was used for each step and what model was applied to it. But this Figure is not helping me with that. I would suggest a simplified version, for example in a diagram or schematic without images but that explains the framework, perhaps more like a workflow.

This is very valuable feedback. We have simplified Figure 1 to a simple transmission diagram that includes the data components in text form. We highlight which pieces are modeled with sub-models and which pieces are used as raw data or point estimates.

1.8 – Throughout the manuscript the authors mention "the model" to refer to the methodology or framework they develop. This is confusing because there are many models within 'the model' and the model is not actually a model per se… I suggest updating this terminology for clarity.

We have changed our terminology so that we use the term "framework" when referring to the overall strategy and "sub-model" or just "model" for a single statistical model fit to a single data type (e.g., titer curves).

1.9 – The results seem to fully hang on the definition and estimation of physiological competence. The implications should be well explained.

It seems that our language was not clear. While it is true that different choices for how to collapse the continuous titer curves (pictured in Appendix 1-Figure 1) into a single metric to compare species (whether that be AUC, max, or days above some threshold) could affect estimates of physiological competence (though we do not expect the choice to change results much given that these metrics are all very correlated), the continuous titer curves themselves are used in the next generation matrix model for calculating species' contributions to transmission and not any collapsed form of these curves. Thus, our choice of AUC has no bearing on host-to-mosquito or host-to-host transmission. To clarify this point, we have:

A) Stated this directly on lines 445-449 (for hosts) and lines 474-478 (for mosquitoes)---here we also explain that our rationale for choosing AUC was to simultaneously capture the amount and duration of titer in one metric.

B) Added an explicit statement into the Figure 2 and Figure 3 captions stating that the full titer curves, mosquito infection, and transmission curves and not the AUC summaries are used in calculating transmission; AUC is simply used to collapse a continuous curve (which are pictured in Appendix 1- Figure 1) into a metric to allow comparison among hosts.

1.10 – Why was "study" not included as random effect in the models?

A good point. We would optimally also include a random effect of study, but we had problems fitting a random effects structure with both "species" and "study" given that many studies only focused on one species. This was especially the case for mosquito transmission probability. We have added an additional caveat in the Discussion (lines 324-325) that one problem with attributing a given response to a given mosquito species with little data is a potential conflation of a mosquito species intrinsic response with a given laboratory.

1.11 – Why do a single Bayesian model for the mosquito feeding behaviour but not for the other models, and why leave out the rest of the transmission cycle? The approach is piece-meal like with multiple predictions from different GLMMs that are disconnected with one another and their summary statistics are put together at the end. Could these be put into a single Bayesian framework that allows uncertainties and data be propagated throughout the different model components?

This comment raises an important point about uncertainty propagation through the model, which we have thoroughly revised the manuscript to address. First, we would like to clarify that while we did use a piece-meal approach, we did propagate the uncertainty from each sub-model (titer curves, mosquito infection probability, mosquito transmission probability, and mosquito biting preferences), using the full density distributions for all parameters (in one of three ways---details given on lines 545-560 and in Table 1), for use in the calculations of all final results (e.g., host to mosquito and host to host transmission). We restate the point about uncertainty propagation in the methods recap sentences in the Results sections "Half-transmission cycle " and "Complete-transmission cycle " that uncertainty from sub-models was propagated.

– We decided on the piece-meal approach instead of a full Bayesian model for a few reasons. First, this approach makes it generally more user-friendly for empiricists to understand and fit a given sub-model to the data they collect. Second, this approach makes it easier to extend to other systems where given modular components can be more easily adjusted or substituted. Finally, because the sub-models relied on completely different data sets and are independent from one another, we felt like we didn't gain much from putting them all together given that we were able to extract and propagate uncertainty anyway.

1.12 – Bayesian model not defined, what was the likelihood and the priors? Provide reference for priors too.

We have expanded our explanation of the priors in the second half of the first paragraph in the Appendix 1 section “Mosquito vectors: feeding behavior”. Though we describe our rationale for the form of the priors, we do not have specific references for the exact parameters used for each. We did, however, include a citation for the general use of a Dirichlet prior for the multinomial model.

– We define the Bayesian model on lines 601-617 (including the model equation on line 608) and describe it in further detail with the two paragraphs in Appendix 1: “Mosquito vectors: feeding behavior”. We decided to keep much of the detail of this model in the Appendix in order to keep the manuscript as streamlined as possible.

1.13 – With the size of the human confidence intervals, can we make any statements about the role of human to community transmission?

We have toned down our language in the Discussion and Abstract focused on the importance of humans. However, despite wide 95\% CIs, it is still true that 32% of the density of their host-to-host transmission results lie above the upper 95% of any other species which is evidence of their importance (see Appendix 2-Figure 1). We have added this more nuanced description of the CIs and our confidence that humans are among the most competent vertebrate hosts on lines 152-155 of the Results.

1.14 – First paragraph of discussion is repetition of introduction.

We have cut this paragraph approximately in half but have kept some of it, which we like to set the tone for the Discussion.

1.15 – Table S1. Add disease associated with study.

Most of these studies are reviews and/or discussions of hypothetical frameworks and are not associated with a single disease system. We have modified the table title and caption to reflect this.

1.16 – Data does not seem available in the GitHub and the model codes need to be organised. Is the Bayesian model missing?

We appreciate you catching this omission. The missing data have been uploaded.

– The Bayesian model is present in the GitHub (mosq_bite_random.stan)

– As a road map for readers interested in looking at the code, the readme in the GitHub repository that prints to the main page states to open top_level_script.R and that script will guide the user through the other scripts. It also provides a description of all of the items in the repository. To increase ease of use we have renamed every script with a number in front to indicate the order in which files are sourced in top_level_script.R.

Reviewer #2 (Recommendations for the authors):

The work presents a formidable amount of data and I particularly liked the distinction between half cycles and full cycles, and the explicit differentiation between physiological competence and exposure/behavioral processes. The finding that physiological competence may not align with epidemiological salience is an important one for vector-borne pathogens in general, and a useful lesson going forward. The paper falls short a bit on its description of the modelling process in relating them to the results, and the novelty of the finding that physiological competence does not fully explain epidemiological importance needs to be clarified, but overall this is a solid and useful contribution. The following points should be addressed

We appreciate this thoughtful and positive feedback.

2.1 If the paper is going to be in results-first format, the results need a bit more description of the models to facilitate understanding. I found myself wondering how exactly these results were being produced, and needing to flit back and forth from the results to the methods, which isn't ideal. Including the right level of detail (i.e. without recapitulating the methods) is tricky in a results-first Results section, but I think this paper could definitely do with including more information. A few sentences at the start of each section giving a bit of background on the model formulation and how the answers were produced would do the job.

Thank you for this valuable suggestion. We have opted for one to two recap/highlight sentences at the start of each Results section to give the reader an overview of the most important methods used in each Results sub-section. We have also added a ‘Box 1” in the Introduction (and tie it to the revamped Figure 1) which also provides some expanded detail which will hopefully help with the Results-first format.

2.2 The models are doing a very heavy lift here (particularly the NGM's), and I appreciate the authors' decision to include model limitations fairly prominently in their discussion. I have no personal experience with models like these, but they appear to have been conducted well.

Thanks.

2.3 Although the authors' reference to host species' ecology as an important determinant of their epidemiological role is useful, the discussion particularly could do with zeroing in more precisely on what aspects of the hosts' ecology are driving the disparity between physiological competence and epidemiological importance (i.e., behaviors, vector biting preferences, and population dynamics). The findings in the paper are definitely novel, but without clarifying what might be going on to drive this disconnect there's a risk of reinventing the wheel of behavioral competence a bit. These processes are all decomposed in Cynthia Downs and colleagues' 2019 competence paper in Trends in Parasitology (for example), and I think the paper should be a bit more explicit in the introduction and discussion that there are well-appreciated reasons that we might not expect physiology to paint the whole picture of competence. For example, it is relatively unsurprising that very rare hosts are epidemiologically unimportant in the system irrespective of their physiological competence. Is it possible that humans' role in these dynamics could be reduced to "there are lots of humans, so they end up being important"? The fact that physiological competence is only one component is in the introduction (lines 54-58), but a more detailed outline of the traits (behavioral and demographic) that could override it is necessary. This will then set the stage for the findings of the models.

This is a great point. First, we have rephrased and expanded paragraph 3 of the Discussion (lines 237-260) to open with a statement that others have thought about changes between physiological and ecological ability citing Downs et al., 2019 and Simpson et al., 2012. In the next sentence we cite 4 papers that have found strong discrepancies between physiological and ecological importance. Then, near the end of the paragraph we talk somewhat broadly that one strength of the three step approach is that it can help reveal what ecological factor matters for a host or vector species' changing importance. We have also expanded our Discussion paragraph 4 (lines 261-277) as a continuation of this topic to describe with more nuance what is driving human and bird species transmission importance in RRV. As you say, humans are indeed abundant so increase in importance. However, this is not the full story as they are a moderately preferred sources for blood meals by moderately competent mosquitoes and also have a stronger physiological response than they have previously been given credit for. We contrast this scenario with birds, which are much less abundant but have a very high feeding preference for one of the most competent vectors.

– Finally, we rephrased the second paragraph of the Introduction to clarify that our framework is not presenting an entirely new strategy for thinking about competence and vector-borne disease transmission (e.g., lines 37-39). Rather, that it is a data-driven synthesis and analysis method that streamlines previously proposed ideas.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

1 – Figure 1 is much more informative now but perhaps could be expanded to also provide an indication of the type of model used in each parameter to show the workflow.

We now note in the bottom right of the figure the type of model used for each of the black-boxed terms (the terms fit with statistical models).

2 – On previous comment 1.1. I appreciate the revisions and the more honest presentation of the results. All much clearer. I also welcome the explanation of propagation of uncertainty, this is an important feature that I have missed. One of my key points here though was that regardless of how uncertainty is estimated, a formal quantitative way of comparing the densities estimated (such as z-test, Wilcoxon or Kolmogorov tests?) could strengthen the outcomes.

We respectfully disagree that a statistical test is necessary to strengthen the outcomes. Given that we propagated uncertainty and calculated 95% CI on the resulting densities we prefer relying on the simple heuristic of overlapping CI, which we find easy to interpret and visualize.

3 – On previous comment 1.11. I don't agree with the reasons mentioned for not using a single Bayesian model instead of the piece-meal approach – 2/3 are considered advantages of Bayesian modelling and the less-friendly approach is debatable (indeed the framework proposed in this manuscript does not seem very user-friendly for empiricists either). That said, although I would approach it differently, I cannot say it is incorrect – though I raise here the concern about how Bayesian and ML-type outcomes are combined in the NGM. The outcomes of a Bayesian model are fundamentally different from those so-called frequentist approaches. All your parameters are 'frequentist' apart from feeding behaviour. Are these being incorporated in a comparable way?

We appreciate your concern about combining parameter estimates in a piece-meal approach. We describe our strategy for estimating probability densities for each parameter fit using Frequentist methods on lines 548-555. By constructing these probability density functions (using the variance-covariance of the estimates and assuming multivariate normality) we were able to combine the Frequentist and Bayesian estimates by sampling both the Frequentist probability density functions of the parameter estimates and the Bayesian posteriors. Finally, while we agree that fitting the whole model using our approach may not be more empiricist-friendly than fitting a Bayesian model, using a piece-meal model does allow an empiricist to use a single code chunk to fit a single model which is not possible with one giant Bayesian model.

4 – Titre profiles. I have two standing concerns with these. Not sure if problems or if they just need to be explained and discussed.

Thank you for raising these points. We gave considerable thought on how to combine experimental data from different studies that reported only summary statistics. We believe our method appropriately captures the variation among the individuals of a given species and the number of individuals exposed (e.g., more variation and a smaller sample size produces wider uncertainty bands as shown in Appendix Figure 1). We clarify further below.

4.1. Titration values tend to be highly variable within species, let alone between species. Given the small differences in the results, how confident can we be in that the results are meaningful at the ecological scale used?

It is true that individuals vary strongly in their response to infection. However, this variation is captured in our statistical model (see Appendix 1 Figure 1; note that this variation is very large given that the y-axis is log10 titer). Given that the uncertainty in host responses is propagated through the model (to the probability that a mosquito will become infected given a bite, which then affects the number of second generation hosts they will infect), any clear results (e.g., that rats have the highest physiological competence) are clear despite this uncertainty. We also note that including the proportion of individuals that generated a viremic response is, by itself, an advancement from all previous models for RRV which only included the viremic response of a single individual and did not consider any variation within species.

– It is also true that the variation among individuals within and between species may cause the dynamics of a local emerging epidemic to not be predicted well with a model using the mean response of all species. However, this is not the primary goal of our research. Given that our primary aim was to estimate the relative importance of each species on a larger scale on average, we believe our findings are meaningful at the scale we chose. Finally, we are careful not to overstate confidence (e.g., the caveats in the Discussion) given our various simplifications/assumptions regarding ecological scale (e.g., homogeneous mixing and species abundances from surveys over a small area perfectly representing their abundances over Brisbane at large).

4.2. The analysis are based on peak titre but considering the different profiles this could be problematic. E.g. in the example shown (Appendix 2 Figure 7: A) where one species has a marked high peak and the other has a constant titre lower than the peak of the other species. How can we link this to an infection/transmission probability? Especially the short time frame of the experiments.

We agree very strongly with this point, and it supports the approach that we did use, which was not based on peak titer. We use the full time-dependent titer curve to calculate mosquito infection probability by translating the actual time-dependent titer values (y-axis on Appendix 1 Figure 1) to mosquito infection probability (x-axis in Appendix 1 Figure 3). This is stated in the Equation.1 term pj | θid which indicates that mosquito infection probability (pj) is a function of the titer in the host species i that the mosquito bites on day d of that host’s infections period (θid) (described on lines 497 and 498). We have now modified lines 437-438 to more explicitly state that we modeled “continuous functions of titer over time”. Further, we have now adjusted the caption of Figure 2 to read “…time-dependent [previously “raw”] titer profiles (pictured in Appendix 1 Figure 1) are used in half-cycle and complete-cycle transmission.” As you note, a high but short-lived viremia may not lead to more infections in mosquitoes than a low but long-lived viremia—by using the full titer curve and translating the values into mosquito infection probabilities we capture this possibility. The only place we discuss peak viremia is in the Methods section on the titer curves (lines 425-450) where we describe in detail how we translated peak and duration as reported in empirical studies into continuous titer curves. On lines 436-439 we explicitly state: “For non-human species, only means and standard deviations for peak titer and duration of detectable titer were reported. We transformed these summary measures into continuous titer profiles (continuous functions of titer over time that are needed to quantify mosquito infection probability) by modeling titer profiles as quadratic functions of time since infection, based on observed patterns in the data.” Finally, to quantify host physiological competence we calculate the area under the viremia curve rather than peak viremia alone (as explained on lines 91-96 and 445-450).

5 – Appendix 1 Figure 4: Still don't understand how these curves are generated. I appreciate that a baseline curve is being generated from all the data combined (all studies and species). But then how can you differentiate among species. A species random effect is added but, on most occasions, there is 1 transmission value for 1 time point for a single species. Did the model pass validation diagnostics? There is also some confusion with the use of dose: the legend first mentions that dose was used as a fixed effect and then a few lines down says it couldn't be used as fixed effect. This also raises issue about the meaning of the transmission probability, especially post ~15 days as all is driven by a single species (no data for all others).

First, thanks for catching the discrepancy with dose; the text has been corrected. As we correctly state in the manuscript’s Methods, dose was not used as a fixed effect in the transmission probability regression model.

– Second, while it is true that few species have data points after day 15, this has relatively little effect on the logistic regression because many high probability values (i.e., a high proportion of mosquitoes were found to be able to transmit) were recorded prior to day 15. Given that the logistic regression assumes a response will eventually reach the asymptotic probability of 1 (though we note that transmission probability is scaled by survival probability so a transmission probability of 1 is never realized by a mosquito), a number of high probability values prior to day 15 strongly constrains the location of the inflection point and thus the probabilities after day 15. That is, since a number of species approach a high transmission probability prior to day 15 (e.g., Ae. camptorhynchus, Ae. notoscriptus, Ae. aegypti, Ae. vigilax, Cq. linealis), the model is able to estimate (albeit with relatively large uncertainty) the inflection point of the logistic curve and thus the probability at higher values (> 15) of the covariate of “day”.

– Finally, to the point about generating the curves, the model is fit with lme4 and passed validation diagnostics. To generate the curves for each species we:

A) extracted the fixed effects estimates and variance–covariance matrix for the “grand mean” intercept and slope over day.

B) extracted the species-level deviates (i.e., the conditional modes of the random effect). A fitted lme4 model object in R stores the estimates for the conditional modes (here how much each species’ intercept and slope over time deviate from the fixed effects (grand mean) estimates) and the uncertainty on those estimates (in their own variance-covariance matrix). To say it in another way, these are the samples from the zero-centered multivariate Normal distribution that describes the variance in the random effects (intercept and slope over time).

We state that we use these “conditional modes and conditional covariances of the random effect” in two places: in the Appendix 1 text (at the end of the section “Mosquito vectors: infection and transmission probability”) and in the Appendix 1 Figure 4 caption. In both spots we direct the reader to the code hosted on GitHub for more detail. The R code contains commenting about steps A and B and also step C detailed below. We also state, albeit in abbreviated form, that we construct density distributions using the variance covariance matrix of the estimated coefficients on lines 550-554.

C) In brief, a mixed regression model—here the transmission probability of mosquito species i on day j—can be written: logit(yij) = α + ai + (β + bi)*xj (where α and β are the fixed effects intercept and slope estimates, ai and bi are the estimated species-level intercept and slope deviates, and xj is day j). To obtain uncertainty on yij we combine 1000 samples for the intercept and slope estimates (α and β) (using the covariance matrix for these terms) with 1000 samples from the conditional modes (ai and bi) (using the covariance matrix for the conditional modes of the random effect).

The grand mean intercept and slope over day are estimated reasonably well in the model, while the estimate for individual species (the ai and bi terms) are sometimes estimated with little uncertainty and sometimes with a lot of uncertainty (e.g., see Ae. alternans). Specifically, in the case that the data point for a given species resides far from the overall estimated mean response (e.g., Ae. alternans has a lower than average transmission probability on day 10), the uncertainty gets very large, which is sensible —the model essentially is saying: it doesn’t seem like this species is responding similar to the average, but it is very unclear what its response is. A curve can still be drawn, but as can be seen in the Ae. alternans panel, the 95% intervals at day 20 span almost the whole range of probabilities (nearly 0-1).

6 – Appendix 1 Table 1: swap word 'reservoir' for 'host' to be consistent with main text.

Done

7 – Figure 2 add sample sizes from each species.

These are modeled outputs and not raw data associated with specific sample sizes. For example, Figure 2B and 2C are showing the total number of mosquitoes or hosts a single viremic host can go on to infect, while Figure 2A represents the viremic potential for a single infected host compared to other infected hosts. We have added the term “estimated” to the description of each panel to help clarify this point. The sample sizes for the raw data used in all of the statistical models that contribute to this figure are located in the Appendix Tables; we do not feel that they should be re-reported here.

– L98-100. cattle, horses and flying foxes seem all pretty similar to me

Indeed. Thanks for catching this mistake in our language. We have updated this sentence to read that birds had a stronger viremic response that flying foxes, horses, and cattle.

https://doi.org/10.7554/eLife.67018.sa2

Article and author information

Author details

  1. Morgan P Kain

    1. Department of Biology, Stanford University, Stanford, United States
    2. Natural Capital Project, Woods Institute for the Environment, Stanford University, Stanford, United States
    Contribution
    Conceptualization, Software, Formal analysis, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing
    Contributed equally with
    Eloise B Skinner
    For correspondence
    kainm@stanford.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-0605-7289
  2. Eloise B Skinner

    1. Department of Biology, Stanford University, Stanford, United States
    2. Centre for Planetary Health and Food Security, Griffith University, Gold Coast, Australia
    Contribution
    Conceptualization, Data curation, Formal analysis, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing
    Contributed equally with
    Morgan P Kain
    For correspondence
    ebskinn@stanford.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-9032-2710
  3. Andrew F van den Hurk

    Public Health Virology, Forensic and Scientific Services, Department of Health, Brisbane, Australia
    Contribution
    Conceptualization, Data curation, Supervision, Investigation, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-6262-831X
  4. Hamish McCallum

    Centre for Planetary Health and Food Security, Griffith University, Gold Coast, Australia
    Contribution
    Supervision, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3493-0412
  5. Erin A Mordecai

    Department of Biology, Stanford University, Stanford, United States
    Contribution
    Conceptualization, Supervision, Funding acquisition, Writing - original draft, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4402-5547

Funding

National Science Foundation (DEB-1518681)

  • Erin A Mordecai

National Science Foundation and Fogarty International Center (DEB-2011147)

  • Erin A Mordecai

National Institute of General Medical Sciences (R35GM133439)

  • Morgan P Kain
  • Eloise B Skinner
  • Erin A Mordecai

Stanford University

  • Morgan P Kain

Stanford University (Natural Capital Project)

  • Morgan P Kain

Stanford University (King Center on Global Development)

  • Erin A Mordecai

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank the Mordecai and McCallum labs for feedback on model construction and our presentation of results. We further thank the Mordecai lab for feedback on our first draft of the manuscript. We thank Leon Hugo (QIMR) for advice relating to vertebrate titre profiles, John Mackenzie for advice on human titres, Charles El-Hage for advice on horse populations, and Cameron Webb for advice on mosquito ecology. We also thank the Brisbane City Council for mosquito abundance data. Finally, We thank our reviewers, whose comments improved our manuscript. EAM was supported by the National Science Foundation and the Fogarty International Center (DEB-1518681 and DEB-2011147), the King Center for Global Development, and the Terman Award. EAM, MPK, and EBS were supported by the National Institute of General Medical Sciences (R35GM133439). MPK was supported by the Natural Capital Project. AVDH states that: ‘The opinions, interpretations and conclusions are those of the author and do not necessarily represent those of the organization’.

Senior Editor

  1. Miles P Davenport, University of New South Wales, Australia

Reviewing Editor

  1. Thomas S Churcher, Imperial College London, United Kingdom

Reviewer

  1. Gregory Albery, Department of Biology, Georgetown University

Publication history

  1. Preprint posted: January 28, 2021 (view preprint)
  2. Received: January 28, 2021
  3. Accepted: August 19, 2021
  4. Accepted Manuscript published: August 20, 2021 (version 1)
  5. Version of Record published: September 22, 2021 (version 2)

Copyright

© 2021, Kain et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 561
    Page views
  • 86
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Ecology
    Line K Bay, Emily J Howells
    Insight

    The ability of corals to adapt to global warming may involve trade-offs among the traits that influence their success as the foundational species of coral reefs.

    1. Ecology
    2. Evolutionary Biology
    Motoko Iwashita, Masato Yoshizawa
    Research Article

    Social behaviour is a hallmark of complex animal systems; however, some species appear to have secondarily lost this social ability. In these non-social species, whether social abilities are permanently lost or suppressed is unclear. The blind cavefish Astyanax mexicanus is known to be asocial. Here, we reveal that cavefish exhibited social-like interactions in familiar environments but suppressed these interactions in stress-associated unfamiliar environments. Furthermore, the level of suppression in sociality was positively correlated with that of stereotypic repetitive behaviour, as seen in mammals. Treatment with a human antipsychotic drug targeting the dopaminergic system induced social-like interactions in cavefish, even in unfamiliar environments, while reducing repetitive behaviour. Overall, these results suggest that the antagonistic association between repetitive and social-like behaviours is deeply shared from teleosts through mammals.