Associations of topicspecific peer review outcomes and institute and center award rates with funding disparities at the National Institutes of Health
Abstract
A previous report found an association of topic choice with racebased funding disparities among R01 applications submitted to the National Institutes of Health (‘NIH’) between 2011 and 2015. Applications submitted by African American or Black (‘AAB’) Principal Investigators (‘PIs’) skewed toward a small number of topics that were less likely to be funded (or ‘awarded’). It was suggested that lower award rates may be related to topicrelated biases of peer reviewers. However, the report did not account for differential funding ecologies among NIH Institutes and Centers (‘ICs’). In a reanalysis, we find that 10% of 148 topics account for 50% of applications submitted by AAB PIs. These applications on ‘AAB Preferred’ topics were funded at lower rates, but peer review outcomes were similar. The lower rate of funding for these topics was primarily due to their assignment to ICs with lower award rates, not to peerreviewer preferences.
Introduction
Data recently reported by Hoppe et al., 2019 from the National Institutes of Health ('NIH') suggest that part of the welldocumented funding disparity (Ginther et al., 2011) affecting AfricanAmerican Black ('AAB') principal investigators ('PIs') may be related to the topic of their applications. The authors of that report (including the first author of this report) found that topic choice accounted for over 20% of the disparity and wondered whether biases on the part of peer reviewers might explain why some application topics fare less well when submitted to the NIH for consideration of funding. In other words, peer reviewers might prefer certain topics over others, and specifically may be biased against applications focused on topics preferred by AAB PI’s.
However, peer review outcomes are not the only determinant of funding. Across the agency, grants are not simply awarded in order of peer review scores (Taffe and Gilpin, 2021). One reason is that applications submitted to the NIH are not just submitted to NIH; they are assigned to one of 24 grantissuing institutes or centers ('ICs') that in turn make decisions about which applications to fund. The proportion of applications funded (or 'award rate') varies across ICs; therefore, we can think of the NIH process as not one competition hinging entirely on peer review but rather as 24 separate competitions. The variability of award rates relates to differences in number of applications each IC receives, available funds, and IC priorities.
Hoppe et al., 2019 did not account for IC assignment or variation in ICspecific award rates. It is possible that the apparent link between topic choice and funding disparities may reflect differences in IC assignment, since ICs receive applications according to alignment with their stated mission. For example, applications focusing on cancer epidemiology are more likely to be assigned to the National Cancer Institute while those focusing on basic human biology are more likely to be assigned to the National Institute of General Medical Sciences. If award rates at the National Institutes of General Medical Sciences are higher than at the National Cancer Institute, it might appear that NIH 'prefers' basic human biology over cancer epidemiology. While the former topic does fare better with a higher likelihood of funding, this may be largely because of different IC award rates as opposed to differences in how the topics are received by peer reviewers.
We therefore reanalyzed the data from Hoppe et al., 2019 focusing on topicspecific peer review outcomes according to AAB PI preferences and on the possible role of IC assignment in application outcomes. To minimize welldocumented biases related to repeated looks by the peer review system on individual proposals (from resubmissions [Lauer, 2017] or competing renewals [Lauer, 2016]) we focus on de novo applications submitted to the NIH for the first time.
Materials and methods
These analyses are based on R01 applications submitted to NIH between 2011 and 2015. Hoppe et al., 2019 described in detail NIH peer review processes and the 'Word2vec' algorithm Mikolov et al., 2013 used to designate a topic for each application. Briefly, each application is assigned to a peer review group; approximately 2% of applications are withdrawn prior to review for administrative reasons. After a preliminary premeeting review, approximately half are deemed to be potentially meritorious and are therefore discussed during a formally convened meeting. After the meeting, each discussed application receives a priority score ranging from 10 (best) to 90 (worst); many, but not all, applications also receive a 'percentile ranking' to account for differences in how individual peer review groups calibrate their scores.
Applications are not only assigned to peer review groups; they are also assigned to ICs. The Center of Scientific Review (CSR) Division of Receipt and Referral makes assignments (CSR, 2021) based on IC interests, specific Funding Opportunity Announcements (as appropriate), and sometimes on applicant request. Some applications receive a dual assignment (that is, two ICs with one primary and one secondary), but changes in primary IC assignment occur less than 2% of the time. ICs ultimately make decisions about which applications to fund, with funding decisions based on peer review scores, strategic priorities, and availability of funds (NIH, 2021c).
Data and descriptive measures
Request a detailed protocolTo eliminate biases due to prior reviews and to maximize the likelihood that the analysis data frame reflects comparable observations with independently distributed variables, we focused on applications coming to the NIH for the first time; in other words, we excluded resubmissions (Lauer, 2017) and competing renewals (Lauer, 2016) which are known to systematically fare better on peer review. Each application was designated to one of 148 topics identified by the Word2vec algorithm (Mikolov et al., 2013). Hoppe et al., 2019 stated there were 150 topics, but subsequently discovered data processing errors for two topics yielded 148 for the reanalysis.
For each topic, we counted the number of applications of submitted by AAB PIs. We assessed AAB preference for each topic t by calculating the proportion of all AAB applications within that topic, that is
where i refers to individual applications, n is the number of applications in topic t, aab is an indicator of whether application i had an African American or Black contact PI (1 if yes, 0 if no), and N is total number of applications (across all topics). For each IC, we calculated award rates A as number of applications funded divided by number of applications assigned, that is
where m is the number of applications received by the IC and a is an indicator of whether an application i was awarded (1 if yes, 0 if no). Similarly, for each topic we calculated awards rates as
where again a is an indicator of whether an application i was awarded (1 if yes, 0 if no) and n is the number of applications in the topic. We estimated a predicted award rate for each topic by designating the probability for funding for each application i as its IC award rate (in other words, if an application i was assigned to NEI, the probability of funding was 0.15). Thus, the predicted topic award rate was
where $\widehat{a}$ is the award rate for the IC to which each application i is assigned and n is the number of applications in the topic.
For each topic, we calculated two peer review outcomes. The proportion of applications making it to discussion (or 'discussed') was
where d is an indicator of whether an application i was discussed (1 if yes, 0 if no) and n is the number of applications in the topic. For those applications that were discussed and therefore received a priority score, we calculated a topicspecific mean score as
where l is the number of applications in that topic that made it to discussion and s is the priority score of each application i.
Regression analyses
Request a detailed protocolFigure 1 shows the association of cumulative percentage of AAB applications by the cumulative percentage of topics. Consistent with Hoppe et al., 2019, there was a nonrandom distribution whereby 10% of topics (or 15 topics) accounted for 50% of AAB applications, while 25% of the topics accounted for 70% of AAB applications. We designted a topic as 'AAB Preferred' if it was among the 15 topics that accounted for 50% of AAB applications.
To further assess the association of topic choice with peer review outcomes, we focused on applications that were discussed and therefore received a priority score. We constructed a plot of topicspecific mean peer review scores according to number of applications in each topic. We expected to find a greater variance of mean scores for topics receiving fewer applications ('regression to the mean'). We generated a linear regression model to estimate a predicted mean score for each topic based on topic size (namely the number of submitted applications that were discussed for each topic), and calculated a residual for each topic by subtracting from each topicspecific mean score the modelbased predicted mean score. We compared the distribution of residuals for AAB preferred topics and other topics.
We next performed a series of probit regression analyses of application funding as
where $P{I}_{AAB}$ is an indicator variable for whether application i had an African American or Black PI (yes or no), ${T}_{AABP}$ is an indicator variable for whether application i focused on an AAB Preferred topic (yes or no), ${A}_{IC}$ is the IC award rate (Equation 2), and $\mathbf{\mathbf{X}}$ corresponds to variables including 'Early Stage Investigator' status, whether application i designated more than one PI, and whether the proposed research in application i involved animal and/or human subjects. Early Stage Investigators were those PIs within 10 years of their terminal research degree or completion of clinical training. Akaike Information Criteria, Bayesian Information Criteria, and Log Likelihood values informed model strength.
All analyses used R (R Foundation, 2021) packages, including tidyverse (Wickham et al., 2019), ggplot2 (Wickham, 2016), finalfit (Harrison, 2020), and texreg (Leifeld, 2013).
Results
Of 157,405 applications received, there were, after exclusion of prereview withdrawals, resubmissions and competing renewals, 96,697 applications considered and reviewed by NIH for the first time. Of these 8381 were funded, for an overall award rate of 9%. The vast majority of applications (88%) were reviewed in the NIH Center for Scientific Review (or CSR). There were 1643 applications, or 2%, submitted by AAB PIs. Table 1 shows ICspecific values for applications received, applications funded, award rates (Equation 2), and percent applications coming from AAB PIs. Of note, award rates varied from 6% to 15%, while the proportion of applications with AAB PIs ranged from <1% to nearly 15%. We also show 2015 appropriations for each IC (NIH, 2021b).
Table 2 shows review and funding outcomes for applications according to whether the assignment was to an IC in the top quartile of proportion of applications with AAB PIs ('Higher AAB'). These ICs were the National Institute of Allergy and Infectious Diseases, the National Institute of Dental and Craniofacial Research, the National Institute of Child Health and Human Development, the National Institute of Minority Health and Disparities, the National Institute of Nursing Research, and the Fogarty International Center. Applications assigned to Higher AAB ICs were three times more likely to come from AAB PIs. Review outcomes – proportion discussed and, for those applications that were discussed at peer review meetings, priority scores and percentile rankings – were similar in both groups. Despite the similar review outcomes, these applications were 10% less likely to be funded. These descriptive data show that two ICdefined cohorts of applications with similar review outcomes can have substantively different funding outcomes.
Review outcomes according to topic
Figure 2 shows the topicbased peer review outcomes according to the proportion of total AAB applications linked to each topic. There was a nonsignificant trend whereby applications focusing on topics that accounted for a greater proportion of AAB applications (Equation 1) were more likely to make it to discussion (Equation 5)(Panel A, p=0.12). When we focused on applications that were discussed and therefore received a priority score, there was no association of mean score for each topic (Equation 6) with proportion of total AAB applications linked to each topic (Equation 1) (Panel B, p=0.87).
To gain greater insights into possible peer review biases against topics preferred by AAB PIs, Figure 3, Panel A, shows the mean priority score by topic (Equation 6; again, only discussed applications receive priority scores) according to the topic size, namely the number of submitted applications that were discussed for each topic. As would be expected topics of smaller size showed greater variability, a manifestation of regression to the mean.
The line in Figure 3, Panel A, is based on a linear regression of mean score according to topic size. Although the slope was slightly negative (coeffecient −0.0002634), the association was not significant (p=0.63). Among AAB preferred topics (orange dots), there were 5 more than 1 point above the line (meaning with scores worse than predicted), while there were 3 more than 1 point below the line (meaning with scores better than predicted). The remaining seven topics had mean scores that were within 1 point of the predicted value.
For each topic, we calculated a residual by subtracting from the topicspecific mean priority score (Equation 6) the predicted mean priority score; we weighted the residuals by the topic size, as the larger topics contribute more information. Figure 3, Panel B, shows the distribution of the weighted residuals according to topic type. Residuals were more positive (i.e. worse) for AAB preferred topics. However, the absolute differences were small, much less than one priority score point (over a possible range of 10–90, with topicspecific mean values ranging from 35 to 45).
Funding outcomes according to topic: descriptive findings
Table 3 shows peer review and funding outcomes according to whether applications were focused on the 15 (or 10% of) topics that made up 50% of all applications with AAB PIs ('AAB Preferred' topics). Consistent with the absence of association with proportion of AAB applications linked to each topic (Figure 2), peer review outcomes were similar in the two groups. However, applications focusing on AAB Preferred topics were 7% less likely to be funded. Table 4 shows similar findings when focusing on the 37 (or 25% of) topics that made up 70% of all applications with AAB PIs.
Why do applications on AAB Preferred topics have worse funding outcomes despite similar peer review assessments? Table 3 shows that applications on AAB Preferred topics (those 10% of topics that accounted for 50% of all AAB applications) were 42% more likely to be assigned to Higher AAB ICs. The scatter plot in Figure 4 shows IC award rate according to the proportion of applications assigned to it that focus on AAB Preferred topics. ICs with receiving a higher percentage of AAB Preferred topic applications tended to have lower award rates (r = −0.35, p=0.08).
The association between IC award rates and application success is demonstrated in Figure 5, Panel A, which shows that observed topicspecific award rates (Equation 3) were strongly correlated with predicted award rates (Equation 4), that is award rates that would be expected based solely on ICspecific award rates (r = 0.62, p<0.0001). Figure 5, Panel B shows a tendency toward lower predicted topicspecific award rates (Equation 4) with increasing proportion of AAB applications (Equation 1) linked to each topic (p=0.06). Panel C shows that the actual topicspecific award rates (Equation 3) were nonsignificantly negatively associated with increasing proportion of AAB applications (Equation 1) (p=0.76).
Funding outcomes according to topic: probit regression models
Table 5 shows the association of an application with an AAB PI with the probability of funding. Consistent with Hoppe et al., 2019 and prior literature (Ginther et al., 2011), AAB PI applications had a lower likelihood of funding (Model 1, negative regression coefficient for AAB Principal Investigator, p<0.001). Adjusting for the topic (AAB Preferred or Other) reduced the absolute value of the regression coefficient for race by 5% (Model 2); similarly adjusting for IC assignment (Higher or Lower AAB) reduced the absolute value of the regression coefficient by 5% (Model 3). However, adjusting for the award rate of the assigned IC reduced the absolute value of the regression coefficient for race by 14% (Model 4).
Table 6 focuses on topic and funding outcomes. Without consideration of other variables, an AAB preferred topic was associated with a lower probability of funding (Model 1, negative regression coefficient for AAB preferred topic, p<0.01). However, after adjusting for IC award rate alone (Model 2, regression coefficient for AAB preferred topic close to zero, p=NS) as well as IC award rate and other characteristics (whether the PI is an early stage investigator, whether the application included more than one PI, and whether the proposed research included human and/or animal subjects), there was no association between AAB preferred topics and funding (Model 3, regression coefficient for AAB preferred topic close to zero, p=NS). The IC award rate was strongly associated with likelihood of funding (Models 2 and 3, regression coefficient positive, p<0.001).
Resubmission and singlePI applications
We repeated the main analyses, but this time focusing on resubmission applications, which as a general rule are assigned to the same IC as the original submission. The findings were similar, except that, as expected, the absolute award rates were higher. We also conducted a separate series of analyses which repeated our primarily analyses but focusing solely on singlePI applications. Again, findings were similar (see Appendix 1 and Appendix 2).
Discussion
Among over 96,000 R01 applications submitted to NIH between 2011 and 2015, 2% were submitted by AAB PIs. Their applications were skewed towards a relatively small group of 'AAB Preferred' topics (Figure 1); 10% of 148 topics accounted for 50% of AAB applications. Applications on AAB Preferred topics had similar review outcomes as those on other topics (Figure 2 and Table 3) but were less likely to be funded. The lower award rates for applications on AAB Preferred applications (Table 5, Model 1) were no longer seen after accounting for IC award rates (Table 6, Models 2 and 3).
These observations reflect that there are two factors at play in determining whether an application submitted to NIH will be funded. The first, well known to all involved with NIH system, is peer review; those applications that receive better scores are more likely to be funded. But there is a second factor, namely the funding ecology of the IC to which the application is assigned. As shown in Table 2 applications with similar peer review outcomes are less likely to be funded if they are assigned to ICs with lower overall award rates. AAB PIs are more likely to submit applications to ICs with lower award rates, and applications (whether submitted by AAB or other PIs) that focus on AAB Preferred topics tended to be assigned to ICs with lower award rates (Figure 4).
Hoppe et al., 2019 found that topic choice accounted for over 20% of funding disparities that adversely effect AAB PIs. We confirm this, but find that IC assignment (which, of course, is linked to topic) explains the disparities just as well, and that IC award rates explain the disparities even better (Table 5, 14% vs 5% reduction in absolute value of regression coefficient for AAB PI). Furthermore, after accounting for IC award rate, we found no association between topic choice and funding outcome (Table 6).
There is variability in how well different topics fare at peer review, but inspection of Figure 3 suggests that much of this variability reflects instability of estimates stemming from smaller sample sizes. Many topics that are not favored by AAB PIs receive better (lower) priority scores than the overall average, but many other such topics receive worse scores (Figure 3, Panel A). An inspection of weighted residuals suggest that AAB Preferred topics may fare a bit worse (Figure 3, Panel B), but to a lower degree than the difference of award rates among assigned ICs (Table 1). Furthermore, it should be noted that applications on these topics were more likely to make it past the first hurdle of peer review, that is reaching the point of formal discussion (Figure 2 and Table 3, see line 'Discussed'); thus, if anything, peer reviewers may be slightly biased in favor of AABpreferred topics.
Our primary analysis focused on firsttime submissions of de novo applications in which the award rates were low (<10%). Nonetheless, our analyses of resubmission applications yielded similar findings (Appendix 1). It is also important to note that this was an analysis of applications, not persons. This is an issue when considering multiPI applications, since all PI’s, not just the contact PI, plays a role in choosing the topic of their proposal. Nonetheless, an analysis confined to single PI applications yielded similar findings (Appendix 2).
Only 2% of PIs in this study were African American or Black and thus it is not possible that this small percentage of scientists cover the entire breadth of what NIH funds. The small number of AAB PIs is a source of error for describing AAB scientific interests as reflected in grant applications. Nonetheless, in recent years, NIH has increasingly recognized the importance of enhancing research in targeted topics, as exemplified by a recently announced Common Fund program to enhance transformative research in health disparities and health equity (NIH, 2021a) and by the large investment in RADxUP, a COVID19 diagnostic testing program in underserved communities (Tromberg et al., 2020).
Conclusion
The lower rate of funding for applications focused on AAB Preferred topics is likely primarily due to their assignment to ICs with lower award rates. These applications have similar peer review outcomes as those focused on other topics (Table 3). When AAB preference for each topic (Equation 1) was considered as a continuous variable, there was similarly no association between AAB preference and peer review outcomes (Figure 2). Topic choice does partially explain racebased funding disparities (Table 5, Model 2, absolute value of regression coefficient reduced by 5%), but ICspecific award rates explain the disparities to an even greater degree (Table 5, Model 4, absolute value of regression coefficient reduced by 14%). After accounting for ICspecific award rates, we find no association between topic choice and funding outcomes (Table 6, Models 2 and 3).
Appendix 1
Resubmission applications
In our primary analyses, we excluded resubmission applications and competing renewals. Since many applications are not funded on the first try, we repeated our analyses focusing on resubmissions. Our findings are similar.
Of 31,900 resubmission applications received and reviewed, 9605 were funded, for an overall award rate of 30%. There were 448 applications, or 1%, submitted by AAB PIs.
Appendix 1—figure 1 shows the association of cumulative percentage of AAB applications by the cumulative percentage of topics. There was a nonrandom distribution whereby 11 percent of topics (or 16 topics) accounted for 50 percent of AAB applications. We designted a topic as 'AAB Preferred' if it was among the 16 topics that accounted for 50% of AAB applications.
Appendix 1—figure 2 shows the topicbased peer review outcomes according to the proportion of total AAB applications linked to each topic. There was a nonsignificant trend whereby applications focusing on topics that accounted for a greater proportion of AAB applications (Equation 1) were more likely to make it to discussion (Equation 5)(Panel A, p=0.34). When we focused on applications that were discussed and therefore received a priority score, there was no association of mean score for each topic (Equation 6) with proportion of total AAB applications linked to each topic (Equation 1) (Panel B, p=0.14)
Appendix 1—table 1 shows peer review and funding outcomes according to whether applications were focused on the 16 topics that made up 50% of all resubmission applications with AAB PIs ('AAB Preferred' topics). Peer review outcomes were similar in the two groups, but applications focusing on AAB Preferred topics were 3% less likely to be funded.
Appendix 1—table 2 shows probit regressions on topic and funding outcomes. Without consideration of other variables, an AAB preferred topic was associated with a nonsignificantly lower probability of funding (Model 1, negative regression coefficient for AAB preferred topic). However, after adjusting for IC award rate alone (Model 2, regression coefficient for AAB preferred topic close to zero, p=NS) as well as other IC award rate and other characteristics, there was no association between AAB preferred topics and funding (Model 3, regression coefficient for AAB preferred topic close to zero, p=NS). The IC award rate was strongly associated with likelihood of funding (Models 2 and 3, regression coefficient positive, p<0.001)
Appendix 2
Single PI applications
In our primary analyses, we included all de novo applications, whether authored by one single PI or more than one PI ('multiPI'). Here we focus on singlePI applications. Our findings are similar.
Of 76,976 singlePI applications received and reviewed, 6556 were funded, for an overall award rate of 9%. There were 1323 applications, or 2%, submitted by AAB PIs.
Appendix 2—figure 1 shows the association of cumulative percentage of AAB applications by the cumulative percentage of topics. There was a nonrandom distribution whereby 11% of topics (or 16 topics) accounted for 50% of AAB applications. We designted a topic as 'AAB Preferred' if it was among the 16 topics that accounted for 50% of AAB applications.
Appendix 2—figure 2 shows the topicbased peer review outcomes according to the proportion of total AAB applications linked to each topic. There was a nonsignificant trend whereby applications focusing on topics that accounted for a greater proportion of AAB applications (Equation 1) were more likely to make it to discussion (Equation 5)(Panel A, p=0.12). When we focused on applications that were discussed and therefore received a priority score, there was no association of mean score for each topic (Equation 6) with proportion of total AAB applications linked to each topic (Equation 1) (Panel B, p=0.84)
Appendix 2—table 1 shows peer review and funding outcomes according to whether applications were focused on the 16 topics that made up 50% of all resubmission applications with AAB PIs ('AAB Preferred' topics). Peer review outcomes were similar in the two groups, but applications focusing on AAB Preferred topics were 7% less likely to be funded.
Appendix 2—table 2 shows probit regressions on topic and funding outcomes. Without consideration of other variables, an AAB preferred topic was associated with a lower probability of funding (Model 1, negative regression coefficient for AAB preferred topic, p<0.01). However, after adjusting for IC award rate alone (Model 2, regression coefficient for AAB preferred topic close to zero, p=NS) as well as other IC award rate and other characteristics, there was no association between AAB preferred topics and funding (Model 3, regression coefficient for AAB preferred topic close to zero, p=NS). The IC award rate was strongly associated with likelihood of funding (Models 2 and 3, regression coefficient positive, p<0.001)
Data availability
The authors have provided the deidentified data frame (in. RData format) along with three R markdown files that will make it possible for interested readers to reproduce the main paper and the two appendices (including all tables, figures, and numbers in the text).
References

Race, ethnicity, and NIH research awardsScience 333:1015–1019.https://doi.org/10.1126/science.1196783

WebsiteTom Drake, and ots Quickly Create Elegant Regression Results Tables and Plots when ModellingAccessed April 25, 2021.

Texreg: conversion of statistical model output in R to LaTeX and HTML tablesJournal of Statistical Software 55:1–24.https://doi.org/10.18637/jss.v055.i08

WebsiteNIH — Office of Budget —Appropriations History by Institute/Center (1938 to Present)Accessed April 25, 2021.

Rapid scaling up of Covid19 diagnostic testing in the united states — The NIH RADx InitiativeNew England Journal of Medicine 383:1071–1077.https://doi.org/10.1056/NEJMsr2022263

BookGGplot2: Elegant Graphics for Data AnalysisNew York: SpringerVerlag.https://doi.org/10.1007/9780387981413

Welcome to the tidyverseJournal of Open Source Software 4:1686.https://doi.org/10.21105/joss.01686
Article and author information
Author details
Funding
No external funding was received for this work.
Copyright
This is an openaccess article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.
Metrics

 2,404
 views

 209
 downloads

 23
 citations
Views, downloads and citations are aggregated across all versions of this paper published by eLife.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading

 Computational and Systems Biology
 Neuroscience
Animal behaviour alternates between stochastic exploration and goaldirected actions, which are generated by the underlying neural dynamics. Previously, we demonstrated that the compositional Restricted Boltzmann Machine (cRBM) can decompose wholebrain activity of larval zebrafish data at the neural level into a small number (∼100200) of assemblies that can account for the stochasticity of the neural activity (van der Plas et al., eLife, 2023). Here, we advance this representation by extending to a combined stochasticdynamical representation to account for both aspects using the recurrent temporal RBM (RTRBM) and transferlearning based on the cRBM estimate. We demonstrate that the functional advantage of the RTRBM is captured in the temporal weights on the hidden units, representing neural assemblies, for both simulated and experimental data. Our results show that the temporal expansion outperforms the stochasticonly cRBM in terms of generalization error and achieves a more accurate representation of the moments in time. Lastly, we demonstrate that we can identify the original timescale of assembly dynamics by estimating multiple RTRBMs at different temporal resolutions. Together, we propose that RTRBMs are a valuable tool for capturing the combined stochastic and timepredictive dynamics of largescale data sets.

 Computational and Systems Biology
 Physics of Living Systems
Explaining biodiversity is a fundamental issue in ecology. A longstanding puzzle lies in the paradox of the plankton: many species of plankton feeding on a limited variety of resources coexist, apparently flouting the competitive exclusion principle (CEP), which holds that the number of predator (consumer) species cannot exceed that of the resources at a steady state. Here, we present a mechanistic model and demonstrate that intraspecific interference among the consumers enables a plethora of consumer species to coexist at constant population densities with only one or a handful of resource species. This facilitated biodiversity is resistant to stochasticity, either with the stochastic simulation algorithm or individualbased modeling. Our model naturally explains the classical experiments that invalidate the CEP, quantitatively illustrates the universal Sshaped pattern of the rankabundance curves across a wide range of ecological communities, and can be broadly used to resolve the mystery of biodiversity in many natural ecosystems.