BuzzWatch: Uncovering Multi-scale Temporal Patterns in Mosquito Behavior Through Continuous Long-term Monitoring

Reviewer #1 (Public review):

Summary:

This paper describes a behavioral platform "BuzzWatch" and its application in long-term behavioral monitoring. The study tested the system with different mosquito species and Aedes aegypti colonies and monitored behavioral response to blood feeding, change in photoperiod, and host-cue application at different times of the day.

Strengths:

BuzzWatch is a novel, custom-built behavioral system that can be used to monitor time-of-day-specific and long-term mosquito behaviors. The authors provide detailed documentation of the construction of the assay and custom flight tracking algorithm on a dedicated website, making them accessible to other researchers in the field. The authors performed a wide range of experiments using the BuzzWatch system and discovered differences in midday activity level among Aedes aegypti colonies, and reversible change in the daily activity profile post-blood-feeding.

Weaknesses:

The authors report the population metric "fraction flying" as their main readout of the daily activity profile. It is worth explaining why conventional metrics like travel distance/activity level are not reported. Alternatively, these metrics could be shown, considering the development and implementation of a flight trajectory tracking pipeline in this paper.

The authors defined the sugar-feeding index using occupancy on the sugar feeder. However, the correlation between landing on the sugar feeder and active sugar feeding is not mentioned or tested in this paper. Is sugar feeding always observed when mosquitoes land on the sugar feeder? Do they leave the sugar feeding surface once sugar feeding is complete? One can imagine that texture preference and prolonged occupancy may lead to inaccurate reporting of sugar feeding. While occupancy on the sugar feeder is an informative behavioral readout, its link with sugar feeding activity (consumption) needs to be evaluated. Otherwise, the authors should discuss the caveats that this method presents explicitly to avoid overinterpretation of their results.

Throughout the manuscript, the authors mentioned existing mosquito activity monitoring systems and their drawbacks. However, many of these statements are misleading and sometimes incorrect. The authors claim that beam-break monitors are "limited to counting active versus inactive states". Though these systems provide indirect readouts that may underreport activity, the number of beam-breaks in a time interval is correlated with activity level, as is commonly used and reported in Drosophila and mosquitoes and a number of reports in mosquitoes an updated LAM system with larger behavioral arenas and multiple infrared beams. The authors also mentioned the newer, camera-based alternatives to beam-break monitors, but again referred to these systems as "only detecting activity when a moving insect blocks a light beam"; however, these systems actually use video tracking (e.g., Araujo et al. 2020).

The fold change in behavior presented in Figure 4D is rather confusing. Under the two different photoperiods, it is not clear how an hourly comparison is justified (i.e., comparing the light-on activity in the 20L4D condition with scotophase activity in the 12L12D condition). The same point applies to Figure 4H.

The behavioral changes after changing photoperiod (Figure 4) require a control group (12L12D throughout) to account for age-related effects. This is controlled for the experiment in Figure 3 but not for Figure 4.

Reviewer #2 (Public review):

Summary:

This study establishes a platform for studying mosquito flight activity over the course of several weeks and demonstrates key applications of such a paradigm: the comparison of daily activity profiles across different Aedes aegypti populations and the quantification of responses to physiological and environmental perturbations.

Strengths:

(1) Overall, the authors succeed in setting up a low-cost, scalable tracking system that stably records mosquito flight activity for several weeks and uses it to demonstrate compelling use cases.

(2) The text is organized well, is easy to read, and is understandable for a broad audience.

(3) Instructions for constructing housing and for performing tracking with a dedicated GUI are available on an accompanying website, with open-source (and well-organized) code.

(4) A complementary pair of methods (one testing for activity signals at specific times of the day, and the other capturing broader daily patterns) is used effectively.

Weaknesses:

(1) In the interval-based GLMM results, since each time interval is tested independently, p-values should be corrected for multiple hypotheses (for instance, through controlling the false discovery rate).

(2) The accompanying GUI application needs some modifications to fully work out of the box on a sample video.

Reviewer #3 (Public review):

Summary:

The authors in this paper introduce BuzzWatch, an open-source, low-cost (200-300 Euros) platform for long-term monitoring of mosquito flight and behavior. They use a Raspberry Pi with a Noirv2 Camera set up under laboratory conditions to observe 3 different species of mosquitoes. The system captures a variety of multimodal data, like flight activity, sugar feeding, and host-seeking responses, with the help of external modules like CO2 and fructose-soaked cottons. They also release a GUI in addition to automated tracking and behaviour analysis, which doesn't run on Pi but rather on a personal laptop.

Four main use cases are demonstrated:

(1) Characterizing diel rhythms in various Aedes aegypti populations.

(2) Differentiating behaviors of native African vs. invasive human-adapted subspecies.

(3) Assessing physiological (blood-feeding) and environmental (light regime) perturbations.

(4) Testing time-of-day variation in responses to host-associated cues like CO₂ and heat.

Description (Strengths):

(1) The authors introduce a low-cost, scalable system that uses flight tracking in 2D as an alternative to 3D multi-camera systems.

(2) Due to the low pixel quality required by the system, they can record for weeks at a time, capturing long temporal and behavioral activities.

(3) They also integrate external modules such as lights, CO2, and heat as a way to measure responses to a variety of stimuli.

(4) They also introduce a wiki as a guide for building replication and a help in using the GUI module.

(5) They implement both GLMM hourly and PCA of behavior data.

Limitations - Major Comments:

(1) Most experiments are only done with single replicates per colony. If the setup is claimed to be cheap and replicable, there should be clearer replicates across experiments.

(2) No external validation for the flight tracking algorithm using manual annotation or comparison with field data. The authors focus early on biological proof of principle, but the validity of the tracking algorithm is not presented. How accurate is the algorithm at classifying behaviours (e.g., vs human ground truth)? How reliable is tracking?

(3) Why develop a custom GUI instead of using established packages such as rethomics (https://rethomics.github.io/) that are already available for behavioral analysis?

(4) Why use RGB light strips when perceptual white light for humans is not relevant for mosquitoes? The choice of lighting should be based on the mosquito's visual perception. - https://pmc.ncbi.nlm.nih.gov/articles/PMC12077400/ .

(5) Why use GLMMs instead of GAMs (with explicit periodic components)? With GLMMs, you do not account for temporal structure, which is highly relevant and autocorrelated in behavioral time series data.

(6) What is the proportion of mosquitoes that stay alive throughout the experiments? How do you address dead animals in tracking? No data are available on whether all mosquitoes made it through the monitoring period. No survival data is mentioned in the paper, and in the wiki, it is not clear how it is used or how it affects the analyses - https://theomaire.github.io/buzzwatch/analyze.html#diff-cond .

(7 )The sugar feeding behavior is not manually validated.

(8) Figure 4d is difficult to understand - how did you align time? Why is ZT4 aligning with ZT0? Should you "warp" the time series to compare them (e.g., from dawn to dusk)?

(9) No video recordings are made available for demonstration or validation purposes.

Appraisal

(1) The core conclusions---that BuzzWatch can capture multiscale mosquito behavioral rhythms and quantify the effect of genetic, environmental, and physiological variation - show promise but require stronger validation.

(2) Statistical approaches (GLMM, PCA) are chosen but may not be optimal for temporal data with autocorrelation.

(3) The host-seeking module shows a differential response, which is a potentially valuable feature.

Author response:

We were pleased to read the positive comments regarding our manuscript and thank the reviewers and editors for the constructive feedback which we believe will be very helpful to improve the current version of the manuscript.

Prior to addressing all comments in a full response, we provide a response to three issues that were raised in this provisional plan for revision: validation of the tracking algorithm, biological replicates, and mosquito survival.

(1) Validation of the tracking algorithm:

Reviewer 2 mentions that there is "No external validation for the flight tracking algorithm using manual annotation". We will address this comment in our full response by creating a manually labelled dataset to validate our detection algorithm.

However, we would like to point out two important points:

i) Quantifying the accuracy of a detection algorithm using a manually annotated set is indeed common practice in deep/machine learning algorithms in which manually annotated data are used to train the algorithm, and another set of manually annotated data is used to validate it. However, our detection and tracking algorithm is based on conventional computer vision techniques (not using any deep learning) that have been in use for several decades. Given that these algorithms are completely transparent and deterministic (as opposed to deep learning algorithms that are difficult to dissect and are created using partly stochastic processes) it is not common practice to use human annotations for validation. However, to address Reviewer 2's comment we will provide validation metrics in our full response.

ii) We furthermore would like to note that our main metrics of interest (e.g. fraction of mosquitoes flying) only depends on accurately detecting mosquitoes and quantifying movement, its accuracy is not affected by potential identity swaps (the typical bottleneck in tracking algorithms).

(2) Replicates:

Reviewer 3 states that "Most experiments are only done with single replicates". This statement is not accurate: In Figure 2 we used 3 independent biological replicates for 4 colonies, 2 of which are Aaa and 2 are Aaf. We indeed provide additional data for 6 more colonies using a single replicate. Combined this data set comprises 588 days of continuous recordings. For Figures 3 and 4 we have 2 replicates for each perturbation experiment. For Figure 5 we provided 3 replicates for the host-seeking experiments. As outlined, the vast majority of our experiments has multiple replicates. We realize this may not have been described clearly enough in the manuscript, we will clarify this in the revised manuscript.

(3) Mosquito survival:

Below we provide survival data for the data shown in Figures 1 - 4, we will include this data as supplementary material. Overall we note here that mortality for all experiments was similar to the 'baseline' mortality we observe in our standard colony maintenance procedures. After three weeks, we typically observed that 70% of mosquitoes were still alive.

Author response image 1.

Survival curves for the data presented in Figures 1 - 4 of the main text. Day 0 indicates the day on which the BuzzWatch experiment started