The complex swarming dynamics of malaria mosquitoes emerge from simple minimally-interactive behavioral rules Authors: Antoine Cribellier, Bèwadéyir Serge Poda, Roch K. Dabiré, Abdoulaye Diabaté, Olivier Roux, Florian T. Muijres, Tobias Bollenbach, David Fisher, Tobias Bollenbach, Tobias Bollenbach, Tobias Bollenbach, Tobias Bollenbach Published: 2026 ABSTRACT Swarming is a widespread collective behavior in animals, often thought to emerge from complex interactions among individuals. Here, we show that mating swarms of malaria mosquitoes can emerge from simple behavioral rules with minimal interaction between individuals. We analyzed two published experimental datasets with three-dimensional flight tracks of Anopheles coluzzii mosquitoes swarming in the lab under simulated sunset conditions and above a visual marker. We found that individuals alternate between straight flight and rapid turning maneuvers known as saccades. These saccades are triggered at the edge of the swarm, and their directions are biased along the sunset direction. This behavior was found consistent between both laboratory datasets, including swarming of solitary individuals; this indicates that inter‑individual interactions may play a limited role within the small-to-medium sized swarms that we studied. We developed a simple agent-based model incorporating three behavioral rules: attraction to the swarm center, alignment normal to the sunset horizon, both driven by environmental cues, and repulsion through short-range collision avoidance. This parsimonious model reproduced key features of the observed laboratory mosquito swarms, including looping flight paths, central density peaks, and directional alignment. Even in the absence of direct inter‑individual interactions, the model generated realistic emergent swarm dynamics, suggesting that social coordination is not required for swarm emergence. Although our lab-based findings remain to be confirmed across more naturalistic conditions, they suggest that the most parsimonious hypothesis behind male mosquito swarming is that it is primarily guided by environmental perception rather than social interaction. This minimal framework offers a new perspective on insect swarming and may apply broadly to the many other insect species that form mating swarms. Understanding the behavioral rules of mosquito swarming could inform strategies for vector control and improve the effectiveness of interventions targeting mosquito reproduction. Introduction Few behaviors are as mesmerizing as the coordinated movements of flocking, schooling or swarming animals. Such collective behaviors can serve the purpose of foraging, mating, migrating, or protecting against predators [1–6]. Studying swarms offers insights into social learning, information transfer and animal ecology [7,8]. One key discovery about these apparently complex behaviors is that they seem to emerge from simple behavioral rules [9,10]. The “boids” model suggests that animals follow three rules based on distance: attraction to distant individuals, alignment with nearby neighbors, and repulsion to avoid collisions. Our understanding of bird flocks and fish schools is largely based on these rules and models that integrate them [11–15], but it is still unknown if flying insects in swarms follow these rules or how these swarming insects achieve coordination [16–18]. Insects exhibit a wide range of collective behaviors, from coordinated migratory swarms of locusts [19], to task-oriented aggregations of eusocial species such as honeybees [20], and the mating swarms of mosquitoes and midges [16,21–23]. Unlike the strong alignment seen in vertebrate groups [24–27], many insect swarms appear comparatively chaotic [16,18,21,24–31]. Mating swarms of mosquitoes are temporary aggregations formed by males near visual markers at dusk or dawn, where females visit to find mates [28,29]. These swarms act as leks, spatially fixed arenas for sexual selection, rather than migratory collectives. Extensive research has characterized their spatial structure and environmental triggers [16,21–23,30,31], yet whether individuals interact strongly with conspecifics remains debated [16,18,23,32]. Some studies report flight-path correlations suggesting attraction or alignment [17,23,32], while others attribute these patterns to simultaneous responses to external cues such as swarm markers and light gradients [21,30,31]. Quantitative models are needed to resolve these uncertainties [16,21–23,32–41]. Mating swarms have been described to exhibit looping flight paths [21,22,28,33]. This led to the modeling of mosquito swarming as damped harmonic oscillator coupled to the swarm centroid [23,33], and to the claim that male mosquitoes follow the boids model behavioral rules [17]. The proposed behavioral rules are based on the assumption that individual mosquitoes are attracted to the perceived center of the swarms, and that swarming mosquitoes align their flight direction with those of their neighbors. However, while there are quantitative evidences of collision avoidances between mosquitoes [18,23], mutual attraction and alignment remain unproven. The argued attraction and alignment among swarming mosquitoes are based on observed correlations in their flight paths [17,23]. However, these correlations might instead arise from simultaneous responses to environmental cues like a swarm marker and the sunset [21]. This is further supported by the importance of environmental cues for swarming of many insects [34], and the limited sensory capabilities of male mosquitoes. Most mosquitoes swarm in low light conditions at dusk, and their hearing is tuned to detect female conspecifics rather than other swarming males [35]. Thus, given that male mosquitoes likely cannot accurately detect other males beyond a few centimeters [18,36], their capacity for mutual attraction or alignment is limited. Here, we aim to identify the behavioral rules that govern swarming in male Anopheles coluzzii mosquitoes. Building on our previous work that revealed the environmental cues enabling stable swarm formation [32], we combine detailed swarm kinematics analyses with agent-based modeling to test whether complex swarm dynamics can emerge from simple, minimally interactive rules. We base our swarm‑kinematics analyses on two published laboratory datasets containing three‑dimensional flight tracks of individual swarming mosquitoes [32,37], and use the empirically derived kinematic patterns to parameterize and test a minimal agent‑based model. Specifically, we investigate the relative roles of environmental cues and conspecific interactions in shaping swarm organization. To address this, we adopt a bottom-up, process approach: (1) use the two published swarming datasets to analyze and characterize the three-dimensional flight kinematics of individual mosquitoes and the emergent swarm structure [32,37]; (2) infer candidate behavioral rules from these data and prior literature [18,31,32,37,38]; (3) implement these rules in a physics-based agent-based model; and (4) evaluate whether this minimal rule set reproduces the experimentally observed swarm-level dynamics. This integrated framework allows us to directly compare empirical observations from literature with model predictions and assess the necessity of social interactions for swarm cohesion. Materials and methods In this study, we combine analyses of published three‑dimensional swarming datasets with agent‑based modeling to identify the behavioral rules that generate mosquito mating swarms. Our workflow consists of two components: (i) extracting and characterizing the flight kinematics and emergent swarm structure from the existing empirical datasets [32,37], and (ii) implementing these empirically derived patterns in a computational agent‑based model to test whether simple rules are sufficient to reproduce observed swarm dynamics. For our swarm‑kinematic analyses, we used two empirical datasets published by Poda et al. 2024 [32], and Feugère et al. 2022 [37], hereafter referred to as the Poda dataset and the Feugère dataset, respectively. Specifically, we use the Poda dataset as the primary empirical reference and the Feugère dataset to assess whether broad swarming features are observed across different laboratory conditions. Both datasets contain reconstructed three‑dimensional trajectories of individual swarming mosquitoes. We used both datasets to identify and characterize the flight patterns of individual swarming mosquitoes and the emergent structure of the swarm. We then used our primary high‑fidelity Poda dataset to develop and parameterize the agent‑based model; this dataset provides the most continuous and detailed swarming trajectories under conditions closely aligned with our modeling assumptions (e.g., being above of a flat and square swarm marker). Finally, we evaluated the model by comparing its output to the kinematics observed in the Poda dataset. We then compared the model output more broadly with the Feugère dataset to assess whether similar large-scale spatial and directional patterns occur under an independent laboratory condition. Experimental animals The empirical datasets analyzed in this study were generated using laboratory‑reared male Anopheles coluzzii mosquitoes, as described in Poda et al. 2024 [32] and Feugère et al. 2022 [37]. In both studies, mosquitoes originated from recently established laboratory lines derived from wild females, were reared under controlled insectary conditions, and were 4–6‑day‑old at the time of swarming recordings. The Poda dataset contains swarming recordings of males only, whereas the Feugère dataset includes separate recordings of male and female swarms. Further details on rearing conditions, environmental parameters, and experimental procedures are provided in the original publications [32,37]. Experimental setup The Poda and Feugère datasets were both generated in controlled laboratory environments, in flight arenas equipped with a black ground marker to induce swarming under simulated sunset illumination. The most relevant experiment-specific details are summarized below. In the Poda dataset [32], swarms were recorded in a transparent plexiglass flight arena measuring 2.0 × 0.7 × 1.8 m and placed within a larger experimental room (5.1 × 4.7 × 3.0 m). A 40 × 40 cm black cloth on the arena floor acted as the swarm marker. Sunset was simulated by gradually dimming overhead lights while maintaining a dedicated sunset light behind a black horizon panel. Two side‑mounted near‑infrared cameras (Basler acA2040‑90umNIR) equipped with IR band‑pass filters recorded at spatial and temporal resolutions of 2048 × 2048 pixels and 50 fps, respectively; four near‑infrared illuminators provided minimally-invasive videography illumination. In the Feugère dataset [37], swarming was recorded in a larger and more open flight arena, thus reducing potential wall‑interaction effects. Illumination differed from the Poda setup: infrared sources were placed above and behind the swarm marker rather than laterally. The swarm marker itself also differed in geometry, consisting of a 30‑cm‑diameter slanted vertical cylinder instead of a flat square marker. Acoustic stimuli were delivered during recordings, although the original study showed little effect of this on overall swarm dynamics [37]. Three‑dimensional flight tracks were reconstructed at the same 50 Hz recording framerate, allowing the same kinematic‑analysis pipeline to be applied as for the Poda dataset. Experimental conditions In the Poda dataset [32], swarming behavior was recorded across six laboratory swarming events, each consisting of three 5‑minute recording periods corresponding to the onset, peak, and end of swarming activity. For each event, 30 or 50 male An. coluzzii mosquitoes were introduced into the flight arena one hour before dusk to allow acclimatization. This produced swarms ranging from a single initiator to up to 26 simultaneously swarming males, enabling analysis of swarm‑size effects within the experimentally observed range. Swarming began when the overhead lights were fully dimmed and only the simulated sunset light remained illuminated. In the Feugère dataset [37], swarming recordings were obtained under comparable dusk‑transition light conditions, using the same dusk‑onset timing to induce swarming. Separate experiments were conducted for male and female swarms, providing independent datasets for both sexes. As in the Poda study, swarms were recorded for several minutes once stable swarming was established. Acoustic stimuli were presented during recordings, although analyses in the original publication showed minimal influence of these stimuli on overall swarm dynamics [37]. Flight kinematics analyses The swarming flight path reconstruction resulted in three large databases of flight tracks, being the Poda dataset and the Feugère male and female datasets. Each dataset consisted of a series of flight tracks, each defined using a space-time array (x, y, z, t), where t is the time stamp array throughout the flight track, and (x,y,z) is the position array of the mosquito relative to the center of the swarm marker ((x,y,z) = (0,0,0)). The x-, y- and z-axes are directed parallel to the sunset horizon, normal to the sunset horizon, and vertically up. For convenience, the z-coordinate is generally referred to as height, and the y-axis as the sunset axis (pointing towards the sunset). We further analyzed these flight kinematics data using Matlab 2022b (MathWorks) (see S1 Code and S1 Data [39]). First, we smoothed the track positions using a Savitzky-Golay filter. Then, we estimated for each flight track the velocity arrays (u, v, w) and acceleration array (ax, ay, az), using a central finite difference scheme of second-order accuracy. From the velocity array, we estimated the angular speed throughout each flight track as in [40]. Finally, we determined the temporal dynamics of angle of view of the visual swarm marker along y-axis (α) [32]. Diptera commonly fly in relatively straight flights, interspersed with rapid turning maneuvers, commonly called body saccades [41,42]. To test whether our swarming mosquitoes also do so, we used an analysis approach inspired by the study of van Breugel et al. (2022) on saccadic flight maneuvers in fruit flies [43]. To quantify when, where and how often mosquitoes were flying straight or performing rapid turning maneuvers, we identified where their angular speeds were close to zero (= straight flight) and where it peaked (= turning maneuver). For that, the temporal positions of peaks and valleys (i.e., local maxima and minima, respectively) of angular speed were detected using the Matlab function findpeaks, with a minimum prominence of 300° s-1 (Fig 1C). This threshold was the same as the one used in the previous Diptera flight study [43]. Furthermore, we performed a sensitivity analysis in which we systematically varied the threshold value (Fig A in S1 Appendix). This showed that the chosen threshold allows for a good separation between turning maneuvers and straight flight paths, and that the set threshold value has little impact on the main results (Fig A in S1 Appendix). The characteristics of male Anopheles coluzzii mosquito swarms. (A-E) Experimental swarming data are from the Poda dataset [32]; see Fig L in S1 Appendix for equivalent data from the Feugère datasets [37]. (A) All 938 3D flight tracks of swarming male mosquitoes above a black ground marker, and inside the flight arena. (B,C) Side view (B) and top view (C) of the positional density distribution of mosquitoes flying within the swarm, defined as the total number of recorded mosquito positions per cell (2 × 2 cm) for all replicates (n = 6 swarms × 3 recordings per swarm = 18). The contour of the swarm comprising 95% of the data points is shown in white. The swarm marker is shown as a black line or square, and the sunset direction is shown on the right. (D,E) Side view (D) and top view (E) of the median location and quartiles of each of the six recorded swarms. (F) Legend and schematics explaining how the three-dimensional shape of the swarm above the square marker is defined using three thresholds of the marker’s visual angle α (from [32]). The temporal positions of angular speed peaks and valleys were then used to divide each flight track into straight flight sections and saccadic turns. For each straight flight section between two saccadic turns, we computed the following flight kinematics parameters: flight distance traveled d, Euclidian distances dE, flight path tortuosity (𝜏 = d/dE), duration (Δt), flight speed (U), climb angle (γ), and flight heading relative to sunset axis (azimuth angle Ψ = 0° defines flight towards the sunset) (Fig 1D-1I). For each saccadic turning maneuver, we estimated the turn angle (Ω) as the difference in flight direction between the straight flight paths before and after the turn (Fig 1D,1G). For the Poda dataset, we determined the distance to the nearest walls (dwall,x and dwall,y), to test for potential wall effects. We used these metrics to characterize the individual flights of swarming mosquitoes, and test the effect of environmental cues and those of conspecifics on the swarming flight kinematics. Spatial distribution of swarming behavior To determine the spatial distribution of the flight kinematics within the swarm, we computed a series of kinematic parameters and visualized them in combined side-view and top-view projections [40]. For this, the experimental volume was divided in smaller sub-volumes of cross-section equal to 2 × 2 cm (Fig 2). In each sub-volume, we calculated the relevant flight kinematics parameters and visualized the results within the two projections: a side view and a top view. For these visualizations, we used heatmaps for scalar parameters and vector fields for vectorial parameters. Swarming flight tracks consist of series of straight flight paths and saccade turns. Experimental swarming data are from the Poda dataset [32]. see Fig M in S1 Appendix for equivalent data from the Feugère datasets [37]. (A,B) Temporal dynamics of flight speed (A) and angular speed (B) of an example swarming flight trajectory. These metrics are oscillating over time in a seemingly periodic way. Local maximums and minimums (i.e., peaks and valleys) of angular speed are identified to highlight the part of the track when the mosquito was flying straight (angular speeds close to zero), or exhibiting rapid saccadic turning maneuvers (high angular speeds). (C) Top view of a swarming flight trajectory, including three consecutive saccadic turns (1,2,3; defined in orange), the straight flight path between saccades (green vector) and the saccadic turn angle (Ω). (D,E,F) Histograms showing the tortuosity (D), distance traveled (E), and time duration (F) of the straight flight phases between successive saccadic turns (n = 19,572 straight flight phases within 938 trajectories). (G) Polar histogram of the saccadic turn angle Ω (n = 19,572 saccades). (H,I) Polar histograms of the climb angle (i.e., angle from the horizontal plane; H) and azimuth angle (i.e., angle on horizontal plane; I) during the straight flight phases. Azimuth angles of -90° and +90° correspond to directions away and towards the sunset, respectively (see sunset symbol in I). Mosquitoes horizontal flight directions showed a significant alignment with the sunset axis (Rayleigh test on axial data, (p < 0.001, with r = 0.426, showing moderate alignment). The scalar flight kinematics parameters of which we estimated the spatial distribution are: flight position density, linear and angular flight speed, saccade-to-straight flight ratio, straight flight path length, straight flight duration, straight flight heading (azimuth angle Ψ), straight flight climb angle γ, saccadic turn angle Ω, and distance to the closest neighbor. The vectorial flight kinematics parameters were the local mean flight acceleration. To test whether mosquito flight directions were aligned with the sunset axis, we analyzed flight azimuths Ψ using circular statistics in Matlab with the CircStat toolbox. Flight directions Ψ were first expressed relative to the sunset direction by subtracting the sunset azimuth (90°) from each flight heading. Because mosquitoes exhibited bidirectional flight either toward or away from the sunset, the data were treated as axial circular data by doubling the relative angles, thereby collapsing directions separated by 180° onto the same orientation axis. A Rayleigh test was then applied to the transformed angles to test for non-uniform alignment with the sunset axis. The mean resultant length r ranges from 0 to 1 and quantifies directional concentration, where values of r ≈ 0.05–0.2 indicate weak but detectable alignment in animal orientation data, r ≈ 0.2–0.5 indicate moderate directional preference, and r ≈ 0.5–1 reflect strong clustering around a preferred direction or axis. The spatial distribution of flight position density was computed as in Poda et al [32]. These mean track densities were normalized by the maximum number of data point in one sub-volume. To visualize the acceleration vector field, we computed the mean acceleration components within each sub‑volume and used these to draw a single representative vector in the xy‑plane (top view) or yz‑plane (side view). Agent-based modeling. To test which behavioral rules are needed to reproduce the swarming dynamics observed in the Poda dataset, we developed a minimal agent‑based model informed by the empirical flight kinematics extracted from this datasets. The model treats mosquitoes as self‑propelled agents whose motion is governed by a small set of behavioral rules derived from measured straight‑flight and saccadic‑turn statistics. Below, we describe how empirical kinematic distributions were used as model inputs, how the behavioral rules were formulated, and how the resulting in silico swarms were simulated and analyzed. Experimental data as input for the agent-based model We used the Poda dataset to estimate the distributions of climb and azimuth angles of the straight flight paths between two saccadic turns, relative to the position of the preceding turn (Fig B in S1 Appendix). We visualized these correlations using vector fields (Fig Ba,g in S1 Appendix) and as probability density distributions of flight heading and climb angle versus angular position of the turn (Fig Bb,h). We then fitted linear trends to these density distributions (Fig Bc,I in S1 Appendix) and computed the distributions of flight directions around these linear fits (Fig Bd,j in S1 Appendix). These combined linear‑trend functions and residual distributions were used as direct inputs for the agent‑based model. In simulations, after each saccade, the new flight direction was drawn by first evaluating the linear trend at the saccade location and then sampling from the corresponding residual distribution (Fig Be,k in S1 Appendix), which accurately reproduced the empirical post‑turn directional distributions (Fig Bf,l in S1 Appendix). This heading and climb‑angle relationship is the only component of the model directly fitted to the Poda dataset; all other parameters were taken from previous empirical observations in literature and were not tuned to improve agreement with the empirical data. This ensured that the correspondence between simulated and observed swarm dynamics arises from the minimal behavioral rule set rather than parameter optimization. Identifying agent-based model rules By combining the results of our analyses of the Poda and Feugère datasets (Figs 1-3, and Figs B, L-N in S1 Appendix) with those of previous research [18,32,38,44], we identified candidate behavioral rules that may underly swarming of malaria mosquitoes. The three behavioral rules identified are: Spatial distributions flight kinematics parameters within the swarm. Experimental data are from the Poda dataset [32]; see Fig O in S1 Appendix for equivalent data from the Feugère datasets [37]. Side view (top row) and top view (bottom row) of the spatial distributions of the major flight kinematics parameters of swarming mosquitoes (n = 938 trajectories), including (A,B) angular speed, (C,D) the saccade-to-straight flight ratio, and (E-H) the azimuth angle (E,F) and climb angle (G,H) for the straight flights between two saccades. The thin grey line shows the contour of the volume comprising 95% of the three-dimensional tracking data, the black bar and square represent the swarm marker, and the sunset location is indicated on the right. Data was shown at a cell size resolution of 2×2 cm. No data was shown for cells with less than an average of one track per recording period and per swarm (i.e., 3 recording periods × 6 swarms = 18). Behavioral rule 1. Swarming mosquitoes preferentially fly straight through the swarm center, where the marker viewing angle is maximal. They then continue flying straight, until the marker viewing angle falls below a threshold value, which triggers a turning maneuver [32]. Behavioral rule 2. The turning maneuver consists of a rapid saccadic turn within the horizontal plane, with biases towards the swarm center and along the sunset axis (towards and away from the sunset position) [32]. After the turn, the mosquito continues to fly in a straight line (rule 1). Behavioral rule 3. Mosquitoes avoid collision when the distance to a conspecific decreases to within a few body lengths [18]. Agent-based model development To test how well this set of behavioral rules results in the observed complex emergent swarm dynamics, we developed an agent-based model in Matlab 2022b (see S1 Code and S2 Data [39]). The model is similar to the ones developed for flocking birds and schooling fish [11–18], but here agents primarily interact with environmental cues (the swarm marker and sunset position) rather than with conspecifics. Note that we do not aim to identify the exact sensory cues used by swarming mosquitoes, but focus on the response dynamics to positional and directional environmental cues. We structured the model description following the ODD (Overview, Design concepts, Details) protocol [45], to improve reproducibility and clarity. Specific model parameters were taken from the Poda dataset and its accompanying publication [32]. We used the Poda dataset to evaluate the model in detail, and the Feugère dataset as an independent comparison to assess whether the broad inferred patterns were also present under different recording conditions. We applied the behavioral rules to the model using the following model rules (Fig 4): The behavioral rules used in our agent-based mosquito swarming model. (A) Legend and schematics describing the attraction and alignment behavioral rules of swarming male malaria mosquitoes. Swarming mosquitoes alternate between (i) straight flight across the center of the swarm volume and (ii) a saccadic turn at the swarm boundaries. During each saccade, they redirect their flight towards the swarm center (attraction) and along the sunset axis (alignment). (B) Legend and schematics defining the repulsion rule used in our model to prevent collision between close mosquitoes. Model rule A - Initialization and process overview. A swarming mosquito starts at a random location above the marker, flying in a random horizontal direction at a constant speed of 0.5 m/s. This flight speed matches the mean reported in the Poda dataset [32]. Model rule B - Submodel: Saccadic turning. During this straight flight section, the simulated mosquito is constantly sampling the marker viewing angle, normal to the sunset horizon α (Fig 1F). Based on this, the swarming mosquito performs a saccadic turn maneuver if (Fig 1F) [32,43]: the marker viewing angle α becomes larger than 55° or smaller than 24°, i.e., when the mosquito reaches the bottom or top boundary of the swarm, respectively. the marker viewing angle α becomes larger than 55° or smaller than 24°, i.e., when the mosquito reaches the bottom or top boundary of the swarm, respectively. the marker viewing angle deviation Rα is larger than 4.5%. The marker viewing angle deviation is defined as Rα = (αmax−α)/αmax ∙ 100%, where α is the current viewing angle and αmax is the maximal viewing angle for the current flight height (i.e., right above the marker). This lateral trigger corresponds to the elliptical cone‑shaped sides of the swarm. the marker viewing angle deviation Rα is larger than 4.5%. The marker viewing angle deviation is defined as Rα = (αmax−α)/αmax ∙ 100%, where α is the current viewing angle and αmax is the maximal viewing angle for the current flight height (i.e., right above the marker). This lateral trigger corresponds to the elliptical cone‑shaped sides of the swarm. Model rule C - Stochasticity and submodel: Turn angle selection. The new flight azimuth and climb angles after each saccadic turn were randomly sampled from distributions based on the empirical trend‑plus‑residual distributions derived from the Poda dataset (Fig B in S1 Appendix). Model rule D - Submodel: Collision avoidance. During straight flight, each mosquito constantly predicts the distance to its nearest neighbor dpredict at a future time step Δt, using linear extrapolation from current position x and velocity u (xpredict = x + u·Δt). We set the prediction time step equal to the simulation time step (Δt = 0.02 s). If dpredict falls below one body length (~6 mm), both mosquitoes execute a rapid evasive maneuver. Motivated by observed banked evasions [46], we implement this as an instantaneous 180° reorientation (velocity‑vector inversion). Model rule E - Process scheduling. After both the saccadic turn or the evasive maneuver, the simulated mosquito continued to fly in a straight line at a constant speed of 0.5 m/s, until the next saccadic turn or evasive maneuver was triggered. Agent-based model simulations We used an agent-based model of swarming mosquitoes to simulate mosquito swarming. A complete list of model parameters and their default values is provided in S1 Table [39], using ODD terminology [45]. The in-silico swarm experiments were done in the same conditions as for our in-vivo experiments. All in‑silico swarm simulations were performed under the same arena and marker dimensions as in the Poda dataset. For all simulations, we simulated ten simultaneously swarming mosquitoes, matching the typical number of simultaneously swarming individuals observed in the Poda dataset [32]. We initiated flight of these ten mosquitoes at the same time, where each had a different random initial position and flight heading (model rule A). The simulations were run at a temporal resolution of 50 Hz, which is the same as the empirical video frame rate. At each simulation time step, we updated the position of mosquitoes based on their previous position, and current flight speed and direction. We continued the simulations for two minutes, resulting in 6000 sampling points per flight track. We replicated each in-silico experiment 20 times, resulting in 200 simulated flight tracks per condition. Using this approach, we performed two sets of simulations. one in which all model rules were active, and one in which the collision‑avoidance rule (Model rule D) was disabled. This allowed us to test for the relative importance of this repulsion rule in mosquito swarming. Comparing the agent-based simulation output with the Poda dataset results. To evaluate how well the agent‑based model reproduces the swarming kinematics observed in the Poda dataset, we compared model output directly to the empirical trajectories using a unified analysis pipeline. Both the empirical data and the simulated trajectories were processed with the same codebase, which included smoothing flight tracks with a Savitzky–Golay filter and computing velocities and accelerations via a second-order central finite difference scheme. All kinematic variables were derived using identical methods for both datasets, ensuring consistency in filtering and metric computation. The smoothing procedure attenuates peak accelerations, particularly during saccadic or evasive maneuvers involving rapid turning, which results in simulated turning dynamics that are more comparable to the empirical Poda trajectories. Our primary validation step consisted of a detailed comparison between the simulated swarms and the Poda dataset, focusing on the spatial distribution of the main swarm‑kinematic parameters. We visualized the in‑silico and Poda trajectories using combined top‑ and side‑view projections, employing heatmaps for scalar parameters and vector fields for vectorial parameters, as described above. A second test of model fidelity involved comparing heading and climb‑angle behavior between the simulations and the Poda dataset, based on the dependence of straight‑flight headings on azimuthal position within the swarm and the dependence of climb angle on flight height (Fig B in S1 Appendix). Because the saccadic‑turn rule (Model rule C) was parameterized from the Poda dataset, these comparisons test whether the implemented distributions correctly reproduce the empirical relationships (experimental results in Fig Ba–d,g–j in S1 Appendix; simulation results in Fig Be,f,k,l in S1 Appendix). Next, we compared the temporal dynamics of speed and acceleration along straight‑flight segments between successive saccades, for both the Poda dataset and the agent-based model simulations. This analysis tests whether the model captures the characteristic temporal modulation of speed and acceleration observed in empirical straight‑flight segments. Finally, we compared simulations with and without the collision‑avoidance rule (Model rule D) to the Poda dataset, assessing whether short‑range repulsion improves agreement with empirical nearest‑neighbor dynamics (Fig 5I–5L). For both simulation variants, and for the Poda dataset, we estimated the nearest‑neighbor distance and the acceleration directed toward the nearest neighbor at each time step. We then evaluated how this acceleration depends on neighbor distance, testing the hypothesis that collision‑avoidance behavior produces negative (repulsive) accelerations at very short range [18]. Comparison of the spatial characteristics within the experimental swarms and simulated swarms from the Poda dataset. Throughout, empirical results (n = 938 tracks) and simulated results (n = 938 tracks) are highlighted by blue and red dashed boxes, respectively. (A-D) Side view (A,C) and top view (B,D) of the positional density distributions, within experimental swarms (A,B) and simulated swarms (C,D). The simulated data was generated using the behavioral rules described in Fig 4. (E-H) Side view (E,G) and top view (F,H) of the acceleration vector fields, within experimental swarms (E,F) and simulated swarms (G,H). The thin grey line shows the contour of the volume comprising 95% of the three-dimensional tracking data, the black bar and square represent the swarm marker, and the sunset location is indicated on the right. Data was shown at a cell size resolution of 2 × 2 cm. No data was shown for cells with less than an average of one track per recording period and per swarm (i.e., 3 recording periods × 6 swarms = 18). (I) Schematic for defining the acceleration towards the nearest neighbor as the component of the acceleration vector of a mosquito directed towards its closets neighbor. (J) Histogram of distances between nearest neighbors for the experimental and simulated swarms. (K,L) The accelerations towards their nearest neighbors versus the distance between the two neighbors, for the empirical Poda dataset (K) and the simulated swarm (L). (K) In the experimental swarms, neighboring mosquitoes tended to accelerate away from each other (negative accelerations) only at very close distances (< 1.4 cm), most likely to avoid collision. (L) Adding a simple collision avoidance mechanism to our simulations produced very similar negative accelerations between conspecifics at close distances (< 1.4 cm). Without such collision avoidance mechanisms, no negative accelerations between neighbors were produced. (K,L) At larger distances, neighboring mosquitoes tended to accelerate towards each other for both the experiments and simulations, although no mutual attractions were modeled. This shows that highly correlated swarming behavior can emerge from responses to environmental cues only, and thus no mutual attractions and alignments between conspecifics are needed. In addition to the detailed comparisons performed with the Poda dataset, we carried out a general, lower‑resolution comparison with the Feugère dataset to assess whether broad spatial and directional patterns were shared across independent laboratory conditions. This comparison was not used for parameter fitting and does not constitute a full predictive validation of our model. The Feugère dataset differs from the Poda dataset in key experimental features, most notably marker geometry. These differences likely alter how mosquitoes perceive and respond to the visual environment, and therefore limit direct transfer of the model across datasets. Results Empirical swarming results We analyzed two published laboratory datasets containing three‑dimensional trajectories of Anopheles coluzzii swarms [32,37]. The Poda dataset serves as the primary empirical reference, while the Feugère dataset is used to assess the generality of the observed spatial and kinematic patterns. In the Poda dataset, male mosquitoes swarmed above a black visual ground marker under simulated sunset illumination, yielding 938 reconstructed 3D flight tracks from 18 recordings across six swarming events that spanned the onset, peak, and end of swarming (1–26 individuals; Fig 1A). Consistent with previous analysis [32], these swarms displayed a characteristic upside‑down elliptical cone shape above the marker (Fig 1A,1B). The major axis of the ellipse was oriented normal to the sunset horizon, and the minor axis was oriented parallel to it. Swarm density peaked at the center and decreased gradually toward the boundaries (Fig 1B,1C). Following Poda et al. [32], the boundaries of the swarm volume can be described using the marker viewing angle α and its relative deviation from the local maximum value Rα. Mosquitoes remained within the swarm height range when 24° ≤ α ≤ 55°, and within the lateral boundaries when Rα ≤ 4.5% (Fig 1F). Together, these criteria delineate the three‑dimensional region in which swarming males are most likely to occur. The Feugère dataset includes swarms of both male and female An. coluzzii, recorded under a different arena geometry and using a cylindrical ground marker [37]. Despite these differences, the Feugère data exhibit similar large‑scale swarm structure, including an elevated elliptical swarm volume and a central density peak (Fig La–d in S1 Appendix). These similarities indicate that the primary geometric features of laboratory mating swarms are robust across experimental setups and do not depend on specific marker shape or arena configuration. Flight kinematics of individual swarming mosquitoes In our previous analysis of the Poda dataset [32], we focused primarily on swarm‑level structure and temporal dynamics. Here, we extend that work by quantifying the flight behavior of individual mosquitoes within the swarm (S1–S4 Videos [39]). Flight‑track durations ranged from less than one second to 232 s, with an average of 15 s [32]. Swarming males exhibited stereotypic flight behavior, as expressed by the temporal dynamics of their linear and angular flight speeds (typical example in Fig 2A,2B). They alternated rapidly between two flight phases: relatively straight flight paths and rapid sharp turns. These rapid turns are often referred to as saccadic maneuvers, and are typically exhibited by flying Diptera [41,42]. We identified straight‑flight segments and saccadic turns using the local minima and maxima of angular speed, respectively (Fig 2B). On average, swarming mosquitoes exhibited saccadic maneuvers every 0.66 ± 0.32 s (mean±standard deviation) (Fig 2F). The resulting saccadic turn frequency of 1.5Hz is similar to the turn frequencies found in previous research on swarming mosquitoes [21,33]. During a single saccadic turn, swarming mosquitoes changed their flight direction with 165° (mode), and so they turned almost completely in the opposite direction. These turn angles during swarming are much larger than the typical turn angles of saccadic maneuvers previously described in foraging flies and host-seeking mosquitoes [41,47]. During the flight sections between consecutive saccadic maneuvers, swarming mosquitoes were approximately flying straight, as the tortuosity of these flight trajectories was close to one (𝜏 = 0.84 ± 0.14; Fig 2D). We therefore define these flight sections between saccadic maneuvers from here-on as straight flight sections. The distance traveled by mosquitoes in these sections was on average 31 ± 14 cm, which is slightly less than the swarm marker size (of 40 cm) (Fig 2E). These straight segments were primarily in the horizontal plane, and directed along the sunset axis (Fig 2H,I; respectively). The climb angles γ during the straight flight sections were consistently close to zero (mode: γ = 0°); Fig 2H). To analyze the directionality of the straight flight sections, we determined the mean flight direction of each section, as defined by the azimuth angle Ψ. These flight directions combined show a bimodal distribution (Fig 2I): straight flight sections that followed a saccadic maneuver distally from the sunset were oriented primarily towards the sunset (ysaccade>0: Ψ = 91.8° (mode); Fig 2I), whereas mosquitoes flew mostly away from the sunset after a saccade proximally to the sunset position (ysaccade<0: Ψ = -91.8° (mode); Fig 2I). Together, these patterns generated the characteristic back‑and‑forth looping flights significantly aligned with the sunset axis (Rayleigh test on axial data, (p < 0.001)). Analysis of the Feugère dataset [37] showed similar individual‑level kinematics for both male and female swarms (Fig M in S1 Appendix). Mosquitoes exhibited comparable saccade frequencies, turn angles, tortuosity values, and straight‑flight climb angles (Fig Me,b,d in S1 Appendix). Straight‑flight distances were shorter (Fig Mc in S1 Appendix), likely due to the smaller cylindrical marker used in that study. As in the Poda dataset, straight‑flight azimuths in the Feugère data showed a preference aligned with the sunset axis (Rayleigh test on axial data, (p < 0.001), Figs 2I and Mg in S1 Appendix, respectively). These similarities indicate that the observed kinematic patterns are robust across different experimental geometries and recording laboratory conditions. Variations in flight kinematics within the swarm volume Next, we investigated how flight kinematics of the swarming mosquitoes varied with their position in the swarm. In the Poda dataset, we previously found that mosquitoes flew the fastest and accelerated the least in the swarm center [32]. In contrast, mosquitoes flew the slowest and accelerated most rapidly near the swarm boundaries, particularly in the regions adjacent to and distally from the sunset horizon [32]. We observed the same spatial trends in both datasets: the spatial distribution of angular speed within the swarm is very similar to the distribution of flight accelerations, whereby angular speeds were lowest near the vertical central axis of swarm, and highest on the swarm sides, adjacent to and distally from the sunset horizon (Poda dataset: Fig 3A,3B; Feugère datasets: Figs N-O in S1 Appendix). When looking at azimuth angles between successive saccades, the swarms can be divided in two halves: a region proximal to the sunset horizon, and one distally from it (Poda: Fig 3E,3F; Feugère: Figs N-O in S1 Appendix). After a saccade located near the sunset, mosquitoes primarily flew away from the sunset, and opposite. Similarly, mosquitoes near the top of the swarm tend to fly downward after a saccade, whereas near the bottom of the swarm, mosquitoes flew on average upwards after a saccade (Poda: Fig 3G,3H; Feugère: Figs N-O in S1 Appendix). All together, these results shows that swarming mosquitoes, fly fast, horizontally and straight when crossing the swarm center, and perform rapid turns at the swarm boundaries (Poda: Fig 3C,3D; Feugère: Figs N-O in S1 Appendix). These kinematics are similar to those observed in swarming midges [44]. The influence of experimental conditions on the swarming flight kinematics In any behavioral experiment, it is important to test potential effects of the experimental conditions on results. Because the Poda dataset was recorded inside a plexiglass flight arena with limited dimensions, we quantified the distance between each saccadic turn and the nearest arena wall to assess whether turning events occurred close to the arena walls (Fig K in S1 Appendix). For turns at the swarm edge, distances to the side walls were 45 ± 8.5 cm and to the front/back walls 23 ± 7.1 cm, corresponding to ~75 and ~38 body lengths, respectively. These values are far greater than the short‑range distances at which collision‑avoidance behavior occurs (~2.5 body lengths; Fig 5K) [18], and exceed reported wall‑interaction distances of one to a few body lengths [46,48]. As an additional check, we examined the Feugère dataset, which was recorded in a substantially larger and more open arena. In this dataset, the same turning patterns, boundary‑localized saccades and elevated angular speeds at the swarm periphery, were observed (Figs N-O in S1 Appendix), despite the much greater distances to any physical boundaries. Together, the large wall distances in the Poda dataset and the recurrence of the same kinematic structures in the more open Feugère recordings provide indirect evidence that the spatial organization of turning behavior is not tightly coupled to arena proximity. Variations in flight kinematics with swarm size In the Poda dataset, swarm sizes ranged from a single initiator male (S1 Video [39]) to groups of up to 26 simultaneously swarming individuals (S2–S3 Videos [39]). This range corresponds to the modal sizes reported for natural Anopheles mating swarms in the field, which commonly consist of several to a few dozen males [49–51], although much larger swarms, occasionally reaching hundreds to thousands of individuals, have been documented in some locations. Within the Poda dataset, single initiators displayed the same looping straight‑flight and saccadic‑turn sequences as individuals in larger groups. To evaluate whether individual‑level flight behavior varied over the swarm‑size range present in the dataset, we quantified all saccadic‑turn and straight‑flight metrics for each swarm and compared them across group sizes (Fig H in S1 Appendix). No systematic differences were detected across swarm sizes, in saccadic turn angle, straight‑flight distance, or straight‑flight duration. Thus, within the small‑to‑medium swarm sizes captured in the Poda dataset individual kinematics were invariant with respect to group size. Behavioral rules of malaria mosquitoes swarming The combined kinematic patterns observed in the Poda and Feugere datasets (Figs 1–3, and H in S1 Appendix, N-O in S1 Appendix) show that male An. coluzzii follow a stereotyped sequence of straight‑flight segments and rapid saccadic turns while swarming (Fig 4; S1–S4 Videos [39]). Mosquitoes fly straight through the swarm center, where the visual angle of the ground marker is maximal (αmax), and continue this straight flight until the viewing angle decreases below a threshold, at which point a turning maneuver is triggered (Fig 1F) [32]. These turns are sharp saccadic rotations within the horizontal plane, biased toward the swarm center and along the sunset axis. Saccades also include vertical adjustments: individuals near the lower boundary of the swarm tend to turn upward, whereas those near the upper boundary tend to turn downward, consistent with the spatial pattern of marker viewing angles (in the Poda dataset: 24° and 55°, respectively). After each turn, mosquitoes resume straight flight until the visual‑angle threshold is crossed again, initiating the next saccade. Previous work further shows that swarming mosquitoes perform collision‑avoidance maneuvers at close range [18]. Modeling and simulation results Based on the Poda dataset results, we developed an agent-based model for simulating mosquito swarming. We used this model to test how well these simple behavioral rules can reproduce the complex emergent swarm dynamic observed in the Poda datasets (see Material and Methods). Agent-based models are powerful tools that simulate the interactions of individual agents to understand complex collective behaviors. By modeling each animal’s simple rules of movement and interactions, we can uncover how these lead to complex patterns, providing insights into social dynamics. The agent‑based model reproduced key aspects of the kinematics observed in the Poda dataset, with simulated mosquitoes exhibiting swarm dynamics consistent with the empirical trajectories (S1–S4 Videos [39], Figs 5 and C–G in S1 Appendix). Simulated mosquitoes primarily performed saccadic maneuvers at the swarm edges and flew mostly horizontally and straight through the swarm center, with flight directions oriented primarily along the sunset axis (Figs D-E in S1 Appendix). Still, these kinematics are directly resulting from the behavioral rules, and thus these strong similarities are expected. More striking is the fact that the model generated emergent properties not directly tied to the specific rules. As in the Poda dataset, simulated mosquito density was highest in the swarm center (Fig 5A-5D), and distributions of straight flight path distances and durations were similarly distributed to those observed experimentally (Fig G in S1 Appendix). The model also replicated the radial acceleration field observed in real swarms, where mosquitoes accelerate towards the swarm center, with maximum acceleration at the edges and minimum at the center (Figs 5E-5H, and as in [31,52]). This field arises from saccadic turns at the swarm boundary followed by straight flights towards the vertical centerline. Analysis of the Poda dataset revealed that neighboring mosquitoes accelerate away from each other at distances below ~1.4 cm, likely due to evasive maneuvers [18]. In contrast, at larger distances, conspecifics tend to accelerate towards each other (Fig 5K). These positive accelerations towards each other may be the by-product of simultaneous flight behaviors of conspecifics, not mutual attractions. We tested this in the model by comparing versions with and without a collision avoidance rule (Fig 5L). Both models reproduced the positive acceleration towards neighbors at distances over 1.4 cm, suggesting that synchronized movements in swarms can emerge from reaction to environmental cues alone, without direct interactions between conspecifics (Fig 5K, 5L). Still, only the model with collision avoidance captured the repulsive accelerations of real mosquitoes at close distances (Fig 5K), showing that a minimal amount of interaction is needed to capture the emergent swarming behavior of mosquitoes. A lower‑resolution comparison with the Feugère dataset showed that the large‑scale spatial and directional patterns generated by the model were consistent across datasets with different marker geometry and arena size (Figs 5 vs. Figs L-O in S1 Appendix). This provides additional evidence that the modeled behavioral rules capture robust features of mosquito swarming. Discussion A minimal behavioral rule model for mosquito swarming According to the principle of parsimony (or Occam’s razor), the hypothesis that requires the fewest assumptions, here minimally-interactive behavioral rules, should be tested before more complex hypothesis. This is what we did in our mosquito swarm study, where our combined analyses of empirical data from literature [32,37] with agent-based modeling show that the complex emergent swarming of malaria mosquitoes can be modeled based on a small set of minimally-interactive behavioral rules. Like in flocking bird and schooling fish models, these behavioral rules include attraction, alignment and repulsion rules. But strikingly, here the attraction and alignment behaviors are not towards conspecifics, but instead relative to external environmental cues: the swarming mosquitoes are attracted towards the region above the visual swarm marker, and they align their flight direction along the horizontal sunset axis (axis pointing towards the sunset position). In our simple agent-based model, the simulated mosquitoes used the angular view of the marker and the sunset direction as sensory cues for attraction and alignment, respectively. This resulted in very similar simulated flight dynamics than in real swarms (Fig 5 and [16,22,23,32,37,38,44]). In nature, mosquitoes might use additional environmental cues than in our model, or might even use different ones [29]. For example, they might show different or even no directional preferences in function of available environmental cues, as suggested by the weaker and more variable directional bias observed in the Feugère dataset compared to the Poda dataset (Figs 2 and Fig M in S1 Appendix). Uncovering how exactly swarming mosquitoes use environmental sensory cues, such as the swarm marker and sunset light, will require experiments whereby the environmental cues are systematically manipulated. For example this could be done by varying marker shape, size and contrast with the environment, and will inform us on what makes some marker in the field more attractive than others. Still, our model shows that it is possible to generate swarming with realistic characteristics using simple minimally-interactive rules, and consequently the hypothesis that such minimally-interactive rules govern mating swarms should be preferred before assuming more complex rules. Out of the three identified behavioral rules, we found that only the collision avoidance rule required to be interactive among conspecifics. Thus, collision avoidance can be regarded as a social response, but these occur very rarely (Fig 5J), and only at very short distances between individuals (<2.5 body lengths; Fig 5K) [18]. In contrast, birds and fish display high level of direct interactions, at comparable distances [26]. This shows that insect mating swarming might be inherently different from flocking and schooling, and that the previously suggested apparent long-distance interactions between swarming mosquitoes or midges [16,22,23,44] might in fact not be interactions, but instead the result of simultaneous responses to external environmental cues. This does not show that other types of interactions between swarming mosquitoes are not present, but it suggests that such potential interactions are not directly needed to produce stable mating swarms. Testing the minimal behavioral rule model To test whether the identified behavioral rules capture the core dynamics of mosquito swarming, we developed an agent-based model implementing these rules, and compared simulation outputs in detail to the Poda dataset (Figs 5 and Figs C-G in S1 Appendix) [32]. We then used the Feugère dataset for a broader comparison of spatial and directional patterns across a different laboratory condition (Figs L-O in S1 Appendix) [37]. In the model, simulated mosquitoes fly along straight trajectories while continuously sampling the marker viewing angle α to assess whether they are about to cross the swarm boundary [32]. When a boundary-crossing is predicted, they execute a saccadic turn back toward the marker centerline, with new azimuth and climb angles drawn from kinematic distributions measured in the Poda dataset (Fig B in S1 Appendix). Thus, the model treats the swarm boundary as a binary threshold (inside vs. outside), producing sharply defined spatial zones where saccadic maneuvers occur versus straight-flight regions (Figs 5C, 5D and Fig Dc, d, g, h in S1 Appendix). In the empirical datasets, these zones are less sharply delineated (Figs 5A-5B, and Figs Da,b,e,f, and La-d), likely due to individual heterogeneity and perceptual noise in boundary estimation. For model parsimony, we did not incorporate such variability. Despite this minimalism, the model reproduces key empirical features: high mosquito density near the swarm center, acceleration vectors directed toward the swarm’s central axis, a concentration of saccades at the periphery and straight flights in the core, as well as a dominant flight orientation along the sunset axis (Figs 3-4, and Figs C-G, M and O in S1 Appendix). Together, these parallels indicate that the simple interaction rules are sufficient to generate the principal characteristics of observed mosquito swarms. Moreover, analysis of the Poda dataset showed that the swarming dynamics were largely invariant across swarm sizes, with the number of individuals (1–26), falling within the ecologically relevant range for Anopheles mating swarms [49–51], and having no measurable impact on key flight-kinematics metrics (Fig H in S1 Appendix). This confirms earlier findings that swarming remains highly stereotyped independent of group size [32], and supports the observation that even a single mosquito can display swarming behavior in the absence of conspecifics [21,53]. The low densities of mosquito swarms likely explain this invariance. Even in our largest groups, nearest-neighbor distances averaged ~20 body lengths, far above the values typically observed for individuals in bird flocks (~3–7 body lengths) or fish schools (~1–2 body lengths) rarely come close enough for neighbor-mediated interactions. Studies indicate that larger field swarms consist not only of more individuals, but are also of larger volumes [30,34,54], thereby maintaining similarly low densities. Together, these observations explain the lack of measurable density effects on swarm structure and kinematics, and suggest that potential interactions among males contribute little to the initiation or execution of saccadic turns under natural swarming conditions. It is to be noted that even if current analyses show minimal male-male interactions, observed collision avoidances behavior may suggest that males exhibit territorial behaviors, typical in leks, potentially to increase their chance to encounter females and exhibit traits likely to be chosen by females [55,56]. In that case, some locations in the swarm might be more favorable for males. Such territorial effects would likely be easier to observe in very large swarms that can comprise thousands of mosquitoes [57], and thus would require dedicated studies. Similarly, further analysis of male-female interactions during swarming will be needed to better understand how the here-described behavioral rules affect male competitiveness and mating success. Conceptual and experimental limitations of our approach In this study, we used a bottom‑up process approach to test whether simple sensory–motor rules at the individual level are sufficient to generate the characteristic structure of mosquito swarms. This contrasts with top‑down approaches that infer effective collective forces or statistical correlations from group‑level data, including models of laboratory insect swarms that reproduce cohesion without specifying short‑range behavioral rules [44,58]. Recent statistical‑inference methods, such as that of Iacomelli et al. (2025), estimate interaction strengths from correlated trajectories in large laboratory swarms (>80 individuals) [59]. These correlation patterns suggest collective effects at high densities but do not resolve whether such effects are necessary for swarm formation more generally. Our bottom‑up model is therefore complementary to these established approaches: statistical and effective‑force frameworks quantify group‑level patterns [44,58,59], whereas our parsimonious model tests how these patterns can emerge from explicit and minimal individual‑level behavioral rules [43,57,58]. We based our computational analysis on the Poda and Feugère datasets, both recorded under controlled laboratory conditions. Laboratory swarms inevitably differ from those observed in nature, but detailed three‑dimensional kinematic recordings of natural swarms are not yet available. These two datasets therefore represent the highest‑quality empirical material currently accessible, even though their recording environments differ from the field, and our results will need to be confirmed in more realistic natural conditions. The main differences relative to natural swarms are the following: (1) In both datasets, dusk conditions were created using artificial light sources, and the videography relied on infrared illumination. This differs from natural sunset light, which is parallel and partially polarized. (2) Swarming occurred in finite flight arenas. The Poda dataset was collected in a relatively small enclosed arena, whereas the Feugère dataset used a larger and more open space. As a consequence, potential wall‑proximity effects are most relevant for the Poda recordings. (3) The number of released mosquitoes produced small to medium swarm sizes, ranging from single initiators to maximum group sizes of 26 and 6 individuals in the Poda and Feugère datasets, respectively. We addressed these limitations as far as possible using the available data and within the scope of a computational study. Below, we discuss the main limitations of our modeling approach in more detail. (1) In both the Poda and Feugère datasets, the preferred flight direction and the major axis of the swarm were along the sunset axis (y-axis), rather than with the geometry of the flight arenas. The recurrence of this directional structure across two independent datasets, recorded in arenas of different sizes and with different marker shapes, indicates that the alignment is robust to variation in the recording environment (Figs 1-3, and Figs L-O in S1 Appendix). However, the present data cannot establish whether sunset cues directly cause the observed asymmetry in swarm shape. Demonstrating causality will require experiments in which the properties or position of the sunset cue are systematically altered. Our results do show that when a consistent external directional cue is present, whether sunset light or another environmental gradient, mosquitoes align their straight‑flight segments along that axis. This alignment generates asymmetry in both individual flight dynamics and the emergent swarm structure. (1) In both the Poda and Feugère datasets, the preferred flight direction and the major axis of the swarm were along the sunset axis (y-axis), rather than with the geometry of the flight arenas. The recurrence of this directional structure across two independent datasets, recorded in arenas of different sizes and with different marker shapes, indicates that the alignment is robust to variation in the recording environment (Figs 1-3, and Figs L-O in S1 Appendix). However, the present data cannot establish whether sunset cues directly cause the observed asymmetry in swarm shape. Demonstrating causality will require experiments in which the properties or position of the sunset cue are systematically altered. Our results do show that when a consistent external directional cue is present, whether sunset light or another environmental gradient, mosquitoes align their straight‑flight segments along that axis. This alignment generates asymmetry in both individual flight dynamics and the emergent swarm structure. (2) To assess whether the walls of the Poda arena influenced turning behavior, we quantified the distance to the nearest wall at the onset of each saccadic turn (Fig K in S1 Appendix). Saccades at the swarm boundary occurred at approximately 75 body lengths from the side walls and 38 body lengths from the front and back walls. These distances are far greater than the short‑range interaction distances associated with collision‑avoidance behavior (~2.5 body lengths; Fig 5K) [18], and exceed reported wall‑interaction distances of one to a few body lengths [46,48]. As an independent comparison, the Feugère dataset, recorded in a much larger arena with a different marker geometry, showed similar swarm structure, turn dynamics, and directional tendencies (Figs L-O in S1 Appendix). This cross‑dataset consistency indicates that wall proximity is unlikely to be a major driver of the observed flight behavior. Natural swarms of anthropophilic Anopheles mosquitoes also often form near buildings or other vertical structures [57], so occasional proximity to walls is ecologically realistic. Taken together, these observations suggest that although wall effects cannot be entirely excluded, their influence on the behavioral patterns central to this study is likely limited. (2) To assess whether the walls of the Poda arena influenced turning behavior, we quantified the distance to the nearest wall at the onset of each saccadic turn (Fig K in S1 Appendix). Saccades at the swarm boundary occurred at approximately 75 body lengths from the side walls and 38 body lengths from the front and back walls. These distances are far greater than the short‑range interaction distances associated with collision‑avoidance behavior (~2.5 body lengths; Fig 5K) [18], and exceed reported wall‑interaction distances of one to a few body lengths [46,48]. As an independent comparison, the Feugère dataset, recorded in a much larger arena with a different marker geometry, showed similar swarm structure, turn dynamics, and directional tendencies (Figs L-O in S1 Appendix). This cross‑dataset consistency indicates that wall proximity is unlikely to be a major driver of the observed flight behavior. Natural swarms of anthropophilic Anopheles mosquitoes also often form near buildings or other vertical structures [57], so occasional proximity to walls is ecologically realistic. Taken together, these observations suggest that although wall effects cannot be entirely excluded, their influence on the behavioral patterns central to this study is likely limited. (3) We compared the swarm sizes in the two laboratory datasets with values reported from natural Anopheles mating swarms [49–51]. The laboratory swarms fall within the ecologically relevant range, where field surveys typically report modal and median swarm sizes of a few dozen males. Within this size range, our analysis of the Poda dataset showed that swarm kinematics were largely invariant: key metrics such as straight‑flight distance, turn angle, and turn frequency showed no measurable dependence on group size (Fig H in S1 Appendix). This agrees with earlier findings that swarming remains highly stereotyped across small to medium swarm sizes [32]. Field observations do, however, show a broad size distribution with a heavy upper tail, and swarms can occasionally reach very large sizes [49–51]. Our results should not be extrapolated to such extreme cases [59]. Future work that tracks individual flight trajectories in these larger, high‑density swarms would be valuable for testing how well the minimal behavioral rules identified here generalize to denser and more populous swarms. (3) We compared the swarm sizes in the two laboratory datasets with values reported from natural Anopheles mating swarms [49–51]. The laboratory swarms fall within the ecologically relevant range, where field surveys typically report modal and median swarm sizes of a few dozen males. Within this size range, our analysis of the Poda dataset showed that swarm kinematics were largely invariant: key metrics such as straight‑flight distance, turn angle, and turn frequency showed no measurable dependence on group size (Fig H in S1 Appendix). This agrees with earlier findings that swarming remains highly stereotyped across small to medium swarm sizes [32]. Field observations do, however, show a broad size distribution with a heavy upper tail, and swarms can occasionally reach very large sizes [49–51]. Our results should not be extrapolated to such extreme cases [59]. Future work that tracks individual flight trajectories in these larger, high‑density swarms would be valuable for testing how well the minimal behavioral rules identified here generalize to denser and more populous swarms. (4) In our study, we tracked individual swarming mosquitoes but without maintaining unique identity over extended periods, as overlapping trajectories might cause tracks to be split. As a result, we could not reliably quantify individual residence time within swarms, and track duration may be underestimated in denser groups due to tracking limitations rather than biological effects. This prevents a controlled assessment of how individual residence time or swarm persistence depend on swarm size. Consequently, our analysis is restricted to instantaneous flight kinematics and does not address persistence-related dynamics. This limitation is unlikely to affect the identification of the behavioral rules proposed here, which are based on instantaneous flight kinematics. Future work using more advanced tracking approaches that maintain long-term identity would allow direct assessment of persistence-related dynamics and their relationship to swarm kinematics. (4) In our study, we tracked individual swarming mosquitoes but without maintaining unique identity over extended periods, as overlapping trajectories might cause tracks to be split. As a result, we could not reliably quantify individual residence time within swarms, and track duration may be underestimated in denser groups due to tracking limitations rather than biological effects. This prevents a controlled assessment of how individual residence time or swarm persistence depend on swarm size. Consequently, our analysis is restricted to instantaneous flight kinematics and does not address persistence-related dynamics. This limitation is unlikely to affect the identification of the behavioral rules proposed here, which are based on instantaneous flight kinematics. Future work using more advanced tracking approaches that maintain long-term identity would allow direct assessment of persistence-related dynamics and their relationship to swarm kinematics. Overall, the swarm dynamics observed in our analyses are consistent with field descriptions of Anopheles mating swarms [54], but important limitations remain. Addressing these limitations will require dedicated laboratory, semi‑field and field experiments that quantify and manipulate key environmental variables, including light polarization, environmental clutter, and swarm size and composition. This would allow refining of the agent-based model, and testing whether additional behavioral rules are needed under more complex natural conditions. Mating swarms in nature It is plausible that many other insect species follow similar rules in mating swarms, although this requires further quantitative validation in a comparative framework (e.g. swaRmverse [60]). For example, An. gambiae mosquitoes have been observed to swarm near ground markers along an east-west axis [61], suggesting a comparable sunset-oriented alignment. In contrast, Culex mosquitoes display predominantly vertical movements during swarming [21], likely reflecting the influence of other environmental cues. Our approach could be used to systematically test such interspecific differences, particularly across the many species of swarming Diptera. It could also be used to explore how swarming adapts to changing environmental conditions such as wind or light perturbations [33,52,62], potentially revealing additional behavioral rules. Our results do not address interactions between males and females within these mating swarms. Males are known to be strongly attracted to females entering the swarm [22,63], and it has been hypothesized that the high density of males in the swarm center allows them to quickly intercept females at the periphery [30]. Understanding the behavioral rules governing female interception is crucial for a complete understanding of dipteran mating swarms. Our finding that swarming males fly straight through the swarm center and perform sharp saccadic turns near its boundaries provides a basis for how males may optimize encounters with incoming females. Moreover, the combined experimental and modeling framework developed here offers a powerful foundation for future investigations of male–female interactions within swarms. Contrary to bird flocks and most fish schools, mosquito swarms act as leks in which males aggregate in a spatially fixed swarm that females visit to locate mates. These mating swarms also differ fundamentally from the collective behavior of eusocial insects such as honeybees, and from cooperative mass-migratory swarms seen in locusts [64,65]. In stationary mating swarms like those studied here, individuals are not required to achieve consensus decisions or cooperative task allocation as in eusocial groups, and their interactions can thus be governed by simple, local behavioral rules. Nevertheless, external perturbations, such as wind gusts or predator attacks, may still trigger coordination similar to the collective responses observed in social insects, birds, or fishes [15,16]. Importantly, eusocial insects represent only a tiny fraction of insect diversity across non-social taxa (e.g., Diptera, Ephemeroptera, Lepidoptera), likely exceeding the number of eusocial species by orders of magnitude [53]. Consequently, our results not only clarify the minimal interaction rules sufficient to generate mosquito swarming, but also provide a framework that could be applied to a large number of insect species that mate in swarms. This understanding also has direct implications for disease vector control. By identifying the behavioral rules and flight dynamics that give rise to realistic swarming, our results provide a basis for assessing how males locate, join, and maintain position within mating swarms. Such insights are particularly valuable for improving the quality and competitiveness of laboratory-reared males used in vector control programs. Reduced mating success of released males, whether sterile [63,66,67] or carrying gene drive [68,69], remains a major limitation to the efficacy of these approaches [70]. A deeper understanding of the flight characteristics, spatial positioning, and sensory cues that determine swarming behavior and mating success can guide the selection, rearing, or conditioning of males that perform more effectively in the field. More broadly, because mosquito swarms are spatially constrained, easily reproduced under laboratory conditions, and responsive to controlled perturbations, they represent a great experimental model for studying collective animal behavior in general and for translating such fundamental insights into applied solutions for vector-borne disease control. Supporting information List of all parameters used in our agent-based model. Here, we provide a complete list of parameters and their default values as used in the agent-based model. The model description follows the ODD protocol (Grimm et al., 2020) in a simplified form. The model code is implemented in Matlab 2022b and in supplementary S1 Code. Available in our Dryad repository [39]. (XLSX) Example of the swarming flight kinematics of a single initiator mosquito. Three-dimensional and two-dimensional views of the flight dynamic of the initiator male Anopheles coluzzii (i.e., the single first swarming mosquito), recorded in the lab above a black swarm marker of 40x40cm. This mosquito was flying in isolation. (MP4) Example of the flight kinematics of a single mosquito flying in a multi-individual swarm. Three-dimensional coordinates, speed acceleration and angular speed of a single swarming Anopheles coluzzii male mosquito recorded in the lab above a black swarm marker of 40x40cm. This mosquito was flying in a multi-individual swarm. (MP4) Example of the swarming flight kinematics of multiple mosquitoes. Three-dimensional and two-dimensional views of the flight kinematics of multiple male Anopheles coluzzii mosquitoes recorded in the lab above a black swarm marker of 40x40cm. (MP4) Example of the swarming flight kinematics of ten simulated mosquitoes. Three-dimensional and two-dimensional views of the flight kinematics of ten simulated male mosquitoes above a black swarm marker of 40x40cm, using an agent-based model and minimally interactive behavioral rules (with collision avoidance). (MP4) Database of the experiment. Flight tracks of Anopheles coluzzii male mosquitoes swarming above a swarm marker in sunset simulated laboratory conditions. A Matlab (all_data.mat) file containing three-dimensional tracks of all flying mosquitoes, described as the time t (s) and the spatial {x, y, z} (m) coordinates of the mosquito at each video frame. The coordinates are in meters, and in the world reference frame as defined in Fig 1, with the z-axis oriented vertically up, and the origin of the coordinate frame at the center of the swarm marker. The trajectories were determined as described in the materials and methods. Available in our Dryad repository [39]. (ZIP) Database of the simulation. Flight tracks of simulated male mosquitoes swarming above a swarm marker. Two Matlab (all_data.mat) files containing three-dimensional tracks of all flying simulated mosquitoes with or without collision avoidance rule, described as the time t (s) and the spatial {x, y, z} (m) coordinates of the mosquito at each video frame. The coordinates are in meters, and in the world reference frame as defined in Fig 1, with the z-axis oriented vertically up, and the origin of the coordinate frame at the center of the swarm marker. The trajectories were determined as described in the materials and methods. Available in our Dryad repository [39]. (ZIP) The complex swarming dynamics of malaria mosquitoes emerge from simple minimally-interactive behavioral rules. (DOCX) Matlab analysis codes. Contains all original Matlab codes that were written to perform the analysis of the article and to generate the figures. Available in our Dryad repository [39]. (MD) 21 Sep 2025 The complex swarming dynamics of malaria mosquitoes emerges from simple minimally-interactive behavioral rules PLOS Computational Biology Dear Dr. Cribellier, Thank you for submitting your manuscript to PLOS Computational Biology. After careful consideration, we feel that it has merit but does not fully meet PLOS Computational Biology's publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript within 60 days Nov 21 2025 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at ploscompbiol@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pcompbiol/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: * A rebuttal letter that responds to each point raised by the editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. This file does not need to include responses to formatting updates and technical items listed in the 'Journal Requirements' section below. * A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. * An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, competing interests statement, or data availability statement, please make these updates within the submission form at the time of resubmission. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter We look forward to receiving your revised manuscript. Kind regards, David Fisher Guest Editor PLOS Computational Biology Tobias Bollenbach Section Editor PLOS Computational Biology Additional Editor Comments : Three expert reviewers have carefully considered your manuscript and returned a range of useful comments and suggestions. Some of these are quite major and directly relate to the interpretation of the experimental data and so what inference we can make from your findings. Since most reviewers agreed the work was potentially of interest, I am willing to suggest “Major revisions”, but the reviewers’ comments must be dealt with through substantial revisions before the MS will be sent back out for review. In particular, please ensure you: Be clearer about the source of the data, and when figures are modified versions of elsewhere. Address concerns about the comparison between leks and other types of collective motion, as it impacts on generality of findings. Re-order to have Methods following Introduction, and be clear on sample sizes throughout. Consider the other interpretations for your findings the reviewers suggest, and in particular deal with concerns of reviewer #2 that your results might be an artefact of your design. Failure to robustly address these and other comments will lead to a desk rejection. Journal Requirements: 1) We ask that a manuscript source file is provided at Revision. Please upload your manuscript file as a .doc, .docx, .rtf or .tex. If you are providing a .tex file, please upload it under the item type u2018LaTeX Source Fileu2019 and leave your .pdf version as the item type u2018Manuscriptu2019. 2) Please provide an Author Summary. This should appear in your manuscript between the Abstract (if applicable) and the Introduction, and should be 150-200 words long. The aim should be to make your findings accessible to a wide audience that includes both scientists and non-scientists. Sample summaries can be found on our website under Submission Guidelines: https://journals.plos.org/ploscompbiol/s/submission-guidelines#loc-parts-of-a-submission 3) Your manuscript is missing the following sections: Introduction, and Methods. Please ensure all required sections are present and in the correct order. Make sure section heading levels are clearly indicated in the manuscript text, and limit sub-sections to 3 heading levels. An outline of the required sections can be consulted in our submission guidelines here: https://journals.plos.org/ploscompbiol/s/submission-guidelines#loc-parts-of-a-submission 4) Please upload all main figures as separate Figure files in .tif or .eps format. For more information about how to convert and format your figure files please see our guidelines: https://journals.plos.org/ploscompbiol/s/figures 5) We have noticed that you have a list of Supporting Information legends for (Supplementary Movies, Data S1, Data S2, and Code S1) in your manuscript. However, there are no corresponding files uploaded to the submission. Please upload them as separate files with the item type 'Supporting Information'. 6) Please amend your detailed Financial Disclosure statement. This is published with the article. It must therefore be completed in full sentences and contain the exact wording you wish to be published. 1) State what role the funders took in the study. If the funders had no role in your study, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." 7) Please ensure that the funders and grant numbers match between the Financial Disclosure field and the Funding Information tab in your submission form. Note that the funders must be provided in the same order in both places as well. Currently, "the Institut de Recherche our le Développement and the government of Burkina Faso" are missing from the Financial Disclosure field. Note: If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. Reviewers' comments: Reviewer's Responses to Questions Comments to the Authors: Please note that one review is uploaded as an attachment. Reviewer #1: The review is uploaded as an attachment. Reviewer #2: I am afraid the manuscript is not ready to be published in Plos Computational Biology. I do not recommend the publication. My concerns are mostly on the interpretation of the experimental data, which I believe are overinterpreted, and on the conclusions the authors draw from the comparison with the boids model. DATA INTERPRETATION Two main features of mosquito behavior are considered crucial in the manuscript. The first is the preferred direction of flight, orthogonal to the sunset horizon. The second is that mosquitoes perform saccades at the border of the swarm, due to a visual interaction (defined by a specific visual range) with the marker. My concern is that both features may be artifacts of the experimental setup. The direction orthogonal to the sunset horizon corresponds to the shortest side of the cage. Therefore, an alternative interpretation of the data, different from that of the authors, is that mosquitoes are not completely free to move in that direction and that the cage is affecting the behavior. For instance mosquitoes may be induced to perform fast 180deg turns (saccades), to avoid the wall and redirect towards the central, open area of the cage. With a similar reasoning, mosquitoes at the borders of the swarm may be induced to perform saccades, not because of a visual interaction with the marker, but because of a physical constraint of the cage itself. To determine which of the two alternatives (the one proposed by the authors and the one described above) is the more realistic, the direction orthogonal to the sunset horizon must be separated from the shortest side of the cage, and the swarm boundaries must be separated from the cage boundaries. These variables cannot be decoupled in the data presented here. Therefore, authors need to perform additional experiments: 1- with the sunset in a different position/direction to test whether the preferred direction is related to the sunset; 2- with different markers (shape and size) to validate the proposed visual interaction. BOIDS MODELS To the best of my knowledge, the dominant term in boids models is the alignment interaction, as in other interacting models, like Vicsek’s. Repulsion and attraction are essentially needed to have realistic simulations and applications in robotics. The low relevance of repulsion is indeed in line with what the authors find in their data, where repulsion events are rare. In regard to the attraction term, the situation is more subtle. Boids model considers attraction towards the average position of the neighbors. In large systems, this means that particles in the bulk will not feel the effect of the attraction. The attraction is instead dominant for particles at the periphery, and attraction is actually needed to avoid a loss of cohesion. In the simulation presented in the manuscript, all the particles are essentially at the periphery of the swarm, since there are only 10 particles in the group. As a result, they are all influenced by the attraction toward the average position of the other ones, which due to the limited setup, always coincides with the position of the landmark. It is then not possible to discriminate between the presence of an individual attraction toward the marker, as suggested by the authors, or an attraction toward the other members of the group. This ambiguity should be discussed within the manuscript. MINOR COMMENTS 1- The number of mosquitoes in the analyzed swarms should be clearly stated. At the beginning of the Experimental section, authors mention that they reconstructed 938 tracks. In SI, section “Experimental conditions”, they describe their experimental procedure and refer to the number of mosquitoes released in the cage (30 or 50 depending on the experiment). It is only in the description of the simulations, SI Section “Agent-based model simulations”, that they finally state that the data in the manuscript refer to swarms of approximately ten mosquitoes. It would be better to clarify from the beginning of the manuscript that the data are 938 reconstructed tracks, corresponding to 18 swarms of approximately ten mosquitoes. 2- The discussion section appears too general and superficial. For instance, the last sentence on vector control should be removed or expanded. In principle any study on mosquitoes may be considered useful for vector control. The authors need to be specific on how their findings could be useful for applications in vector control. Another example in this direction is the paragraph at line 10 pag.7 (second last paragraph) of the discussion section, about male-female and male-male interaction. This paragraph does not add anything to the overall message of this work, and should be removed or revised. 3- Authors do not address potential effects of the laboratory environment on mosquito behavior. 4- Rules defining the models should be at least mentioned in the main text. Otherwise the reader is forced to go to SI. 5- The comparison between model and data shows that simulations produce saccades symmetric on the plane xy and yz, against a clear asymmetry in the data. The authors should comment on this discrepancy. The same symmetry/asymmetry is present in the comparison of the straight flight. Reviewer #3: This work describes the swarming of malaria mosquitoes based on individual 3D tracks in the lab and investigates the underlying behaviour rules of their swarms using an agent-based model. I find the paper clear and concise, and a very nice fit to the journal. It is well-written and with nice visualization of results. I have however a few remarks and questions that I think should be addressed before the paper’s publication, and some comments that I think can increase the paper’s contribution and impact. Please find them below: - Why there are no methods in the main text? Some experimental conditions should be mentioned in the results text, especially how many mosquitoes were tracked and for what duration. Same for some main aspects of the simulations. - I understand this paper is an extension of the Poda et al. 2024 work. Given the overlap, I think that the authors should more clearly state (perhaps even in the introduction/results) and discuss in the end the differences/extensions in empirical (especially for sections ‘Flight kinematics of individual swarming mosquitoes’ and ‘Variations in flight kinematics within the swarm volume) and modelling findings (since a different model was used in the previous study). - Fig. 1A is exactly the same as in ref [37], Poda et al. 2024, without a reference in the caption. Not sure what’s the journal’s policy is on this. - The comparison with other species could be made a bit more specific, e.g., pg.7 line 5, what different environmental cues? Are there any relevant findings from other swarming insects? - I find the number 4.5% reduction of the viewing angle as the trigger to a turning manoeuvre a bit too specific, perhaps the authors could elaborate more in their Discussion about it. - Model description: As I mentioned already above, some more details about the simulations should be added in the main text. A table with all the parameters and their default values should also be added in the Supplementary. Additionally, despite the model description being clear, I would recommend that the authors could at least use more terminology from the ODD protocols (to not write a full one) [1]. This will support both the reproducibility of the agent-based model as well as strengthen its link to the literature. - Pg 7 line 19: ‘mosquito swarms as an ideal system for studying animal collective behavior’ I found this opposing the conclusion that mosquito swarms are ‘inherently different from flocking and schooling’. Perhaps the authors meant collective behaviour of insect swarms? Saying that, a paragraph on comparisons with findings in other insect swarms would be nice addition to the Discussion. - Maximum swarm size from Fig S7 seems to be 26, is this the maximum over the experiments or just the one used for the plot? In general, are the swarm sizes here reflecting the ones found in the field? Perhaps the authors could expand a bit on their discussion pg 6 line 44. - Fig. S4 f and h: there seem to be larger differences between empirical data and simulations here, perhaps the authors can provide some clarification of why this is the case in the caption or text. - Pg. 6, line 39: ‘previously suggested apparent long distance interactions’ I think that this statement needs references. Minor comments: - pg. 3 line 40: unnecessary second ‘and’ ? - pg. 4 line 19: ‘remain’ - Ref 37: bioRxiv references instead of the published version - title: ‘emerge’ rather than ‘emerges’ if it refers to the dynamics and not the swarming. References: [1] Grimm, Volker, Steven F. Railsback, Christian E. Vincenot, Uta Berger, Cara Gallagher, Donald L. DeAngelis, Bruce Edmonds et al. "The ODD protocol for describing agent-based and other simulation models: A second update to improve clarity, replication, and structural realism." Journal of Artificial Societies and Social Simulation 23, no. 2 (2020). ********** Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: No: Unless I have missed, the code should be available via Code S1 and S1 was not provided. Reviewer #2: Yes Reviewer #3: Yes ********** PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No Reviewer #3: No [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] Figure resubmission: While revising your submission, we strongly recommend that you use PLOS’s NAAS tool (https://ngplosjournals.pagemajik.ai/artanalysis) to test your figure files. NAAS can convert your figure files to the TIFF file type and meet basic requirements (such as print size, resolution), or provide you with a report on issues that do not meet our requirements and that NAAS cannot fix. After uploading your figures to PLOS’s NAAS tool - https://ngplosjournals.pagemajik.ai/artanalysis, NAAS will process the files provided and display the results in the "Uploaded Files" section of the page as the processing is complete. If the uploaded figures meet our requirements (or NAAS is able to fix the files to meet our requirements), the figure will be marked as "fixed" above. If NAAS is unable to fix the files, a red "failed" label will appear above. When NAAS has confirmed that the figure files meet our requirements, please download the file via the download option, and include these NAAS processed figure files when submitting your revised manuscript. Reproducibility: ?> Submitted filename: Review.pdf 21 Nov 2025 Submitted filename: Response to reviewers.pdf 19 Dec 2025 PCOMPBIOL-D-25-00797R1 The complex swarming dynamics of malaria mosquitoes emerge from simple minimally-interactive behavioral rules PLOS Computational Biology Dear Dr. Cribellier, Thank you for submitting your manuscript to PLOS Computational Biology. As with all papers, your manuscript was reviewed by members of the editorial board. Based on our assessment, we have decided that the work does not meet our criteria for publication and will therefore be rejected. If external reviews were secured, reviewers' comments will be included at the bottom of this email. We are sorry that we cannot be more positive on this occasion. We very much appreciate your wish to present your work in one of PLOS's Open Access publications. Thank you for your support, and we hope that you will consider PLOS Computational Biology for other submissions in the future. Yours sincerely, Tobias Bollenbach Section Editor PLOS Computational Biology Additional Editor Comments (if provided): While two of the reviewers are now mostly satisfied with the revisions, reviewer #2's major concerns remain largely unresolved. Here, we tend to side with reviewer #2. In particular, we agree that: - Artifacts caused by the experimental setup (e.g., cage geometry) cannot be ruled out based on the presented data. Proper controls for this would be essential. Ultimately, the conclusions based on the current data are too strong. - The dynamic range of swarm sizes in the experiments is insufficient to support the current conclusions. Resolving these issues will require either additional experimental evidence or drastic changes to the main conclusions (see the specific suggestions of reviewer #2). The latter option would require changes that go beyond what can be done in a major revision and would fundamentally alter the manuscript. If you believe you can address these concerns, we are willing to consider a substantially revised manuscript as a new submission. [Note: HTML markup is below. Please do not edit.] Reviewers' Comments (if peer reviewed): Reviewer's Responses to Questions Comments to the Authors: Please note here if the review is uploaded as an attachment. Reviewer #1: I thank the authors for substantially improving the manuscript. The revisions clarify the conceptual framework and highlight the manuscript’s contributions to the literature much better. I find the addition of individual-trajectory comparisons highly useful in demonstrating that the swarming behavior is scale-free. While I do not entirely agree that the absence of group-size effects conclusively indicates little or no contribution of social cues, the authors are entitled to argue for parsimony (i.e., that environmental cues alone drive this behavior) given the breadth of analysis presented. Echoing another reviewer’s comments, I believe further work is needed to elucidate the precise relationship between environmental factors and swarming dynamics. I therefore look forward to the authors’ future contributions on this topic. Reviewer #2: Without additional experiments, the conclusions of this work, particularly the claim of mosquito alignment with the sunset direction, are not sufficiently supported by the data. Therefore, I do not recommend publication of this work in its current version. Below, I provide comments on the authors’ responses. To clearly distinguish my remarks from those of the authors, each paragraph is prefaced with “Reviewer” or “Authors”, respectively. Authors: Reply 2.2: We thank the reviewer for raising this important point. We agree that the observed flight direction and saccadic maneuvers could, in principle, be influenced by the physical constraints of the flight arena. To address this, we performed a dedicated analysis quantifying the proximity of mosquitoes to the arena walls during saccadic turns (see new Fig. S11). Reviewer: The distance from the wall at which saccades occur is indeed an interesting quantity; however, the analysis presented here cannot objectively exclude the possibility that the cage interferes with mosquito flight dynamics. In Figure S11, authors show the distributions of these distances along the two axes x and y, and in the discussion they focus on mean values. I see two weaknesses in this approach. First, the most informative part of these two distributions is not the mean but the left tail, which measures the shortest distances from the wall. Second, as discussed in detail below, the comparison with the distance of collision avoidance is not motivated to assess whether there is an influence of walls on mosquitoes dynamics. The two histograms shown in Fig.S11 show that the distributions are quite different between the two axes. Along the x-axis the left tail indicates that mosquitoes do not get closer than 10/15cm to the wall. Along the y-axis, instead, saccades are found at very short distances from the walls, where it is unlikely that the effects of confinement are negligible. A substantial fraction of saccades is also found in the region (highlighted in red) that defines the distance of the swarm from the swarming arena. These histograms would be conclusive only if the left tails were comparable along both axes, indicating similar wall-related behavior in the x and y directions. We cannot tell from this data, whether this is a genuine effect of the alignment with the light, or instead it is the effect of the confinement. The only way to discriminate between these two potential interpretations is to perform experiments changing the position of the light source and checking swarming behavior under these different conditions. Authors: Our results show that mosquitoes turned at average distances of 45 ± 8.5 cm from the side walls and 23 ± 7.1 cm from the front/back walls—equivalent to 75 and 38 body lengths, respectively. These distances are an order of magnitude greater than the inter-individual distances that trigger collision avoidance (~2.5 body lengths; Fig. 5K, [8]). Reviewer: The comparison between the distance from the walls at which saccades occur and the distance that triggers collision avoidance with other insects is not motivated. Such a comparison seems to imply that mosquitoes avoid flying toward a wall using the same mechanism involved in collision avoidance with other insects, which is unlikely. A wall imposes a hard constraint on the mosquitoes’ flight phase space, whereas the presence of another nearby insect does not. This argument would have been compelling if the authors could objectively prove that the presence of a wall would influence mosquitoes dynamics in the same way as another insect flying nearby, which is not shown here. Otherwise, the comparison remains quantitative but strongly speculative. Authors: Additionally, the preferred flight direction—orthogonal to the sunset horizon—aligns with the short axis of the arena (y-axis), which is counterintuitive if wall effects were dominant. If cage geometry were driving the behavior, we would expect alignment along the long axis (x-axis), which is not observed. Reviewer: This argument is too speculative and not supported by the data. Intuitiveness or counterintuitiveness cannot be used to assess the validity of an interpretation, particularly in biology. Therefore, I invite the authors to either revise or remove this part of the manuscript, or to support this conjecture with appropriate references from the literature. Authors: Together, these findings strongly suggest that wall avoidance does not explain the observed saccadic maneuvers or flight orientation. Instead, the data support our interpretation that these behaviors are visually mediated interactions with the marker. We acknowledge the reviewer’s suggestion for further experiments to disentangle these effects. While such tests (e.g., varying sunset direction or marker shape) would be valuable for refining our understanding of swarm geometry, they fall outside the main scope of this study. Our primary aim is to identify the behavioral rules underlying mosquito swarming, rather than to characterize the exact shape or orientation of the swarm. Moreover, our newly-added analysis based on current data already provide strong evidence that the observed behaviors are not artifacts of the experimental setup. We have added this new analysis and expanded the discussion accordingly (lines 420-426 and 564-570; Fig. S11). Reviewer: I do not think that further experiments will be relevant to refining the understanding of swarm geometry. They are necessary to confirm that the preferred direction is related to the light, which is one of the behavioral rules highlighted in the manuscript, and to actually prove that the same behavior is observed under a different setup. If this is not the case, i.e, if authors believe that the fact that the preferred direction of flight is relevant as a result of their study but only saccades are relevant, they should revise the manuscript and remove the claims about the direction of light from the abstract as well as from the result section, leaving it at a pure speculative level in the discussion. Authors: Reply 2.3: We appreciate the reviewer’s insightful comment. Indeed, distinguishing between attraction to conspecifics and attraction to an environmental marker is challenging in collective behavior studies. Two key findings support that the marker is the primary swarm attractor: 1. Stable swarm position: Swarms consistently remained above the marker for extended periods. If attraction were primarily to conspecifics, we would expect positional drift over time—especially given that our chamber design allows for such movement. This drift was never observed. 2. Swarming by single individuals: We recorded flight kinematics from the onset of swarming (single mosquito) to fully developed swarms (up to 26 individuals; Fig. S8). Even solitary males exhibited sustained swarming above the marker, with kinematics comparable to those in larger swarms (see Reply 1.5 for details). This suggests that attraction to the marker is independent of swarm size. Based on these results, attraction to the marker remains the most parsimonious explanation for swarm cohesion. To test this, we developed an agent-based model using the marker as the sole attractor, with collision avoidance as the only conspecific interaction. This minimal model successfully reproduced swarm kinematics across group sizes, indicating that explicit attraction between individuals is not required. While we cannot entirely exclude conspecific attraction, our data show it is not necessary to explain mosquito swarming. We have updated the manuscript to include these analyses and expanded the discussion on the model’s assumptions and limitations (lines 429-437 and 537-549) Reviewer: I agree that the attraction to the marker is the most parsimonious choice. However, I would not use the argument of a single mosquito flying on the marker to assess its validity. The crucial role of environmental markers in insect swarms is already well established in the literature, and inter-individual interactions should be viewed not as an alternative to marker attraction, but as an additional ingredient. The absence of positional drift, or the ability of a single individual to remain above the marker, does not discriminate between an underlying interaction driving swarming behavior and a response to external cues; by this logic, the present work, as well as much of the existing swarming literature, would not be needed. Despite this remark, if the authors wanted only to show that stable swarms can be reproduced with very simple rules, I think the manuscript would benefit from a clearer comparison with existing models. For example, models developed in Ouellette’s lab are able to reproduce stable swarms without explicitly incorporating inter-individual interactions. In this context, it is not clear in what sense the present model represents an advance over the current state of the art. Authors: Reply 2.4: As requested by the reviewer, we now clarify in detail the number of swarming experiments performed (six swarms with three recordings per swarm), and the number of swarming mosquitoes per swarm (from 1 to 26 individuals) thorough the manuscript. Moreover, we now explicitly state the sample sizes used for each figure in its figure legend. We also discussed how the number of swarming mosquitoes did not seem to affect their swarming behavior (lines 537-549). Reviewer: I thank the authors for adding details on the number of swarming experiments. However, I would use much more caution, all over the manuscript, in claiming that the swarm size does not seem to affect swarming behavior, since the data presented here refer to a quite limited size (from 1 to 26 mosquitoes), in particular when compared to field swarms, which may be composed of thousands of individuals, as shown in [1]. In this regard, the section “Variations in flight kinematics with swarm size” (lines 419-437) should be revised or removed. [1] Ouédraogo TFX, Sawadogo SP, Millogo AA, Niang A, Ouedraogo J, Ouattara SB, Cribellier A, Namountougou M, Dabiré RK, Muijres FT, Diabaté A. Characterization of factors influencing swarm dynamics and mating efficiency in Anopheles coluzzii. Parasit Vectors. 2025 Nov 27. doi: 10.1186/s13071-025-07151-w. Epub ahead of print. PMID: 41310880. Authors: Reply 2.5: As requested by the reviewer, we extended the discussion section in the revised manuscript. Hereby, we including a comparison of the studied mosquito swarming with other group behaviors (lines 599-612), we revising the paragraph discussing how our study informs us on male-female interactions in mosquito mating swarms (lines 590-598), and we expended the part of the discussion about vector control (lines 613-625). We hope that the reviewer is satisfied with the extended discussion section. Reviewer: In general, I am not particularly in favor of general discussions that are not directly related to the results. I appreciate the authors' effort to broaden the discussion of the relevance of their work, but I do not consider it necessary, nor do I think it adds anything substantial to the main message of the manuscript. The relevance of this study seems primarily related to fundamental biology, which is as valuable as applied biology, and I would prefer a discussion oriented at fundamental biology. Authors: Reply 2.6: As mentioned in our Reply 2.2, we have added a dedicated analysis addressing potential effects of the laboratory environments on the recorded mosquito behaviors. We discuss this in the methods, results and discussion sections of the revised manuscript (lines 420-426 and 561-570; see also Fig. S11). Reviewer: As already discussed above, I do not think that this analysis is robust enough and does not prove that mosquitoes behavior is not affected by the experimental setup. Authors: Reply 2.5: We have now included the material and methods section in the main text, thus resolving this issue. Reviewer: I thank the authors for this change. Authors: Reply 2.6: We believe that the reviewer refers to results presented in Fig. S4. In that figure, we compare the 2D heat maps of mean angular speed or of saccade-to-straight flight ratio in the XY and YZ plans. When comparing the heatmaps from the experimental data and from simulated data, we see that both real and simulated mosquitoes have on average higher angular speeds on the boundaries of the swarm, thus showing high proportion of saccadic maneuvers at the edges, and high proportion of straight flights in the middle of the swarm. These 2D distributions of both experimental and simulated data follow a clear axis symmetry around the central vertical axis of the swarm (Fig. S4). Therefore, it is unclear to us which asymmetry the reviewer refers to. There are, however, key differences between the distributions of the experimental data and simulated data. Most importantly, in the simulations we observe high angular speed and high proportion of saccades around the swarm boundaries (Fig. S4c,d and Fi. S4g,h). This is due to a simplification of our agent-based model where mosquitoes will only trigger a saccade if they are going to exactly cross the swarm boundaries (i.e. binary threshold), while in reality they can trigger a saccade a bit before or after crossing the boundary. This results for the experimental swarms in more diffused distributions of saccades in space (Fig. S4c,d and Fi. S4g,h) and less defined swarm boundaries (as visible on Fig. 5C,D). We have now included a comparison of the experimental and simulation results in the discussion (lines 520-536 and 571-578), highlighting these similarities and differences Reviewer: I apologize for not having specified that I was referring to Figure S4, as correctly guessed by the authors. In Figure S4 (panels b and d), the authors show the distributions on the xy-plane of the angular speed and of the saccade-to-straight flight ratio. In both plots, I see an asymmetry between x and y directions, which does not seem to be in agreement with the symmetry around the central vertical axis claimed by the authors. As stated by the authors, both real and simulated mosquitoes have on average higher angular speeds on the boundaries of the swarm, thus showing a high proportion of saccadic maneuvers at the edges, and high proportion of straight flights in the middle of the swarm. However, the experimental distributions also show a clear axis dependence: at the edges of the swarm along the y-direction, mosquitoes reach higher angular speed than those at the edges of the swarm along the x-direction. Similarly, the saccade-to-straight flight ratio measured at the swarm boundaries exhibits a directional dependence. This asymmetry is not present in the simulations. This discrepancy is not discussed in the manuscript, and I do not think it may be due to the binary threshold for the boundaries used in the simulations. Reviewer #3: The authors did a great job addressing my comments. The updated methods are indeed much clearer and the model description enables model reproducibility. The data and code repository is also well structured and easily accessible. I am happy to recommend this article for publication but please find some minor comments on 3 points below: - I like the additional analysis and text regarding swarm size effects. I was wondering whether density effect can start playing a role in larger aggregations, where there may be more competition for instance for positions closer to the center of the attractor and females. The authors nicely discuss this in lines 537-549 that usually the swarms don’t reach higher densities in nature since the volume of the swarm usually also increases, but if there is any additional ecological insight in the literature (on what the males compete for or on female preference [eg. 1]) could be a nice addition here to solidify the argument (or a mention that this gap of research could be addressed with more data in the future). For instance, under the prism of collective behavior [2], why would interactions among males not matter on their mating success? Minor: Lines 436 (‘confirms previous suggestions that a mosquito mating swarm of a single individual can exist [21, 48]’) and 541( ‘supports the observation that even a single mosquito can display swarming behavior in the absence of conspecifics [21,48]’ ) are rather overlapping, I think the first one in the results is unnecessary and can be omitted (also because a single individual shouldn’t be called a swarm, but rather showing swarming behavior as the authors write later). [1] Ng'habi, K. R., Huho, B. J., Nkwengulila, G., Killeen, G. F., Knols, B. G., & Ferguson, H. M. (2008). Sexual selection in mosquito swarms: may the best man lose?. Animal Behaviour, 76(1), 105-112. https://doi.org/10.1016/j.anbehav.2008.01.01 [2] Rathore, A., Isvaran, K., & Guttal, V. (2023). Lekking as collective behaviour. Philosophical Transactions of the Royal Society B, 378(1874), 20220066. https://doi.org/10.1098/rstb.2022.0066 - Comparisons with other species: I find that the new discussion paragraph nicely puts the results in a wider context (lines 599-612). I only find unclear what the authors mean by framework (line 611) here (a collection of metrics? Or the behavioral algorithm of the model?). Perhaps they could add a clarification. A nice link here could also be with a recently developed framework to compare collective motion across species (‘swarm verse’ [2]) that has been focused only on vertebrates. Mating swarms of insects seems like an ideal space to adjust this framework to, expanding towards invertebrates. [3] Papadopoulou, Marina, Simon Garnier, and Andrew J. King. "swaRmverse: An R package for the comparative analysis of collective motion." Methods in Ecology and Evolution 16.1 (2025): 29-39. https://doi.org/10.1111/2041-210X.14460 - In the author summary, the sentence: ‘how male malaria mosquitoes form mating swarms’ may make readers think about long range attraction to environmental cues and one another to ‘initiate’ a swarm rather than how the swarming characteristics arise. Perhaps the authors should revise/clarify this. ********** Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes ********** PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No Reviewer #3: No [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] 18 Mar 2026 Submitted filename: Response to reviewers 2.pdf 7 May 2026 The complex swarming dynamics of malaria mosquitoes emerge from simple minimally-interactive behavioral rules PLOS Computational Biology Dear Dr. Cribellier, Thank you for submitting your manuscript to PLOS Computational Biology. After careful consideration, we feel that it has merit but does not fully meet PLOS Computational Biology's publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jul 07 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at ploscompbiol@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pcompbiol/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: * A letter that responds to each point raised by the editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. This file does not need to include responses to formatting updates and technical items listed in the 'Journal Requirements' section below. * A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. * An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. We look forward to receiving your revised manuscript. Kind regards, Tobias Bollenbach Section Editor PLOS Computational Biology Additional Editor Comments: The reviewers have raised several issues that still need to be addressed. In particular, please address Reviewer #2's concerns about the compatibility of the behavioral rules with the new dataset. Additionally, ensure that the conclusions are not overstated, particularly in the abstract and summary sections of the main text. Journal Requirements: 1) Please ensure that the funders and grant numbers match between the Financial Disclosure field and the Funding Information tab in your submission form. Note that the funders must be provided in the same order in both places as well. Currently, "Sectorplan Biology of the Dutch Ministry of Education, Culture and Science (OCW)" is missing from the Financial Disclosure field. Note: If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. Reviewers' comments: Reviewer's Responses to Questions Comments to the Authors: Please note that one review is uploaded as an attachment. Reviewer #1: . Reviewer #2: I really appreciate the effort made by the authors to expand the analyzed data and extend the discussion. However, I have some concerns regarding the compatibility of the Feugere male dataset with the model introduced in the present work. At this stage, I am unable to recommend publication of the manuscript. For a model to be meaningful, it must demonstrate predictive power on datasets not used to fit the parameters. Given the availability of multiple datasets, the authors could have followed a standard cross-validation procedure: using the Poda dataset to fit the model parameters, as they do, and then quantitatively assessing the extent to which the Faugere datasets reproduce the predicted behavior. This step is, in my view, not properly addressed and it lacks a quantitative analysis. More specifically, my concerns regarding the compatibility of the model with the Feugere male dataset involve both Behavioral rule 1 and Behavioral rule 2, as detailed below. Behavioral rule 1: Swarming mosquitoes preferentially fly straight through the swarm center, where the marker viewing angle is maximal. They then continue flying straight, until the marker viewing angle falls below a threshold value, which triggers a turning maneuver. Behavioral rule 1 implies that the swarm center corresponds to the marker center, and that there is symmetry in the structure of the swarm with respect to the marker position. Although the Feugere male dataset exhibits a structural symmetry around its own vertical axis, it is shifted in the y-axis towards the negative direction, see Fig.S12a-b. As a result, part of the swarm is consistently out of the marker, and mosquitoes do not appear to show a preference for the region where the viewing angle is maximized, namely the center of the marker. This discrepancy suggests that behavioral rule 1 fails to predict and reproduce the position of the Feugere male dataset, and therefore highlights the lack of generality of the model. Maximizing the viewing angle of the marker may therefore not be the mechanism underlying swarm formation and spatial structure. Behavioral rule 2. The turning maneuver consists of a rapid saccadic turn within the horizontal plane, with biases towards the swarm center and along the sunset axis (towards and away from the sunset position) [32]. After the turn, the mosquito continues to fly in a straight line (rule 1). In my opinion this is the most critical aspect. There is a lack of quantitative assessment of the existence of a preferred direction of flight, and authors rely only on a qualitative evaluation of the plots shown in Fig.S13g and Fig.S15f. In detail, my concerns are the following: 1- In Fig.15f, I do not see clear evidence of a preferred flight direction. The distribution is indeed radially symmetric, in contrast with Fig.S15h (female dataset) and Fig.S4 (Poda dataset) 2- I am struggling to understand how distributions in Fig.S13g are normalized. By definition, the integral of each distribution should be equal to 1. But the distributions in this panel seem to cover areas that are not comparable with each other. To be more clear, let’s focus on the dark red and dark blue distributions. The dark red has a well pronounced peak at -90deg and it is quite narrow (the radii of the distributions are very small except for a very narrow portion close to -90deg). The dark blue is instead much broader, covering the entire [-180deg:180deg] interval with a peak at -90deg. By definition of probability distribution, its peak at -90deg cannot have the same radius of the dark red, as instead it has in Fig.S13g: if both distributions are properly normalized, the broader distribution should exhibit a lower peak. This makes the area of the dark blue much larger than the area of the dark red, giving the misleading impression that the two peaks (the one of the dark red and the one of the dark blue) have similar statistical significance. Authors are probably scaling one of the two distributions. 3- The distributions in Fig.S13g for the Feugere male dataset are computed separately for y>0 and y<0, as done for all the other datasets. This choice would have been reasonable if the swarm were centered with the marker (see comment above). But in this specific situation y=0 does not correspond to any physically meaningful reference position. The distributions should instead be defined with respect to the actual swarm center. Choosing y=0 as a splitting point, likely causes the presence of a significant portion of the dark blue distribution in the region of positive angles, as well as the imbalance between the dark and light blue distributions. Also note that, because of the color code and of the transparency of the plot, the portion of the dark blue distribution in the positive angles region is barely visible, unless when zooming in the figure. Its visibility could be improved using different colors. 4- To better assess the presence of a preferred flight direction, it would be useful to include, for the Feugere male dataset, a figure showing the distribution of the azimuth angle between two saccades as a function of mosquito positions, analogous to Fig.3c-d and Fig.S5a-b for the Poda dataset. 5- It would also be important to introduce a quantitative metric to characterize the degree of alignment in the flight directions, rather than relying solely on qualitative interpretation of the distributions. Reviewer #3: I think the authors did a great job addressing my previous comments, I do not have any follow up. In relation to the larger discussion with the comments of Reviewer 2, I wanted to mention the following points: 1) I think the newly analysed published dataset is an excellent addition to the paper that indeed provides further support to the findings. 2) Comment 2.2: I agree with the authors’ arguments and I find the extended discussion clear and effective in communicating the response to the concerns of Reviewer 2. 2) Comment 2.3: Inter-individual distances are often used as a unit metric, so I don’t find it peculiar or confusing in the manuscript. I see that it may be misleading for an interdisciplinary audience and thus the clarification the authors added is a good improvement. 3) Comment 2.6: I agree with Reviewer 2 in relation to the markers and inter-individual attraction. I think a stronger clarification on the comparison with existing models would be beneficial, and I think the authors did a good job explaining this in their new discussion. As the authors mention, testing minimal individual-level rules necessary to reproduce collective patterns is an invaluable insight from such agent-based models with important theoretical contributions despite the possibility of more complex behaviours of individual organisms. 4) Concerning the wall effects overall, as the authors note at Reply 2.9, experimental setups may always have an effect on behaviour and usually cannot be ruled out, it is the cost of having a more standardised alternative to a study in the field. For instance, wall effects are present in most studies with laboratory experiments in fish schools. I think the authors did the best they could in testing these effects, and the new dataset further supports that their conclusions are not overstated. In my opinion, the new limitations section of the Discussion covers everything quite extensively. Overall, I think that more cautious interpretation is always a plus, especially when studying a complex multi-agent system (where causality is very hard to prove empirically) and when simulations are used, given other potential underlying rules that may be able to reproduce the same emergent properties. Thus, I find the revised manuscript highly improved and the claims of the authors now well balanced and justified. I am happy to recommend its publication. Minor remarks: - Lines 560, I think there is a typo in the first new sentence. Perhaps a reference is also needed in the first sentence of that paragraph. - There is some repetition in the new part of the discussion, eg. lines 645 vs lines 655. Perhaps the first short paragraph is not necessary and can be merged with the following 2 for conciseness. ********** Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes ********** PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No Reviewer #3: No [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] Figure resubmission: Reproducibility: ?> Submitted filename: Comments.pdf 1 Jun 2026 Submitted filename: Response to reviewers 3.pdf 4 Aug 2026 Dear Dr. Cribellier, We are pleased to inform you that your manuscript 'The complex swarming dynamics of malaria mosquitoes emerge from simple minimally-interactive behavioral rules' has been provisionally accepted for publication in PLOS Computational Biology. Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests. Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated. IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript. Should you, your institution's press office or the journal office choose to press release your paper, you will automatically be opted out of early publication. We ask that you notify us now if you or your institution is planning to press release the article. All press must be co-ordinated with PLOS. Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Computational Biology. Best regards, Tobias Bollenbach Section Editor PLOS Computational Biology *********************************************************** Reviewer's Responses to Questions Comments to the Authors: Please note here if the review is uploaded as an attachment. Reviewer #1: I am thoroughly satisfied with the current state of the paper and can only recommend it for publication. Reviewer #2: This manuscript is now at its fourth round of revision. Despite the new figure and the introduction of the Rayleigh test, I find that the data interpretation is still flawed and that there is an overall lack of scientific rigor. Therefore I cannot recommend this work for publication. My suggestion to the authors is to focus their analysis on the saccades, which represent a robust feature across the datasets, instead of attempting to define a specific behavioral model. Authors present data from three datasets: the original Poda dataset on males, and two datasets (one on males, one on females) from Feugere. They define a behavioral model, designed and fitted on the Poda dataset, based on three simple behavioral rules. My main concerns are about the incompatibility of the male Feugere data with the model and the authors’ justifications for this incompatibility, which in my opinion compromise the validity of the entire work. Behavioral rule 1 (BR1) concerns swarm structure, specifically involving mosquitoes’ viewing angle of the marker. The authors do not provide a quantitative measure to test whether, and to what extent, the Feugere datasets follow this rule. However, based on the figures they provide, the female swarm appears almost centered over the marker, hence in principle compatible with BR1. In contrast, the male swarm is shifted along the y direction, thereby showing a lack of compatibility with BR1. If the viewing angle were the dominant factor shaping the structure of the swarm, mosquitoes above the center of the marker would be in the ideal condition, with maximum viewing angle; therefore they should not exhibit a tendency to turn as they do, and the swarm should be centered on the marker. The authors justify this incompatibility in the male data by pointing to the different geometry of the marker (square and flat for Poda, elevated, cylindrical and asymmetric for Feugere). However, this justification is directly contradicted by the female data. The female behavior demonstrates that the Feugere setup does not physically or visually preclude the emergence of BR1-like swarm structures. Therefore, the failure of the male dataset cannot be attributed to the experimental geometry alone, but rather indicates that the model's behavioral assumptions lack the necessary robustness to generalize across datasets. Behavioral rule 2 (BR2) concerns the direction of flight related to saccadic turns, which is biased towards the sunset direction. Through circular probability distribution and the Rayleigh test, the authors evaluate at both a qualitative and quantitative level whether the Feugere data are compatible with BR2. The female distribution is quite similar to that of the Poda dataset, with two large peaks corresponding to the direction of the sunset and a pronounced minimum in the orthogonal direction. In contrast, after correcting the severe methodological error of scaling the plot, the male distribution is only slightly skewed in the direction of the sunset and lacks a pronounced minimum, indicating that all directions are almost equivalent. This visual observation is confirmed by the Rayleigh test, which yields r=0.371 for females and r=0.138 for males, expressing a moderate and weak alignment respectively. Compared with the r=0.426 of the Poda dataset, these values suggest that the female data are compatible with both Poda and BR2, while the male data are not. In this regard, the manuscript presents three major issues. First, depending on the context, the authors refer either to p-values or to r to assess the significance of the alignment they found. But these two parameters are inherently different. The p-value is a measure of the statistical significance of the test, and it is heavily dependent on the sample size; a low p-value can be completely driven by a large number of samples, even in the case of a negligible alignment effect. The actual measure of the alignment strength is given by r. The authors ultimately contradict themselves in one single sentence in the caption of Fig.S12, where they write “Both males and females horizontal flight directions show significant alignment with sunset axis (Rayleigh test on axial data, (p < 0.001, r = 0.138 and 0.371, respectively showing weak and moderate alignment)”. Moreover, going from r = 0.426 (Poda) to r = 0.138 (male Feugere data) we move from a highly structured alignment to near-total uniformity. Claiming that an r = 0.138 validates the model, even in the broad sense the authors refer to, is weak and scientifically not rigorous. Second, throughout the manuscript and their replies, the authors consistently treat the two Feugere datasets as equivalent, despite males and females exhibiting radically different behaviors. Third, the authors attempt to justify the weak alignment found in the male Feugere data by pointing to differences in the experimental setup and 3D marker geometry. Again, as for BR1, if these were truly the cause of the model's failure, they would have disrupted the alignment of both males and females, making both datasets incompatible with the Poda data and BR2. To justify the incompatibility of the male Feugere data with the model, the authors also use the argument of aiming for a parsimonious sufficiency model, rather than a predictive one. However, I find this defence conceptually weak. Even a sufficiency model must demonstrate that its minimal rules are robust enough to explain the phenomenon under investigation. If a biological factor has a dominant effect in one dataset and is negligible in another, it is not the driving mechanism of the phenomenon. Consequently, the model cannot be generalized to wider swarming behavior - such as the mating swarms in nature mentioned in lines 730–774 - and the model loses its biological relevance. Reviewer #3: I believe that the authors did a great job replying to all comments and revising the manuscript. I think it is a strong piece of work that perfectly fits in the journal. As the authors note, I believe that this bottom-up approach with a model based on empirical data is an important contribution to our mechanistic understanding of this system, without proving the predictive power of the model. If future empirical work provides a solid comparison between swarms across a range of experimental set ups in terms of marker characteristics, there will be good insights to then expand the model and gain a more universal understanding in future work. The revised manuscript is very clear concerning the conditions from which it draws conclusions and doesn't overstate its findings. ********** Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes ********** PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No Reviewer #3: No PCOMPBIOL-D-25-00797R3 The complex swarming dynamics of malaria mosquitoes emerge from simple minimally-interactive behavioral rules Dear Dr Cribellier, I am pleased to inform you that your manuscript has been formally accepted for publication in PLOS Computational Biology. Your manuscript is now with our production department and you will be notified of the publication date in due course. The corresponding author will soon be receiving a typeset proof for review, to ensure errors have not been introduced during production. Please review the PDF proof of your manuscript carefully, as this is the last chance to correct any errors. Please note that major changes, or those which affect the scientific understanding of the work, will likely cause delays to the publication date of your manuscript. Soon after your final files are uploaded, unless you have opted out, the early version of your manuscript will be published online. The date of the early version will be your article's publication date. The final article will be published to the same URL, and all versions of the paper will be accessible to readers. For Research, Software, and Methods articles, you will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing. Thank you again for supporting PLOS Computational Biology and open-access publishing. We are looking forward to publishing your work! With kind regards, Sharmila Kamatchi PLOS Computational Biology | Carlyle House, Carlyle Road, Cambridge CB4 3DN | United Kingdom ploscompbiol@plos.org | Phone +44 (0) 1223-442824 | ploscompbiol.org | @PLOSCompBiol