Both genotype and environment affect behaviour. However, even in an experimentally controlled environment, a genotype also interacts with the inevitable environmental stochasticity – temporal fluctuations in temperature and atmospheric pressure, unexpected vibrations and air movement, or constantly changing molecular environments in the cell. Such small stochastic changes in the environment can add up. They primarily affect individuals during development, resulting in the ubiquitously observed normally distributed variance of measured phenotypes, even among genetically identical individuals raised in a shared environment (Ayroles et al., 2015; Klingenberg, 2019; Little, 1933). This Gaussian variance has been reliably used as a read-out of the extent of individuality expressed within a genotype (Ayroles et al., 2015; de Bivort et al., 2022). We asked if the distribution of expressed behaviour within a genotype is constant or altered in the presence of learning if all individuals shared the same life history and the same experience of the environment while the behaviour is tested.

We tested learning-dependent and spontaneous behaviours in all individuals of a genotype at the same time. This way every individual experienced the same environment (and its temporal stochasticity) during development and while behaving. To do so, we designed a platform that implements an operant conditioning assay that can be parallelised across 64 freely behaving individual Drosophila melanogaster (Figure 1B and E). The system takes input from real-time video tracking of individuals. The tracking is then used for closed-loop coupling of visual stimuli with a mild foot shock that is delivered to individual flies upon an undesired behaviour (Figure 1C and D; Figure 1—figure supplement 1A). The aversion to shock prompts the individual to learn to avoid it. In the assay, each individual fly freely moves in a Y-shaped arena (‘Y-maze’). The floor of the arena is illuminated such that each arm of the maze projects a colour. The conditioned task for a fly is to reduce the proportion of time spent on a particular colour (or arm) that is associated with shock (task performance or T_shocked, where smaller value represents better task performance).

We used this assay to first test the performance of known learning- and sensory-deficient flies (Dudai et al., 1976; Harris and Stark, 1977) and one wild-type genotype. We validated the assay using several learning paradigms, such as ‘green place learning’, ‘blue place learning’, ‘green colour learning’, and ‘blue colour learning’ (measured N=192, n=64 individuals per genotype and per paradigm, Supplementary files 2–4, Figure 1C–D, F, Figure 1—figure supplement 1A –B, D–E), and in a multiday experiment where paradigms were swapped each day for individual flies that were tracked over multiple days (Figure 1—figure supplement 2, Supplementary file 5). In the main experiment, we focused on the paradigm where flies are conditioned to avoid green colour that is always associated with the same place (‘place learning green’) and tested 88 genetically diverse wild-type genotypes (N=5632, n=64 individuals per genotype, Supplementary file 6; Mackay et al., 2012; Figure 1F–G). We ensured strict multi-generational control for each genotype and their two biological replicates. Replicates were raised independently in separate vials, but otherwise at the same time and under the same environmental conditions (Figure 2B, Supplementary files 6–8, Figure 2—figure supplement 1). Individuals from each replicate (vial) were distributed evenly into conditioned and control groups and their behaviour was tested simultaneously (Supplementary file 6, Supplementary file 8). Unlike in the conditioned group of flies, in the control group, the shock was never applied, allowing the flies to behave spontaneously, without a need to learn to avoid any colour (Figure 1C). Note that this also means that in the control group, task performance reflects innate colour preference, while in the conditioned group it can reflect innate preference, learned colour avoidance, as well as shock avoidance. We tracked their momentary task-relevant behaviours, including their locomotor activity, handedness, and approximately half a million individual decisions (i.e. switches from one arm in the Y-maze to another arm; N=497,450 decisions; Figure 1D).

Figure 2 with 8 supplements see all

Download asset Open asset

As expected, we found that, on average, conditioned flies spent significantly less time on the colour associated with the shock (23.3%), compared to the control flies (29.9%; N=2611 and N=2628, respectively, after filtering; t-test p<2.2e^–16; Figure 2A, Figure 2—figure supplement 2A, Supplementary file 9). This change was lacking in learning- and vision-impaired genotypes (Figure 1F, G, Figure 1—figure supplement 1B, D, E). Given that the memory mutant dnc¹ is known to be able to sense shock (Dudai et al., 1976), the observed lack of change in task performance pointed to impaired learning. Oppositely, the majority of wild-type genotypes (87.5%) showed change in average task performance towards avoidance of the shocked colour (Figure 1G; ANOVA p<2.2e^–16, Sum Sq.=3.143; Supplementary file 9). This change was consistent across sexes, over time, and across paradigms (Figure 1—figure supplement 1; Supplementary file 10).

Most other task-relevant behaviours were also significantly different between control and conditioned flies (Figure 1—figure supplement 2). These differences were consistent over the time course of the assay, across learning paradigms and replicates, and between sexes (ANOVA p=0.537, Sum Sq.=0.004, Figure 1—figure supplement 2, Supplementary files 10 and 11).

Interestingly, we found an extensive variability in expressed individual behaviour within the wild-type genotypes. Distributions of individual task performance were broader in some genotypes than in others (Figures 1E, F, 2A, Figure 1—figure supplement 2A, Figure 2—figure supplement 1A). This variability appeared more prominent in conditioned wild-type flies. Conversely, we also observed smaller inter-individual variability in task performance in the conditioned learning mutants dnc¹ (Figure 1—figure supplement 2A).

Curiously, not only the means and variances but also the shapes of the distributions of individual behaviours varied reproducibly between genotypes, particularly in the conditioned group (Figure 2B; Figure 2—figure supplement 1A). Unlike control flies, the individual behaviour of conditioned flies often occupied non-Gaussian distributions (Figure 2B and C; Figure 2—figure supplement 1A). We reasoned that the process of conditioning, and specifically learning, may be inducing this additional expression individuality within genotype.

However, one obvious difference in the environment between conditioned and control flies is the presence of the shock stimulus, which could be perceived individually, thus affecting the shape of the distribution. We tested whether individual sensitivity to shock could account for the observed changes in individual task performances in the two groups. We defined shock response as the difference between the velocity of an individual fly measured during the 3 min of bias testing prior to the start of the task and its velocity in the first quarter of the task, when the shock becomes implemented in the conditioned group. This way, instead of using yoked shock response measurement, we preserve within-individual measurement of shock response. Shock response in individual flies did not correlate with their task performance, whether it was measured at the individual (Figure 2—figure supplement 3C) or the genotype level (Figure 2—figure supplement 4A). It was, however, expectedly higher in the conditioned flies that actually experienced shock (mean shock response_control=0.294 mm/s, mean shock response_conditioned = 0.451 mm/s, t-test p<2.2e^–16, n=5239, Figure 2—figure supplement 4B, C). There was also no correlation between shock response and other task-relevant behaviours in the conditioned group (Figure 2—figure supplements 3 and 4A). Therefore, shock response alone is unlikely to account for the change in the distribution of individual task performance. This leaves learning as the more likely explanation for the observed variation in the shape of the distributions.

Because the shapes of the distributions were not always normal, we could not use the classical measurements of expressed individuality, namely the genotypes’ mean and variance, to compare between groups. Instead, to compare the differences in expressed individuality between any group of flies, we use Hellinger distance which captures the differences in shapes of distributions. This way we need not assume Gaussian probability distributions of individual behaviour. Further, we introduce a measure of differential entropy (H) to capture the extent of expressed individuality within a group of flies (see Materials and methods for details).

The entropy of task performance was significantly higher in conditioned flies (Figure 2—figure supplement 5) but correlated weakly, though significantly, between conditioned and control flies across genotypes (Pearson product-moment correlation between H_control and H_conditioned r=0.232, p=0.03, n=88 genotypes, Figure 2—figure supplement 4A). This indicated that a significant portion of the observed variation in individual behaviour within genotype was sourced from life history processes that preceded the splitting of flies to control and condition group – before the behaviour was measured. On the other hand, the difference in distributions between replicates of the same genotype was equally low in both conditioned and control flies (mean Hellinger_replicates = 0.06, mean Hellinger_replicates = 0.05, respectively, t-test p=0.256, n=88, Figure 2—figure supplement 4C). This confirmed that the controlled environmental conditions were matched across vials. The observed individual variability is thus primarily sourced from stochastic processes that occurred during development and life prior to the behaviour assay.

The entropy of task performance was strongly dependent on the genotype (resampling test, parametric p<2.2e^–16 resampled p=1e^–06, Figure 2—figure supplement 4B). Yet, the extent of expressed individuality (entropy) in spontaneous behaviour was not translated linearly to individuality in conditioned behaviour (Figure 2B, Figure 2—figure supplements 1A and 4A). In other words, the extent of expressed individuality of a genotype differed significantly between innate colour preference and learned colour preference. Higher entropy can reflect a difference in interaction of genotypes with conditioned compared to control task environment; a simple change in the distribution’s mean and variance could result in higher entropy. However, the evident changes in higher-order features of the distributions, such as their shape, indicated an additional reorganisation of individual behavioural expression between control and conditioned flies.

To capture and quantify this additional expression of individuality, we introduced a measure of ‘residual individuality’ (H_resid). H_resid is calculated as Shannon entropy of a probability distribution after scaling the distribution within its own range (Figure 2D, Figure 2—figure supplement 6). By scaling the distributions within their range, this measure nullifies the variability induced by factors shared across all groups of flies, including stochasticity in rearing environments, genotype, and their interaction, which are normally read out from measures of central tendency and dispersion (Figure 2—figure supplement 6). With such scaling, we could measure the residual changes in the shapes of the distribution, independent of their ranges and means.

Besides the measure of residual individuality within a distribution, we used three different measures of shape divergence between distributions (ΔH_resid) to dissect the source of the residual individuality. Learning divergence (ΔH_resid Learning) quantifies the change in expressed residual individuality of a genotype that arises due to learning (H_resid conditioned vs. H_resid control; Figure 2D). Experimental divergence (ΔH_resid Experimental) reflects the change in H_resid that may have been caused by some experimental factor affecting flies individually but systematically differently across vials (H_resid replicate 1 vs. H_resid replicate 2). Genetic divergence (ΔH_resid Genetic) captures the genetic contribution to the expressed residual individuality observed during the behaviour assay (H_resid genotype A vs. H_resid genotype B) (see Materials and methods for details on all three measures).

We found no difference in experimental divergence between conditioned and control groups (ΔH_resid Experimental_control=2.08, ΔH_resid Experimental_conditioned=1.85, t-test p=0.206, n=88, Figure 2I). Technical conditions of the experiment thus did not cause significant changes in the shape of the distributions of conditioned flies. On the other hand, Genetic divergence revealed that genotype contributed to the shapes of the distributions of task performance. The shapes were significantly more divergent across genotypes in the conditioned flies (ΔH_resid Genetic_control = 0.70, ΔH_resid Genetic_conditioned = 0.83, t-test p<2.2e^–16, n=88 genotypes; Figure 2H). This was also true for the learning divergence (ΔH_resid Learning), which differed significantly across genotypes (resampling test, parametric p<2.2e^–16, resampled p=1e^–06; Figure 2E).

These results were unexpected; in a strictly controlled environment shared between individuals, genetic differences are expected to be the only deterministic variables that can reproducibly affect expressed individuality. More strictly, in isogenic populations in a shared environment, the effect of a genotype on individuality is assumed to be absent except for its effect on phenotype’s variance. This is primarily a result of the genotype-specific buffering of stochastic processes during development. In both cases, individuality is expressed and read out only through changes in the mean and variance of a distribution, but not its shape (Ayroles et al., 2015; Linneweber et al., 2020). However, here, the mean and variance were accounted for, and yet we still find systematic differences in expressed residual individuality in task performance. We found this to be true for other behaviours as well (Figure 2—figure supplement 1E–F, Figure 2—figure supplement 7). We could not explain this variation with residual heterozygosity (Figure 2—figure supplement 4A) or response to shock (either individual or genotype-specific; Figure 2—figure supplements 3 and 4A). However, we found that learning divergence also evolved within each genotype with every made decision (Figure 2F). This revealed that the dynamics of decision-making may help pinpoint the source of this additional individuality.

Despite the efforts to fully control and parallelise experience, perfect control over freely behaving biological agents is by definition not possible. In non-uniform environments, momentary experience of a freely behaving individual will depend on its previous decisions. As a result, the sequence of decisions made in a natural environment is too diverse to be tractable. The Y-maze arena, however, minimises the diversity of unique experiences and binarises the decisions a fly can make. This allowed us to track and enumerate decision-making (Figure 3B and C).

Figure 3 with 3 supplements see all

Download asset Open asset

We aligned discrete decision sequences across thousands of individuals to ask whether momentary experience drives behavioural divergence. This uncovered that even minimal, single-moment differences in individual experience, such as the initial conditions of the first decision, can strongly shape learning trajectories. Because flies move freely, and the task onset occurs at an arbitrary but common timepoint, each individual can begin the task from any arm. A fly that finds itself on the green arm will immediately experience shock in the very first second, at the start of the task. In contrast, a fly that starts in the blue arm will experience its first shock only after making its first incorrect decision. Thereafter, the consequences of subsequent decisions are identical across individuals. This bears an effect on individual task performance dynamics and amplifies individuality over time (Figure 3A and D, Figure 3E and F; Figure 3—figure supplement 1D).

Early stochastic differences in the first decision provide a plausible mechanism by which residual individuality diverges between control and conditioned flies and across genotypes, despite the otherwise shared environment. Although the initial arm choice appeared random at the individual level, starting positions were normally distributed across genetically identical flies, with genotype-dependent means (Figure 3—figure supplement 1A, B, Figure 3—figure supplements 2 and 3). Each decision can therefore be conceptualised as a single draw from an individual-specific probability distribution of colour preference, partially determined by genotype. With each subsequent decision, this distribution is resampled and, if learning occurs, updated as a function of experience. Consistent with this framework, we found that learning progressively outweighed innate starting bias as the determinant of individual task performance (Figure 3F, Figure 3—figure supplement 1E, F, Figure 3—figure supplements 2 and 3).

Among conditioned flies sharing the same initial biases, an early stochastic event could trigger an earlier or qualitatively different learning trajectory in one individual but not another. In control flies, which do not learn, decision rules and their consequences on task performance would be insensitive to initial conditions. Accordingly, in control flies, time-averaged distribution of task performance converged to the same shape despite differences in first decisions. In conditioned flies, residual individuality originating from the first decision persisted across many subsequent choices (Figure 3E). We found that the distributions of individual behaviour were broader and their shapes changed substantially over the course of the experiment for the conditioned flies, and not for the control flies (Figure 3—figure supplements 2 and 3).

To further characterise the potential causes of this effect, we grouped flies by initial colour bias and by final learning score. We then measured change in residual individuality (ΔH_resid Start) in task performance between individuals starting the task from different arms. Divergence was maximised at the first decision (Figure 3E and F). In control flies, this divergence rapidly collapsed despite distinct initial conditions or initial biases. In conditioned flies, divergence persisted and was further amplified in individuals with high learning scores and unfavourable initial biases (Figure 3E). Alignment of decision sequences further revealed an early alternating-choice pattern in conditioned flies, particularly pronounced in individuals with an initially unfavourable bias who ultimately achieved high learning scores (Figure 3C). This suggested that the clash between learning and innate bias likely contributes to residual individuality. These dynamics indicate that the learning experience of the first decision can initiate a butterfly effect: because learning is non-ergodic, the effects of early stochastic differences accumulate over time, progressively diversifying experience and amplifying individuality within a genotype.

While we do not track behaviour throughout the entire life of a fly, the design of our experiment allows us to dissect the effects of life experience prior to the task from the effects of momentary experience during the task. We asked how extensive is the effect of the entire life’s experience on individual momentary behaviour, compared to the effect of genetics or learning during the task? To quantify the relative contributions of different sources of individuality, we turned to modelling of observed behaviours. To test how well our models capture and represent the observed behaviour, we used them to simulate behaving flies and then compared their simulated behaviour to the behaviour of observed, real flies. Simulated flies incorporated a range of behavioural parameters (Supplementary file 12). Based on them, the simulated flies decided how long they want to stay in the current arm of a simulated Y-maze, and whether they want to turn left or right when leaving the current arm. The left-right turns were sampled from a Bernoulli distribution with a time-dependent rate. The decision to leave the arm was made by sampling the escape duration (time to leave one arm) from a log-normal distribution with a time-dependent mean and stationary variance. The mean of their escape duration distribution depended on the colour of the arm and the presence of shock at the time point $t$. Importantly, the mean escape duration and the rate of turning left or right could change with a reinforcement learning rule, which depends on the presence of shock.

To assess the effects of genetics, learning, and individual life experience, we systematically exclude one or none of them in four different behaviour models. First, all flies of the same genotype have identical parameters, but different genotypes have different parameters (no individuality model), thereby omitting the effect of any individual experience obtained prior to expressing a behaviour in the task. Second, the ith parameter of an individual simulated fly is sampled from a normal distribution with genotype-dependent mean $(m_i^{(g)})$ and standard deviation $(s_i^{(g)})$. These simulated flies have individuality in the sense that their initial behavioural policies differ from one another, because no two flies have the exact same parameters (individuality model). In other words, simulated flies start the task as if in addition to their genetics and learning, their life experience prior to the task (including stochastic processes during development) had shaped their individual behaviour. Third, parameters are sampled as in the individuality model, but the learning rate is set to zero, such that the presence of shock or colour induce no lasting change on their behaviour beyond the immediate effect of changing the escape duration, simulating only the shock response (no learning model). Fourth, parameters are sampled per individual, but there is no genotype dependence of the means and standard deviations of the parameter distributions. This way the effect of genetics is omitted, including its effect on developmental variability (no genetics model).

Best fit was found for individuality model, followed by no learning and no genetics model. No individuality model failed to recapitulate the observed behavioural individuality (Figure 4A and B). Particularly striking was the mismatch in the distribution of the number of decisions made by flies simulated without individuality compared to the actually measured distributions of the number of decisions made by real flies (Figure 4—figure supplement 1). Compared to whole-life experience or genetics, the effect of momentary learning on individual behaviour was much smaller, and for some genotypes seemed irrelevant (Figure 4A). However, we found that the model without learning could not capture the temporal dynamics of many observed behaviours, highlighting its significance despite apparently modest effect size (Figure 4C). Interestingly, the excess of observed non-Gaussian distributions of individual behaviour in conditioned flies was recreated only if the simulated flies could learn, despite the effect of shock on activity being simulated in no learning model, and despite all parameters of the model being drawn from normal distributions (Figure 2C). Lastly, repeated simulations revealed that the genotype distribution of simulated task performance matches the observed behaviours better for the model with learning (Figure 4D, Figure 4—figure supplements 1 and 2). This reaffirmed that residual individuality most plausibly emerges from the very process of learning.

Figure 4 with 3 supplements see all

Download asset Open asset

The fitted models allowed us to estimate how the momentary task-relevant behaviour of a given fly differs from other flies, or how it compares to other moments in time. We found that the individual behaviour had changed significantly during the experiment (Figure 4E and F) and that this change was significantly bigger in the flies that learned (Figure 4F, Figure 4—figure supplement 3). A visualisation of these behavioural differences with multidimensional scaling revealed that the behaviour of individual flies does not cluster according to genotype throughout the experiment (Figure 4E). This lack of clustering was also confirmed with a classification analysis: using a long-short term memory (LSTM) classifier applied to the behavioural data of individual flies, we could predict the genotype in only 3.4% of the cases (chance level: 1.1%, fourfold cross-validation). Although our models confirmed that genetic variation is essential for generating the observed behavioural diversity, they demonstrate it determines an almost stochastic fraction of any given individual’s behaviour.

Altogether, these results decisively demonstrate that learning is not just an emergent but a fundamental driver of behavioural individuality.

Behavioural arenas were 3D-printed in black Tough PLA (Ultimaker B. V., Utrecht, Netherlands) using desktop Fused Deposition Modeling (FDM) 3D printer (Ultimaker S5, Ultimaker BV, Utrecht, Netherlands), with a z-axis resolution of 0.1 mm and 100% grid infill pattern. The mazes are in the form of the letter ‘Y’, with equal arm lengths. The arms are connected at the centre with an angle of 120° from each other, and length = 10 mm and width = 2 mm. This specific arm width was chosen so that the flies are not able to turn around in the middle of the arm. This way the fly must traverse the whole arm and reach the end, where it is provided with a circular turning area of a diameter = 4 mm. This is done so that the fly, after making a decision, remains in the arm for long enough to associate the electric shock with the colour of the arm. For the same reason, the fly only gets shocked once it has entered an arm with its full body length, and not at the junction of the three arms. The same requirement is implemented for the removal of electric shock – the fly needs to exit the shocked arm but also enter the non-shocked arm with the full body length in order to escape the shock (Figure 1C). The total height of the maze within which the fly is restricted is 1.2 mm. The wall of the maze does not erect straight up to the ceiling, but in a cross-section forms a step at 0.8 mm. This is done so that the fly, once resting on the floor of the maze, is not able to climb up to the ceiling of the maze and avoid getting shocked. The ceiling of the maze is a regular 25 mm×75 mm glass slide coated on one side (facing the maze) with dry polytetrafluoroethylene (PTFE) aerosol (KONTAFLON 85, Kontakt Chemie, CRC Industries Europe BV, Zele, Belgium). The coating makes them hydrophobic, and hence, slippery for the flies to walk on. Each slide covers two Y-mazes horizontally. Eight Y-mazes (4 rows × 2 columns) are 3D-printed together, with each Y-maze restricted within an area of 30 mm × 30 mm, which corresponds to an individual shocking area on the transparent shocking grid (described in the next section). A single plate consisting of 8 Y-mazes is then laid over, and glued onto a transparent shocking grid. Eight such arenas (64 Y-mazes) are placed on top of the LCD for the assay.

Indium tin oxide (ITO) coated (400 nm coating corresponding to 4–6 Ω/sq) glass plates (150 mm × 100 mm × 1 mm) were laser etched to form a grid of 8 (4 rows × 2 columns) isolated 30 mm × 30 mm shock delivery pads for each Y-maze. The shocking pads were patterned in the form of an interdigitated grid, with a central pad connected to the ground and 8 isolated pads connected to an array of normally off electromagnetic relays. Each isolated pad was etched in an interdigital pattern with 15 digits of 1 mm width on each side (ground pad and active pad connected to the relay) separated by a distance of 20 μm between each digit. Holes (diameter 0.8 mm) were laser cut into the glass plate at positions that correspond to the centre point of the circular ends of the Y-maze arms. The holes are used to deliver CO₂ to anesthetise flies for fly retrieval at the end of the assay. The designed plates were manufactured at Diamond Coatings Ltd, UK. The 3D-printed Y-mazes are assembled with the glass plate using adhesive tape or superglue, making sure the glue does not touch the ITO coating. Each such assembled plate is placed on a 3D-printed holder fixed on top of an LCD screen (LG 25BL56WY, 25", 1920×1200 pixels). The holder can hold 8 glass plates and is fitted with grooves for electrical wiring and place marks for copper tape (12 mm width). Copper tape is applied to the glass plates to make electrical contact between the ITO pads and the copper tapes on the holder. Custom-made printed circuit boards fitted with 5.1 mm round head gold spring contacts (P/N 5099-D-2.0N-AU-1.3C, PTR Hartmann, Germany) electrically connect the copper tapes to the electromagnetic relays, ultimately connecting it to each isolated shocking area. Electrification of the 8 isolated Y-mazes is controlled by an array of 8 relays, which in turn is controlled by an 8-bit shift register (TI 74HC595). The 64 mazes are controlled by 64 individual relays, which are controlled by 8 shift registers. The registers are controlled by the computer program (testY_LCD.py) via an Arduino Mega 2560 Rev. 3 (ATmega2560) (loaded with Arduino_yMaze_shock_595.ino) in closed loop with the position of the fly inside the maze and the visual stimulus presented to it via the LCD screen. Visual stimuli can be, but are not limited to, colours, colour gradients, patterns, or light intensities. The whole system is made to be modular, and the plates holding the Y-maze arenas wireless, so that it is not fixed to the LCD screen. This way it can be moved around for the purpose of fast loading the arenas with flies. When electrified, the floor of each maze delivers a constant 30 V DC through the flies' legs. The current is limited to 500 mA distributed between 64 arenas (on average 7.8 mA per fly) and powered using a benchtop power supply (RND 320-KA330-5P, RND Lab, Switzerland). The electronics circuits are powered using a 5 V, 90 W switch mode power supply (LRS-100-5, Mean Well Enterprises Co. Ltd., New Taipei City, Taiwan). This current was chosen through optimisation experiments in which the goal was to maintain reproducible shock avoidance in flies across several randomly chosen wild-type genotypes, while the memory mutant dnc¹ flies, that usually experience most shocks due to their inability to learn, do not get injured during the experiment.

Each Y-maze is associated with an Augmented Reality University of Cordoba (ArUco) marker (Garrido-Jurado et al., 2014), which is identified by the control software (getBoxes_aruco_64.py) to acquire the position of each maze. This is done both to automate the experiment and to maintain the accuracy in case of an accidental misalignment of the plate within the tolerance of the plate holders when setting up the experiments. The position of each moving fly is captured using a custom background subtraction algorithm (unit_fly.py) and fed into the control algorithm (testY_LCD.py, unitMaze.py). In turn, the algorithm changes the visual stimulus presented to each fly via coloured patterns on the LCD screen (LCD.py). The algorithm also controls the electric shock delivery for each fly via an Arduino Mega 2560 Rev. 3 in real time (Arduino_yMaze_shock_595.ino). The platform is illuminated with an array of 12×8 infrared (940 nm) LED (Figure 1B). The video is captured at 6 fps using a greyscale industrial USB 3.0 camera without an infrared filter (DMK 38UX253, The Imaging Source, Germany), fitted with an infrared long pass filter (dia. 40.5 mm thread, 0.5 mm, 093 IR Black 830, Schneider Kreuznach GMBH, Germany). The filter allows the program to ignore any background illumination changes inside the room that can interfere with the accuracy of the fly detection algorithm. The program is generally robust to light interference caused by movement around the setup, but to minimise any potential light interference, as well as to minimise influencing flies’ behaviour, the experimenter left the room during the assay or wore black or dark clothes. Different associative learning protocols can be implemented with this system, whose codes can be found on our GitHub repository https://github.com/jaksiclab/GeneticsOfLearningIndividuality (copy archived at jaksiclab, 2026). At the end of each protocol, the program outputs a folder containing comma-separated values (.csv) files reporting the data acquisition times (in nanoseconds starting from the time of acquisition of the first fly position), the coordinates of the fly, the arm in which the fly resides at each moment, whether the arena is electrified or not, the correct, incorrect, and total number of choices the fly has made at that point of time. The program also automatically outputs the final task performance scores of the flies at the end of the experiment, the time at which the experiment was performed, sped up infrared video recordings of the arena (.mp4), and setup metadata that needs to be input by the user at the start of the experiment. For further automation, this information can directly be accessed by the control software immediately after an experiment.

All genotypes were acquired from the Bloomington Drosophila Stock Centre (BDSC) in the year 2020. 88 Drosophila Genetic Reference Panel (DGRP) lines, norpA^EE5 (BDSC stock number 5685) mutants, and dnc¹ (BDSC stock number 6020) mutants were used in this study. DGRP genotypes are denoted by a number (nnn) or a label DGRP-nnn, where nnn represents the DGRP genotype. For example, in Figure 2D and H, genotype was denoted as 208 or DGRP-208, respectively, and both indicate that the genotype used was DGRP-208 line (synonym RAL-208; BDSC stock number 25174). Flies were maintained and raised on cornmeal agar medium: 6.2 g agar-agar, 58.8 g cornmeal, 58.8 g brewer’s yeast, 0.1 L grape juice – Migros M-Classic, 4.9 mL propionic acid, 26.5 mL methylparaben Moldex, 1 L of tap water (Lausanne, Switzerland). Environmental temperature was 26°C (±0.5°C), humidity 50% RH (±5%), and circadian light condition was set to 12 hr light/dark cycles (instant switch, no ramping). The maternal generation (F1) was generated in three vial replicates from eggs laid by two sets of F0 flies taken from stock vials. F0 stock vials that were maintained at 21°C (±1°C), on the same medium, but with no circadian light control. Stock populations varied between 20 and 60 flies per vial with no larval density control. Nine F0 females and 5 F0 males were selected from stock population, housed together, and allowed to lay eggs for 3 days, after which they were discarded. The vials were checked periodically to ensure the development was healthy and progressing at an expected pace. While we set up three replicate vials initially, only two were used – the extra vial served as backup for the cases when a vial needed to be discarded. Any vial showing signs of distress, such as dry food, biofilm, too few larvae or pupae, or overcrowding, was discarded, and the rearing protocol was restarted from the beginning unless a backup vial was available. This was assessed always by the same experimenter. The F1 progeny were collected 12 days after the last day of egglay. The experimental F2 generation was set up from each vial replicate the same way as the F1 generation. However, after discarding the F1 adult flies, and after 9 days of development, the vials were checked for eclosion daily. As the first few flies eclosed, they were discarded. The next day, the peak eclosion would follow, and the newly eclosed flies were collected. Around 40 males and 40 females were then separated into two vials with fresh medium and prepared for the behaviour assay that would follow 3 days later, when the adult flies were 5 days of age. All the preparatory flywork, including collection, sorting, and counting flies across the generations, was performed swiftly and using light CO₂ anaesthesia (5 L/min flow through standard CO₂ pad) at a laboratory room temperature of 21°C (±0.5°C). All flies were handled by a single experimenter (by RM for the single day experiments and by GVB for multiday experiments).

We use two operant conditioning paradigms: the associative visual place learning (Figure 1C) and the associative colour learning (Figure 1—figure supplement 1A). In both paradigms, the unrestrained fly can explore the Y-maze and make sequential binary decisions at the junction of the three arms. Conditioning is initiated after initial measurement of unconditioned behaviour. The decisions during conditioning can result in the absence, introduction, or removal of foot shock.

For each genotype, we used two biological replicates, progeny of two different sets of males and females, that were reared independently in different vials. Each replicate consisted of 32 flies, equally split between males and females and between two treatment groups, conditioned and control. Environment during conditioning was kept the same as for rearing. Flies were very lightly anaesthetised (10 L/min CO₂ flow on 30 cm×40 cm custom-made CO₂ pad) and loaded into the Y-maze arenas using an aspirator. In the conditioned group, flies underwent operant conditioning in the presence of shock, while in the control group the flies were allowed to spontaneously explore the maze experiencing the same visual and environmental stimuli, but without experiencing any shock. All flies were tested during the same ‘non-siesta’ circadian time frame (20% ZT 0–1, 30% ZT 1–2, 30% ZT 2–3, 15% ZT 3–4) with exception of around 4% of flies extended beyond ZT 3 (3% ZT 4–5, 1% ZT 5–6, 1% ZT 6–7) due to technical issues (Supplementary file 7). We found no difference in task performance given the zeitgeber time frames (ANOVA: task performance ∼ experiment time (ZT) * treatment, p=0.531).

During this time, flies tend to make on average 76 decisions per 20 min, with more decisions made by males (N_male = 82, N_female = 62), and fewer decisions made in the presence of conditioning (N_control = 89, N_conditioned = 63). In the place learning assay, one of the three arms of the Y-maze is coloured differently from the other two, and an electric shock is associated with that colour. The colours chosen are blue and green, as flies have been shown to be able to distinguish well between the two in previous associative learning studies (Salcedo et al., 1999; Schnaitmann et al., 2013). At the junction of the Y-maze, the fly is presented with a binary choice between two arms. When conditioning to avoid the colour green, the fly gets shocked when it enters a green arm, while the shock is stopped when it leaves the green arm and enters a blue arm. No conditioning occurs when it enters a blue arm from another blue arm. Hence, we are implementing a combination of positive punishment (adding aversive stimulus following incorrect behaviour) and negative reinforcement (removing aversive stimulus following correct behaviour). The shocked arm was associated with a colour assigned to different arms in a clockwise way, column-wise across the 64 mazes (Figure 1E), and kept fixed for the entire place learning assay. In the colour learning assays, the shock arm was associated with a colour, but the colour of the arms changed with every decision, making the shock and the colour uncoupled from a position in space. At the start of any assay, the entire arena is illuminated with red colour for 2 s while the floor becomes electrified. This moves the flies from a state of rest to a state of arousal. The video recording is initiated and for the next 3 min, the fly is allowed to explore the maze and get accustomed to it without delivering any shock. This time is also used to assess prior individual biases and initial fly behaviours. The operant conditioning assay is then initialised and recorded for 20 min.

In the place learning paradigm, wild-type conditioned flies spent an average of only 10.5% of the total assay time in the shocked arm, while the control group spent significantly more time in this arm in the absence of shock (T_shocked = 32.5 %, Student’s t-test p=2.52e^–10). The learning-deficient (dnc¹) showed poor relative task performance at avoiding the shocked arm (conditioned T_shocked = 27.8%, control T_shocked = 30.8%; p=0.20). Similarly, relative task performance of visually impaired (norpA^EE5) was low (conditioned T_shocked = 25.4%, control T_shocked = 25.0%; p=0.90; Figure 1F), suggesting that in flies visual context such as colour is informative for place learning (Whishaw, 1998). Alternatively, the norpA^EE5 mutation may also induce learning impairment beyond perceptual deficiency (Melnattur et al., 2014).

While average genotype performance was consistent across learning paradigms, across different colour stimuli and across consecutive measurements of the same populations over multiple days, we found that individual learning task performance was variable and evolved over time (Figure 2, Figure 1—figure supplements 1 and 2, Figure 2—figure supplement 1).

In the multiday learning task, genotypes DGRP-362 and dnc¹ were reared based on the standard rearing explained above but in 12 replicate vials per genotype (3 replicate vials per experiment type, except for dnc¹ for which we had 2 instead of 3 replicates for the last two experiments because two vials were discarded due to poor development). 256 individuals from the two genotypes were collected (total N=512 flies) to be tested in four types of experiments. Before the test, the collected flies were housed in fly vials in groups of around 40 individuals of mixed sexes. On the day of each first test, 32 females and 32 males were selected for the test. In each of the four experiments, a set of 64 individual flies were tested once per day with the exception of the second experiment of DGRP-362 that contained 59 flies (32 males and 27 females) due to experimenter error (fly identities were accidentally and intractably swapped). In total, N=507 flies were tested successfully. Flies were tested on 2 consecutive days followed by 1 day without a test, then another set of 2 test days. All tests were done under conditioned treatment. Before each conditioning session, flies were allowed to explore the arena for 5 min in the absence of shock. A 20 min conditioning session ensued, followed by a 5 min session in the absence of shock to assess change in bias due to learning and short-term memory. This regime was repeated once per day, at the same time of the day (zeitgeber 0–2). In the first experiment (PPCC_gbgb), flies were tested in a place learning paradigm the first 2 days (‘PP’), then in the colour learning paradigm the last 2 days (‘CC’). The colours associated with the shock across 4 days alternated across tests between green and blue (‘gbgb’). In the second experiment (PPCC_bgbg), they alternated between blue and green. In the third (CCPP_gbgb) and fourth experiment (CCPP_bgbg), on the first 2 days place learning was tested, and on the last 2 days, we tested the flies in a colour learning paradigm. The colours associated with the shock alternated the same way for the third and fourth experiment as in the first and second, respectively. Between tests, the identity of each fly was tracked by housing them individually in a custom 3D-printed 64-well single housing plate, placed on food and isolated with a dome to avoid the food from drying up. The flies were transferred to the single housing upon completion of each test. Each fly was always tested in the same Y-maze on the platform over the days. The position in the well plate corresponded to the position on the platform to facilitate quick setup of the experiment. Males were placed on the top 4 rows of the platform and females on the 4 bottom ones. All flies were handled by a single experimenter (by GVB), except for the third experiment (where the assay was set up by AMJ).

The single housing plate used in the multiday learning task consists of 64 cube-shaped wells, organised in 8 rows and columns. The floor of the plate is a 0.2-mm-thick printed mesh. The mesh allows the flies to reach food and eat when the plate is placed on a tray filled with fly medium, while keeping them as dry and clean as possible throughout the experiment. Care should be taken to only gently press the well plate on the surface of the medium and prevent food from entering the wells. The mesh also helps with anaesthetising the flies by simply placing the plate onto the CO₂ pad. The top of the 64-well plate is covered with a 3D-printed lid, shaped as an inverse imprint of the 64-well plate. This way, when it covers the well plate, the grid of the lid protrudes by 0.2 mm into each well, preventing the flies from wiggling their way between the lid and the plate and crossing into another well. A thin, fine filter mesh, usually used for Drosophila egg collection, was printed into the top printed layers of the lid. To manufacture this, the top four 0.06 mm layers of the lid are printed first. Then, the printer is paused, the filter mesh is laid over the printed part and tightly affixed to the printer’s build plate with an adhesive tape. The printing is resumed so that the next layers of print material ‘bakes’ the mesh into the print. After printing, the mesh remaining on the outer part of the print is cut with a precision knife. The .stl models can be found in the GitHub repository. The plate and lid were printed using red PLA filament (to avoid colour imprinting or colour bias) on an Ultimaker S5 3D printer. Black filament can be used as well, but the low contrast with the fly makes picking and placing flies from the well more challenging and prone to errors. The 3D models were sliced with Ultimaker Cura, using default ‘Fine resolution’ print setting (0.06 mm layer height). To create the bottom printed mesh, we use Ultimaker Cura slicer to slice a 0.2-mm-thick block with width and length matching the size of the dimensions of the 64-well plate. We use the 60% ‘grid’ infill and omit the bottom and top layer print. Once printed, the mesh is glued to the bottom of the 64-well plate using cyanoacrylate adhesive. This has to be prepared at least 24 hr before the experiment, in a well-ventilated room, to allow the volatile compounds of the adhesive to fully air out. Occasional slight warping of the lid would make it possible for flies to cross to other wells, thus losing track of their identity. Thus, testing the lids before an experiment is advised.

All basic statistical tests were performed within the R environment (R version 4.0.2). Pearson correlations and t-tests were performed using base cor.test() and t.test() functions. t-Tests were two-sided. Resampling test was performed using the function lmperm() from the package permuco (Frossard and Renaud, 2021). Exact general independence test was performed using the function independence_test() from the package coin (Hothorn et al., 2008; Strasser and Weber, 1999). Generalised linear mixed models were fitted using lmer() function from the lmerTest package, and the analysis of variance for fitted models was performed using the anova() function. Least square means and the pairwise difference between them were calculated using emmeans (pbkrtest.limit=100000000) and pairs() function from lsmeans R package. Where multiple tests were compared, p-values were corrected using Bonferroni correction for multiple test using base p.adjust(method = ‘bonferroni’) function. Density distributions are estimated as kernel densities using density() function and default parameters.

Task performance (T_shocked) of a fly is the fraction of time a fly has spent in the conditioned arm during the length of the experiment (20 min). It is measured the same way regardless of whether the shock is applied or not (e.g. in the control experiment). Handedness of a fly is the absolute value of the total number of right turns subtracted from the total number of left turns, divided by the total number of turns taken by the fly at any given time frame. Developmental noise is the absolute handedness of a fly measured during the 3 min prior to the conditioning. Velocity in mm/s for a given time frame is calculated by averaging the displacement of a fly per second over the said time frame. Relative velocity is the difference in average velocity of a fly between the shock-associated arm of the Y-maze and the neutral arm of the Y-maze at a given time frame. Shock response is the absolute value of the difference between the average velocity of a fly measured during the 3 min prior to the start of the conditioning part of the experiment and the average velocity of the fly during the first quarter (5 min) of the conditioning experiment (when shock becomes implemented in the conditioned group). colour bias of a fly is defined as the total fraction of the time the fly spends on the colour during a set time frame. Percent correct decisions is the proportion of made decisions to enter the non-shocked arm out of all made decisions to enter any arm. Learning score is computed by splitting the experiments into 10 bins of 2 min and computing the negative of the average of two Spearman’s rank correlations: (1) the rank correlation of the interval index $i$ with the relative time spent in the shock arm during interval i and (2) the rank correlation of the interval index $i$ with the relative number of visits to the shock arm during interval i.

Before calculating mean task performance for a genotype, we first filter out individual flies based on their lack of activity. Flies are removed from the analysis if they remain at the shocked arm for 3/4 of the experiment or if they never enter the shocked arm/colour (in either conditioned and control group of flies). Based on observations in trial and final experiments, this outlier behaviour was due to the fly getting accidentally injured by mis-manipulation during assay setup (usually by slightly misplacing them outside the maze boundary and pinching them with the glass slide). The filtering accounted for 3.78% of all tested wild-type flies. A linear mixed modelling is then used to estimate the effect of genotype on the response to shock and to regress out nuisance factors that may affect the measures of learning. Those included sex, replicate, sex-specific locomotor activity during and in the absence of shock, developmental noise, position of the Y-maze, position of the shocked arm, position of the arm in which the fly initialises the task. The model with the T_shocked as the response variable, genotype interacting with the treatment (conditioned or control), sex, developmental noise (absolute bias locomotor handedness), shock response and the starting arm (shock or neutral left, or neutral right) as the fixed effects, and the replicate number, and the position of the fly on the 64-maze arena as the random effects, is fitted to the data to capture individual components of variance of the data (Supplementary files 9 and 10). The least squared means of the genotype × treatment effect was extracted. Relative task performance of a genotype was measured as the Tukey’s honestly significant difference in least square mean of T_shocked for a genotype in the conditioned group and least square mean of T_shocked for a genotype in the control group.

Extent of expressed individuality in a behaviour within a group of flies is measured using the differential entropy (Cover and Thomas, 2005) of the probability distribution function of the behaviour for that particular group, estimated over the theoretical range of the behaviour (Supplementary file 12). For Gaussian distributions, entropy scales with the logarithm of the standard deviation and is thus consistent with classical measures of variability in isogenic populations, often called individuality (Figure 2—figure supplement 8). The difference in expressed individuality between two groups of flies is calculated as Hellinger distance (Hellinger, 1909) between the two distribution functions. Specifically, we calculate Hellinger distance between two continuous probability density functions of the distributions, over the theoretical range of the concerned behaviour.

To assess the extent of residual individuality (H_resid), the distributions of expressed individual behaviours within genotypes, groups, or replicates are first independently divided into a histogram of 10 bins within the individual range specific to each distribution. The probabilities of finding a measurement in each of the 10 bins are then calculated to obtain the binned probability distribution, based on which the Shannon entropy (Hausser and Strimmer, 2008; Shannon, 1948) is calculated. The H_resid correlates negatively with the range of T_shocked (Figure 2—figure supplement 8). This is a result of scaling the distributions within their own ranges and is consistent with our definition of H_resid being free from the effects of environmental and developmental stochasticity. That is, assuming we are able to capture the full range of the expressed behaviour (evidenced by consistency of the distributions between replicates), a group of flies having a narrow range of T_shocked will be scaled to an approximately uniform distribution within the range of the particular distribution (Figure 2—figure supplement 6), and hence have a higher entropy. To capture the difference in expressed residual individuality between two groups, the same independent binning approach is taken for each group-specific probability distribution separately, and the Kullback-Leibler (KL) divergence (Drost, 2018; Kullback and Leibler, 1951) of the two distributions is calculated. Since each distribution is binned independently within their own range, every distribution becomes a list of 10 probability values (summing up to 1), with each value having an equal weight for the calculation of the entropy as well as the divergence. This procedure scales each probability distribution to a common range (Figure 2—figure supplement 6), and ensures that the entropy and the divergence measures are only capturing the shapes without being confounded by the central tendencies or the ranges of the distributions. Effectively, we capture residual individuality that is expressed independently of the variability caused by developmental stochasticity or environmental plasticity. The KL divergence is a non-metric, asymmetric measure of how different one distribution is from a reference distribution. Therefore, when we measure the change in the residual individuality as a result of conditioning (Learning divergence, ΔH_resid Learning), we measure the KL divergence of the conditioned (shocked) population with respect to the control (non-shocked) population. In other cases where no reference distribution is obvious, KL is calculated in both directions (such as in the comparison between genotypes within a treatment, genetic divergence, ΔH_resid Genetic), or the compared distribution pairs are randomised, and the pairs are kept consistent between tested environments (such as in the comparison between replicates within a genotype and treatment, experimental divergence, ΔH_resid Experimental).

To provide an alternative measure of differences in the shapes of distributions that are symmetric or less sensitive to outlier values, we also calculated Hellinger distance (Hellinger, 1909) and Jensen-Shannon divergence (Lin, 1991) between the distributions of T_shocked in control and conditioned treatment, which were centred around the median. We also applied the KL divergence on non-binned, continuous distributions of T_shocked (Figure 2—figure supplement 8). We found that these measures all correlate strongly and significantly with each other and with either the range, variance, and entropy of the T_shocked distributions and/or the relative task performance. While they do uncover various differences between distributions, these measures are not free from variance and central tendency effects. Hence, we refrain from using them for measuring the changes in Residual individuality.

To model individual fly behaviour, we extract for each fly a sequence of arm identities $x_i\in \{1,2,3\}$, arm colour $c_i\in \{\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n},\mathrm{b}\mathrm{l}\mathrm{u}\mathrm{e}\}$, presence of shock $s_i\in \{\mathrm{s}\mathrm{h}\mathrm{o}\mathrm{c}\mathrm{k}\mathrm{e}\mathrm{d},\text{not shocked}\}$, arm escape-duration $\mathrm{\Delta }{t}_i\in {\mathbb{R}}^{+}$, and turn events $y_i\in \{\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t},\mathrm{l}\mathrm{e}\mathrm{f}\mathrm{t}\}$. For example, the sequence S = ((x₁ = 3, c₁ = green, s1 = shocked, Δ t₁ = 1.0 s, y_i = right), (x₂ = 1, c₂ = blue, s₂ = not shocked, Δ t₂ = 3.5 s, y₂ = left), ...) means that a fly started in arm 3, perceived colour green and got shocked, escaped arm 3 after 1 s by taking a right turn at the centre of the maze, arrived in arm 1 with colour blue and without shock, stayed there for 3.5 s and left it with a left turn, etc. We model the probability of a given sequence $S$ of length $T$ as $P(S)=P(\mathrm{\Delta }{t}_T|{\mu }_T,\sigma )\prod _{i=1}^{T-1}P(y_i|{\rho }_i)P(\mathrm{\Delta }{t}_i|{\mu }_i,\sigma ),$ where $P(y_i|{\rho }_i)$ is a Bernoulli distribution with probability of turning right given by ${\rho }_i$ and $P(\mathrm{\Delta }{t}_i|{\mu }_i,\sigma )$ is a log-normal distribution with mean ${\mu }_i$ and standard deviation $\sigma$. The probability of turning right ${\rho }_i=\sigma (w_{x_i}^{(i)})=1/(1+\exp (-w_{x_i}^{(i)}))$ depends on the weights $w_1^{(i)},w_2^{(i)},w_3^{(i)}$ (one for each state). These weights change according to the following dynamics: $w_{x_i}^{(i+1)}=w_{x_i}^{(i)}+({\eta }_{\mathrm{g}}I(c_{i+1}=\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n})+{\eta }_{\mathrm{s}}/(1+{\varphi }_{\mathrm{w}}{n}_i)I(s_{i+1}=\mathrm{s}\mathrm{h}\mathrm{o}\mathrm{c}\mathrm{k}\mathrm{e}\mathrm{d}))$ · $(I(y_i=\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t})-{\rho }_i)$, where ${\eta }_{\mathrm{g}}$ is the green colour learning rate, $I$ denotes the indicator function with $I(\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e})=1$ and $I(\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e})=0$, ${\eta }_{\mathrm{s}}$ is the shock learning rate, $n_i$ is a shock counter (number of times shock was experienced so far), and ${\varphi }_{\mathrm{w}}$ is a parameter that scales the dependence on the shock counter. This learning rule increases the probability of turning right in state $x_i$ when turning right led to the green arm without shock and ${\eta }_{\mathrm{g}}>0$ (colour preference learning), and it leads to a decrease in the probability of turning right when the negative impact of arriving in a shocked arm outweighs the preference for arriving in a green arm (${\eta }_{\mathrm{g}}+{\eta }_{\mathrm{s}}/(1+\varphi n_i)<0$), and vice versa for turning left. The means of the log-normal distribution are given by ${\mu }_i={\mu }^{(i)}+{\mathrm{\Delta }}_{\mathrm{g}}I(c_i=\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n})+{\mathrm{\Delta }}_{\mathrm{s}}I(s_i=\mathrm{s}\mathrm{h}\mathrm{o}\mathrm{c}\mathrm{k}\mathrm{e}\mathrm{d})$, where the parameters ${\mathrm{\Delta }}_{\mathrm{g}}$ and ${\mathrm{\Delta }}_{\mathrm{s}}$ control the change in escape duration due to being in a green arm or being shocked, relative to the mean ${\mu }^{(i)}$. This mean changes according to the dynamics ${\mu }^{(i+1)}={\mu }^{(i)}+{\eta }_{\mu }/(1+{\varphi }_{\mu }{n}_{\mathrm{s}})I(s_i=\mathrm{s}\mathrm{h}\mathrm{o}\mathrm{c}\mathrm{k}\mathrm{e}\mathrm{d})+{\eta }_{\mathrm{t}}\mathrm{\Delta }{t}_i$. With this update rule, the mean of the log-normal distribution over escape durations can change simply with the passage of time, when ${\eta }_{\mathrm{t}}\ne 0$, but also due to being shocked, when ${\eta }_{\mu }\ne 0$. To model a population of flies, e.g., 64 flies of the same genotype, we place normal priors over the model parameters (Supplementary file 13) and estimate the means and standard deviations of these priors with approximate expectation-maximisation. We considered four different models: no individuality: all flies with the same genotype have exactly the same parameters, i.e., the standard deviation of all priors is zero ($k=11$ parameters per genotype; Supplementary file 13). Individuality: Normal priors with fitted mean and standard deviation over all parameters, except the learning rates ${\eta }_{\mathrm{s}},{\eta }_{\mathrm{g}},{\eta }_{\mu }$ and the factors ${\varphi }_{\mathrm{w}},{\varphi }_{\mu }$, which are shared across individual flies of the same genotype, as in the no individuality model ($k=11+6=17$ parameters per genotype). Sharing the learning rates led to a lower Bayesian information criterion (BIC) than a model with individuality in all parameters (not shown). No learning: Same as individuality, except that ${\eta }_{\mathrm{s}}={\eta }_{\mathrm{g}}={\eta }_{\mu }={\eta }_t=0$ and ${\varphi }_{\mathrm{w}}={\varphi }_{\mu }=1$ ($k=11+5-6=10$ parameters per genotype). No genetics: All flies are fitted with the individuality model ($k=17$ parameters), and shared parameters due to genotype are not modelled.

BIC and marginal log probabilities were estimated with 10⁴ samples. The difference in momentary behaviour between two simulated flies was computed with the Hellinger distance (Hellinger, 1909) between the policies, averaged over all possible states.

We simulated tracks for 64 flies per genotype split equally between treatments based on the genotype-specific or individually fitted parameters from the different models.

To decode the genotype from behaviour, we trained a recurrent neural network with one hidden layer of 32 LSTM cells with relu-activation on the behavioural sequences encoded as $(\mathrm{o}\mathrm{n}\mathrm{e}\mathrm{h}\mathrm{o}\mathrm{t}(x_i),s_i,\mathrm{\Delta }{t}_i)$. The feed-forward baseline model contained 2 hidden layers of 32 relu neurons with the input being the number of decisions encoded as 20 dimensional vectors with Gaussian tuning, i.e., for number of decisions $n$, input neuron $i$’s activity is given by $a_i=\exp (-(n-c_i{)}^2/(2\cdot {10}^2))$ with $c_i=(i-1)\cdot 20+2$.

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

We would like to thank Samuel Bourgeat for helpful discussions and help with fly care, Wulfram Gerstner for the compute time on the cluster, the PCB Lab at the EPFL’s School of Engineering for providing equipment and material for the fabrication of printed circuit boards used in the project, and the members of the 'Project Flybot' WhatsApp chat group (Tomislav Štampar, Stjepan Bukal, Marko Jakšić, and Samuel Bourgeat) for the moral and technical support during design of the platform. This study was funded by the ELISIR scholarship.

1. Preprint posted: March 4, 2026
2. Sent for peer review: March 17, 2026
3. Reviewed Preprint version 1: June 3, 2026
4. Reviewed Preprint version 2: August 13, 2026
5. Version of Record published: September 30, 2026

You can cite all versions using the DOI https://doi.org/10.7554/eLife.111235. This DOI represents all versions, and will always resolve to the latest one.

© 2026, Manna et al.

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