High-resolution analysis of individual Drosophila melanogaster larvae uncovers individual variability in locomotion and its neurogenetic modulation

2.1. The IMBA allows the behaviour of individuals within groups to be studied

In order to evaluate our approach, we compared the results with manual assessments of collisions by an experienced experimenter. In 48 videos with approximately 20 larvae each (for an example video, see electronic supplementary material, movie S1), 1844 collisions were found by the experimenter. We attempted to resolve 1531 by means of the IMBAtracker, and of these, 1511 were correctly resolved (electronic supplementary material, figure S1a). In total, 83.7% of all tracked data belonged to tracks that fulfiled our criterion of being at least half the length of the video after collision resolution and were used in the analysis (electronic supplementary material, figure S1b–d). Notably, without the collision resolution implemented, this fraction dropped to 46.7%, demonstrating the importance of this processing step (electronic supplementary material, figure S1d).

To further evaluate IMBAtracker, we applied it to videos from different sources: (i) to videos gathered with the same setup but with smaller Petri dishes (9 cm diameter instead of 15 cm), used for figures 8–13, (ii) to videos gathered at the University of Leipzig with an independently developed setup, used for figures 5–7, (iii) to videos of the dataset ‘LCD2A’ published at www.uni-muenster.de/Geoinformatics.cvmls/media/datasets.shtml, based on [43], and (iv) to videos published along with [26]. The videos from all these sources could successfully be tracked. Electronic supplementary material, movies S2–S5 show the annotated output of the IMBAtracker for one example of each video type, and demonstrate the range of video types and qualities handled by the IMBAtracker. The dataset (iv) was recorded with the popular FIMtrack setup [25,43] and provides short videos, each depicting a collision of two larvae, sorted into three classes of increasing complexity of the collision. We manually assessed the output of the IMBAtracker on these videos and found high rates of correct assessments in all three classes (electronic supplementary material, figure S1e). This demonstrates that the IMBAtracker is able to reliably track and resolve collisions between larvae.

In the next step, we developed an R-based shiny application to analyse and visualize the behaviour of these individual larvae. This application, called the IMBAvisualizer, calculates a total of 95 attributes for each animal that can be explored and visualized in an interactive manner (for all attributes used in this study, see tables 1–4). For an overview of the process, see electronic supplementary material, figure S2. Both the code and a full documentation of this application can be accessed via GitHub at https://github.com/mthane/IMBA and will be regularly updated and further developed in the future.

The locomotion of Drosophila larvae has been studied in detail in the past decade [5–7]. In brief, the locomotion can be described as sequences of run phases, consisting of relatively straight, forward peristaltic movement and reorientation manoeuvres, consisting of lateral head casts (HCs), which are usually associated with a strong reduction in speed and are followed by changes in direction (figure 1a) [8]. Accordingly, we first defined HC phases by the above-threshold angular velocity of the animal's head vector (HV) (figure 1b,c) (for details, see the Material and methods section). Animals were considered to be in a run phase whenever they were not in a HC phase. A period of 1.5 s before and after each HC was excluded from the run phases to prevent the potential effects of HCs on run behaviour. All these definitions are in line with our previous work [28,44]. For a more detailed analysis of run periods, we defined the cycles of peristaltic movement by means of the animal's forward velocity (figure 1d). Each local maximum of forward velocity during a run was defined as one ‘step’, with the period between two steps (inter-step, IS) capturing one complete cycle of peristaltic movement (for a similar approach, see [19]).

• Download figure
• Open in new tab
• Download PowerPoint

2.2. Attributes of basic locomotion in larvae are variable across individuals

On the basis of these observations, and considering the innate, i.e. experimentally naive, behaviour of 875 animals in an odour gradient, we determined a set of six behavioural attributes to describe the basic locomotion of larvae: the IS speed, indicating the average speed across all peristaltic cycles (figure 2a); the IS distance, indicating the average size of the steps (figure 2b); the IS interval, indicating the average time between two steps (figure 2c); the absolute bending angle, indicating how much the body of a larva was bent throughout the whole observation period (figure 2d); the HC rate, indicating how many HCs an animal performed per second (figure 2e); and the absolute HC angle, indicating the average size of an animal's HCs (figure 2f). To determine the variability in each behavioural attribute, we calculated the coefficient of variation (CV), which indicates the standard deviation as a percentage of the mean and allows the inter-individual variability to be compared across different behavioural attributes. The CV was lowest for the IS distance (15.6%) and highest for the HC rate (47.4%).

• Download figure
• Open in new tab
• Download PowerPoint

One important reason for variability across animals is known from previous studies: the size of animals [19]. We therefore investigated the impact of animals’ size on the six behavioural locomotion attributes and found that both IS speed and IS distance were correlated with the body length of a larva, whereas other attributes were not affected (electronic supplementary material, figure S3). This finding is in line with previous work [19]. We therefore decided to correct for the body length in all following analyses by dividing any distance measured by the average body length of the individual larva in question. Distances are reported in body lengths (bl), speeds and velocities in body lengths per second (bl s⁻¹).

2.3. Bending and head-casting behaviour is more variable within individuals than peristaltic forward locomotion

After investigating the inter-individual variability, we explored the intra-individual variability in behaviour (figure 2g–i). We found that attributes related to the peristaltic forward movement such as the IS speed, IS distance and IS interval were much less variable than attributes associated with directional changes such as the absolute bending and HC angles (figure 2i; as the HC rate is calculated as a single value per individual it is excluded from this analysis). This finding suggests that peristaltic forward movement is a highly stable, regular behaviour with little variation, whereas directional changes are much more variable. Interestingly, the directional changes, but not the peristaltic forward movement, were much more variable within each individual than across individuals (figure 2i).

As animals in this dataset were exposed to an odour source, it is conceivable that their behaviour, in particular the variability in their directional changes, may be affected by the individuals’ distance from the odour as well as their orientation with respect to the odour. To account for these potential effects, we analysed the animals’ behaviour separately for low and high distances from the odour source (below and above 60 mm, respectively) and found the same pattern of results, with low variability for the IS speed, IS distance and IS interval, and high variability for the absolute bending and HC angles (electronic supplementary material, figure S4a,b). Likewise, animals oriented towards the odour source (defined as an absolute bearing angle of below 90°) and animals oriented away from the odour source (absolute bearing angle of above 90°) showed the same result (electronic supplementary material, figure S4c,d). Finally, we tested experimentally naive animals in absence of any odour source and found once more the same result (electronic supplementary material, figure S4e).

2.4. Attributes of peristaltic forward locomotion are stable over time within individuals

Next, we wondered how stable the behaviour of individual animals was over time. To this end, we determined each of the six behavioural attributes for each individual in the first and in the third minute of the 3 min video, and asked whether the behaviour in either minute would be correlated with that in the other. We detected mild correlations of the HC rate as well as the absolute bending and HC angles, and strong correlations of the IS speed, the IS distance and the IS interval (figure 3). This suggests that the behaviour of individual animals, though variable, was nevertheless remarkably consistent over time: the fastest animals in the first minute were also the fastest in the third minute, and the animals making the largest HCs in the first minute tended to do the same in the third minute. Next, we extended this analysis with respect to the distance from the odour source as well as the orientation to the odour source. We found strong correlations of the IS speed, IS distance and IS interval, and weaker correlations of the HC rate, the absolute bending and HC angles, when comparing behaviour while the animals were near or far from the odour source (electronic supplementary material, figure S5a–f). A similar result was observed when comparing behaviour while the animals were oriented towards or away from the odour source (electronic supplementary material, figure S5g–l). Finally, we probed for the stability of the behaviour over time using animals that were tested without any odour source. We detected overall weaker correlations in this dataset, with moderate correlations of the IS speed, IS distance and IS interval and virtually no correlations of the HC rate, the absolute bending and the absolute HC angle (electronic supplementary material, figure S6). Across all these analyses, attributes of peristaltic forward locomotion were more stable within individuals than attributes of directional changes.

• Download figure
• Open in new tab
• Download PowerPoint

To put the newly developed technology to the test, we next performed a number of experiments and detailed analyses to quantify modulations of locomotor behaviour. Our strategy was to start from a previously described or expected effect on locomotion, and challenge our software to uncover novel aspects to the data and deeper insights into the organization of locomotion.

2.5. Previous training modulates HC behaviour

First, we preformed detailed analyses of larval locomotion and its modulations in response to previous associative odour-sugar training. Such associative learning experiments are established for about two decades and used in many laboratories [39–41,45–47]. In brief, groups of larvae were trained to associate an odour with a sugar reward by repeated pairings and subsequently tested with the odour source on one side of the Petri dish. Other groups were trained in an explicit unpaired fashion, with odour and sugar being presented in separate trials (figure 4a). We compared the behaviour of those trained animals with experimentally naive animals tested in the same way. Please note that this ‘innate’ condition is the same data as used in figures 2 and 3. All three experiments were done in parallel for a previous study [28]. As expected, the three experimental conditions differ in their preference for the odour: after paired training the larvae approached the odour, after unpaired training they avoided it, and experimentally naive animals showed an intermediate preference (figure 4b).

• Download figure
• Open in new tab
• Download PowerPoint

Previously, we investigated the modulations of behaviour by learning using group averages, and found that mainly two aspects of behaviour were modulated: the HC rate and the HC direction [28,44,48]. Here, we first replicated these results using individual animals (electronic supplementary material, figure S7). Specifically, we found that after paired training larvae (i) were on average more oriented towards the odour source than after unpaired training (electronic supplementary material, figure S7a), (ii) modulated their HC rate to be lower when heading towards the odour source and higher when heading away from it, compared to larvae after unpaired training (electronic supplementary material, figure S7b,c) (iii) and directed their HCs more towards the odour source than after unpaired training (electronic supplementary material, figure S7d). Experimentally naive animals displayed intermediate behaviour in all these cases. In addition, just like in previous studies, we found a non-significant trend of higher speed when heading towards the odour than when heading away from it after paired training (electronic supplementary material, figure S7e,f) [28,44].

Previous studies suggested that odour preference is controlled by a single motor output that integrates all innate and learned behavioural tendencies and causes constant left-right oscillations that are modulated by the presence of attractive or aversive stimuli such as trained odours [28,49,50]. According to this hypothesis, the behavioural attributes in our study, in particular the modulations of the HC rate and HC direction, should be strongly correlated with each other because they are materializations of the same motor output. We indeed found correlations in all three experimental conditions (electronic supplementary material, figure S7g). However, the correlations were rather weak overall, and relatively strongest between the modulations of the HC rate and HC direction. This rather weak correlation may suggest that the HC rate and HC direction, at least partially, can be modulated independently from each other.

2.6. Previous training reduces intra-individual variability

Next, we wondered how the six basic attributes of locomotion would be modulated by previous training. We found several small but significant differences between naive and trained animals, but only in one case a significant difference between animals after paired and unpaired training. Specifically, we found a decreased IS speed and IS distance, but not IS interval, in trained animals compared to experimentally naive animals (figure 4c,e,g). Furthermore, the intra-individual variability in these three attributes was reduced after training (figure 4d,f,h). Regarding the directional changes, naive animals displayed larger absolute bending angles and HC rates, as well as more variable absolute HC angles, compared to trained animals (figure 4i–m). Thus, experimentally naive animals were overall faster, stronger bent and made more HCs than trained animals, and the behaviour of naive animals was overall more variable than that of trained animals. The kind of training, paired or unpaired, made no difference in these aspects of locomotion behaviour, with the exception of a slightly lower absolute bending angle after paired than after unpaired training (figure 4i). Thus, previous training had two fundamental effects: it guided the animals towards or away from the trained odour source by modulating the respective behaviour in opposite ways after paired and unpaired training (electronic supplementary material, figure S7), and it reduced overall activity and variability independent of the kind of training received (figure 4).

2.7. The Cirl mutation reduces speed, bending and head-casting behaviour

After investigating the modulations by training, we performed detailed analyses of larval locomotion and its modulations upon genetic manipulation. To this end, we first chose the Latrophilin homolog Cirl, which is broadly expressed throughout the larval nervous system. Importantly, Cirl^KO mutant larvae are known to exhibit a strong locomotion deficit compared to wild-type animals as well as a genomic Cirl^Rescue control [42], which has not, however, been studied in detail thus far.

We tracked and analysed the behaviour of 225 mutant and 147 rescue control larvae (figure 5). First, we confirmed the previously observed phenotype in crawling speed (figure 5b) [42]. Then, we analysed mutant and control animals regarding the six behavioural attributes described above. We found that the IS speed of the mutants was less compared to the control (figure 5c). Notably, a lower speed can come about in two ways: the larvae may make a slower peristaltic movement, resulting in a higher IS interval, or they may move less per peristaltic cycle, resulting in a lower IS distance. In the present case, we found only a small difference in the IS distance and a large difference in the IS interval (figure 5d,e). An even closer look into the run behaviour revealed that mutant animals are disturbed in their regular peristaltic cycle (see below). In addition to the expected phenotype in speed, we detected a reduced absolute bending (figure 5f), which was accompanied by a reduced HC rate, as well as a reduced absolute HC angle, in the Cirl^KO mutant larvae (figure 5g,h). That is, the mutants made fewer and smaller HCs, and were bending less. This suggests that mutant larvae are generally impaired in the bending of the body and in directional changes, possibly due to impaired functioning of the chordotonal organs. This result was surprising, as in a previous study a qualitative description of the Cirl^KO mutant locomotion implied an increased number of HCs [42].

• Download figure
• Open in new tab
• Download PowerPoint

We also analysed the variability of each of the behavioural attributes within individuals, and observed increased variability in the IS speed and the IS interval in Cirl^KO mutant larvae compared to the controls (electronic supplementary material, figure S8).

2.8. A machine-learning approach reveals additional modulations of locomotion in Cirl^KO mutants

So far, our analysis had focused on the six basic locomotion attributes. But the loss of Cirl may affect behaviour in further ways. We therefore extended our analysis to include additional behavioural attributes, including angular velocities of the head and tail, accelerations of head and tail, as well as the coefficients of variation whenever applicable. A full list of the 30 behavioural attributes used can be found in tables 2 and 3. We then employed an unbiased, machine-learning-based approach to determine the most decisive attributes that differentiated mutants and control larvae. Random forests work by building a large number of decision trees, each of them trying to classify each individual animal according to either one genotype or the other. Importantly, random forests are not only useful in classifying data, but also in estimating which attributes are most important for the classification. For more details, see the Material and methods section. Applying this approach to our dataset, we were able to discriminate mutant and control larvae with an accuracy of 0.91 (figure 6a). Our assessment of which of the behavioural attributes were the most important in discriminating the behaviour of the two genotypes found that most of the top behavioural attributes were those already covered in our previous analysis: the IS speed, the IS interval, the absolute bending and HC angle, as well as the CV of the IS speed and the IS interval (figure 6b). This result confirmed our analyses so far. However, the random forest method also revealed four additional behavioural attributes: the distance travelled by the larva, the absolute angular speeds of the HV and tail vector (TV), as well as the CV of the forward velocity. A closer look revealed that the Cirl^KO mutant larvae travelled shorter distances (figure 6c), consistent with their reduced speed. Furthermore, the mutants had decreased absolute HV and TV angular speeds (figure 6d,e), consistent with the general observation of decreased bending and turning rates in the mutant. Finally, we did not observe any clear difference in intra-individual variability in the tail forward velocity between mutant and rescue (figure 6f). The last result implies that the random forest can use attributes to differentiate individual animals that do not display obvious differences when plotted as box plots.

• Download figure
• Open in new tab
• Download PowerPoint

To confirm that these top ten behavioural attributes were indeed the critical ones for describing the behaviour of Cirl^KO mutants and control animals, we ran the random forest algorithm again, but this time with only those ten attributes. The algorithm was able to discriminate the two genotypes with an accuracy of 0.91, just as high as with the full set of behavioural attributes. The ten attributes in question were thus fully sufficient to discriminate between the genotypes. In turn, when we ran the random forest algorithm with the remaining 20 attributes and excluding the top ten, the accuracy was still 0.83. This suggests some redundancy in our analysis, with multiple attributes providing related information such that the algorithm can employ substitute attributes to discriminate between the genotypes even when the most decisive attributes are excluded.

2.9. Cirl^KO mutant larvae have an altered rhythmic peristaltic cycle

In order to understand the origin of the Cirl^KO phenotype in crawling speed, we analysed the temporal behavioural patterns of individual animals (figure 7). Just like wild-type larvae (figure 1d), Cirl^Rescue controls displayed a very regular rhythm in their forward velocity, indicated by a smooth, wave-like curve with local maxima of very similar heights that we detected as steps, and always one local minimum between two steps (figure 7a, top; electronic supplementary material, movie S6). Cirl^KO mutants, in contrast, regularly exhibited a pattern of two local minima with one very small local maximum between two steps (figure 7a, bottom; electronic supplementary material, movie S7), resulting in IS intervals that were longer and more variable than those of Cirl^Rescue controls (figure 5e, electronic supplementary material, figure S8c). For further analyses, we defined each period between two steps with only one local minimum in the tail forward velocity as normal stepping, and each period with two local minima as ‘stumble’ stepping. A quantification revealed that stumble stepping was much more common in mutant than in control larvae, and that their proportion increased with increased IS intervals (figure 7b,c).

• Download figure
• Open in new tab
• Download PowerPoint

Next, we directly compared the animals’ behaviour specifically during normal stepping and stumble stepping. We found that larvae were slower and had longer IS intervals, but unchanged IS distances during stumble stepping compared to normal stepping (figure 7d–f). In other words, stumble steps took longer to complete, but were not larger. Moreover, neither the animals’ bending nor the angular speed of their HV was different between normal and stumble steps; only the angular speed of the TV was slightly decreased (figure 7g–i). This suggests that the phenotype in bending and turning that we observed was largely independent of the irregular stepping behaviour of the Cirl^KO mutants.

In summary, we found that the Cirl^KO mutants’ phenotype in crawling speed can be at least partially attributed to an unusual stepping pattern that causes a slower peristaltic forward locomotion, and a largely independent phenotype in bending and turning.

2.10. Optogenetic activation of dopamine neurons triggers stopping and bending

Next, we investigated how the acute activation of genetically defined sets of neurons affected the locomotion of individual larvae. First, we applied our unbiased approach to the behaviour of Drosophila larvae upon optogenetic activation of dopamine neurons. To this end, we used larvae that expressed the highly sensitive effector Channelrhodopsin2-XXL (ChR2-XXL) [51] in a broad subset of dopaminergic neurons, as defined by the TH-Gal4 driver strain (experimental genotype). Groups of larvae were allowed to move freely over a Petri dish for 30 s in darkness, followed by 30 s of blue light in order to trigger the activation of the dopaminergic neurons, and another 60 s of darkness. The behaviour of the animals was compared to that of heterozygous driver and effector controls that did not express ChR2-XXL (figure 8). Upon light stimulation, animals of all three genotypes increased their absolute bending, suggesting that this was a response to the sudden light. However, the experimental genotype did so to a much greater extent than the genetic controls, suggesting that this stronger reaction was caused by the activation of the dopaminergic neurons (figure 8b, electronic supplementary material, movie S8).

• Download figure
• Open in new tab
• Download PowerPoint

In order to evaluate the reaction of the animals to the light stimulation and the light-induced neuronal activation, we normalized each individual animal's behaviour during the 30 s of light stimulation by subtracting its behaviour in the 30 s of darkness before. The normalized values of 30 behavioural attributes (called Δ-values) were fed into the random forest. The random forest classified the three genotypes with an accuracy of 0.90 and provided the top ten most important attributes that differed between the three genotypes in their change in response to the light stimulation (figure 8c). In detail, we found a more strongly reduced distance travelled by larvae of the experimental genotype than of the genetic controls and accordingly a more strongly reduced tail forward velocity and midpoint speed (figure 8d–f). Interestingly, the intra-individual variability in the midpoint speed, as determined by the CV, was increased in all three genotypes, but more in the experimental genotype (figure 8g). In addition, the random forest approach provided three behavioural attributes regarding the animals’ bending. For the bending angle, which is positive if a larva is bent to the left and negative if bent to the right, we found no significant differences, but a strongly increased scatter of the data in the experimental genotype (figure 8h). Fittingly, the absolute bending angle, not taking the direction into account, was increased in all three genotypes, but much more so in the experimental genotype (figure 8i). The intra-individual variability in the absolute bending angle was increased only in the genetic controls, but not in the experimental genotype (figure 8j). Finally, the angular speeds of the animals’ head and TVs also differed between the genotypes: both angular speeds were increased more strongly in the experimental genotype than in the controls (figure 8k,l). Regarding the TV angular speed, the intra-individual variability was also more strongly increased in the experimental genotype (figure 8m).

• Download figure
• Open in new tab
• Download PowerPoint

In summary, activating a broad set of dopaminergic neurons caused the animals to slow down markedly and to bend and rotate their bodies (for a time-resolved view of the three genotypes’ behaviour, see electronic supplementary material, figure S9).

2.11. Individuals reduce speed consistently upon repeated activation of dopamine neurons

In wild-type animals we observed that behavioural attributes associated with peristaltic forward crawling were more consistent within each animal than attributes associated with bending and HC behaviour (figures 2 and 3; electronic supplementary material, figures S3–S6). We therefore wondered whether this would also be true for the dopamine-induced changes in behaviour that we had observed. To tackle this question, we repeated the previous experiment with the experimental genotype expressing ChR2-XXL in TH-Gal4 neurons, but this time we activated the neurons three times for 10 s, each time followed by 60 s of darkness (figure 9). Then we determined the changes in the four behavioural attributes that had previously been found to be most important by the random forest method (figure 8c): tail forward velocity, midpoint speed, absolute bending angle and absolute angular speed of the HV. On average, the animals reacted very similarly to the three neuronal activations (figure 9). As regards the changes in behaviour within each animal, however, we found differences between the behavioural attributes. We observed mild to moderate correlations across the three neuronal activations for the tail forward velocity and the midpoint speed (figure 9b,d; electronic supplementary material, figure S10a–d), but no correlations for the absolute bending angle and the absolute angular speed of the HV (figure 9f,h; electronic supplementary material, figure S10e–h).

Thus, we indeed found that the reduction in speed upon activation of dopamine neurons was consistent within individual animals, whereas the increases in bending and lateral head movement were not. This further stresses the distinction between forward crawling and bending behaviour with respect to intra-individual variability.

2.12. Optogenetic activation of ‘mooncrawler’ neurons triggers backward crawling

In order to probe the utility of our approach further, we applied it to an analogous experiment to that described above, but activating a very different set of neurons. This time we chose the driver strain R53F07-Gal4, which reportedly covers a rather broad set of interneurons in the ventral nerve cord (VNC), the insects’ functional analogue of the spinal cord, including the so-called moonwalker descending neurons (MDNs) [52]. Optogenetically activating these neurons (henceforth called ‘mooncrawler’ neurons) had caused larvae to crawl backwards in a previous study [52], but an in-depth analysis of their behaviour is lacking. Therefore, we crossed R53F07-Gal4 to the UAS-ChR2-XXL effector strain, performed the same experiment as described above, and compared the animals’ behaviour with the respective genetic controls (figure 10). Light stimulation did indeed cause backward crawling, as indicated by a negative forward velocity, in animals of the experimental genotype but not of the genetic controls (figure 10b, electronic supplementary material, movie S9).

• Download figure
• Open in new tab
• Download PowerPoint

We applied the same strategy as before, fed the Δ-values of behavioural attributes into the random forest (accuracy 0.83), and obtained the top ten attributes that differed between the three genotypes in their change in response to the light stimulation (figure 10c). Similarly to the case of dopaminergic neurons, we found the distance travelled, the tail forward velocity and the midpoint speed to be more strongly decreased in the experimental genotype than in the genetic controls, and the intra-individual variability of the midpoint speed to be more strongly increased (figure 10d–g). By contrast to the previous experiment, however, the IS speed and the IS interval were also among the top ten attributes: we found the IS speed to be strongly reduced, and the IS interval to be strongly increased, in the experimental genotype but not the genetic controls (figure 10h,i). Again, we found the animals’ bending to be important, yet in a different way from the previous experiment: the absolute bending angle was reported as an important attribute by random forest, but the experimental genotype was different only from the effector control, not from the driver control (figure 10j). Thus, there was no evidence that the activation of the R53F07-Gal4 neurons were responsible for the change in the absolute bending angle. Interestingly, the intra-individual variability in the absolute bending angle was increased in the genetic controls but decreased in the experimental genotype (figure 10k). Although the change in the average bending was not significantly different from the genetic controls, thereby activating the R53F07-Gal4 neurons made the bending less variable. Finally, the angular speed of the HV as well as the HC rate were reported by the random forest. However, for the angular speed we found a significant difference in the experimental genotype only in relation to the driver, but not the effector control (figure 10l), and no significant differences in the HC rate at all (figure 10m).

2.13. Individuals differ in the timing of their switching gears from backward to forward locomotion

A closer look at the temporal patterns of the behavioural attributes induced by the two sets of neurons activated revealed striking differences. The activation of dopaminergic neurons triggered a number of behavioural changes that quickly reverted back to normal as soon as the light stimulation ended (electronic supplementary material, figure S9). By contrast, activating the R53F07-Gal4 neurons triggered behavioural changes of very different, more complex dynamics (electronic supplementary material, figure S11). For example, the tail forward velocity dropped quickly to the negative upon light onset and started to become more positive gradually over the course of 90 s (electronic supplementary material, figure S11a). The midpoint speed and IS speed, in contrast, continued to drop slowly until a minimum was reached about 30–40 s after light onset (electronic supplementary material, figure S11b,c). The absolute bending angle was increased in all three genotypes during light stimulation. But whereas it reverted back to normal upon light offset in the genetic controls, it remained high for another 30 s in the experimental genotype before gradually decreasing (electronic supplementary material, figure S11e). What is the cause of these complex temporal patterns?

We compared the temporal pattern of the tail forward velocity across individual animals (electronic supplementary material, figure S12). Whereas all the individuals switched to backward crawling within a few seconds after light onset, the reverse switch to forward crawling was highly variable, with some animals not reverting at all within the recording time (figure 11a). In other words, most larvae continued to crawl backwards even after the optogenetic activation of the R53F07-Gal4 neurons had finished, and each reverted to forward crawling in its own time. On average, those animals that did revert to forward locomotion within the recording time crawled backwards for about 51 s (figure 11b). This also means that our previous analysis was confounded by mixing unknown numbers of forward and backward crawling individuals at any given time point. Therefore, we determined the switching points in each individual animal (electronic supplementary material, figure S12) and compared the individual animals’ behaviour while the animals were crawling backward (i.e. between the switching points) with their behaviour while crawling forward (figure 11c).

• Download figure
• Open in new tab
• Download PowerPoint

We found that the midpoint speed was reduced, but its intra-individual variability was increased, while animals crawled backwards (figure 11d,e). Taking only run phases into account, the IS speed was reduced and the IS interval was accordingly increased, during backward crawling (figure 11f,g). The absolute bending angle was increased, with reduced intra-individual variability, while larvae crawled backwards (figure 11h,i). Finally, we found no differences between forward and backward crawling regarding the angular speed of the HV, but a reduced HC rate during backward crawling (figure 11j,k).

In summary, backward crawling was characterized by longer IS intervals, leading to a reduced and more variable speed, and by increased but less variable bending of the larvae, compared to forward crawling.

2.14. Switches between forward and backward locomotion are characterized by changes in speed, and peaks in bending and head-casting behaviour

Finally, we wondered how specifically the animals’ behaviour was altered during the switches between forward and backward locomotion. To this end, we aligned the data of all the animals either to the switch to backward crawling, or to the switch to forward crawling (figure 12a). This alignment revealed that both midpoint speed and IS speed rapidly dropped at the switch to backward crawling, followed by a slow further decrease in the following seconds. Both speeds decreased until a few seconds before the switch to forward crawling and quickly rose again after the switch (figure 12b,c). Fittingly, the IS interval increased after the backward switch and decreased again after the forward switch (figure 12d). The absolute bending angle rapidly increased precisely at the backward switch and stayed constant afterwards. At the forward switch, the absolute bending angle peaked and then dropped quickly back to normal levels (figure 12e). The absolute angular speed of the HV did not change at all at the backward switch but peaked at the forward switch (figure 12f). Finally, the HC rate peaked at both switching points, albeit on different absolute levels (figure 12g).

• Download figure
• Open in new tab
• Download PowerPoint

In order to determine how the direction of crawling changed at each switching point, we calculated the change in the TV in the 10 s before and after each switch (figure 12h). We found that in both cases the larvae changed their crawling direction by about 100°, suggesting that either switch was followed by a movement approximately orthogonal to the previous direction.

We characterized backward switches by a decrease in speed and an increase in absolute bending angle, and forward switches by an increase in speed, and peaks in absolute bending angle and absolute angular speed of the HV. How stable were these changes for repeated switches across and within individuals? To address this question, we stimulated the R53F07-Gal4 neurons three times for 10 s and aligned all animals either to the backward or to the forward switch (figure 13). We found that the average behaviour of the animals was astonishingly consistent across all three backward and forward switches. Looking at individual behaviour with regard to the changes in tail forward velocity and midpoint speed, we observed mild to moderate correlations across repeated backward switches (figure 13a,c; electronic supplementary material, figure S13a,c), but not across repeated forward switches (figure 13b,d; electronic supplementary material, figure S13b,d). A notable exception was a significant moderate correlation in the midpoint speed between the second and third forward switch (electronic supplementary material, figure S13d). By contrast, no correlations were found for the absolute bending angle and the absolute angular speed of the HV (figure 13e,f,h; electronic supplementary material, figure S13e,f,h). The IS speed, IS interval and HC rate were not included in this analysis, as steps and HCs occur too rarely within the short observation periods used for this analysis.

• Download figure
• Open in new tab
• Download PowerPoint

In addition, we determined the duration of each of the three backward crawling phases upon each activation of R53F07-Gal4 neurons in individual animals, and found positive correlations that were significant between the first and the second, as well as between the second and third backward phases (electronic supplementary material, figure S13g). Between the first and third backward phase too, a positive correlation was observed; this remained non-significant due to the sample size of only seven animals that completed both backward phases (figure 13g).

In summary, we again found that changes in speed upon neuronal activation were more consistent within individuals than changes in bending or lateral head movement, just as we found to be the case with respect to dopamine neurons.

3.1. Bending and head-casting behaviour is variable; peristaltic locomotion is stable

Our analysis revealed that peristaltic forward locomotion is characterized by very low intra-individual variability and high consistency over time (figures 2 and 3; electronic supplementary material, figures S4–S6). Bending and head-casting behaviour, in contrast, is much more variable, and less stable. We note that, while similar findings have led to the suggestion that peristaltic forward locomotion would thus be best suitable for high-throughput phenotyping [19], we found reliable effects of transgenic manipulations on bending and HC behaviour, too (figures 5f–h, 8i–m and 10j–m). In any event, when we repeatedly activated the same set of neurons in the same individual animals, we found mild yet significant correlations for these repetitions for peristaltic forward locomotion, but not for bending and head-casting behaviour. This was true for both dopamine neurons and mooncrawler neurons (as covered by the R53F07-Gal4 driver) (figures 9 and 13). This suggests that individual larvae not only have a fairly stable, inherent rhythm of peristaltic locomotion, but they also modify this rhythm in a relatively idiosyncratic manner upon repeated neuronal activation. Bending and head-casting behaviour, on the other hand, appear highly variable and are modulated not only by previous experience (electronic supplementary material, figure S7) but also in an unpredictable manner upon repeated neuronal activation.

The reason for this apparent difference in variability between peristaltic locomotion on the one hand and bending and head-casting behaviour on the other hand is not known. Interestingly, however, previous studies have found that peristaltic locomotion originates in both the anterior and posterior segments of the VNC, whereas head-casting behaviour and bending originate more anteriorly [53,54]. In both cases, input from the brain is reportedly not required. Input from the brain is required, in contrast, for organizing peristalsis, bending and head-casting behaviour in a goal-directed manner [53]. Consistently with this, we repeatedly found that ‘higher’ brain functions, such as associative memories, almost exclusively modulate head-casting behaviour, but not peristaltic runs (electronic supplementary material, figure S7) [28,44,48]. Overall therefore it seems that the characteristics of peristaltic runs that are particularly relevant for speed are relatively stable features of a given individual animal, ‘hard-wired’ in the VNC, whereas bending and head-casting behaviour, and thus changes in direction, are more flexible features that can be adaptively modulated by external stimuli and internal goals, modulations that apparently take place via the brain. We are curious whether such a distinction between speed-related, hard-wired, stable motor outputs from the VNC and direction-related, flexible, brain-modulated behaviours may turn out to be a general principle of behavioural control.

We note that intra-individual variability itself can be affected by specific experimental manipulations: for example, we found that the Cirl^KO mutants show an increased variability in their IS interval (electronic supplementary material, figure S8). The optogenetic activation of select sets of neurons can also elicit changes in the variability of behaviour (figures 8g,j,m and 10g,k). One might therefore expect there to be cases in which experimental manipulation specifically affects the variability of a behavioural measure, but not its average. This was indeed the case when comparing innate locomotion with locomotion after a previous associative training (figure 4). We found a couple of differences, for example in the speed and the HC rate that likely reflected some level of fatigue after the training procedures. Importantly, in four out of five locomotor attributes, trained larvae displayed a lower intra-individual variability than experimentally naive animals, including in the IS interval and the absolute HC angle which did not differ in their average (figure 4g,h,l,m). Such decreased levels of variability may reflect a loss of degrees of freedom by a training during which animals learn how to pursue their goal (i.e. finding food). Similar effects on the variability of behaviour would be interesting in particular in relation to processes for which behavioural variability is of particular value, e.g. for generating unpredictable actions during escape, for generating novel actions in pursuit of desired outcomes in operant learning, or for adapting to changed environmental contingencies during reversal learning.

Our findings suggest that individual larvae may have behavioural traits that are different between, but consistent within, individual animals. However, given that we only included videos of a maximal length of 4 min, it is possible that even the attributes that are relatively stable within animals, such as those related to peristaltic locomotion, reflect transient internal states or ‘moods’ rather than permanent behavioural idiosyncrasies. To detect such idiosyncrasies, longer recordings and ideally repeated recordings of the same individual over several days or even across metamorphosis would be required. In adult Drosophila, repeated behavioural measurements demonstrated such behavioural idiosyncrasies even in strongly inbred fly populations [16,17]. In principle, inter-individual behavioural variation can originate from genetic, developmental and environmental factors. A previous study found highly significant correlations between the expression patterns of many genes and a broad array of behavioural attributes [17]. A particularly striking example of a genetic factor causing inter-individual differences in behaviour is the polymorphism of the rover gene that causes different foraging strategies in larvae [15]. On the other hand, non-heritable, random asymmetries in the wiring of a specific set of visual interneurons were found responsible for inter-individual variability in a visual task studied in adult Drosophila [16]. Whether the inter-individual differences we report in this study reflect behavioural idiosyncrasy, and what would be their causes, remains to be determined.

3.2. Cirl^KO mutants exhibit ‘stumble’ steps

We discovered a new, complex behavioural phenotype of Latrophilin/Cirl mutants. Cirl constitutes an aGPCR (adhesion G protein-coupled receptor) expressed in larval mechanosensory organs crucial for proprioception [42]. Previously, it was known only that mutants were slower, as was confirmed by our analysis (figure 5b). Here we found that the reason for the reduced speed was the abnormal ‘stumble’ stepping of the mutant animals, in which the peristaltic cycle was interrupted, and the time required to complete one cycle was therefore prolonged (figure 7). Interestingly, this different way of stepping was in itself rather regular, such that some mutant individuals always performed stumble steps (figure 7a).

The stumble stepping explained parts of the observed behavioural phenotype: the IS interval was increased, and the IS speed was thus decreased during stumble stepping compared to ‘normal’ stepping in the same individual mutant animals (figure 7d,f). However, the normal stepping phases of mutant and control animals also differed in these behavioural attributes, suggesting that the peristaltic cycles that we classified as ‘normal’ were partially impaired as well. Moreover, we found that mutant animals bent and turned less than control animals, a phenotype that was completely independent of the kind of stepping the animals performed (figure 7g,h).

The molecular mechanisms through which aGPCRs exert their function remain incompletely understood. However, it is known that chordotonal neuron-located Cirl suppresses cAMP levels to shape larval proprioception and thus behaviour [42,55]. If and how Cirl has an impact in other cells that are part of the neuronal circuits that modulate locomotion remains elusive. Our analysis suggests that the correct functioning of the receptor is required for a normal, uninterrupted peristaltic cycle, but that even without the receptor larvae can eventually crawl forward. This is in line with published accounts showing decreased but existent proprioceptive capability [42]. It will be interesting to study larval locomotion in animals that express genetically modified Cirl proteins or specific Cirl isoforms [56], and the ability of Cirl mutants to successfully reach goals such as odours or tastants despite their impairment in crawling.

3.3. Activation of dopamine neurons makes larvae stop and turn

Across the animal kingdom, the dopaminergic system is closely associated with locomotion and the generation of movement [57,58]. In humans, the progressive death of dopaminergic neurons is widely accepted as a cause of Parkinson's disease, which results in difficulties in initiating movement [59–61]. Based on studies mostly of rodents, locomotion has been found to be controlled by interactions of several dopamine receptors in the ventral striatum, with some of them increasing, others inhibiting forward locomotion [62]. Turning behaviour can be induced by the unilateral administration of dopamine receptor agonists [63]. The administration of cocaine, which strongly activates the dopaminergic system, causes complex, concentration-dependent modulations of locomotion, including increased forward locomotion upon administration of lower concentrations of cocaine, and circling behaviour as well as a reduced forward movement upon repeated, or highly concentrated cocaine administration [64–66].

In Drosophila too, the dopaminergic system is linked to locomotion. Genetic fly models for Parkinson's disease show similar impairments in locomotion, as well as the death of dopaminergic neurons as found in mammals [67]. Cocaine administration leads successively to a reduction in forward movement, circling behaviour and rotating, before the animals stop moving at all, similar to the case in mammals [68]. In a pioneering optogenetic study, activation of the dopamine neurons covered by TH-Gal4 resulted in increased locomotion in flies that were little active before, and decreased locomotion in flies that were highly active [69]. Recently, a study has found the activity of specific dopamine neurons to be correlated with speed and turning of flies, suggesting that dopamine neurons both report and change specific movements [70].

Much less is known about the impact of dopamine on locomotion in Drosophila larvae. In this study, we optogenetically activated most dopamine neurons, and found that larvae stop forward locomotion and start strong bending (figure 8). Such behaviour would correspond to the behaviour of adult flies and mice after cocaine administration [64–66,68], and to what was observed in highly active flies upon activation of dopamine neurons [69]. Our results are also in line with one of the very few studies that link dopamine function with locomotion in larvae: a transgenic knockdown of the dopamine receptor Dop1R1, one of the two Drosophila homologues of the vertebrate D1-type dopamine receptors, was reported to increase locomotion in larvae [71]. Our experimental treatment, in contrast, acutely increased dopamine signalling, leading to reduced forward locomotion.

Interestingly, the effects of the Dop1R1-knockdown were also observed when it occurred locally only in the mushroom body, one of the central brain regions in insects [71]. The dopamine neurons innervating the mushroom body play a crucial role in signalling reward or punishment information in associative learning [30,72–76]. The mushroom body-innervating neurons covered by TH-Gal4 are mostly linked to punishment signalling, with the reward-signalling dopamine neurons notably missing in the expression pattern [72,77,78]. This raises the question of what the activation of these neurons ‘means’ to the animals. On the one hand, stopping and turning can be interpreted as an avoidance behaviour in response to a sudden punishment signal (for a similar interpretation regarding the activation of dopamine-downstream neurons, see [79]). On the other hand, an excessive release of dopamine, as probably happens in our experiment, may cause malfunctions in the motor control pathways, similar to what was observed upon cocaine administration in adults [68]. The fine genetic tools available make the Drosophila larva an ideal study case to answer this question. Dopamine neurons can be manipulated in small subsets and even one by one [73–75] to find out which neuron causes which behavioural reaction.

3.4. Switching from forward to backward locomotion, and back again

In free, undisturbed locomotion, Drosophila larvae normally crawl forward, and backward crawling is very rarely observed. They do, however, perform short bouts of backward locomotion on encountering an obstacle or receiving a noxious stimulus to their head [80,81]. Thus, backward locomotion can be seen as part of an avoidance reaction. In recent years, several Gal4 driver strains have been described that cover neurons which induce backward locomotion upon activation [52,81,82]. One of them, R53F07-Gal4, we used in the present study [52]. We described many facets of the behaviour of individual animals over extended phases of backward locomotion, including the bending and head-casting behaviour that had not been looked at in previous studies [52,81,82]. We found, for example, that during backward locomotion bending was increased in amplitude, but decreased in variability, and that the HC rate was decreased compared to forward locomotion (figure 11h–k). Moreover, we found that larvae did not revert to forward locomotion immediately after the optogenetic activation terminated. Rather, they continued to crawl backward for various time periods, and then quickly switched to forward locomotion (electronic supplementary material, figure S12). Both kinds of switches were accompanied by characteristic changes in speed and bending behaviours (figures 12 and 13). The reason for the inter-individual variability in reverting to forward locomotion is unknown. In this specific case, in addition to general genetic variation, differences in transient internal states or behavioural idiosyncrasies (see above), also differences in transgene expression strength may contribute to the inter-individual variability.

At the neuronal level, both forward and backward locomotion are characterized by waves of activity in the motor neurons and interneurons of the segmented VNC, which travel from anterior to posterior during forward locomotion, and the other way around during backward locomotion [8,53,83,84]. More specifically, two pairs of command-like neurons have been identified in the rather broad expression pattern of R53F07-Gal4, called mooncrawler descending neurons (MDNs), which inhibit forward and induce backward locomotion [82]. It is likely that the effects of activating R53F07-Gal4 neurons come about through these MDNs. Two main effects of activating the MDNs have been proposed: on the one hand, one of their main downstream partners, called Pair1, stops forward locomotion by inhibiting the forward movement-inducing A27h neuron; on the other hand, another downstream partner, A18b, induces backward locomotion specifically in the most anterior segment of the VNC [11,14,82]. Remarkably, both the MDNs and Pair1 neurons persist through metamorphosis, and induce analogous behavioural effects in adult flies [82,85,86].

It is a straight-forward assumption that the switches between forward and backward locomotion that we describe here may be caused by the backward movement-inducing A18b and the forward movement-inducing A27h neurons, respectively. The observation that larvae did not revert to forward locomotion as soon as the optogenetic activation ended suggests that activating the MDN → A18b pathway leads to backward locomotion that continues until forward locomotion is triggered again, potentially by A27h. Our behavioural approach, together with the available tools for fine circuit analysis, is suited to test this hypothesis.

3.5. Utility for analysing individual behaviour

The IMBA software introduced here can serve to track, visualize and analyse the behaviour of individual Drosophila larvae within groups. The IMBA is compatible with well-established, standard Petri dish assays and requires no specific additional hardware beyond reasonable light conditions and a camera. In this study, we only used videos of up to 4 min, as many standard Petri dish assays are performed in a similar time frame. Longer videos can be processed by the IMBA, but larvae tend to follow the Petri dish walls where they are very difficult to track. Therefore, attempts to use longer videos require precautions such as larger behavioural arenas, or borders that do not reflect light and therefore allow for the tracking of the larvae. In any case, the present study features data from setups located both at the Leibniz Institute for Neurobiology and the University of Leipzig, which use different means of lighting, cameras and video quality. We also successfully tracked videos from previous publications from other laboratories [26,43]. In total, we used five kinds of videos with different resolutions, cameras and light settings which we could successfully track. Electronic supplementary material, movies S1–S5 show the results of these tracking attempts and demonstrate the wide variety of video qualities that could be handled. Such robustness makes the IMBA a useful tool for studying a broad range of research questions using derivatives of this standard assay. In the present study, we focused on locomotion and used the IMBA to compare innate and learned behaviour in odour gradients (figures 2–4, electronic supplementary material, figures S4–S7), phenotype a mutant (figures 5–7) and study the effects of optogenetic neuronal activation (figures 8–13). Other future applications may include the screening of pharmacological effectors [87] and libraries for RNAi knockdown, and comparisons across inbred strains of Drosophila or across species.

A critical property of the IMBA that allows for such broad use is that we included a large number of behavioural attributes that can be adapted and extended by the user (the full list is available online under https://github.com/mthane/IMBA). Among them are 22 attributes describing behaviour relative to a local stimulus. This makes the IMBA useful for experiments on behaviour towards stimuli such as odours, tastants, light or temperature, as well as for analysing stimulus-related behaviour after various types of learning experiment (as we exemplified in electronic supplementary material, figure S7). Importantly, we combine this large number of behavioural attributes with an approach to selecting the most relevant ones for each particular dataset in an unbiased way. Random forest analysis provides a ranking of behavioural attributes according to their importance for differentiating animals in the particular experimental conditions of the particular dataset under study (compare figures 8c and 10c). This includes attributes that do not differ in their centres of distribution (and thus result in non-significant standard statistical tests) but rather in the distribution of the data (figures 6f and 8h).

The IMBA is versatile for analysing relatively fine-grained temporal patterns of behaviour. This has led us to discover stumble stepping in Cirl^KO mutants (figure 7). Indeed, the IMBA allows for the analysis of even more extreme kinds of ‘silly walks’ that it was not designed to study, such as the backward locomotion elicited by activation of MDN neurons (figures 10–13). Furthermore, the IMBA lends itself to protocols that involve repeated experimental stimulation, as in the case of the repeated activation of select sets of neurons (figures 9 and 13). This should make it possible to establish larval versions of paradigms such as prepulse inhibition or repetition priming, which are workhorse paradigms in the experimental neuropsychology of rodents and humans. Last but not least, the IMBA allows data to be aligned to changes in behaviour (that is, to the time of action initiation rather than stimulus application), revealing effects that would otherwise be masked by differences in action-timing between individuals (figures 12 and 13).

A crucial step for the above-mentioned applications is the ability to resolve collisions. Obviously, some of the results we reported here could also have been obtained at the level of sub-individual tracks or group averages. But many of our findings required to analyse the behaviour of individuals, such as the comparison between inter- and intra-individual variability, the repeated neuronal activation experiments or the alignment to the switches between forward and backward locomotion. Without collision resolution, such analyses would either require recording animals one by one, or would need to be made on a much smaller database, restricted to those few animals that happen to not collide with others, which is potentially not representative of the whole population of the experiment (electronic supplementary material, figure S1d). We are not aware of any other ready-to-use tracking software that can perform similar analyses for animals tested in groups. The arguably most complete and widely used tracking approach for larvae so far, the FIMtrack, does not yet include such collision resolution [24,25,31]. In addition, the FIMtrack software provides basic attributes like bending and velocity, but no more detailed attributes like the inter-step distance and interval, HC angles or angular speeds that we employed here. Modifying the FIMtrack is possible but requires at least basic programming skills. In table 5, we compare the capabilities between the FIMtrack and the IMBA.

In summary, the IMBA is an easy-to-use toolbox providing an unprecedentedly rich view of behaviour and behavioural variability across the multitude of biomedical research contexts in which larval Drosophila can be used.

4.1. Datasets

We used seven different datasets for our analysis:

4.2. Behavioural experiments

4.3. Tracking

We developed the IMBAtracker to track multiple larvae on a single Petri dish while keeping track of their identity. Although several tracking solutions have been developed in recent years [18–28], the development of new tracking software seemed necessary to ensure two important aspects: first, increasing and controlling the overall measurement precision according to the needs of the analysis; second, reducing the data fragmentation caused by collisions or by the larvae leaving the region of interest. This attempt to reduce fragmentation is especially important since it paves the way for the individual-level analysis.

4.3.1. Architecture

The tracker is designed to process a region of interest (the Petri dish area), which is defined per video. We do not require the region to contain all the objects of interest at all times (i.e. larvae can exit the region of interest). Videos should be made from above or below with a minimum resolution of 30 pixels per animal (although higher resolutions improve accuracy). Either a light background with dark objects or a dark background with light objects is supported, and any colour information is discarded as the image is converted to greyscale. In order to deal efficiently with the issue of fragmentation, it is crucial to separate the tracking into three distinct phases (see below). The general architecture is depicted in figure 14.

• Download figure
• Open in new tab
• Download PowerPoint

4.4. Data analysis

The main output of the IMBAtracker is a time series of 12 spine points that reach from the head of the animal to its tail, as well as 24 contour points (figure 1). These data are the basis for all further data analyses and visualizations with the IMBAvisualizer. To prevent multiple tracks from the same individual animal from entering the analysis, we discarded all tracks shorter than half of the video length after the collision resolution.

The spatial resolution in our video recordings was not sufficient to resolve the peristaltic wave along individual body segments. However, we noted that the rhythmic velocity changes of the tail point provided a reliable indicator for the occurrence of one peristaltic forward movement (figure 1). We therefore defined each local maximum in the tail forward velocity as one step, and the period between two maxima, i.e. one complete peristaltic cycle, as an inter-step (IS) period (figure 1d). Since the local-maxima detection algorithm we employed also produced false maxima during intervals of little or no forward motion, we discarded maxima detected whenever the tail speed in a forward direction was less than 0.6 mm s⁻¹.

Lateral HCs are a very prominent feature during larval chemotaxis on a plain agarose surface. We quantified these head movements based on the angular speed of the HV, where the HV was defined as the vector between the ninth and the eleventh spine point (figure 1b). In order to detect HCs to the left (or right) we first detected all intervals of angular speed smaller than −35° s⁻¹ (or larger than 35° s⁻¹) (figure 1c). The start and endpoints of these intervals correspond to salient movements of the head part of the animal to the left or right side. We next introduced two criteria in order to discard incorrectly assigned HCs. (1) We discarded any initially detected HC whenever the angular speed of the TV during head casting exceeded 45° s⁻¹. This criterion takes account of the visual observation that whenever the tail parts of the animal reorient, these periods most likely do not coincide with an active lateral head casting movement, but rather a stretching out of the body bending angle when stepping is resumed. (2) We discarded any initially detected HC whenever more than one step was detected during head casting. This criterion takes account of the observation that any active lateral head casting interval can coincide with at most one forward stepping detection point.

After the detection of runs, steps and HCs, we calculated a number of behavioural attributes. In the following we provide a detailed description of all the attributes used in this study. A full list of attributes implemented by the IMBAvisualizer is available online (https://github.com/mthane/IMBA). Please note that for some attributes two versions exist, using either mm or individual body lengths (bl) as the unit. Unless explicitly mentioned in the results, it was always the version using body lengths that was applied.

4.5. Statistical tests

Two-tailed non-parametric tests were used throughout. When multiple comparisons were performed within one analysis, a Bonferroni–Holm correction was applied to keep the experiment-wide error rate below 5% [106]. Values were compared across multiple independent groups with Kruskal–Wallis tests (KW tests). In the case of significance, subsequent pair-wise comparisons used Mann–Whitney U-tests (MW tests). In cases of within-animal comparisons we used Wilcoxon signed-rank tests (WS tests). For correlations, we employed Spearman's rank correlation test (SC tests).