Systematic exploration of unsupervised methods for mapping behavior

Abstract

To fully understand the mechanisms giving rise to behavior, we need to be able to precisely measure it. When coupled with large behavioral data sets, unsupervised clustering methods offer the potential of unbiased mapping of behavioral spaces. However, unsupervised techniques to map behavioral spaces are in their infancy, and there have been few systematic considerations of all the methodological options. We compared the performance of seven distinct mapping methods in clustering a data set consisting of the x-and y-positions of the six legs of individual flies. Legs were automatically tracked by small pieces of fluorescent dye, while the fly was tethered and walking on an air-suspended ball. We find that there is considerable variation in the performance of these mapping methods, and that better performance is attained when clustering is done in higher dimensional spaces (which are otherwise less preferable because they are hard to visualize). High dimensionality means that some algorithms, including the nonparametric watershed cluster assignment algorithm, cannot be used. We developed an alternative watershed algorithm which can be used in high-dimensional spaces when the probability density estimate can be computed directly. With these tools in hand, we examined the behavioral space of fly leg postural dynamics and locomotion. We find a striking division of behavior into modes involving the fore legs and modes involving the hind legs, with few direct transitions between them. By computing behavioral clusters using the data from all flies simultaneously, we show that this division appears to be common to all flies. We also identify individual-to-individual differences in behavior and behavioral transitions. Lastly, we suggest a computational pipeline that can achieve satisfactory levels of performance without the taxing computational demands of a systematic combinatorial approach.

A neural circuit basis for context-modulation of individual locomotor behavior

“…Individual measures of fly behavior (Ayroles et al, 2015;Bierbach et al, 2017;Buchanan et al, 2015;Freund et al, 2013;Kain et al, 2012;Linneweber et al, 2019;Schuett et al, 2011;Todd et al, 2017;Werkhoven et al, 2019a) are typically stable over days. This was indeed the case for the specific turn biases that flies showed in the dark and light conditions, as well as their individual LDM values ( Fig 1I).…”

The structure of behavioral variation within a genotype

“…We did not assess the day-to-day persistence of behavioral modes identified in the unsupervised analysis, so the observed variation across flies could reflect moods rather than permanent personalities. However, previous supervised (Kain et al, 2013) and unsupervised (Todd et al, 2017) analyses of spontaneous microbehaviors similar to those identified by unsupervised approaches have found such behaviors to persist across days.…”

“…Genetic model systems hold particular promise for the mechanistic dissection of this variation, and intragenotypic variability (IGV) in behavior has been characterized in mice (Freund et al, 2013), zebrafish (Pantoja et al, 2016) and Drosophila. In flies, IGV of many behaviors has been studied, including: phototaxis (Kain et al, 2012), locomotor handedness and wing-folding , spontaneous microbehaviors (Kain et al, 2013;Todd et al, 2017), thermal preference and objectfixated locomotion (Liu et al, 2018). Mechanistic studies of these behavioral phenomena have addressed two major questions: 1) what biological mechanisms underlie the magnitude of behavioral variability (e.g., genetic variation (Ayroles et al, 2015), or neural state variation (Kain et al, 2012;), and 2) what specific differences within individual nervous systems predict individual behavioral biases (Liu et al, 2018;Mellert et al, 2016).…”

Quantifying behavior to solve sensorimotor transformations: advances from worms and flies

“…Converting postures (e.g., ‘straight forearm’ or ‘extended elbow’) into actions (e.g., ‘reaching’) requires some way of clustering together stereotyped sequences of postures that are repeatedly seen with only minor variations. Methods to accomplish this range from embedding postures into a low-dimensional space to find ‘clumps’ of similar sequences[18] to identifying when one posture is predictable from previous postures[19], though other methods exist[15,17,20,21]. What is exciting about these ‘unsupervised’ algorithms is that they offer not only ever-more-precise quantification of what an animal is doing at each moment in time, but also reveal the underlying structure of behavior (e.g., which behaviors are sub-programs of other behaviors or which behaviors co-occur), and offer the potential for discovering completely new behaviors.…”