Powerful and interpretable behavioural features for quantitative phenotyping of Caenorhabditis elegans

Abstract

Behaviour is a sensitive and integrative readout of nervous system function and therefore an attractive measure for assessing the effects of mutation or drug treatment on animals. Video data provide a rich but high-dimensional representation of behaviour, and so the first step of analysis is often some form of tracking and feature extraction to reduce dimensionality while maintaining relevant information. Modern machine-learning methods are powerful but notoriously difficult to interpret, while handcrafted features are interpretable but do not always perform as well. Here, we report a new set of handcrafted features to compactly quantify Caenorhabditis elegans behaviour. The features are designed to be interpretable but to capture as much of the phenotypic differences between worms as possible. We show that the full feature set is more powerful than a previously defined feature set in classifying mutant strains. We then use a combination of automated and manual feature selection to define a core set of interpretable features that still provides sufficient power to detect behavioural differences between mutant strains and the wild-type. Finally, we apply the new features to detect time-resolved behavioural differences in a series of optogenetic experiments targeting different neural subsets.This article is part of a discussion meeting issue 'Connectome to behaviour: modelling C. elegans at cellular resolution'.

Keywords: C. elegans; computational ethology; phenotyping; worm tracking.

Figure 1.

Schematics of the core features. Each of the features is summarized and expanded according to the transformations in figure 2. See the electronic supplementary material for detailed definitions.

Figure 2.

Operations that expand and summarize each of the core features. See electronic supplementary material, figure S1 for a more detailed description on how the core time series features are subdivided.

Figure 3.

Results of recursive feature elimination. (a) The features reported in this paper (Tierpsy) achieve a higher accuracy than the ones used in Yemini et al. [20] (original). (b) The addition of transformations using subdivisions by motion type and derivatives over time are necessary to achieve the high accuracy. On the other hand, the eigenprojections, the normalization by length and the ventral/dorsal sign have little or no effect.

Figure 4.

Boxplots of N2 and three mutants using the Tierpsy_8 subset of features. The small set of features facilitates a visual comparison between different strains.

Figure 5.

Example of two-dimensional histograms of short pulses for different strains and features.

Figure 6.

Comparisions between the changes of behaviour among the different strains under blue light stimulation for selected features. The heatmaps show the p-values of the Jensen–Shannon divergence between ATR and control plates for the short pulse experiments (a), and the long pulse experiments (b). The p-values were corrected for multiple comparisons within strains using the Benjamini–Hochberg procedure.

Figure 7.

Cluster maps of the median value of the Jensen–Shannon divergence between different strains among the ATR plates.