Fig. 1: Machine learning model predicts behavioural states from high-throughput tracking data of a predatory nematode.

a, Scanning electron microscopy image of P. pacificus (background nematode) with C. elegans larvae (foreground nematode). b, myo-2p::YFP expression in C. elegans compared with myo-2p::RFP expression in P. pacificus. c, Predatory P. pacificus animal surrounded by larval C. elegans prey. d, Schematic of the machine learning pipeline used to classify behavioural states. ML, machine learning; UMAP, uniform manifold approximation and projection. e, UMAP embedding of behavioural features. Colours indicate the six behavioural states identified by hierarchical clustering. f, Performance metrics of the behavioural state classifier on new, unseen data using weighted metrics. g, Distribution of key behavioural features in each state. Each point in the box plots corresponds to the mean value per state and per tracked animal. Box plots follow Tukey’s rule with the box from first to third quartiles, and a line at the median. The whiskers denote 1.5× interquartile range. h, Feature importance (Shapley additive explanations (SHAP) values) for each of the basic behavioural features involved in the model. Features derived from the same base feature are averaged. Colours correspond to the behavioural states represented in e. cwt, continuous wavelet transform. See the Methods for statistics. Scale bars, 20 μm (a), 50 μm (b), 500 μm (c).