Recurrent Neural Networks with Interpretable Cells Predict and Classify Worm Behaviour

An important goal in behaviour analytics is to connect disease state or genome variation with observable differences in behaviour. Despite advances in sensor technology and imaging, informative behaviour quantification remains challenging. The nematode worm C. elegans provides a unique opportunity to test analysis approaches because of its small size, compact nervous system, and the availability of large databases of videos of freely behaving animals with known genetic differences. Despite its relative simplicity, there are still no reports of generative models that can capture essential differences between even well-described mutant strains. Here we show that a multilayer recurrent neural network (RNN) can produce diverse behaviours that are difficult to distinguish from real worms behaviour and that some of the artificial neurons in the RNN are interpretable and correlate with observable features such as body curvature, speed, and reversals. Although the RNN is not trained to perform classification, we find that artificial neuron responses provide features that perform well in worm strain classification.