Behavioural motifs of larval Drosophila melanogaster and Caenorhabditis elegans
MetadataShow full item record
I present a novel method for the unsupervised discovery of behavioural motifs in larval Drosophila melanogaster and Caenorhabditis elegans. Most current approaches to behavioural annotation suffer from the requirement of training data. As a result, automated programs carry the same observational biases as the humans who have annotated the data. The key novel element of my work is that it does not require training data; rather, behavioural motifs are discovered from the data itself. The method is based on an eigenshape representation of posture. Hence, my approach is called the eigenshape annotator (ESA). First, I examine the annotation consistency for a specific behaviour, the Omega turn of C. elegans, and find significant inconsistency in both expert annotation and the various Omega turn detection algorithms. This finding highlights the need for unbiased tools to study behaviour. A behavioural motif is defined as a particular sequence of postures that recurs frequently. In ESA, posture is represented by an eigenshape time series, and motifs are discovered in this representation. To find motifs, the time series is segmented, and the resulting segments are then clustered. The result is a set of self-similar time series segments, i.e. motifs. The advantage of this novel framework over the popular sliding windows approaches is twofold. First, it does not rely on the ‘closest neighbours’ definition of motifs, by which every motif has exactly two instances. Second, it does not require the assumption of exactly equal length for motifs of the same class. Behavioural motifs discovered using the segmentation-clustering framework are used as the basis of the ESA annotator. ESA is fully probabilistic, therefore avoiding rigid threshold values and allowing classification uncertainty to be quantified. I apply eigenshape annotation to both larval Drosophila and C. elegans, and produce a close match to hand annotation of behavioural states. However, many behavioural events cannot be unambiguously classified. By comparing the results to eigenshape annotation of an artificial agent’s behaviour, I argue that the ambiguity is due to greater continuity between behavioural states than is generally assumed for these organisms.