An Unsupervised Framework to Decompose the Morphokinetic Behaviour of Cell Migration

Abstract: We propose an informative, unsupervised machine learning framework that deconstructs complex T cell migration into a vocabulary of fundamental "dynamic motifs". By clustering short sequences of cell tracks, our method automatically identifies a set of canonical, repeated behaviors without any prior assumptions. We demonstrate that representing whole-video dynamics through the frequency of these motifs and the probabilities of transitions between them provides a highly sensitive and robust phenotypic fingerprint.