Uncovering Interpretable Fine-Grained Phenotypes of Subcellular Dynamics through Unsupervised Self-Training of Deep Neural Networks

Abstract

Live cell imaging provides unparallel insights into dynamic cellular processes across spatiotemporal scales. Despite its potential, the inherent spatiotemporal heterogeneity within live cell imaging data often obscures critical mechanical details underlying cellular dynamics. Uncovering fine-grained phenotypes of live cell dynamics is pivotal for precise understandings of the heterogeneity of physiological and pathological processes. However, this endeavor introduces formidable technical challenges to unsupervised machine learning, demanding the extraction of features that can faithfully preserve heterogeneity, effectively discriminate between different molecularly perturbed states, and provide interpretability. While deep learning shows promise in extracting useful features from large datasets, it often falls short in producing such high-fidelity features, especially in unsupervised learning settings. To tackle these challenges, we present DeepHACX (Deep phenotyping of Heterogeneous Activities of Cellular dynamics with eXplanations), a self-training deep learning framework designed for fine-grained and interpretable phenotyping. This framework seamlessly integrates an unsupervised teacher model with interpretable features to facilitate feature learning in a student deep neural network (DNN). Significantly, it incorporates an autoencoder-based regularizer, termed SENSER (SENSitivity-enhancing autoEncoding Regularizer), designed to prompt the student DNN to maximize the heterogeneity associated with molecular perturbations. This approach enables the acquisition of features that not only discriminate between different molecularly perturbed states but also faithfully preserve the heterogeneity linked to these perturbations. In our study, DeepHACX successfully delineated fine-grained phenotypes within the heterogeneous protrusion dynamics of migrating epithelial cells, uncovering specific responses to pharmacological perturbations. Remarkably, DeepHACX adeptly captured a minimal number of highly interpretable features uniquely linked to these fine-grained phenotypes, each corresponding to specific temporal intervals crucial for their manifestation. This unique capability positions DeepHACX as a valuable tool for investigating diverse cellular dynamics and comprehensively studying their heterogeneity.