Generalizable Morphological Profiling of Cells by Interpretable Unsupervised Learning

Abstract

The intersection of advanced microscopy and machine learning is revolutionizing cell biology into a quantitative, data-driven science. While traditional morphological profiling of cells relies on labor-intensive manual feature extraction susceptible to biases, deep learning offers promising alternatives but struggles with the interpretability of its black-box operation and dependency on extensive labeled data. We introduce MorphoGenie, an unsupervised deep-learning framework designed to address these challenges in single-cell morphological profiling. Enabling disentangled representation learning integrated with high-fidelity image reconstructions, MorphoGenie possesses a critical attribute to learn a compact, generalizable and interpretable latent space. This facilitates the extraction of biologically meaningful features without human annotation, additionally overcoming the "curse of dimensionality" inherent in manual methods. Unlike prior models, MorphoGenie introduces a systematic approach to mapping disentangled latent representations to fundamental hierarchical morphological attributes, ensuring both semantic and biological interpretability. Moreover, it adheres to the concept of combinatorial generalization--a core principle of human intelligence-- which greatly enhances the models capacity to generalize across a broad spectrum of imaging modalities (e.g., quantitative phase imaging and fluorescence imaging) and experimental conditions (ranging from discrete cell type/state classification to continuous trajectory inference). The framework offers a new, generalized strategy for unbiased and comprehensive morphological profiling, potentially revealing insights into cellular behavior in health and disease that might be overlooked by expert visual examination.