Organellomics: AI-driven deep organellar phenotyping of human neurons

Abstract

AbstractSystematic assessment of organelle architectures in cells, known as the organellome, could provide valuable insights into cellular states and disease pathologies but remains largely uncharted. Here, we devised a novel pipeline combining self-supervised deep learning and transfer learning to generate a Neuronal Organellomics Vision Atlas (NOVA). Analyzing over 1.5 million confocal images of 24 distinct membrane-bound and membrane-less organelles in human neurons, we enable a simultaneous evaluation of all organelles. We show that organellomics allows the study of cellular phenotypes by quantifying the localization and morphological properties embodied in multiple different organelles, using a unified score. We further developed a strategy to superimpose all organelles, which represents a new realization of cellular state. The value of our approach is demonstrated by characterizing specific organellar responses of human neurons to stress, cytoplasmic mislocalization of TDP-43, or disease-associated variations in ALS genes. Therefore, organellomics offers a novel approach to study the neuro-cellular biology of diseases.HighlightsAI-driven organellomics without cell segmentation or multiplexed imaging.Analysis of 24 membrane-bound and membrane-less organelles in more than 1.5 million images of human neurons.Quantitative organelle-level description of neuronal response to chemical and genetic perturbations.Organelles ranked on a single metric scale and integrated organellome view via superposition of multiple organelles.