Self-supervised deep learning uncovers the semantic landscape of drug-induced latent mitochondrial phenotypes

Abstract

Imaging-based high-content screening aims to identify substances that modulate cellular phenotypes. Traditional approaches screen compounds for their ability to shift disease phenotypes toward healthy phenotypes, but these end point-based screens lack an atlas-like mapping between phenotype and cell state that covers the full spectrum of possible phenotypic responses. In this study, we present MitoSpace: a novel mitochondrial phenotypic atlas that leverages self-supervised deep learning to create a semantically meaningful latent space from images without relying on any data labels for training. Our approach employs a dataset of ~100,000 microscopy images of Cal27 and HeLa cells treated with 25 drugs affecting mitochondria, but can be generalized to any cell type, cell organelle, or drug library with no changes to the methodology. We demonstrate how MitoSpace enhances our understanding of the range of mitochondrial phenotypes induced by pharmacological interventions. We find that i) self-supervised learning can automatically uncover the semantic landscape of drug induced latent mitochondrial phenotypes and can map individual cells to the correct functional area of the drug they are treated with, ii) the traditional classification of mitochondrial morphology along a fragmented to fused axis is more complex than previously thought, with additional axes being identified, and iii) latent spaces trained in a self-supervised manner are superior to those trained with supervised models, and generalize to other cell types and drug conditions without explicit training on those cell types or drug conditions. Future applications of MitoSpace include creating mitochondrial biomarkers for drug discovery and determining the effects of unknown drugs and diseases for diagnostic purposes.

Toward generalizable phenotype prediction from single-cell morphology representations

“…Alternatively, foundation models potentially offer consistent feature representation extraction across datasets, which would circumvent the alignment step. 10 For example, initiatives like "Mitospace", which focuses on extracting a common feature space of mitochondria 47 , the masked autoencoder Phenom-Beta, which is a vision transformer foundational model for embedding microscopy images 10 , "BioMorph", which links morphology to organelle processes 48 , and the Allen Cell Explorer, which uncovers cell phenotypes organelle-by-organelle 49,50 , illustrate promising future directions for annotating universal cell representations with generalizable single-cell phenotypes. Nevertheless, these foundation models still require phenotype interpretation to analyze cells and perturbations on a uniform, biologically-interpretable basis.…”