4D spatiotemporal landscape of mitochondrial phenotypes across cellular states unlocked through representation learning.

Abstract

Mitochondria are four-dimensional (4D: x, y, z, and time) organelles essential for cellular function. Characterizing their 4D phenotypic landscape across diverse cellular states requires both 4D imaging and analytical frameworks. We present MitoSpace, a self-supervised deep learning model trained without labels on terabytes of single-cell lattice light-sheet microscopy data of mitochondria under mechanistically distinct perturbations. MitoSpace learns latent representations that outperform predefined features in drug classification and capture interpretable variation in mitochondrial morphology and dynamics. Regression probes predict mitochondrial membrane potential from the learned representations (R² = 0.91), establishing a quantitative mapping between form and function at the single-cell level. MitoSpace also generalizes zero-shot to unseen perturbations and human lung organoids. Dimensionality ablation reveals that representation quality improves monotonically from 2D to 3D to 4D, demonstrating the importance of volumetric and temporal information. The model, dataset, and interactive explorer are publicly available, providing a foundation for 4D phenotypic screening.