DynaCLR: Contrastive Learning of Cellular Dynamics with Temporal Regularization

TL;DR

DynaCLR is a self-supervised contrastive learning method that creates robust, temporally regularized embeddings of cellular dynamics from time-lapse images, enabling various biological analyses with minimal annotations.

Contribution

It introduces DynaCLR, a novel contrastive learning framework that incorporates temporal regularization and tracking for analyzing cellular dynamics in microscopy images.

Findings

Embeddings generalize well to new datasets.

Effective for classifying cell states and behaviors.

Enables cross-modal and asynchronous response analyses.

Abstract

We report DynaCLR, a self-supervised method for embedding cell and organelle Dynamics via Contrastive Learning of Representations of time-lapse images. DynaCLR integrates single-cell tracking and time-aware contrastive sampling to learn robust, temporally regularized representations of cell dynamics. DynaCLR embeddings generalize effectively to in-distribution and out-of-distribution datasets, and can be used for several downstream tasks with sparse human annotations. We demonstrate efficient annotations of cell states with a human-in-the-loop using fluorescence and label-free imaging channels. DynaCLR method enables diverse downstream biological analyses: classification of cell division and infection, clustering heterogeneous cell migration patterns, cross-modal distillation of cell states from fluorescence to label-free channel, alignment of asynchronous cellular responses and broken cell tracks, and discovering organelle response due to infection. DynaCLR is a flexible method for comparative analyses of dynamic cellular responses to pharmacological, microbial, and genetic perturbations. We provide PyTorch-based implementations of the model training and inference pipeline (https://github.com/mehta-lab/viscy) and a GUI (https://github.com/czbiohub-sf/napari-iohub) for the visualization and annotation of trajectories of cells in the real space and the embedding space.