Learning Cellular Dynamics with Cell–Cell Interaction–Aware Optimal Transport

ICLR 2026 Conference Submission20800 Authors

Keywords: cell dynamics, cell cell interaction, optimal transport

Abstract:

Inferring dynamics from population snapshots is a core challenge in machine learning and biology. In scRNA-sequencing (scRNA-seq), destructive measurements yield irregular, high-dimensional samples of cell states, obscuring how populations evolve. Existing trajectory inference methods either use graph heuristics or cast alignment as an Optimal Transport (OT) problem. However, they treat cells as independent points, ignoring intercellular interactions.
In this work, we ask whether incorporating cell–cell interactions can improve the reconstruction of cellular dynamics from scRNA-seq snapshots. We introduce IADOT (Interaction-Aware Dynamic Optimal Transport), which integrates cell-cell interaction networks into an OT objective and then learns a time-continuous vector field via Conditional Flow Matching. Across a synthetic task and diverse scRNA-seq datasets, we find that incorporating interaction structure can improve snapshot alignment and inference of cellular dynamics versus feature-only baselines. IADOT also supports in-silico ligand–receptor perturbation analyses: we show on lung cancer data that inferred trajectories are sensitive to edits of the ligand–receptor catalog, consistent with known effects of targeted pathway inhibition.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 20800