FreeBridge: Variational Schr\"odinger Bridges for Cellular Transition Dynamics

FreeBridge is a novel framework utilizing a Schrödinger Bridge formulation for modeling cellular transition dynamics from endpoint-only observations in high-content imaging assays. It employs instance-segmented single-cell representations to define atomic states and learns stochastic transport within a fixed cellular manifold, achieving competitive endpoint fidelity and reduced intermediate support violations across datasets like BBBC021, RxRx1, and JUMP. This approach emphasizes the significance of geometric grounding in understanding biologically interpretable perturbation dynamics, which is crucial for practitioners working with single-cell data in AI applications.