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

FreeBridge, a new computational method based on Schrödinger Bridges, addresses a fundamental challenge in cellular biology by inferring continuous cell transition pathways from static snapshots. The approach constrains predicted intermediate cell states to geometrically valid regions observed in real data, improving both accuracy and biological interpretability in perturbation modeling across multiple imaging datasets.

FreeBridge solves a critical problem in computational biology: inferring how individual cells transition between states when only endpoint observations exist. High-content imaging captures cellular responses to chemical or genetic treatments, but chemical fixation prevents continuous trajectory observation. Previous generative models achieved strong endpoint alignment without guaranteeing biologically plausible intermediate states—a limitation that could propagate errors through downstream analysis and drug discovery pipelines.

The technical innovation centers on anchoring predictions to a fixed cellular manifold derived from instance-segmented single-cell representations. By applying empirical latent support regularization, FreeBridge constrains stochastic transport within geometrically valid regions, preventing the model from extrapolating to morphologies never observed in training data. This geometric grounding directly addresses a pathological failure mode: multiple mathematically valid stochastic processes can connect identical marginal distributions while traversing biologically implausible intermediates.

Validation across three major high-content imaging benchmarks—BBBC021, RxRx1, and JUMP—demonstrates competitive or improved endpoint fidelity while reducing intermediate support violations. The unified evaluation protocol enables fair comparison with existing methods. For computational biologists and drug discovery teams, this represents meaningful progress in mechanism-of-action retention, the ability to distinguish drug mechanisms based on phenotypic signatures.

The work exemplifies how geometric constraints improve generative modeling in scientific domains. Future applications extend beyond perturbation modeling to disease progression trajectories, cellular differentiation pathways, and any biological system where continuous dynamics must be inferred from discontinuous observations. The approach's emphasis on interpretability positions it as infrastructure for hypothesis-driven discovery rather than black-box prediction.

• →FreeBridge infers biologically plausible cell transition pathways by constraining predictions to observed morphological geometry using empirical support regularization
• →The method achieves competitive endpoint alignment while reducing intermediate states that violate observed cellular morphologies across multiple benchmark datasets
• →Geometric grounding addresses a fundamental limitation of prior generative models that could predict mathematically valid but biologically implausible intermediate states
• →Validation on BBBC021, RxRx1, and JUMP demonstrates improved mechanism-of-action retention for drug discovery applications
• →The approach generalizes beyond perturbation modeling to any biological system requiring continuous pathway inference from discrete observations