Modeling Complex System Dynamics with Flow Matching Across Time and Conditions

Learn about Multi-Marginal Flow Matching (MMFM), a novel computational method for modeling complex system dynamics from temporal snapshot data in this research presentation. Discover how this approach addresses critical limitations in existing Schrödinger Bridge and Flow Matching methods by effectively combining data from multiple time points and different experimental conditions. Explore the technical framework that uses smooth spline-based interpolation across time points and conditions, regressed with neural networks using classifier-free guided Flow Matching to share contextual information about dynamics across multiple trajectories. Examine applications to real-world scenarios where observations from certain combinations of time points and experimental conditions are missing due to experimental costs or sensory failure, including gene regulation, climate change, and financial market fluctuations. Review performance evaluations on both synthetic and real-world datasets, particularly a single-cell genomics dataset with approximately one hundred chemical perturbations across time points, demonstrating significant improvements in imputing data at missing time points compared to existing methods.