Research map / Probabilistic and scientific modeling / Dynamical systems
Modeling Cell Dynamics and Interactions with Unbalanced Mean Field Schrödinger Bridge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes the Unbalanced Mean-Field Schrödinger Bridge (UMFSB) and the CytoBridge neural solver to infer unbalanced stochastic cell dynamics with cell-cell interactions from snapshot scRNA-seq data.
Problem
Inferring continuous cell dynamics from sparse time-resolved snapshots while jointly modeling stochasticity, growth/death, and cell-cell interactions, which existing unbalanced Schrödinger bridge and OT methods do not account for.
Approach
Extends the mean-field Schrödinger bridge to unnormalized densities (adding a growth term) and reformulates it with a Fisher-information regularization so the SDE constraints become ODEs; CytoBridge parameterizes velocity, growth, score, and interaction potential with neural networks, uses random batch methods for fast interacting-particle simulation, and trains with energy, reconstruction (unbala
Key results
- On synthetic attractive-interaction data at t=1..4, CytoBridge achieves the lowest W1 (e.g., 0.015, 0.014, 0.018, 0.038) versus DeepRUOT (0.044, 0.045, 0.053, 0.057).
- On mouse hematopoiesis at t=2, CytoBridge gets W1 6.644 and TMV 0.078, versus DeepRUOT's 6.757 and 0.260.
- On held-out mouse hematopoiesis time point, CytoBridge W1 6.847 vs DeepRUOT 6.868.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add explicit cell-cell interaction modeling | CytoBridge compared with DeepRUOT · Synthetic gene regulatory data with attractive interactions | Wasserstein distance (W1): 0.044 ±0.002 -> 0.015 ±0.001 (t=1); 0.045 ±0.002 -> 0.014 ±0.001 (t=2); 0.053 ±0.002 -> 0.018 ±0.002 (t=3); 0.057 ±0.003 | Helped |
| Add explicit cell-cell interaction modeling | CytoBridge compared with DeepRUOT · Synthetic gene regulatory data with Lennard-Jones interactions | Wasserstein distance (W1): 0.036 ±0.000 -> 0.030 ±0.001 (t=1); 0.042 ±0.001 -> 0.030 ±0.001 (t=2); 0.049 ±0.002 -> 0.027 ±0.001 (t=3); 0.054 ±0.002 | Helped |
| Include an interaction network when ground-truth interactions are absent | CytoBridge compared with DeepRUOT · Synthetic gene regulatory data without interactions | Wasserstein distance (W1): 0.035 ±0.001 -> 0.036 ±0.001 (t=1); 0.054 ±0.001 -> 0.053 ±0.001 (t=2); 0.042 ±0.001 -> 0.044 ±0.001 (t=3); 0.046 ±0.002 | Mixed |
| Add explicit cell-cell interaction modeling | CytoBridge compared with DeepRUOT · Mouse hematopoiesis held-out time point | Wasserstein distance (W1): 6.868 ±0.003 -> 6.847 ±0.003 | Helped |
Surprising findings
- On synthetic data without interactions, CytoBridge does not outperform DeepRUOT at most time points, attributed to difficulty for the neural network in producing all-zero interaction outputs.
- The learned interaction effect in the A549 EMT dataset appears very weak.
Limitations
- Minimizes an upper bound of the energy rather than the original UMFSB objective.
- Interaction potential parameterized by RBF expansion, which may limit expressiveness.
- Multi-stage training and lack of ligand-receptor biological priors.
Experimental setup
Datasets: synthetic three-gene gene regulatory network mouse hematopoiesis scRNA-seq embryoid body scRNA-seq pancreatic beta-cell differentiation A549 EMT scRNA-seq zebrafish spatiotemporal transcriptomics
Benchmarks: synthetic three-gene gene regulatory data mouse hematopoiesis embryoid body zebrafish spatiotemporal transcriptomics
Metrics: W1 Wasserstein distance Total Mass Variation (TMV) Moran's I
Baselines: SF2M Meta FM MMFM Metric FM UOT-FM MIOFlow uAM UDSB TIGON DeepRUOT
Builds on: regularized unbalanced optimal transport DeepRUOT mean field Schrödinger bridge random batch methods conditional flow matching
Closest papers by meaning
- Learning stochastic dynamics from snapshots through regularized unbalanced optimal transport (ICLR 2025 · similarity 0.88)
- Inferring stochastic dynamics with growth from cross-sectional data (NeurIPS 2025 · similarity 0.87)
- Momentum Multi-Marginal Schrödinger Bridge Matching (NeurIPS 2025 · similarity 0.85)
- Joint Velocity-Growth Flow Matching for Single-Cell Dynamics Modeling (NeurIPS 2025 · similarity 0.85)
- Learning non-equilibrium diffusions with Schrödinger bridges: from exactly solvable to simulation-free (NeurIPS 2025 · similarity 0.84)
- Curly Flow Matching for Learning Non-gradient Field Dynamics (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.