Research map / Probabilistic and scientific modeling / Dynamical systems
Meta Flow Matching: Integrating Vector Fields on the Wasserstein Manifold
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 Meta Flow Matching, which amortizes flow matching over initial populations by embedding each population with a GNN so the vector field generalizes to unseen populations, and applies it to patient-specific single-cell drug-response prediction.
Problem
Flow-based population dynamics models are limited to a single initial population and predefined conditions, so they cannot predict the evolution of unseen populations or account for interactions between samples.
Approach
Condition the flow matching vector field on a GNN embedding of the initial population (the set of control samples), trained jointly with the vector field, so the learned model defines density-dependent vector fields on the Wasserstein manifold that generalize to unseen initial distributions without requiring population identity labels.
Key results
- On the organoid patient-split experiment, MFM (k=50) reaches W1 4.087 ± 0.122 and MMD 9.26e-3, beating non-OT baselines FM (4.171) and CGFM (4.189) and ICNN (4.157).
- On the synthetic letters test sets of unseen letters X and Y, MFM (k=50) achieves W1 about 0.21 and 0.20 while FM and CGFM reach 0.23-1.06, indicating generalization to unseen silhouettes.
- On replica-2 organoid split, MFM (k=0) gives r^2 0.918 vs 0.865 for FM and 0.899 for CGFM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| condition the flow model on the input population embedding using MFM rather than an index-conditional model | Flow Matching · Synthetic letters Test X | W1: 0.334 \pm 0.000 -> 0.201 \pm 0.006 | Helped |
| condition the flow model on the input population embedding using MFM rather than an index-conditional model | Flow Matching · Synthetic letters Test Y | W1: 0.327 \pm 0.000 -> 0.204 \pm 0.005 | Helped |
| use MFM rather than FM | Flow Matching · Organoid replicate-2 test | W1: 4.340 \pm 0.078 -> 4.100 \pm 0.093 | Helped |
| use MFM rather than FM without OT | Flow Matching · Organoid patients test | W1: 4.171 \pm 0.107 -> 4.087 \pm 0.122 | Helped |
| use MFM rather than FM with OT | Flow Matching · Organoid patients test | W1: 4.064 \pm 0.152 -> 4.087 \pm 0.122 | Hurt |
| vary the GCN neighbor count k (ablation) | MFM · Synthetic letters unseen test populations | W1: 0.202 \pm 0.002; 0.205 \pm 0.008; 0.201 \pm 0.006; 0.206 \pm 0.008 | No clear effect |
| use k=0 instead of k=100 (ablation) | MFM · Organoid replicate-2 test | W1: 4.100 \pm 0.093 -> 4.209 \pm 0.007 | Hurt |
| use k=100 instead of k=0 (ablation) | MFM · Organoid patients test | W1: 4.135 \pm 0.094 -> 4.112 \pm 0.148 | Helped |
| use a Gaussian source distribution instead of the original source (ablation) | MFM · Organoid replicate-2 test | W1: 4.209 \pm 0.007 -> 4.177 \pm 0.042 | No clear effect |
| use OT couplings for MFM (ablation) | MFM k=50 · Synthetic letters test X | W1: 0.189 \pm 0.003 -> 0.189 \pm 0.006 | No clear effect |
Surprising findings
- Adding OT couplings to FM gave the best patient-split results among OT methods, and OT did not significantly change MFM performance on synthetic letters.
- The best MFM variant depended on the split: k=0 was best on replicate splits while k=100 was best on patient splits.
Limitations
- Does not incorporate optimal transport couplings in the main method, which the authors found improves FM and CGFM on biological data.
- Only the highest drug dose is considered in the organoid experiments.
- Gains over FM with OT on patient split are not consistent; FM w/ OT outperforms MFM on patient split metrics.
Experimental setup
Datasets: Synthetic letters (constructed) Organoid drug-screen single-cell mass cytometry dataset
Benchmarks: Synthetic letters Organoid drug-screen dataset (replicate and patient splits)
Metrics: 1-Wasserstein 2-Wasserstein MMD r^2
Baselines: FM CGFM ICNN FM w/ OT CGFM w/ OT CellOT
Builds on: Flow Matching Conditional Flow Matching CellOT DeepSets Meta Optimal Transport
Closest papers by meaning
- GENOT: Entropic (Gromov) Wasserstein Flow Matching with Applications to Single-Cell Genomics (NeurIPS 2024 · similarity 0.88)
- Wasserstein Flow Matching: Generative Modeling Over Families of Distributions (ICML 2025 · similarity 0.87)
- Fisher Flow Matching for Generative Modeling over Discrete Data (NeurIPS 2024 · similarity 0.87)
- Joint Velocity-Growth Flow Matching for Single-Cell Dynamics Modeling (NeurIPS 2025 · similarity 0.86)
- Curly Flow Matching for Learning Non-gradient Field Dynamics (NeurIPS 2025 · similarity 0.86)
- Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.86)
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.