# Meta Flow Matching: Integrating Vector Fields on the Wasserstein Manifold

**Authors:** Lazar Atanackovic, Xi Zhang, Brandon Amos, Mathieu Blanchette, Leo J. Lee, Yoshua Bengio, Alexander Tong, Kirill Neklyudov  
**arXiv:** [2408.14608](https://arxiv.org/abs/2408.14608)  
**Date:** 2024-08-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.14608v1/

## Key points

- Understanding dynamics of systems composed of interacting entities evolving over time is crucial for predicting population evolution in various scenarios.
- Current flow-based models provide a way to learn these dynamics at a population level by modeling the evolution of entire sample distributions.
- Representing multiple phenomena as vector fields on the Wasserstein manifold of probability densities is essential to capture the complexity of natural processes.
- Meta Flow Matching (MFM) integrates vector fields on the Wasserstein manifold by amortizing the flow model over initial populations, using a Graph Neural Network (GNN) to embed sample populations and train a Flow Matching model.
- MFM can generalize over initial distributions, enhancing predictive capabilities and has been effective in improving predictions of individual treatment responses in personalized medicine applications.

## Summary

In the realm of biological and physical sciences, understanding the dynamics of systems composed of interacting entities evolving over time is crucial for predicting how populations will evolve in various scenarios. Current flow-based models offer a way to learn these dynamics at a population level by modeling the evolution of entire sample distributions. However, these models are typically limited to a single initial population and predefined conditions that dictate different dynamics. Recognizing the complexity of natural processes, researchers argue that representing multiple phenomena as vector fields on the Wasserstein manifold of probability densities is essential. This approach acknowledges that changes in population at any given moment are influenced by interactions between samples within the system. This concept is particularly significant in personalized medicine, where disease development and treatment responses are intricately linked to the unique microenvironment of cells specific to each patient. To address this challenge, a practical solution known as Meta Flow Matching (MFM) has been proposed. MFM integrates vector fields on the Wasserstein manifold by amortizing the flow model over initial populations. This is achieved by embedding sample populations using a Graph Neural Network (GNN) and leveraging these embeddings to train a Flow Matching model. Unlike previous methods, MFM possesses the ability to generalize over initial distributions, enhancing its predictive capabilities. The effectiveness of MFM has been demonstrated through its application in improving predictions of individual treatment responses using a large-scale multi-patient single-cell drug screen dataset. By integrating along vector fields on the Wasserstein manifold, MFM offers a promising avenue for advancing our understanding of complex biological systems and personalized medicine applications.