Elucidating the Design Space of Flow Matching for Cellular Microscopy

2026-03-25

AI Summary

Existing flow-matching models for cellular microscopy image generation suffer from an expansive and underexplored design space, often resulting in structural redundancy and suboptimal performance. This work systematically investigates key architectural choices and proposes a streamlined, stable, and scalable modeling framework, establishing the first large-scale foundational generative model in this domain. By eliminating several popular yet redundant or even detrimental components and integrating pretrained molecular embeddings with transfer learning, the model achieves a two-order-of-magnitude increase in scale. On the task of simulating cellular responses to unseen molecular perturbations, it demonstrates a 2× improvement in Fréchet Inception Distance (FID) and a 10× improvement in Kernel Inception Distance (KID), substantially outperforming existing approaches and setting a new state of the art.

Abstract

Flow-matching generative models are increasingly used to simulate cell responses to biological perturbations. However, the design space for building such models is large and underexplored. We systematically analyse the design space of flow matching models for cell-microscopy images, finding that many popular techniques are unnecessary and can even hurt performance. We develop a simple, stable, and scalable recipe which we use to train our foundation model. We scale our model to two orders of magnitude larger than prior methods, achieving a two-fold FID and ten-fold KID improvement over prior methods. We then fine-tune our model with pre-trained molecular embeddings to achieve state-of-the-art performance simulating responses to unseen molecules. Code is available at https://github.com/valence-labs/microscopy-flow-matching

Problem

• flow matching
• cellular microscopy
• generative models
• design space
• biological perturbations

Innovation

• flow matching
• cellular microscopy
• generative modeling
• foundation model
• molecular embedding