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 can add unnecessary complexity and may even hurt performance. We develop a simple, stable, and scalable recipe which we use to train our generative microscopy model. We scale training compute to an order of magnitude larger than prior methods, saturating existing public benchmarks with approximately two-fold FID and ten-fold KID improvements over prior methods. We report an extended battery of metrics, finding that improved evaluation encoders can alleviate the saturation problem. Our methods represent progress towards high-fidelity virtual microscopy assays for tasks such as virtual screening, batch effect correction, and data enrichment.

翻译：暂无翻译