From Pixels to Proteins: Flow Matching for Biomedical Applications

Generative models have recently shown remarkable potential across both computer vision and the life sciences. In this talk I take a closer look at one such model, flow matching, which offers a simulation-free training objective and, through its ODE-based formulation, often faster inference than its diffusion-based counterparts. I then showcase two biomedical applications built on this common framework. In the first, we cast medical image segmentation as the conditional generation of signed distance functions, yielding smooth mask representations together with pixel-wise uncertainty maps that discriminative models cannot provide. In the second, we take a variational perspective on protein fitness optimization: sequences are embedded in a continuous latent space, and a learned flow-matching prior over mutations is combined with a fitness predictor to guide sampling toward high-fitness variants. Together, these applications show how a single generative framework can be employed for a range of biomedical problems.