Skip to content

Flow matching in computational biology

Adjacent methods resource for the FM-to-Virtual-Cells talk. Morehead, Atanackovic, Hegde et al., "Flow matching for generative modelling in bioinformatics and computational biology", Nat Mach Intell 2026, 8:517–534.

What it is

Flow matching is a generative-modelling technique — a training objective for learning a continuous transformation from a simple noise distribution to a complex data distribution. It is the close cousin of diffusion models, but typically simpler to train and faster to sample from. This paper is a review: it surveys how flow matching is being applied across bioinformatics and computational biology — molecular structure generation, protein design, single-cell trajectory and perturbation modelling, and more.

How it connects to the talk

The talk's framing of a virtual cell is increasingly a generative one — a model you can sample from to ask "what would this cell look like under perturbation X." Several of the models in the paper map are generative by design:

  • TranscriptFormer is a generative cross-species sc-FM.
  • Generative Virtual Cells is explicitly a generative design pattern.
  • The broader virtual-cell goal in Bunne et al. is a model you can run forward to predict cell state.

Flow matching is one of the main engines under that hood. Knowing the technique matters for reading those papers critically — and for the small-lab Lane 8 (FM as a generative data-augmentation engine), where the generative method's sampling quality is the whole ballgame.

Why it's adjacent, not core

It is a technique review, not a foundation model and not a single algorithm — so it does not earn a node in the paper map. But it is the right background read before you evaluate any claim that a generative virtual cell "samples realistic cells": the realism depends on the generative objective, and flow matching is now one of the default choices.

Read more

Last updated 2026-05-14.