Two years ago, flow matching was a theoretical framework. Today it is the dominant method in computational structural biology, reshaping drug discovery, protein design, and the long-running effort to build a working digital replica of a living cell.
The biggest bet in biomedical AI right now is the virtual cell — a computer simulation of living biology accurate enough to test a drug before it is ever synthesised. The mathematical framework most likely to make that possible is not the one most people are watching. It is called flow matching, and a new review in Nature Machine Intelligence maps how it has taken over computational biology in the past two years.
Flow matching is an AI training technique. Most people who follow AI have heard of diffusion models — the technology behind image generators like Midjourney and Stable Diffusion.
Diffusion works by adding noise to data until it becomes random, then training a model to reverse that process. It works well for images. For biology, it is slower and more constrained than it needs to be.
Flow matching does something simpler. Instead of routing through a noisy intermediate state, it learns a direct path between two distributions. In plain terms: you tell the model where you are starting and where you want to end up, and it learns the most efficient route between them. Starting point and ending point can be almost anything — a diseased cell and a healthy cell, a random molecular cloud and a valid drug candidate, a known protein and a hypothetical variant with new properties. The path can be made mathematically straight, which is faster to train and cheaper to run. Biological constraints can be woven into the path itself rather than imposed as corrections after the fact.
The review, from researchers at UC Berkeley, University of Missouri, and collaborating institutions, synthesises how flow matching is now being deployed across four domains simultaneously: small molecule drug design, protein structure generation, DNA and RNA modelling, and single-cell biology. In each area it is producing faster training, better samples, and more biologically accurate outputs than the diffusion models it is replacing.
The drug discovery implication is the most immediate. A meaningful discovery programme evaluates millions of candidates, and cost per sample matters. Flow matching approaches consistently produce valid, drug-like structures faster and at lower cost per candidate. That changes which compounds get screened and how many. Downstream, it changes how early in a programme computational screening can replace wet lab synthesis.
For protein design, flow matching generates novel structures with specified target properties. This is relevant to enzyme engineering, therapeutic protein design, and the synthetic biology goal of building molecular machines that do not exist anywhere in nature.
For the virtual cell, the implications are larger and longer in development. The vision is a simulation accurate enough that you can introduce a drug molecule, a mutation, or a pathogen into a digital cell and watch what happens without touching a pipette. Google DeepMind, the Chan Zuckerberg Initiative, and the Wellcome Sanger Institute all have substantial programmes aimed at this goal.
The review argues that flow matching is the natural mathematical substrate for this kind of simulation — because flow matching is built to learn mappings between states of complex systems, which is exactly what a virtual cell needs to do every time it simulates a biological event.
The limitations matter. This is a review paper, not an experimental result. Applications range from well-validated to early-stage. Small molecule generation is mature. Virtual cell simulation at meaningful biological resolution is not close. Biological datasets are often too small to train the most ambitious models.
What the paper provides is a map of a transition already happening. Flow matching has moved from a theoretical framework to the leading method in computational structural biology in roughly two years. Most researchers outside the field have not tracked that pace. The review closes that gap.
Anyone building in drug discovery, synthetic biology, or AI for medicine needs to understand what flow matching is and why it has displaced its predecessor. This paper is the clearest synthesis currently available.
Sources
Morehead, A., Atanackovic, L., Hegde, A. et al. “Flow matching for generative modelling in bioinformatics and computational biology.” Nature Machine Intelligence 8, 517–534 (2026). DOI: 10.1038/s42256-026-01220-0 Published: 23 April 2026
Jane holds a BSc in Biology from the University of Regina and a Master of Science in Bioscience, Technology and Public Policy from the Univesity of Winnipeg. Her reporting interests include Life Sciences, Physical Sciences and the Cosmos.