AI protein design: from model to laboratory

Designing a protein binder with the help of deep learning

EPFL School of Engineering · 2025-08-27 · Source

A 10-second scientific animation shared in EPFL’s BindCraft news article. It illustrates computational protein design, not a real-time experiment or a clinical outcome.

From predicting structure to designing new proteins

The 2023 RFdiffusion study introduced an approach that generates new protein structures from specified molecular properties. The team experimentally assessed designed proteins’ structures and functions, including cryo-electron microscopy of a designed binder in complex with its target. Its evidence therefore extends beyond a molecular model displayed on a screen.

What did the BindCraft study show?

Published on 27 August 2025, BindCraft presented an open-source approach using AlphaFold2 to design proteins that bind to selected targets. Researchers tested designs against 12 different targets. Experimental binding success varied from 10% to 100% across targets. That range does not guarantee equivalent performance for every new target, nor does it describe treatment success in patients.

An accessible tool still needs experimental scrutiny

EPFL’s news article of 29 August 2025 describes the aim of making protein design more accessible by experimentally testing smaller sets of candidates. Its short video visualises the computational design process behind this research. Institutional research communication should not be read as independent clinical evaluation or product authorisation.

Where do clinical use and regulatory permission begin?

FDA’s clinical research explanation stresses that preclinical findings cannot replace studies in people. Human research requires a defined protocol, safety assessment and the applicable regulatory process. In this framework, laboratory binding does not establish a safe human dose, therapeutic efficacy or marketing authorisation. The design studies discussed here do not, by themselves, establish that any treatment is authorised.

How should we evaluate a research result?

Our editorial recommendation is to keep three separate columns: the property predicted by the model, the property actually measured in an experiment, and the outcome studied in people. Read how many candidates were designed, how many were tested and how success was defined together. Without failed candidates and study scope, a strong example cannot establish generalisability. This is a guide to reading research, not individual treatment advice or evidence for the effectiveness of RAI products.