What Works and Doesn’t Work in Protein-Ligand Co-Folding With AlphaFold 3 and Similar Methods

Picture by the author.

This post is a semi-automated summary of this interesting preprint:

Have protein-ligand co-folding methods moved beyond memorisation? [Link to preprint]

Deepmind started a revolution in the biological sciences, that hasn’t stopped yet. Right now the stars are AI models that are “multimodal”, that is that handle not just proteins but also nucleic acids, ions, and small molecules. Among others, AlphaFold 3, Chai-1, Boltz-1, etc.

Handling small molecules is of special interest because it could power radically new ways to discover new drugs, repurpose existing ones, and improve existing ones. At the core of this application of AI models to drug discovery is the possibility to fold a protein together with the ligand, that is, co-folding them.

In this blog post you will find my summary of what works very well, more or less, or doesn’t work in protein-ligand docking by co-folding or other methods.

I wrote this by summarizing the key points of the above study, which includes among its authors some of the assessors in the latest Critical Assessment of Structure Prediction, CASP, from where AlphaFold came. The study uses a dataset of 2,600…