AlphaFold 3: What DeepMind's Latest Model Actually Does — and Doesn't Do

When DeepMind's AlphaFold 2 solved protein structure prediction in 2020, it was described — accurately — as one of the most significant scientific achievements in decades. The 50-year-old "protein folding problem" had been a fundamental challenge in structural biology, and AlphaFold 2 solved it with an accuracy that surprised even the researchers involved.

AlphaFold 3, published in Nature in May 2024 and expanded since, extends this achievement in an important direction: from predicting the structure of proteins in isolation to predicting how proteins interact with other molecular classes — including DNA, RNA, small molecules (potential drug compounds), and metal ions. For drug discovery, this distinction matters enormously.

What AlphaFold 3 Actually Does

AlphaFold 2 was architecturally based on the Evoformer — a transformer-based system that leveraged evolutionary information from related proteins to predict structure. AlphaFold 3 uses a fundamentally different approach: a diffusion-based architecture that generates atomic coordinates directly, treating structure prediction more like the image generation models that have become familiar in AI.

This architectural shift enables AlphaFold 3 to model joint distributions of multiple molecular species — the key capability needed for drug discovery applications. When a drug molecule binds to a protein target, the question is not just "what does this protein look like?" but "how does this specific molecule orient itself relative to this protein's binding site?" That is precisely what AlphaFold 3 addresses.

According to DeepMind's Nature publication (May 2024), AlphaFold 3 substantially outperforms prior specialised tools on protein-ligand interaction prediction, protein-nucleic acid complex prediction, and antibody-antigen binding prediction. The paper includes head-to-head comparisons against established methods on standardised benchmarks.

The Drug Discovery Application

Structure-based drug design — computationally modelling how potential drug molecules dock into protein binding sites before synthesising them — has been a standard tool in pharmaceutical research for decades. AlphaFold 3 improves the quality of the structural inputs to this process significantly.

One of the more significant applications is identifying "cryptic" binding sites — pockets in protein surfaces that only become accessible in certain conformational states. These sites are difficult and expensive to identify experimentally but can represent valuable drug targets, particularly for proteins previously considered undruggable.

Where It Falls Short

Honest assessment of AlphaFold 3's limitations is important for researchers integrating it into workflows. The model performs notably less well on highly disordered proteins — proteins that do not adopt a fixed three-dimensional structure — and on membrane proteins, which sit in cell membranes and present unique modelling challenges. Both categories include important pharmaceutical targets.

Predicting binding affinity — how tightly a drug molecule binds — remains less reliable than predicting binding geometry. AlphaFold 3 can tell you roughly where a molecule binds better than it can tell you how strongly it binds. For drug discovery applications, affinity prediction remains a separate and harder problem.

The Open Access Question

AlphaFold 2's open release — with model weights freely available — was widely praised as a model for democratising powerful scientific tools. AlphaFold 3's release has been more complicated. Research access is available through the AlphaFold Server, but the full model weights have not been released publicly. Commercial use for pharmaceutical applications requires an agreement with Isomorphic Labs, DeepMind's drug discovery spinout. This has prompted substantial discussion in the structural biology community about the appropriate balance between open science and commercial development of publicly-relevant research tools.