AlphaFold 3 Explained: Predicting Biomolecular Complexes With Diffusion

Quick answer

AlphaFold 3 is one model that predicts the joint 3D structure of biomolecular complexes — proteins together with DNA, RNA, small-molecule ligands, ions, and chemically modified residues — instead of folding a single protein chain at a time. The big architectural change: it drops AlphaFold 2’s hand-built structure module and ends in a diffusion network that generates raw atom coordinates directly. On the interactions that matter most for biology and drug discovery, it beats specialized tools — notably classical docking for protein-ligand binding and dedicated systems for protein-nucleic-acid structure.

From single proteins to full complexes

AlphaFold 2 solved a narrow, important problem: sequence in, single protein fold out. But cells do almost nothing with isolated proteins. Signaling, catalysis, gene regulation, immune recognition, and drug binding are all interaction events — a protein gripping a strand of DNA, an enzyme cradling a substrate, an antibody locking onto an antigen. Predicting those required a patchwork of incompatible tools: one for protein-protein, a docking program for ligands, another model for RNA.

AlphaFold 3’s premise is that the same network should handle all of it. It accepts a mixed system — multiple protein chains, nucleic acids, ligands, ions, and post-translational modifications — and predicts how every atom sits relative to the rest. That unification is the actual contribution; no single benchmark captures it as well as the fact that one model now covers cases that used to need a half-dozen.

A diffusion-based architecture

The headline mechanism is the switch to diffusion. AlphaFold 2 ran an Evoformer over evolutionary alignments, then a geometry-aware structure module that built protein backbones frame by frame. AlphaFold 3 keeps a streamlined evolutionary-representation trunk (a Pairformer) but replaces the structure module with a diffusion model that learns to denoise atom positions into a final structure.

This matters for two practical reasons. First, generating coordinates directly at the atom level is what lets the model treat a ligand or a nucleotide the same way it treats an amino acid — there is no protein-specific scaffold to break. Second, the team had to add training measures to stop a generative model from inventing plausible-looking but wrong geometry, a failure mode diffusion is prone to. That trade is worth naming: the architecture that unlocks generality is also the one that can hallucinate, and the honest reading is that confidence scores matter more here than they did in AlphaFold 2.

Key results

The Nature paper’s value is breadth backed by margins, not one trophy number. For protein-ligand interactions, AlphaFold 3 reports markedly higher accuracy than classical docking tools — without being given the binding pocket as input, which is how docking is usually helped. For protein-nucleic-acid complexes, it outperforms nucleic-acid-specific predictors. For antibody-antigen interfaces it improves over AlphaFold-Multimer v2.3, the prior DeepMind system. Alongside the model, the team launched the AlphaFold Server, a free web tool letting researchers build and predict complexes without running the model themselves.

The honest framing: these are wins across many interaction types from a single system, which is harder and more useful than topping one leaderboard. But “higher accuracy than docking” is a statement about ranking and pose, not about validated binding strength.

Limits and open questions

A predicted structure is a hypothesis, not experimental proof. AlphaFold 3 outputs a static, most-likely conformation; it does not give you binding free energy, kinetics, the conformational ensemble a molecule actually samples, or cellular context. It can be confidently wrong, and on hard or novel targets it still is. Antibody-antigen prediction, despite the gains, remains among the weaker areas.

The release also drew real criticism: the full model and weights were not opened the way AlphaFold 2 was, and access ran through the AlphaFold Server with usage limits, which constrained independent reproduction and slowed scrutiny. The practical question for any lab is not whether to trust AlphaFold 3, but how to position it — as a fast, strong prior that tells you which expensive assays, docking runs, or molecular-dynamics simulations are worth doing, not as a replacement for them.

FAQ

What does AlphaFold 3 predict that AlphaFold 2 could not?

AlphaFold 3 predicts complexes that mix molecule types — proteins bound to DNA, RNA, small-molecule ligands, ions, and modified residues — in a single model. AlphaFold 2 was built to fold individual protein chains (with Multimer extending it to protein-protein assemblies), and could not natively handle ligands or nucleic acids.

How is AlphaFold 3’s architecture different from AlphaFold 2?

AlphaFold 3 removes AlphaFold 2’s geometry-based structure module and ends in a diffusion network that directly generates atom coordinates. The evolutionary-representation trunk is streamlined into a Pairformer. Working at the raw atom level is what lets one network treat amino acids, nucleotides, and ligands uniformly.

Is AlphaFold 3 better than docking software for drug discovery?

For predicting protein-ligand binding poses, AlphaFold 3 reports higher accuracy than classical docking tools, even without being told the binding site. But it predicts geometry, not binding affinity or kinetics, so it complements — rather than replaces — docking, assays, and molecular dynamics in a drug-discovery pipeline.

Is AlphaFold 3 open source?

Not in the way AlphaFold 2 was. At release the full model and weights were not freely published; access ran through the free AlphaFold Server with usage limits, which drew criticism for limiting independent reproduction.

AlphaFold 3 reframes the question from “what shape is this protein?” to “how does this whole molecular system fit together?” — and that shift, more than any single score, is why it matters. Read the original in Nature.