AlphaFold3 and Biomolecular Co-Folding

AlphaFold2 answered a question biology had chased for fifty years: given a protein sequence, what is its fold? But a folded protein alone rarely does anything. It works by binding: to a drug molecule, to a strand of DNA, to a metal ion, to another protein. AlphaFold2 could not see any of that. It spoke one language, amino acids, and everything else in the cell was invisible to it. AlphaFold3 (Abramson et al, Nature 2024) is the model that learned to fold the whole assembly at once: protein plus ligand plus nucleic acid plus ion plus post-translational modification, in a single forward pass. The jump from predicting a structure to predicting an interaction required tearing out the part of AlphaFold2 that made it protein-specific and replacing it with something that does not know or care what an amino acid is.

From a structure module to a diffusion head

AlphaFold2's structure module was built around protein geometry. It reasoned in terms of residue backbone frames and torsion angles, an inductive bias that encodes "this is a chain of amino acids with peptide bonds." That bias is exactly what you cannot afford when the same model must also place a zinc ion, a lipid, or a stretch of double-stranded DNA, none of which have residues or torsions in the protein sense.

AlphaFold3 removes the frames-and-torsions machinery entirely. In its place sits a diffusion module that predicts raw atom coordinates directly by iterative denoising, the same generative recipe that drives image and video models. Training corrupts a known structure with Gaussian noise and asks the network to denoise it; sampling starts from noise and runs the reverse process to assemble a structure atom by atom. Because the target is just 3D coordinates for atoms, the representation is chemistry-agnostic. A carbon in a drug molecule and a carbon in a tyrosine side chain are handled by the same head. This is the single change that makes co-folding possible: coordinates are a universal currency, torsion angles are not.

The diffusion formulation also folds several older components together. AlphaFold2 needed separate loss terms, side-chain packing, and violation penalties to produce physically plausible geometry; the diffusion head learns much of that implicitly from data, operating at multiple length scales so that both local chemistry and global arrangement fall out of the same denoiser.

The Pairformer and the retreat from the MSA

AlphaFold2's Evoformer was a heavy two-track engine that co-evolved a multiple-sequence-alignment (MSA) representation and a residue-pair representation, passing information back and forth. The MSA track was where most of the evolutionary signal lived, and it was expensive.

AlphaFold3 replaces the Evoformer with the Pairformer, and the emphasis shifts decisively. MSA processing is cut down to a lightweight early block; the pair representation now carries the load through the trunk. There is a good reason beyond speed. Ligands and ions have no evolutionary history to align, and many nucleic-acid and antibody targets have shallow or useless alignments. A model that leans as hard on the MSA as AlphaFold2 did cannot generalise to entities that do not have one. Pushing the reasoning into the pair representation, which is defined over all tokens regardless of type, lets a single trunk feed a single diffusion head for every kind of molecule.

The trade is real: for well-conserved single proteins, deep MSAs still help, and AlphaFold3 has not thrown that signal away, only demoted it. What it bought is a model whose backbone does not fall apart the moment you hand it something evolution never wrote down.

One model, many molecule types

The practical result is a unified interface. You describe a complex as a set of tokens: protein chains, nucleic-acid chains, ligands specified by chemical structure, ions, and covalently modified residues (glycosylation, phosphorylation, and the like). The trunk builds a joint pair representation over all of them; the diffusion head places every atom in one shared coordinate frame. The paper reports substantially higher accuracy than specialised docking and structure tools across protein-ligand, protein-nucleic-acid, and antibody-antigen interactions, using one model rather than a pipeline of narrow ones.

Confidence prediction carries over from AlphaFold2 in spirit: per-atom pLDDT for local reliability and predicted aligned error for the relative placement of chains and domains. For interaction work the cross-chain PAE matters most, because it tells you whether the model is confident about where the ligand sits, not just whether each partner folded well on its own.

Why the open reimplementations mattered

AlphaFold3 shipped with a controversy attached. The Nature paper appeared in May 2024 without runnable code, offering only a rate-limited web server restricted to non-commercial use. An open letter of researchers accused the journal of breaking its own code-availability rules; reviewers had assessed the method from pseudocode they could not execute (see Retraction Watch's contemporaneous account). DeepMind released the model weights for academic use roughly six months later, in November 2024, but only under a non-commercial licence and after months of a field unable to build on the result.

That vacuum is what the open community filled. Boltz-1 (Wohlwend, Corso, Passaro et al, MIT CSAIL and the Jameel Clinic, bioRxiv November 2024) reimplemented the AlphaFold3 framework, the diffusion-based structure prediction and the pair-representation trunk, and released training code, inference code, weights, datasets, and benchmarks under a permissive MIT licence. It reports accuracy on par with AlphaFold3 and the commercial Chai-1 on standard co-folding benchmarks, with AlphaFold3 holding a slight edge that falls inside the confidence intervals. The successor, Boltz-2, goes further and jointly models structure and binding affinity. The lesson generalises past this one model: for a method the whole field needs to build on, open weights under a usable licence are not a courtesy, they are what turns a published result into infrastructure.

When it falls down

A predicted complex is a hypothesis, not a measurement. The failure modes are specific and worth internalising before you trust an output.

• Diffusion hallucinates confidently. A generative denoiser will always return a complete, plausible-looking structure, even for a target where it has no real signal. It does not abstain. The danger is a clean-looking complex that is simply wrong, especially for novel folds, shallow-MSA targets, and unusual ligands. Read the confidence, never the picture alone.
• Confidence is calibrated, not infallible. pLDDT and PAE track accuracy on average and are the right first filter, but a high-confidence region can still be wrong, and disordered regions can be reported with spurious order (the reason AF3 uses cross-distillation during training to suppress it). Treat confidence as a prior, not a proof.
• Stoichiometry and assembly state. You largely tell the model what to fold together. It will not reliably discover that a complex is a tetramer, or which of several competing binding partners actually assembles in the cell.
• A static structure is not affinity or dynamics. The model predicts one plausible arrangement. It does not tell you the binding free energy, the on/off rates, the conformational ensemble, or how the complex moves. A confident pose does not mean the ligand binds tightly, or binds at all under physiological conditions. Ranking candidate binders by structural plausibility is not the same as predicting affinity, which is exactly the gap Boltz-2 and dedicated affinity models try to close.
• Out-of-distribution chemistry. Ligands, cofactors, and modifications far from anything in the training data (unusual metals, exotic covalent chemistry) are where accuracy degrades quietly and confidence is least trustworthy.

The correct mental model: AlphaFold3-class models are extraordinary hypothesis generators for experimental structural biology and early-stage screening, and a poor substitute for the assay that tells you whether the hypothesis is true.

Further reading

• Accurate structure prediction of biomolecular interactions with AlphaFold3 - Abramson et al, Nature 2024. The diffusion module, the Pairformer, and the unified handling of proteins, nucleic acids, ligands, ions, and modified residues (open-access PMC mirror of the Nature paper).
• Boltz-1 Democratizing Biomolecular Interaction Modeling - Wohlwend, Corso, Passaro et al, MIT, 2024. The first fully open AlphaFold3-level reimplementation, MIT-licensed, with code, weights, and data.
• jwohlwend/boltz - the Boltz codebase (Boltz-1 and Boltz-2), MIT licence, runnable co-folding you can install and use.
• Nature earns ire over lack of code availability for the AlphaFold3 paper - Retraction Watch, May 2024. The code-release controversy and the open letter, for the context behind why open reimplementations followed.