AlphaFold 3 expands the picture

The claim made on 8 May 2024

On 8 May 2024, Google DeepMind and its drug-discovery sibling Isomorphic Labs described a single deep-learning system able to predict the joint three-dimensional structure of complexes that mix different kinds of biomolecule: proteins together with DNA, RNA, small-molecule ligands, metal ions, and chemically modified residues. The narrow, documented claim is precise. Where earlier models mostly predicted the fold of one protein chain, this model predicts how those different components sit relative to one another in one shared coordinate frame. The peer-reviewed result appeared as Abramson and colleagues in Nature, with the accompanying announcement on the DeepMind blog.

The useful historical question is not whether one model "solved biology," a phrase no careful reading of the paper supports. It is narrower: which constraint the previous generation left in place, how this system removed it, what the reported numbers actually measure, and which competing approaches were pursuing the same target. Those are separable claims, and keeping them separable is the difference between recording an expansion of scope and inflating it into a myth. AlphaFold 3 widened the object being modeled. It did not turn a structural prediction into an experimental fact, and it did not, by itself, produce a medicine.

The constraint AlphaFold 2 left standing

The predecessor set the stakes. In 2020, AlphaFold 2 produced protein structures accurate enough that DeepMind and independent assessors described it as a step change on the CASP14 benchmark, reporting a median Global Distance Test score of 92.4 across targets, roughly a 1.6 angstrom error at the level of atoms. The full method was published as Jumper and colleagues in Nature in 2021, and its predicted structures were released at enormous scale, seeding a public database that grew to hundreds of millions of models and became standard equipment in molecular biology.

That success was also its boundary. AlphaFold 2 was built around single protein chains, leaning heavily on the evolutionary signal in large multiple-sequence alignments, and its later Multimer variant extended the same machinery to protein-protein assemblies. Biology, however, mostly happens at interfaces between unlike molecules: a drug binding a pocket, a transcription factor gripping DNA, an ion completing an active site, a sugar modifying a surface residue. Modeling those cases well previously required a patchwork of specialist tools, each with its own representation for ligands or nucleic acids, and each needing separate assumptions and separate validation. The constraint AlphaFold 3 set out to remove was this fragmentation: the absence of one model that treats every atom of a mixed complex under a common framework rather than stitching protein predictions to separate docking and nucleic-acid pipelines.

How the model was rebuilt: Pairformer and a diffusion head

AlphaFold 3 is not AlphaFold 2 with extra inputs bolted on; the Nature paper describes a substantially reorganized architecture, and an independent walkthrough of that architecture traces the changes block by block. The heavy evolutionary-alignment processing at the core of AlphaFold 2, the Evoformer, is reduced and replaced by a simpler stack the authors call the Pairformer, a run of 48 blocks that do not share weights and that operate mainly on a pair representation of token relationships and a single representation of each token, rather than leaning as hard on the alignment. The triangle-update operations that enforce geometric self-consistency between distances survive from AlphaFold 2, but the alignment now feeds a lighter module. This matters because alignments are abundant for proteins and largely meaningless for an arbitrary drug molecule or a lone metal ion.

The larger change is at the output, and it starts with how inputs are counted. AlphaFold 3 tokenizes a standard residue or nucleotide as one token but represents nonstandard chemistry, a ligand or a modified residue, atom by atom, so a single vocabulary spans polymers and free chemistry at once. AlphaFold 2 then ended in a structure module that reasoned in terms of protein backbone frames and side-chain torsion angles, machinery that assumes you are placing amino acids. AlphaFold 3 discards that geometry-specific head and predicts raw atom coordinates directly with a generative diffusion module. Because the model now speaks in atoms rather than residues, it can represent any chemical entity without a bespoke encoder for each. That single design decision is what actually lets one network span "all of life's molecules," and it is the mechanical heart of the event.

Why generating structure by denoising changes the behavior

The diffusion head borrows the recipe that had reshaped image generation. A denoising diffusion model, as formalized by Ho, Jain and Abbeel in 2020, learns to reverse a process that gradually adds noise: start from random noise and iteratively refine it toward a plausible sample. In AlphaFold 3 the "sample" is a set of atomic positions, and the module denoises noised coordinates across a range of noise scales, learning local stereochemistry at small noise and global arrangement at large noise. Because it works on coordinates rather than residue frames, it removes the need for the hand-built stereochemical constraints and violation losses that the earlier structure head relied on to keep bonds physical.

The trade-off is that a generative model produces a distribution, not a single deterministic answer, so the same input run several times yields several structures. That flexibility introduces a specific failure the authors document candidly: in regions that are genuinely disordered, the model can "hallucinate" confident-looking, spurious structure, because a generative head prefers plausible order to honest disorder. Their mitigation is cross-distillation, training partly on AlphaFold-Multimer predictions so the model learns to render disordered regions as the loose, ribbon-like shapes that signal uncertainty, alongside confidence metrics estimated from the model's own predicted error. It is a revealing detail: the power that lets diffusion place any atom is the same power that must be disciplined so the model does not invent order that is not there.

Reading the reported numbers as task-specific results

The headline figures are best read as performance on named benchmarks, not as a general competence score. For protein-ligand interactions, DeepMind reported that AlphaFold 3 was roughly 50% more accurate than the best traditional docking method on the PoseBusters set, and for some categories of interaction accuracy roughly doubled. The comparison is pointed. The PoseBusters benchmark of Buttenschoen, Morris and Deane, published in Chemical Science, had recently found that no deep-learning docking method reliably beat classical physics-based tools such as AutoDock Vina once physical validity was actually checked, meaning poses were scored on whether bonds and clashes were realistic, not just on proximity. AlphaFold 3 claiming to surpass those classical tools, and crucially doing so from sequence and chemistry alone without being handed the binding pocket, is what made the ligand result notable rather than incremental.

The other categories follow the same pattern. The paper reports higher accuracy on protein-nucleic-acid complexes than nucleic-acid-specialist predictors, and improved antibody-antigen interface prediction over the earlier AlphaFold-Multimer, an accuracy that increases when more diffusion samples are drawn and the best is selected by the model's own ranking. Each of these is a measurement against a specific test set of a specific interaction class, on structures the model had not seen in training. None of them is a statement that the model "understands" binding, and the authors do not make that leap; the numbers describe geometric agreement with experimentally solved structures, which is the only thing they can honestly describe.

The competing approaches of its moment

AlphaFold 3 did not arrive into an empty field. The most direct rival came from the laboratory of David Baker, whose group published RoseTTAFold All-Atom in Science in 2024, pursuing the same generalization from proteins to all-atom biomolecular assemblies including small molecules and covalent modifications, and pairing it with a companion design model. The two efforts represent parallel bets that the future of structure prediction was a single all-atom model rather than a protein model plus separate docking software, and they reached that conclusion within weeks of one another, which is itself evidence that the idea was ripe rather than singular. Where they differed was in emphasis and openness: the Baker group released its method and weights, while DeepMind's plan for release was the source of the controversy described below.

The older competition was the physics-based docking tradition that PoseBusters had used as its yardstick: energy functions and search procedures that place a ligand in a known pocket, refined over decades of computational chemistry. Those tools remained the trusted default in much of the field, and their strength was physical plausibility grounded in real force fields. AlphaFold 3's contribution to that contest was to show a learned model matching or exceeding them on pose accuracy while needing far less problem-specific setup. Placing the event honestly means holding both facts at once: it was a strong result against real, respected baselines, and those baselines were being actively advanced by others rather than standing still while a single lab lapped the field.

The access decision and the reproducibility fight

The scientific claim and the release decision were separate events, and the second became a controversy that reached the pages of the journal itself. At launch AlphaFold 3 shipped not as downloadable code or weights but as the AlphaFold Server, a free web tool for non-commercial research with real limits: a capped number of daily jobs and a restricted menu of ligands and ions rather than arbitrary user chemistry. Isomorphic Labs, the commercial drug-discovery entity, retained the full model for its own pipeline. Publishing a landmark method in a peer-reviewed journal while withholding the means to run it independently triggered a public backlash, including a letter to Nature's editors from researchers who argued the disclosure fell short of the community's standards for reproducibility, and it prompted a public commitment to make code available within six months.

That commitment was met, and the gap is part of the record. Around November 2024, DeepMind released the AlphaFold 3 inference pipeline on GitHub, with the source code under an open license but the trained model parameters gated: available only on request, only for non-commercial use, and only "if received directly from Google." This two-stage pattern, a limited server first and constrained weights months later, is a genuine feature of the event and a signal about how frontier commercial-scientific labs now navigate the tension between the norms of open science and the commercial value of a protected drug-discovery asset. The dispute was not a sideshow to the science; it was a case study in what "published" now means when the method is also a competitive product.

What actually flowed downstream

The influence shows up along two tracks. Commercially, AlphaFold 3 is the structural engine Isomorphic Labs points at drug design. The company had already signed pharmaceutical partnerships with Eli Lilly and Novartis in January 2024, announced before AlphaFold 3's public reveal but predicated on the same AlphaFold line, with modest upfront payments and performance milestones that together could reach billions of dollars. That is a business bet on structure prediction as a step inside a discovery pipeline, not a claim of finished medicines, and the milestone-heavy structure of the deals is itself an admission that a predicted pose is early, not final.

Scientifically, the constrained release paradoxically accelerated open work. With the method described but the weights withheld for months, other groups built and released their own all-atom, diffusion-based structure predictors, so the architecture's ideas propagated faster than its checkpoints did. The clearest example is Boltz-1, released under an MIT license in late 2024, which reproduced and extended the AlphaFold 3 technical report as fully open, commercially usable code; Chai Discovery's Chai-1 arrived in the same window. The broadest downstream marker was recognition. In October 2024, the Nobel Prize in Chemistry 2024 went half to David Baker for computational protein design and half jointly to Demis Hassabis and John Jumper for protein structure prediction. That award honored the AlphaFold line of work broadly rather than AlphaFold 3 specifically, and the distinction is worth preserving: the prize validated the trajectory that AlphaFold 2 opened, and AlphaFold 3 is the chapter that extended it from folds to interactions.

The limits the paper itself states

The honest caveats are load-bearing. A predicted complex is a hypothesis about geometry, not evidence of binding, efficacy, or safety; the model does not output an affinity, a rate, or a functional guarantee, and the DeepMind announcement is careful to frame outputs as predictions to be tested experimentally. The diffusion head's hallucination of order in disordered regions, discussed above, is a documented failure mode, not a rumor. The model predicts largely static structures, so it says little about the conformational dynamics, allosteric motions, and alternative states that govern much of real biochemistry, and it does not model the effect of solvent, temperature, or concentration. Accuracy also varies sharply by target and depends on the available training data, whose Protein Data Bank cutoff was structures released before late September 2021, so genuinely novel chemistry can fall outside what the model has learned.

These limits do not diminish the result; they locate it. A tool that proposes plausible atomic arrangements for mixed complexes, quickly and without pocket priors, is a powerful hypothesis generator that can narrow an experimental search from thousands of possibilities to a handful worth testing at the bench. Treating its confident output as a settled molecular truth is exactly the error the paper's own confidence metrics and caveats are designed to prevent, and it is an error the strongest users are the most careful to avoid.

Placing AlphaFold 3 in AI history

The right frame is expansion of scope, not revolution. The narrow, defensible claim is that AlphaFold 3 unified structure prediction across biomolecule types by moving to a token-plus-atom representation and a diffusion output, and reported strong, benchmark-specific gains on ligand, nucleic-acid, and antibody interactions. The wider judgment, that this reshaped computational structural biology, is supported by the competing all-atom efforts it paralleled, the open replications it provoked, and the Nobel recognition of the surrounding program, but it belongs at a different level of confidence than the paper's measurements, and it is a judgment about a field's direction rather than a single benchmark result.

To navigate outward from this event, three questions help. Which earlier capability did it inherit and depend on? AlphaFold 2's protein accuracy and public structure database, and the diffusion-model machinery that generative imaging had already proven. Which later work reused its idea while changing an ingredient? The open all-atom predictors, Boltz-1 and Chai-1 among them, that followed the withheld weights and swapped closed for open. Which limitation still demands new methods? Dynamics, binding affinity, and the wet-lab validation that no structure predictor can substitute for. Answered together, these keep AlphaFold 3 in its proper place: a decisive broadening of what one model could predict, and a standing reminder that a better picture of a molecule is the beginning of an investigation, not its conclusion.

References

• Abramson et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3.
• Google DeepMind and Isomorphic Labs (2024). AlphaFold 3 predicts the structure and interactions of all of life's molecules.
• Isomorphic Labs (2024). AlphaFold 3 predicts the structure and interactions of all of life's molecules.
• Google DeepMind (2024). AlphaFold 3 inference pipeline.
• Google DeepMind (2024). AlphaFold Server.
• Jumper et al. (2021). Highly accurate protein structure prediction with AlphaFold.
• Google DeepMind (2020). AlphaFold: a solution to a 50-year-old grand challenge in biology.
• Buttenschoen, Morris and Deane (2024). PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences.
• Krishna et al. (2024). Generalized biomolecular modeling and design with RoseTTAFold All-Atom.
• Ho, Jain and Abbeel (2020). Denoising Diffusion Probabilistic Models.
• Simon (2024). The Illustrated AlphaFold.
• Science (2024). Limits on access to DeepMind's new protein program trigger backlash.
• Wohlwend et al. (2024). Boltz-1 Democratizing Biomolecular Interaction Modeling.
• TechCrunch (2024). Isomorphic inks deals with Eli Lilly and Novartis for drug discovery.
• The Nobel Prize (2024). The Nobel Prize in Chemistry 2024.