AlphaFold 3 Architecture: How Diffusion Predicts Biomolecular Structure (2026)

For a decade, computational structure prediction meant one thing: fold a single protein chain from its sequence. The AlphaFold 3 architecture breaks that frame. Instead of a protein-only network that reasons about amino-acid residues and side-chain torsion angles, AlphaFold 3 is a generative model that denoises raw 3D atom coordinates for almost anything in a biological assembly — proteins, DNA, RNA, small-molecule ligands, ions, and chemically modified residues, all in one shot. The change is not cosmetic. DeepMind and Isomorphic Labs threw out two of AlphaFold 2’s most celebrated components, the Evoformer and the invariant-point-attention structure module, and replaced them with a lighter Pairformer trunk and a diffusion head borrowed from image generation. Understanding why they did that — and what it costs — is the difference between trusting a prediction and being fooled by a confident hallucination.

What this covers: the Pairformer, the diffusion module, atom-level tokenisation, the confidence metrics (pLDDT, PAE, pTM, ipTM), cross-distillation, how AlphaFold 3 differs from AlphaFold 2, its real accuracy and failure modes, and the 2024–2026 access story including Boltz and Chai.

Context and Background

AlphaFold 2 was the CASP14 shock of 2020: a network that predicted single-chain protein structures at a median accuracy rivalling experiment, published in Nature in 2021 and followed by a database of over 200 million predicted structures. Its engine was the Evoformer, a deep stack of attention blocks that reasoned jointly over a multiple sequence alignment (MSA) and a pairwise residue representation, feeding a structure module that placed rigid residue frames and predicted side-chain torsions. It was a protein machine, and a superb one — but it could not natively model a bound drug, a strand of DNA, an RNA aptamer, or a metal ion. Real biology is complexes, and AlphaFold 2 saw only part of the picture.

By 2024, several groups had bolted nucleic acids and ligands onto folding networks with mixed success, and RoseTTAFold All-Atom had shown a more general representation was feasible. AlphaFold 3, published by Abramson et al. in Nature (2024), was DeepMind’s answer: one model for the joint structure of an entire biomolecular assembly. It kept the familiar shape — a large trunk building a pairwise representation, then a module that emits atoms — but rebuilt both halves. For the drug-discovery ambitions behind Isomorphic Labs, predicting a protein alone was never the point; predicting how a candidate molecule sits in a pocket was. If you are coming from the experimental side, it pairs naturally with the shift covered in our piece on cryo-EM and AI structure prediction in drug discovery.

The AlphaFold 3 Architecture: Pairformer Plus Diffusion

The AlphaFold 3 architecture has two conceptual halves. A trunk ingests the input sequences, an MSA, and template features, and refines an abstract pairwise-and-single representation of the whole complex through a Pairformer stack. A diffusion module then takes that representation as conditioning and generates explicit atom coordinates by iteratively denoising, starting from random noise. Confidence heads read the trunk and the sampled structure to emit per-atom and per-pair reliability scores. That is the entire model in one sentence: condition, then denoise.

Figure 1: The AlphaFold 3 architecture — inputs are tokenised and embedded, a Pairformer trunk builds single and pair representations, and a diffusion module generates atom coordinates while confidence heads emit pLDDT, PAE, pTM and ipTM.

Figure 1 traces the data flow. Note how thin the MSA path is compared with AlphaFold 2, and how the diffusion module — not a geometry-aware structure module — is the thing that actually produces coordinates. Everything upstream exists to condition that generative step.

Tokenisation: one representation for every molecule

AlphaFold 2 tokenised a complex as amino-acid residues, full stop. AlphaFold 3 needed a vocabulary that could also describe a nucleotide, a lone ion, or the 13 non-hydrogen atoms of citric acid. Its compromise, described in the EBI training materials, is a mixed granularity. A standard amino acid is one token. A standard nucleotide is one token. But every atom of a ligand, every ion, and every atom of a chemically modified residue becomes its own token. A 100-residue protein bound to a 20-atom ligand is therefore 120 tokens, not 100.

This matters downstream. Confidence metrics such as the Predicted Aligned Error are computed per token, so a ligand reports per-atom confidence while a protein reports per-residue confidence. A per-token pair representation of size N×N is what the Pairformer refines, and its quadratic memory cost is exactly why DeepMind did not make everything atomic. The tokenisation is the seam where generality meets a hardware budget.

The Pairformer: a leaner trunk

The Pairformer is the Evoformer’s slimmed successor. In AlphaFold 2, the Evoformer maintained a large MSA representation and passed rich column-wise signal back and forth between the alignment and the pair representation across many blocks. That deep MSA processing was expensive and, it turned out, not always necessary. The Pairformer keeps a smaller, simpler MSA embedding block and lets the pair representation — the running N×N map of relationships between token pairs — carry most of the reasoning, refined across 48 blocks in the published configuration.

Reducing MSA dependence was not only a speed play. DeepMind reports that AlphaFold 3’s improved predictions of antibody–antigen complexes are directly tied to its lighter reliance on co-evolutionary signal, because antibody loops evolve too fast to leave a usable co-evolutionary trace. For ordinary protein folding, though, AlphaFold 3 still leans hard on the MSA; the alignment did not disappear, it was demoted. The Pairformer’s job is to hand the diffusion module a representation rich enough that a comparatively simple generative network can place atoms correctly.

The diffusion module: generating coordinates directly

Here is the sharpest break from AlphaFold 2. The old structure module reasoned in the language of protein geometry — rigid residue frames, invariant point attention, torsion angles — which is elegant but hard to extend to a ligand or an ion that has no backbone. AlphaFold 3 discards it. Its diffusion module predicts raw atom coordinates, treating structure generation the way a text-to-image model treats pixels: start from noise conditioned on a description, and denoise toward a plausible sample.

Because it operates on 3D coordinates rather than protein-specific frames, the same module places a carbon in a drug molecule, a phosphate in a DNA backbone, and a zinc ion with the same machinery. Generality falls out of the representation almost for free. That is the single most important idea in the paper, and the reason the model can claim to predict “all of life’s molecules” rather than just proteins. The trade is that a generative model can also generate confident nonsense, which we return to below.

How the Diffusion Module Works, Step by Step

Diffusion is a denoising game. During training you take a real structure from the Protein Data Bank, corrupt its atom coordinates with Gaussian noise at some randomly chosen level, and train the network to predict the clean coordinates given the noisy ones plus the trunk’s conditioning. At inference you invert it: start from pure noise and walk down a schedule of decreasing noise, letting the network refine the coordinates at each step until a full structure emerges.

Figure 2: The diffusion rollout — the network repeatedly denoises atom positions, conditioned on the Pairformer’s single and pair representations, refining local geometry at low noise and global arrangement at high noise before sampling multiple seeds.

Figure 2 shows the loop. The conditioning — the single and pair representations from the trunk — enters at every step, so the sequence and MSA information continuously steers the geometry.

Multi-scale noise does multi-scale work

A subtle strength is that the noise level controls what the network learns to fix. At high noise, coordinates are almost random and the network can only get the global arrangement right — which chain sits where, roughly how domains pack. At low noise, the structure is nearly correct and the network sharpens local geometry — bond lengths, ring planarity, the fine packing of a side chain against a ligand. One diffusion process therefore does coarse assembly and fine refinement at different points along the schedule, which is elegant and part of why the module can serve such chemically different molecules.

Sampling, seeds, and ranking

Because diffusion is stochastic, running it twice gives two structures. AlphaFold 3 exploits this: by default it samples the diffusion process five times from a single random seed and returns five ranked predictions, and the AlphaFold Server exposes multiple seeds to widen the ensemble. The samples are ranked using the model’s own confidence estimates, so the top-ranked structure is the one the network trusts most. This sampling behaviour is not a quirk to ignore — for a hard complex, the spread across seeds is itself a signal that the answer is uncertain.

The atom-level representation

Coordinates are generated at the atom level even though the trunk reasons largely per token. An embedding step expands token-level conditioning down to individual atoms before diffusion, and the mmCIF output carries every non-hydrogen atom with its predicted position. That atom-level output is what lets AlphaFold 3 describe a covalent modification, a bound cofactor, or a metal-coordination geometry that a residue-frame model simply could not express.

Why diffusion is a better fit than a regression head

It is worth being precise about why diffusion suits this problem, because the choice looks strange at first: image-generation machinery in a structural-biology model. The deep reason is that many biomolecular structures are genuinely multi-modal. A flexible loop, a rotatable side chain, a ligand that can bind two ways — for these there is no single “correct” coordinate to regress toward, and a regression head trained with a mean-squared loss will average the modes into a blurred, physically wrong compromise. A diffusion model instead learns the distribution of plausible structures and samples from it, so each seed lands on one self-consistent mode rather than an average of several. That is also why sampling multiple seeds is diagnostic: the seed-to-seed spread is a crude readout of how multi-modal the true answer is. For rigid, well-determined regions the seeds agree; for flexible or ambiguous ones they diverge, which is exactly the behaviour you want from an honest model.

Training AlphaFold 3: Data, Objectives, and Cross-Distillation

Architecture is only half the story; a diffusion model is only as good as what it denoises toward. AlphaFold 3 was trained primarily on experimentally determined structures from the Protein Data Bank, filtered by release date so that evaluation targets postdate the training cutoff — the standard guard against memorising answers. Because the PDB contains proteins, nucleic acids, ligands, ions, and modified residues, a single all-atom training set can teach one network the whole chemical zoo, which is the data-side justification for the unified tokeniser.

The diffusion objective is denoising: corrupt a real structure’s coordinates with a sampled noise level, then train the network to recover the clean coordinates given the trunk’s conditioning. Crucially, the noise level is drawn across a wide range each training step, so a single network learns both coarse global assembly (from high-noise examples) and fine local geometry (from low-noise examples). The trunk and the diffusion module are trained together, so the Pairformer learns to produce exactly the representation the generator needs.

The most interesting training decision is cross-distillation, introduced above as an anti-hallucination measure but worth stating as a training-data choice. DeepMind augmented the real PDB data with structures predicted by AlphaFold-Multimer v2.3. Those predictions render disordered and unresolved regions as loose, extended coils rather than crisp folds. By including them, AlphaFold 3 is taught what “no confident structure” is supposed to look like, so its generative prior for uncertainty is a ribbon of low-pLDDT coil instead of a confident fabrication. This is a rare case of one model’s predictions being used deliberately to shape another model’s failure mode, and it is a genuinely clever piece of engineering — though, as the limitations section shows, not a complete cure.

AlphaFold 3 vs AlphaFold 2: What Actually Changed

If you already know AlphaFold 2, the cleanest mental model is a two-swap upgrade: Evoformer → Pairformer, and structure module → diffusion module, wrapped in a tokeniser that admits non-protein chemistry.

Figure 3: AlphaFold 3 vs AlphaFold 2 — the deep-MSA Evoformer becomes the lighter Pairformer, the IPA structure module and torsion prediction become a diffusion module over raw atom coordinates, and the scope expands from proteins to all biomolecules.

Figure 3 lines the two pipelines up. The table below makes the differences concrete.

Aspect	AlphaFold 2 (2021)	AlphaFold 3 (2024)
Trunk	Evoformer, deep MSA processing	Pairformer, lighter MSA block
Structure output	Structure module, IPA, residue frames + torsions	Diffusion module, raw atom coordinates
Token =	One amino-acid residue	Residue, nucleotide, or per-atom for ligands/ions
Scope	Single proteins and multimers (AF-Multimer)	Proteins, DNA, RNA, ligands, ions, modifications, complexes
Generative?	No — deterministic regression	Yes — denoising diffusion, sampled
Confidence	pLDDT, PAE, pTM, ipTM (per residue)	pLDDT, PAE, pTM, ipTM (per token)
Anti-hallucination	N/A	Cross-distillation from AF-Multimer

Two consequences deserve emphasis. First, because the output is now a stochastic sample rather than a deterministic regression, reproducibility is per-seed, and the “best” structure is a ranking decision. Second, AlphaFold 3 gives up the structure module’s hard geometric guarantees. Invariant point attention was equivariant by construction; a diffusion network over raw coordinates is not, so it can in principle produce distorted geometry — including wrong chirality — that the old module could never emit. DeepMind mitigates this with training-time penalties and data augmentation, but the guarantee is now statistical, not structural.

Cross-distillation: teaching the model not to hallucinate

A generative model asked to place atoms in a disordered region — one that genuinely has no single structure — will happily invent a crisp, wrong one. That is the diffusion module’s dark side. DeepMind’s countermeasure is cross-distillation: they trained AlphaFold 3 partly on structures predicted by AlphaFold-Multimer (AFM v2.3), which renders disordered and unresolved regions as extended, ribbon-like coils. By learning from that supervisory signal, AlphaFold 3 learns to represent “I don’t know” as a loose, low-confidence coil rather than a confidently folded fabrication. It is an imperfect fix — hallucination in disordered regions remains a real limitation — but it is a deliberate architectural choice, not an accident.

Confidence Metrics: Reading an AlphaFold 3 Prediction

A prediction you cannot audit is a liability. AlphaFold 3 emits the same family of confidence scores as AlphaFold 2, now computed per token, and using them well is half the skill of the tool.

• pLDDT — predicted local distance difference test, per atom, on a 0–100 scale. It is local confidence: how sure the model is about the position of that atom relative to its neighbours. Below roughly 50 is a red flag and, importantly, is often where hallucinated order hides.
• PAE — predicted aligned error, a token×token matrix in ångströms. It answers a relative question: if you align on token i, how far off is token j expected to be? Low off-diagonal PAE between two chains means the model is confident about their relative placement — the key signal for whether a predicted interface is trustworthy.
• pTM — predicted template modelling score, one number for the whole structure’s global fold accuracy.
• ipTM — interface pTM, the accuracy of the predicted relative positions of the components in a complex. For any protein–protein, protein–DNA, or protein–ligand interface, ipTM is the number to check first.

The practical rule: pLDDT tells you whether to trust a region, PAE and ipTM tell you whether to trust an interface. A complex can have every chain individually well-folded (high pLDDT everywhere) yet be assembled wrongly (poor ipTM, high inter-chain PAE). Treating a high per-chain pLDDT as evidence the complex is right is the most common misreading of an AlphaFold 3 result.

Accuracy: Where AlphaFold 3 Is Strong and Where It Is Not

AlphaFold 3’s headline gains, reported in the Nature paper, are on interactions that AlphaFold 2 could not model at all. On the PoseBusters benchmark of protein–ligand complexes, AlphaFold 3 substantially outperformed traditional docking tools, predicting a large majority of ligand poses within the standard success threshold without needing a known pocket as input. On protein–nucleic-acid interactions it beat prior specialist tools. And on antibody–antigen complexes it improved markedly over AlphaFold-Multimer, credited in part to the reduced MSA reliance.

The honest caveats matter as much as the wins. Accuracy is uneven across biomolecule classes: single-protein and protein–ligand predictions are strongest, protein–protein complexes are good but variable, and nucleic-acid and RNA accuracy, while improved, trails protein accuracy. On the community RNA-Puzzles-style targets, AlphaFold 3 did not clearly beat the best dedicated RNA methods. Because the model is generative, its confidence scores are the load-bearing part of every result — a prediction without its pLDDT and PAE read is half a prediction.

Rather than quote exact figures from memory, treat the specific DockQ and LDDT numbers as things to pull from the primary paper and its supplement for your own targets. The safe summary: state-of-the-art and often transformative for ligands, strong for many complexes, and materially weaker for RNA and for anything intrinsically disordered.

Access in 2024–2026: Server, Weights, and the Open Reproductions

The AlphaFold 3 access story is nearly as consequential as its architecture, because it triggered an open-source counter-movement. At launch in May 2024 the model was available only through the free AlphaFold Server, a web tool with daily job limits, a restricted ligand set, and no commercial use. Publishing a landmark method in Nature while withholding the code drew sharp criticism from researchers who could not reproduce or extend it. Under that pressure, DeepMind released the inference code in November 2024 and made model weights available to academic groups on request under a non-commercial licence — a middle path that continued through 2025 and into 2026.

That gap — a state-of-the-art method that most people could not freely run — is exactly the vacuum open reproductions filled. Boltz-1, from MIT’s Jameel Clinic, was the first fully open model to approach AlphaFold 3-level accuracy, released under a permissive MIT licence for both academic and commercial use, with later Boltz iterations and a hosted API extending the line into 2026. Chai-1 offered comparable performance under a free-for-research posture, monetising through pharmaceutical partnerships. ByteDance’s Protenix added a clean, Apache-licensed PyTorch reimplementation. The net effect by 2026 is that AlphaFold 3-class prediction is broadly available, and tools such as ABCFold even let researchers run AlphaFold 3, Boltz, and Chai side by side and compare. The lesson is a familiar one in AI: a closed release of a genuinely useful architecture is an invitation for the community to rebuild it in the open, and here it took only months.

Trade-offs, Gotchas, and What Goes Wrong

The move to diffusion buys generality and pays for it in new failure modes. Knowing them is what separates a useful tool from a confident liar.

Figure 4: Interpreting AlphaFold 3 by task — confidence expectations differ sharply across single proteins, complexes, protein–ligand poses, nucleic acids, and disordered regions, and every branch ends in a pLDDT and PAE audit.

Hallucination in disordered regions. The best-documented failure. A 2025 analysis of intrinsically disordered proteins found a large fraction of residues misaligned with curated disorder databases, with a meaningful share being outright hallucinations — confident, ordered structure where biology has none. Cross-distillation reduces but does not eliminate this. The mitigation in practice: never trust an ordered structure sitting under a low-pLDDT region; treat sub-50 pLDDT as “possibly invented.”

Chirality and physical plausibility. Because the diffusion network is not equivariant by construction, it can occasionally output the wrong chirality at a stereocentre or minor steric clashes — errors the equivariant AF2 structure module could not make. DeepMind added stereochemistry penalties, but you should still validate ligand geometry with a chemistry-aware checker (PoseBusters-style) before believing a pose.

Stoichiometry and dynamics. AlphaFold 3 predicts a single static assembly for the stoichiometry you give it. It does not know how many copies of a chain belong in the complex, does not model conformational ensembles or allosteric states, and does not predict binding affinity. A predicted structure is one snapshot, not a trajectory, and the model will confidently place a ligand even when the real answer is “does not bind.”

Sequence and MSA sensitivity. For folding, deep MSAs still help; orphan sequences and fast-evolving loops degrade. And a well-folded chain in a poorly-assembled complex — high pLDDT, bad ipTM — is a trap that catches careful people. Read the interface metrics, not just the per-chain ones.

Practical Recommendations

Use AlphaFold 3 as a hypothesis generator with a built-in reliability meter, not as an oracle. Start on the AlphaFold Server for interactive complexes, or run the released weights locally when you need scale, custom ligands, or reproducible pipelines. Always sample multiple seeds for hard complexes and inspect the spread; agreement across seeds is corroboration, disagreement is a warning. Validate every ligand pose with a physics/chemistry checker before it informs a decision, and cross-reference disordered regions against experimental disorder data rather than trusting the model’s coil.

A working checklist:

• Match the tool to the task — strongest for single proteins and protein–ligand poses; weakest for RNA and disordered regions.
• Read ipTM and inter-chain PAE first for any complex; per-chain pLDDT does not validate an interface.
• Flag sub-50 pLDDT as possible hallucination, especially where it looks confidently ordered.
• **Sample