AlphaFold 3 — Universal Prediction of Biomolecular Interactions

In May 2024, DeepMind publishes AlphaFold 3. Beyond isolated proteins, the model predicts the structures of protein–DNA, protein–RNA, and protein–ligand complexes with accuracy surpassing all existing methods — opening the door to large-scale computational drug design.

In Plain Terms

A protein is a long chain that folds back on itself, and the number of possible folds is astronomical: that is the problem structure prediction sets out to circumvent. AlphaFold 3, released by DeepMind in May 2024, widens the exercise to assemblies — what a protein forms together with DNA, RNA, a ligand or an ion — rather than isolated proteins alone. To get there, it no longer builds the structure piece by piece: it starts from random noise and gradually strips it away until the atomic coordinates emerge, which lets it propose several plausible shapes for a single complex instead of just one. The intended use is screening candidate molecules on a computer, ahead of the lab bench. That still does not make it a turnkey product: every prediction comes with confidence scores, the least reliable regions remain to be confirmed experimentally, and the model was only submitted to an independent blind evaluation after its release, at CASP16.

Discovery

Parameter	Value
Publication date	May 8, 2024 (Nature, vol. 630, pp. 493–500)
Team	DeepMind / Isomorphic Labs (Josh Abramson, Jonas Adler et al.)
Architecture	Conditioned diffusion model (Pairformer + diffusion head)
Training data	PDB (cutoff September 30, 2021) + cross-distillation from AlphaFold-Multimer v2.3
Targets	Proteins, DNA, RNA, ligands, ions, post-translational modifications
Public server	AlphaFold Server (limited to 5,000 residues per query)
License	Public server at release; weights and source code released in November 2024 for academic use

Technical Explanation

1. Multi-modal input representation. AlphaFold 3 accepts as input a protein sequence (MSA + templates), a nucleotide sequence (DNA/RNA) and/or a ligand in SMILES format. Each entity is tokenized and projected into a shared embedding space of dimension 384. Residue-residue (or residue-atom) pairs are encoded in a pair representation matrix of dimension $N \times N \times 128$ .

2. Trunk: Pairformer (replacing the Evoformer). The Pairformer uses triangle attention and triangle update layers (as in AF2) but drops the explicit MSA representation in favor of a condensed single embedding. This reduces memory complexity from $O(N_{seq} \times N_{res}^2)$ to $O(N_{res}^2)$ , making it possible to handle complexes of > 5,000 residues. The trunk comprises 48 Pairformer blocks with residual connections.

3. Diffusion head: coordinate generation by denoising. Unlike AF2 (which uses an IPA structural module), AF3 generates 3D atomic coordinates through a diffusion-denoising process over 200 steps. The model learns to reverse Gaussian noise applied to PDB atomic coordinates. At each step, the network predicts the noise to be removed, progressively refining the structure. This approach naturally captures multimodality — several possible conformations for the same complex.

4. Confidence head and recycling. The confidence module estimates pTM (predicted TM-score), PAE (predicted aligned error) and pLDDT (predicted local distance difference test). Low-confidence predictions (pLDDT < 50) correspond to intrinsically disordered regions or poorly resolved interfaces.

Why It Worked

The diffusion approach solves two major limitations of AF2: the bias toward a single conformation (AF2's IPA module converges to a unique minimum) and the inability to model small molecules (ligands, ions) that have no MSA. The table below sets the two generations side by side on the architectural points described above.

Point of comparison	AlphaFold 2	AlphaFold 3
Trunk	Evoformer, with explicit MSA representation	Pairformer, condensed single embedding
Memory complexity	$O(N_{seq} \times N_{res}^2)$	$O(N_{res}^2)$
Coordinate generation	IPA structural module	Diffusion-denoising over 200 steps
Conformations produced	Convergence to a unique minimum	Ensemble of structures (multimodality)
Small molecules (ligands, ions)	Not modeled, for lack of an MSA	Supported

The diffusion model generates an ensemble of structures, capturing conformational heterogeneity — critical for protein-ligand binding sites where flexibility determines affinity.

Training combines the PDB with a cross-distillation of AlphaFold-Multimer v2.3 predictions: AF3 learns from its predecessor while surpassing it, creating a cumulative improvement loop.

Causal Chain

Levinthal's paradox (1969, astronomical conformational space) → CASP (1994, biennial benchmark) → AlphaFold 1 (2018, CASP13, GDT 58) → Evoformer + IPA (AF2, 2020, CASP14, GDT 92.4) → Publication of 200M structures (2022) → AlphaFold 3 (2024, diffusion model, multi-target) → In silico virtual screening (2024–) → Accelerated drug design (horizon 2026–2030)

Historical Anecdote

At the time of its publication in May 2024, AF3 had been evaluated only on internal benchmarks — CASP15 had been held in 2022, before its release — which drew criticism from the community over the absence of independent validation. The model has since undergone an independent blind evaluation at CASP16 (2024), published in Proteins.

Limits and Controversies

The results presented correspond to the data published in the cited sources. Historical measurements have been reassessed by later research — modern values are given where they differ. Current statistics come from institutional sources whose methodologies may vary.

Sources

References verified during the August 2026 fact-checking audit: these are the pages
against which this bulletin's claims were checked.