AlphaFold 3

AlphaFold 3 is Google DeepMind and Isomorphic Labs' molecular structure-prediction model, published in Nature on May 8, 2024, that predicts the three-dimensional structures and interactions of proteins, DNA, RNA, and small-molecule ligands within a single unified architecture. The model is distributed through the AlphaFold Server, which provides free access for non-commercial research, and through Isomorphic Labs for pharmaceutical drug-discovery applications. As of April 2026, it remains the most widely cited AI system for biomolecular structure prediction, with the original Nature paper accumulating more than 9,000 direct citations and the broader AlphaFold platform serving over three million researchers across 190 countries.

At a glance

• Lab: Google DeepMind and Isomorphic Labs
• Released: May 8, 2024
• Modality: Biological structure prediction (proteins, DNA, RNA, small molecules, ligands, ions)
• Open weights: Restricted (CC-BY-NC-SA 4.0; model weights require direct authorization from Google; non-commercial use only)
• Context equivalent: Full biomolecular complex, including protein chains, nucleic acids, post-translational modifications, and selected cofactors
• Pricing: Free for non-commercial research via the AlphaFold Server; commercial access through Isomorphic Labs partnerships
• Distribution channels: AlphaFold Server (non-commercial), GitHub (code and restricted weights), Isomorphic Labs (commercial drug discovery)

Origins

Protein structure prediction has occupied structural biologists for more than fifty years, rooted in the observation that a protein's amino acid sequence determines its three-dimensional shape and therefore its function. Predicting that shape computationally, without expensive and time-consuming crystallography or cryo-electron microscopy, was long considered one of the central unsolved problems in biology.

DeepMind entered the field with AlphaFold 1 in 2018. That version used deep learning to analyze correlated mutations across protein sequences to estimate contact and distance maps between amino acid residues. At CASP13, the biennial Critical Assessment of Protein Structure Prediction competition, AlphaFold 1 placed first overall, outperforming all other entrants on the hardest target categories.

AlphaFold 2, presented at CASP14 in 2020 and published in Nature in 2021, was a more fundamental architectural change. It replaced the modular pipeline with a single end-to-end model using the Evoformer, a transformer-based architecture that jointly processes sequence and structural information through attention over pairs of residues. AlphaFold 2 achieved a median GDT score of 92.4 at CASP14, a result described by the assessment organizers as closing the gap with experimental methods. It was subsequently applied to generate predicted structures for virtually all known proteins, eventually covering more than 200 million entries in the AlphaFold Database. Demis Hassabis and John Jumper shared one half of the 2024 Nobel Prize in Chemistry for this work, with David Baker receiving the other half for computational protein design.

AlphaFold 2's success was substantial but bounded: it handled single-chain proteins and homomeric complexes well, but struggled with the interactions between proteins and small molecules, nucleic acids, and other ligands that define most biological processes of pharmaceutical interest. Drug discovery, for example, depends heavily on understanding how a small molecule binds to a protein target. AlphaFold 2 could identify the target's structure but could not reliably predict the binding interaction itself.

AlphaFold 3, co-developed by Google DeepMind and Isomorphic Labs and published in Nature on May 8, 2024 (Abramson et al.), was designed to address that limitation directly. The architecture shifts from AlphaFold 2's Evoformer to a Pairformer module that generates initial representations, followed by a diffusion network that iteratively refines atomic positions to produce the final three-dimensional structure. This diffusion approach generalizes naturally to any molecular entity rather than requiring protein-specific inductive biases, enabling AlphaFold 3 to handle proteins, DNA, RNA, small-molecule ligands, ions, and post-translational modifications within a single model.

Source code was released on GitHub in November 2024. Model weights became publicly available in February 2025, both subject to a non-commercial restriction.

Capabilities

AlphaFold 3's central advance over its predecessor is scope. Where AlphaFold 2 was trained and optimized for protein chains, AlphaFold 3 handles full biomolecular complexes that include protein-ligand, protein-DNA, and protein-RNA interactions alongside the single-chain and multi-chain protein problems the earlier model addressed.

The architecture uses a Pairformer module that processes residue pair representations to generate initial structural predictions, then passes those representations to a diffusion network. The diffusion model starts from a cloud of atoms and converges on a final atomic configuration through iterative refinement, which allows it to represent molecular flexibility and generate plausible structural ensembles rather than a single rigid prediction.

The model handles the following input types in a single prediction run: protein chains, DNA and RNA strands, small-molecule ligands, selected metal ions, and a range of post-translational modifications and covalent chemical modifications. This breadth reflects the actual composition of biological systems: cellular processes involve proteins operating in complex with nucleic acids, cofactors, and small molecules simultaneously, and understanding that full system is often more informative than predicting protein structure in isolation.

In drug discovery, the most immediately relevant capability is protein-ligand binding prediction. Identifying how a candidate drug molecule fits into a protein binding pocket, and how tightly it binds, is one of the rate-limiting steps in early-stage drug design. Traditional physics-based docking tools have been used for this task for decades; AlphaFold 3 outperforms them on standard benchmarks while integrating the full protein structure prediction step into the same model.

A parallel program at Google DeepMind, AlphaProteo, applies related methods to the inverse problem: designing novel protein binders for specific target molecules. AlphaFold 3 and AlphaProteo are complementary tools addressing structure prediction and protein design respectively.

Benchmarks and standing

Benchmarking for molecular structure prediction differs from the standardized leaderboards used for language models. The primary competitive assessment for protein structure prediction is CASP; for protein-ligand interactions and docking, PoseBusters is the relevant benchmark.

AlphaFold 3 was not directly entered in a CASP cycle post-publication, but the Nature paper's results established the following reference points:

• Protein-ligand docking: AlphaFold 3 shows at least 50% improvement in accuracy versus existing computational methods for predicting how proteins interact with small molecules.
• PoseBusters benchmark: AlphaFold 3 is approximately 50% more accurate than traditional physics-based docking tools on this standard benchmark for drug-like molecules.
• Protein-protein interactions: Performance matches or exceeds AlphaFold 2 on standard protein-protein complex prediction tasks.
• Protein-nucleic acid interactions: Substantially improved versus prior specialized tools for protein-DNA and protein-RNA structure prediction.
• Antibody-antigen complexes: Marked improvement versus prior methods on predicting how antibodies bind to their targets, a key task in biologics drug development.

One significant caveat applies across all categories: accuracy drops on test cases with low similarity to the training data. AlphaFold 3's strong benchmark performance is most reliable for molecules and interaction types well represented in its training set. Novel binding modes or highly atypical ligand chemistries can fall into the low-accuracy regime. Some analyses have also raised questions about whether the model's strong performance on certain benchmarks reflects genuine chemical understanding or memorization of training examples that are structurally similar to the test cases.

For protein-only structure prediction, AlphaFold 3 occupies the same tier as AlphaFold 2 on single-chain targets. The meaningful advancement is in the multi-molecular category, where predecessor models and competing tools had a wider performance gap.

Access and pricing

The AlphaFold Server at https://alphafoldserver.com provides free access to AlphaFold 3 predictions for non-commercial research. Researchers can submit protein sequences along with DNA, RNA, and ligand inputs and receive predicted three-dimensional structures without requiring machine learning infrastructure or expertise. The server is the primary access channel for academic and government research.

Source code is publicly available at the AlphaFold 3 GitHub repository under a CC-BY-NC-SA 4.0 license. Model weights are available for download directly from Google subject to an additional terms-of-use agreement that restricts use to non-commercial research. The latest release as of April 2026 is v3.0.2. Running inference requires a GPU; the data preprocessing pipeline can run on CPU.

Commercial access to AlphaFold 3 for drug discovery is managed through Isomorphic Labs rather than directly through Google DeepMind. Isomorphic Labs, a company spun out from DeepMind in 2021 and majority-owned by Alphabet, operates as the commercial application layer for AlphaFold technology in pharmaceutical contexts. Drug companies that want to use AlphaFold 3 in active drug design programs engage through Isomorphic Labs research collaborations rather than through a public API. Isomorphic Labs has announced a research collaboration with Johnson & Johnson, among other pharmaceutical partnerships. The company has also developed proprietary models that extend beyond AlphaFold 3 for drug design applications.

Comparison

AlphaFold 3 operates in a narrower competitive space than general-purpose AI models. Its nearest comparators are other protein and biomolecular structure prediction systems:

• AlphaFold 2. The direct predecessor remains the standard reference for single-chain protein structure prediction and the most widely cited tool in the field. AlphaFold 3 matches or exceeds it on protein-only tasks and substantially outperforms it on protein-ligand, protein-DNA, and protein-RNA prediction. For researchers who need only single-chain protein structures and whose workflows are built around AlphaFold 2, the upgrade path is well-defined; the main barrier is that AlphaFold 3's non-commercial restriction is the same as its predecessor's server offering. The AlphaFold Database, which contains precomputed structures for over 200 million proteins, continues to run on AlphaFold 2-based predictions.
• ESM3 (Meta AI). Meta AI's ESM3 is a protein language model that handles structure prediction and protein sequence generation. ESM3 uses a different architectural approach, treating proteins as sequences and embedding structural information as token representations rather than explicit atomic coordinates. ESM3 is available under an open-weights license for non-commercial research, with commercial access available through a paid API. Its strengths are in generative tasks and sequence-function relationships; AlphaFold 3's strengths are in explicit multi-molecular structure prediction for drug-discovery applications. ESM3 does not natively handle small-molecule ligand binding in the same way AlphaFold 3 does.
• RoseTTAFold All-Atom (Institute for Protein Design, University of Washington). RoseTTAFold All-Atom is an open-source model from David Baker's group at UW that also extends structure prediction to small molecules and multi-molecular complexes, announced in 2024 alongside AlphaFold 3. It operates on the same problem class as AlphaFold 3 and offers fully open weights under a permissive license without the non-commercial restriction that governs AlphaFold 3 model weights. For commercial drug discovery organizations that prefer not to engage through Isomorphic Labs, RoseTTAFold All-Atom is the primary open alternative. On published benchmarks, AlphaFold 3 holds an edge on most protein-ligand docking tasks; the two models are competitive on protein-nucleic acid tasks.
• Chai-1 (Chai Discovery). Chai-1, released in 2024, is a biomolecular foundation model that handles proteins, small molecules, DNA, and RNA in a single architecture, similar in scope to AlphaFold 3. Chai-1 model weights are available under a non-commercial license, and Chai Discovery offers commercial access through a paid server. The model has performed competitively with AlphaFold 3 on several docking benchmarks, positioning Chai Discovery as a direct commercial competitor to Isomorphic Labs in the AI-driven drug discovery space.

Outlook

Open questions for the next 6 to 18 months:

• AlphaFold Server expansion. The current server supports a defined set of molecule types and modifications. Whether Google DeepMind expands the server's input range to cover additional ligand chemistries, covalent modifications, and larger complexes is an ongoing question for the academic research community.
• Commercialization through Isomorphic Labs. Isomorphic Labs' model of channeling AlphaFold 3 commercial access through research collaborations rather than a public API is untested at scale. How that model evolves, and whether a direct commercial API becomes available, will affect adoption in pharma and biotech contexts.
• Successor model. Google DeepMind and Isomorphic Labs have not publicly committed to an AlphaFold 4 timeline. The pace of the field, with competing models from Meta AI, Chai Discovery, and the Baker lab, creates competitive pressure for continued development.
• Training data generalization. The performance drop on low-training-similarity inputs is a fundamental limitation for drug discovery, where novel binding modes are often precisely the cases of interest. Architectural or data-curation advances that improve generalization to out-of-distribution molecular inputs are the key technical problem the field is working toward.
• AlphaFold Database update. The AlphaFold Database currently reflects AlphaFold 2 predictions. An update to AlphaFold 3-quality predictions at scale, particularly for protein complexes, would be a significant resource for the research community.
• Regulatory and clinical path. AlphaFold 3 is explicitly not validated or approved for clinical use. If structure-prediction tools are to contribute to regulatory filings in drug development, a clearer path for validating and documenting AI-generated structural predictions in that context will be needed.

Sources

• Abramson et al., "Accurate structure prediction of biomolecular interactions with AlphaFold 3," Nature, May 8, 2024. The primary publication; describes the Pairformer and diffusion architecture, benchmark results, and access model.
• Google DeepMind: AlphaFold 3. Official technology page; covers capabilities, AlphaFold Server access, and impact metrics.
• AlphaFold 3 GitHub repository. Source code, weights access instructions, and license terms.
• NobelPrize.org: Chemistry 2024 press release. Details on Hassabis, Jumper, and Baker receiving the 2024 Nobel Prize in Chemistry; announced October 9, 2024.
• Isomorphic Labs. Company overview and commercial drug-discovery approach; Johnson & Johnson partnership announcement.
• Wikipedia: AlphaFold. History of AlphaFold 1 through 3, including CASP results and architecture details.
• Wikipedia: AlphaFold 3. Architecture summary, release timeline, access restrictions, and known limitations.