CtrlK
AlphaFold 3 is a biomolecular structure prediction system developed jointly by Google DeepMind and Isomorphic Labs. Published in Nature on May 8, 2024, it extends the capabilities of its predecessor, AlphaFold 2, from single proteins and protein complexes to virtually every class of biological molecule: proteins, DNA, RNA, small molecule ligands, ions, and chemical modifications. The system uses a diffusion-based architecture that replaces the Evoformer module of AlphaFold 2 with a new Pairformer network and a generative diffusion module, allowing it to predict the three-dimensional structure of multicomponent molecular assemblies in a single unified framework.
On the PoseBusters benchmark for protein-ligand docking, AlphaFold 3 achieved roughly 76 percent accuracy under strict physicochemical validity checks, outperforming earlier physics-based docking tools by more than 50 percent. The model was made accessible to academic researchers through the AlphaFold Server shortly after publication, and the inference code and model weights were released for non-commercial use in November 2024. The work contributed to Demis Hassabis and John M. Jumper being awarded the 2024 Nobel Prize in Chemistry.
The AlphaFold project began at DeepMind around 2016 as an attempt to apply deep learning to the protein folding problem, the long-standing challenge of predicting a protein's three-dimensional structure from its amino acid sequence. The first version, AlphaFold 1, competed in the Critical Assessment of Protein Structure Prediction (CASP) competition in 2018 and placed first, but by a modest margin over physics-based competitors.
AlphaFold 2, which competed in CASP14 in 2020, was a complete architectural overhaul. It introduced the Evoformer network, which processes multiple sequence alignment (MSA) information and pairwise distance features simultaneously, along with a structure module that predicted full atomic coordinates using equivariant neural networks. The results at CASP14 were widely described as solving the protein folding problem for many practical purposes. The paper was published in Nature in 2021, and the model was made open source shortly thereafter, with predictions for nearly all known proteins deposited in the AlphaFold Protein Structure Database, maintained jointly with the European Bioinformatics Institute (EMBL-EBI). That database grew to cover more than 200 million protein sequences.
Despite its achievements, AlphaFold 2 had significant limitations. It was designed for single protein chains and, through the AlphaFold-Multimer extension, for homo- and hetero-protein complexes. It could not natively handle small molecule ligands, nucleic acids, or the chemical modifications that are ubiquitous in biology, such as glycosylation, phosphorylation, or methylation. Predicting how a drug candidate would bind to its protein target, for instance, required pairing AlphaFold 2 with separate docking software, introducing additional sources of error at each handoff.
Isomorphic Labs was incorporated in February 2021 and announced publicly in November 2021 as an independent subsidiary of Alphabet, Google's parent company. Demis Hassabis serves as CEO of both Google DeepMind and Isomorphic Labs. The company's founding rationale was that the same capabilities that made AlphaFold 2 effective for protein structure prediction could be extended and retrained to address broader questions in drug discovery, particularly the prediction of how small molecules interact with biological targets.
AlphaFold 3 was developed as a collaboration between Google DeepMind's research team and Isomorphic Labs, with the two organizations contributing complementary expertise in deep learning architecture and drug discovery applications. The paper credits equal contributions from Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, and Lindsay Willmore, alongside a large author list that includes Victor Bapst, Pushmeet Kohli, Max Jaderberg, Demis Hassabis, and John M. Jumper.
AlphaFold 3 replaces the architecture of AlphaFold 2 with two principal new components: the Pairformer module and a diffusion model that operates directly on atomic coordinates. The overall design follows a network that first builds up representations of pairwise relationships between all atoms in the input complex, then uses a diffusion process to sample plausible three-dimensional coordinates consistent with those representations.
The system accepts sequences for protein chains (amino acids), DNA and RNA strands (nucleotides), and small molecule ligands (represented by their SMILES strings or atom-level graphs). Chemical modifications, ions, and covalently attached groups are also accepted. Each entity in the complex is embedded into a per-token representation, where each token corresponds to a single residue for proteins, a single nucleotide for nucleic acids, or a single heavy atom for small molecules.
As in AlphaFold 2, a multiple sequence alignment is used to capture evolutionary information for protein chains. The MSA is processed to produce a set of per-sequence per-residue embeddings, but the MSA module in AlphaFold 3 is simpler than its predecessor and contributes less of the total computational cost than the Evoformer did. Templates (experimentally determined structures of related sequences) can optionally be provided to further constrain predictions, though AlphaFold 3 can also run in template-free mode.
The Pairformer is the central processing module of AlphaFold 3. It maintains two data structures: a single representation (one vector per token) and a pair representation (one matrix element per pair of tokens). Attention operations mix information between these two representations over many layers, allowing the network to reason about the geometry of interactions between all pairs of atoms in the complex simultaneously.
The Pairformer is architecturally similar to the Evoformer of AlphaFold 2 but is simpler in design. It does not perform the outer-product mean operation that coupled the MSA and pair representations in the Evoformer, and it removes the per-sequence dimension from the main processing pipeline. The researchers describe the Pairformer as operating on a "condensed" representation that discards the explicit MSA stack while retaining the evolutionary signal encoded during the earlier MSA processing step. This simplification reduces memory requirements and allows the Pairformer to scale to larger complexes than the Evoformer could handle.
Once the Pairformer has produced per-token and pairwise representations, a diffusion module generates the three-dimensional coordinates of all atoms. AlphaFold 3 uses a denoising diffusion probabilistic model (DDPM) that operates in the space of raw Cartesian coordinates. During training, random Gaussian noise is added to the true atomic positions at varying levels, and the network learns to predict the clean structure from the noisy input conditioned on the Pairformer representations. During inference, the model starts from pure noise and iteratively denoises to produce a final predicted structure.
This is a significant departure from AlphaFold 2's structure module, which predicted backbone torsion angles and applied rigid-body transformations to build atom positions, an approach that enforces correct local geometry by construction but is difficult to extend to chemically diverse ligands and nucleic acids. The diffusion approach treats all atoms uniformly, irrespective of molecular type, and can therefore generate structures for any combination of inputs without molecule-type-specific geometric priors. The trade-off is that the diffusion module must learn correct bond lengths, bond angles, and chirality from data, which is one source of the model's known failure modes.
Because the diffusion process is stochastic, running AlphaFold 3 multiple times on the same input produces slightly different outputs. The authors recommend generating several samples (commonly five) and selecting the one with the highest composite confidence score.
AlphaFold 3 reports several confidence metrics alongside each predicted structure. The pLDDT (predicted local distance difference test) score is a per-atom estimate on a 0-100 scale indicating confidence in local geometry; values above 90 indicate high confidence, while values below 50 suggest the region is probably disordered or incorrectly placed. The PAE (predicted aligned error) is an estimate of the positional error for each pair of tokens after optimal superposition, giving a matrix that characterizes both local accuracy and the confidence in relative orientation between domains or chains. The pTM (predicted template modeling) and ipTM (interface pTM) scores summarize global fold accuracy and interface accuracy, respectively; ipTM values above 0.8 are considered indicative of a reliable prediction, while values below 0.6 suggest a failed prediction.
A defining feature of AlphaFold 3 relative to AlphaFold 2 is its ability to process and co-predict the structures of multiple different molecular classes simultaneously.
AlphaFold 3 predicts protein monomer and complex structures with accuracy broadly comparable to AlphaFold 2 for single chains, and substantially better accuracy for multimers involving antibodies and antigens. For antibody-antigen complexes, AlphaFold 3 significantly outperforms AlphaFold-Multimer v2.3 across standard metrics. For protein monomers, global accuracy gains over AlphaFold 2 are more limited, though local structural accuracy improves.
AlphaFold 3 can predict the structures of DNA duplexes, RNA secondary and tertiary structures, and protein-nucleic acid complexes. On protein-nucleic acid benchmarks, AlphaFold 3 achieves a local distance difference test score in the range of 0.79, compared to approximately 0.65-0.70 for RoseTTAFold-NA, the specialized nucleic-acid-aware version of RoseTTAFold. For RNA multimers and protein-nucleic acid assemblies, AlphaFold 3 substantially surpasses RoseTTAFoldNA in both global and local prediction accuracy, though performance on RNA monomers varies depending on the metric.
Small molecule docking is the area where AlphaFold 3 shows the largest gain over prior tools. On the PoseBusters benchmark, which tests protein-ligand structure prediction with strict physicochemical validity checks (correct chirality, no severe clashes, valid bond lengths), AlphaFold 3 achieves approximately 76 percent accuracy by RMSD (root mean square deviation below 2 angstroms). RoseTTAFold All-Atom achieves approximately 42 percent on the same benchmark. Traditional physics-based docking programs such as AutoDock Vina achieve lower success rates still. The Nature paper describes AlphaFold 3 as achieving "over 50% improvement" relative to the best traditional methods on the PoseBusters benchmark.
Importantly, AlphaFold 3 predicts protein-ligand complexes de novo from sequence and SMILES string alone, without requiring an experimentally determined apo protein structure as input, which distinguishes it from blind docking pipelines that typically require a receptor structure.
AlphaFold 3 handles metal ions and small inorganic species as atomic tokens alongside organic molecules. It also accepts covalently modified residues, including post-translational modifications on proteins (phosphorylation, acetylation, glycosylation) and modified nucleotides in RNA. Performance on these categories is notably lower than for standard protein-ligand prediction: the Nature paper reports around 79 percent accuracy for covalent ligands, 46 percent for single glycan structures, 51 percent for modified protein residues, and 41 percent for modified RNA residues.
The Nature paper presents comprehensive benchmarks comparing AlphaFold 3 against prior methods across molecular categories.
For protein-ligand binding, AlphaFold 3 is the first AI system to surpass physics-based docking tools on the PoseBusters benchmark without providing any experimental structural information. The 76 percent success rate under strict validity filters compares to approximately 42 percent for RoseTTAFold All-Atom and lower figures for physics-based programs.
For protein-nucleic acid complexes, AlphaFold 3 exceeds nucleic-acid-specific predictors by a wide margin, achieving substantially higher TM-scores, lDDT scores, and interaction network fidelity (INF) scores.
For antibody-antigen complexes, the model significantly outperforms AlphaFold-Multimer v2.3 on DockQ scores, an important result for antibody drug development where predicting binding mode is critical.
For protein monomers, the gains over AlphaFold 2 are modest at the global level, since AlphaFold 2 already achieved near-experimental accuracy for many single-chain proteins, but local geometry and loop predictions improve in AlphaFold 3.
Alongside the Nature publication, Google DeepMind launched the AlphaFold Server at alphafoldserver.com. The server is free to use for non-commercial research and provides a web interface for submitting prediction jobs. Users can enter protein sequences, nucleic acid sequences, small molecule SMILES strings, and ions as a combined input, and the server returns predicted structures in CIF or PDB format along with confidence score files.
The server imposes rate limits on submissions to manage compute demand. Individual users receive a limited number of job credits per day, and submissions are restricted to complexes below a defined size threshold. At launch, this meant researchers could not run predictions on very large assemblies or submit large batches programmatically, which drew criticism from scientists who needed high-throughput access.
The server interface was designed in collaboration with EMBL-EBI, which hosts the AlphaFold Protein Structure Database, ensuring continuity with the AlphaFold ecosystem established by AlphaFold 2.
At the time of the original Nature publication in May 2024, Google DeepMind released neither the full source code nor the model weights of AlphaFold 3, providing access only through the web server. This drew immediate criticism from the scientific community. A letter signed by hundreds of researchers argued that the restriction violated the scientific norms of reproducibility and that the lack of open weights prevented independent validation of the results. Nature also published commentary highlighting the tension between commercial interests and open science.
In response, Google DeepMind announced in May 2024 that it would release the code and weights for non-commercial use within approximately six months. On November 11, 2024, the inference code and model weights were made available on GitHub (at github.com/google-deepmind/alphafold3). The release is licensed under Creative Commons Attribution Non-Commercial ShareAlike 4.0 (CC-BY-NC-SA 4.0); commercial use requires a separate agreement with Google DeepMind, and weights can be accessed only upon request after agreeing to usage terms.
The release is therefore not fully open source by the conventional definition used in software and AI communities. The model weights are gated behind a request-and-approval process, and commercial use is prohibited without a license. Nonetheless, the release was welcomed by academic researchers who could now run the model on their own compute infrastructure, perform ablation studies, and fine-tune the model on custom datasets.
On October 9, 2024, the Royal Swedish Academy of Sciences announced the 2024 Nobel Prize in Chemistry. Half the prize was awarded to David Baker of the University of Washington "for computational protein design." The other half was awarded jointly to Demis Hassabis and John M. Jumper "for protein structure prediction."
The Nobel committee cited AlphaFold 2 as the specific work being recognized; AlphaFold 3 had been published only months before the announcement. The prize amount for the AlphaFold half was approximately 5.5 million Swedish kronor, shared equally between Hassabis and Jumper.
Jumper, who holds a PhD in theoretical chemistry from the University of Chicago, had been the lead technical architect of AlphaFold 2 and was a co-author on the AlphaFold 3 paper. Hassabis, who founded DeepMind in 2010, is a co-founder and CEO of Isomorphic Labs in addition to his role at Google DeepMind. The Nobel committee noted that since AlphaFold 2's release, the system had been used by more than two million researchers across 190 countries to predict the structures of virtually all 200 million known proteins.
The award was the first Nobel Prize given primarily for work done by an AI system, prompting broad discussion about the changing relationship between artificial intelligence and scientific discovery.
Several competing systems with similar capabilities to AlphaFold 3 have been published. The table below summarizes the key attributes of the principal alternatives.
| Model | Developer | License | Molecular coverage | PoseBusters ligand accuracy | Notes |
|---|---|---|---|---|---|
| AlphaFold 3 | Google DeepMind / Isomorphic Labs | CC-BY-NC-SA (non-commercial) | Proteins, DNA, RNA, ligands, ions, modifications | ~76% | Gated weights; commercial license required |
| RoseTTAFold All-Atom | Baker Lab, University of Washington | Open source (MIT-style) | Proteins, nucleic acids, small molecules, metals | ~42% | Published Science March 2024; generative design capabilities via RFdiffusion All-Atom |
| Chai-1 | Chai Discovery | Apache 2.0 (code + weights) | Proteins, DNA, RNA, ligands, glycosylations | ~77% | Released September 2024; comparable AF3 accuracy; fully open for commercial use |
| Boltz-1 | MIT (Wohlwend, Corso, Barzilay et al.) | MIT (fully open) | Proteins, nucleic acids, small molecules | Comparable to AF3 | Released December 2024; fully open-source including weights |
| Boltz-2 | MIT / Recursion Pharmaceuticals | MIT | Proteins, nucleic acids, small molecules | Comparable to AF3 | Released 2025; adds binding affinity prediction; 1000x faster than physics-based FEP |
RoseTTAFold All-Atom was published in Science in March 2024, two months before AlphaFold 3, by the Baker Lab at the Institute for Protein Design at the University of Washington. It uses a combined residue-level and atomic-level representation to handle proteins, nucleic acids, small molecules, metals, and covalent modifications in a single network. RoseTTAFold All-Atom also includes a diffusion-based generative component (RFdiffusion All-Atom) that was used to design proteins binding to specific small molecules, a capability that AlphaFold 3 does not natively provide. David Baker's receipt of the Nobel Prize alongside Hassabis and Jumper in 2024 partially recognized this broader body of work.
Chai-1, released by Chai Discovery in September 2024, is a multi-modal foundation model with an architecture that closely follows AlphaFold 3. Chai-1 achieves approximately 77 percent on the PoseBusters ligand docking benchmark, slightly higher than AlphaFold 3's 76 percent figure. It was released under the Apache 2.0 license, meaning both code and weights are fully open for academic and commercial use, which distinguished it sharply from AlphaFold 3's restricted release.
Boltz-1, released by a team at MIT's Jameel Clinic in December 2024, was the first fully open-source model to match AlphaFold 3's performance on biomolecular structure prediction. Lead developers Jeremy Wohlwend and Gabriele Corso, working with professors Regina Barzilay and Tommi Jaakkola, spent approximately four months building Boltz-1. Unlike AlphaFold 3, all weights are freely available under an MIT license with no commercial restriction. Boltz-1 adds a feature for conditioning predictions on user-specified binding pockets and a "Boltz-steering" mechanism to improve the physical plausibility of predicted poses.
Boltz-2, released in 2025 as a collaboration between MIT researchers and Recursion Pharmaceuticals, extends Boltz-1 with joint prediction of binding affinity. It is described as the first deep learning model to approach the accuracy of physics-based free-energy perturbation methods, while running approximately 1000 times faster. Boltz-2 is also released under the MIT license permitting commercial use.
Isomorphic Labs was founded with the explicit goal of applying AI to drug discovery, and AlphaFold 3 is the central tool in its platform. The company's drug design engine combines AlphaFold 3's structure prediction with additional modules for affinity estimation, selectivity prediction, and molecule generation. In January 2024, Isomorphic Labs announced major research partnerships with Eli Lilly and Novartis. The two deals are structured around target identification, lead optimization, and milestone payments, and the companies stated they could be worth nearly three billion dollars in total if all milestones are achieved.
Isomorphic Labs published a case study examining TIM-3, an immune checkpoint receptor involved in cancer immunotherapy resistance. Using AlphaFold 3 predictions of the TIM-3 protein structure and its binding pocket, the team designed small molecule candidates targeting the receptor with high predicted affinity. This work was presented as a demonstration of structure-based drug design that is now practical without requiring experimental structure determination of the target at each step.
By 2025, compounds identified through Isomorphic's AI platform had entered Phase I clinical trials, a milestone cited in reporting about AI-designed drugs reaching human testing.
One application of AlphaFold 3 in drug discovery is target identification: given a newly sequenced protein associated with a disease, researchers can predict its structure and identify potential binding pockets before any experimental structure is available. AlphaFold 3's ability to predict complexes with ligands in a single step means that virtual screening campaigns can be run against predicted structures rather than waiting for X-ray crystallography or cryo-electron microscopy data.
For existing drug programs, AlphaFold 3 can predict the binding poses of candidate molecules to a target protein, providing structural context for medicinal chemistry. Medicinal chemists can use predicted poses to understand which chemical modifications might improve binding affinity, selectivity, or metabolic stability. The model's ability to handle the full chemical diversity of drug-like small molecules without requiring molecule-specific parameterization makes it broadly applicable across therapeutic programs.
AlphaFold 3's improved accuracy on antibody-antigen complexes compared to AlphaFold-Multimer makes it useful for antibody engineering. Predicting how an antibody's complementarity-determining regions contact an antigen epitope can guide engineering efforts to improve binding, reduce off-target interactions, or convert mouse antibodies to human sequences. This complements Google DeepMind's AlphaProteo project, which uses related techniques for de novo protein binder design.
Beyond drug discovery, AlphaFold 3's multimodal capabilities are valuable in structural genomics and basic biology. Researchers studying gene regulation use protein-DNA and protein-RNA complex predictions to model how transcription factors, RNA-binding proteins, and epigenetic regulators interact with nucleic acid substrates. The ability to include modified bases (such as methylated cytosine in DNA) and post-translational modifications in predictions makes the tool relevant to epigenetics research.
As a generative model, AlphaFold 3 can produce plausible-looking but incorrect structures for regions that are intrinsically disordered in solution. AlphaFold 2 also hallucinated structure in disordered regions but typically produced distinctive low-complexity ribbon-like arrangements that were recognizable as model artifacts. AlphaFold 3's diffusion-based outputs in disordered regions can look more like ordered helices or sheets, making them harder to identify as hallucinations by visual inspection alone. Research published in 2025 found that approximately 22 percent of residues in intrinsically disordered proteins were hallucinated by AlphaFold 3, and that 18 percent of functionally important residues in disordered biological process regions were incorrectly predicted as ordered. The pLDDT score below 50 remains the recommended diagnostic for disordered regions.
Like all structure prediction tools based on the Protein Data Bank, AlphaFold 3 predicts a single static conformation rather than an ensemble of conformations or a dynamic trajectory. Biological molecules are flexible; enzymes undergo conformational changes upon ligand binding, and many proteins switch between active and inactive states. AlphaFold 3 does not capture these dynamics, and a predicted pose may correspond to a minority conformation or an induced-fit state that does not reflect the predominant solution structure.
AlphaFold 3 has a chirality violation rate of approximately 4.4 percent on the PoseBusters benchmark, meaning that a small fraction of predicted ligand poses have incorrect stereochemistry at chiral centers. This is a consequence of the diffusion approach learning from data rather than enforcing strict chemical rules by construction. Post-processing steps that check and correct chirality are recommended when using AlphaFold 3 predictions for downstream chemistry.
Accuracy drops substantially for covalently bound ligands and chemically modified residues. Covalent inhibitors (which form a chemical bond to the target) are predicted correctly in approximately 79 percent of cases. Single glycan structures are predicted correctly only about 46 percent of the time, and modified RNA residues only about 41 percent of the time. These categories remain substantially harder than standard protein-ligand docking.
Despite the November 2024 code release, AlphaFold 3 is not freely available for all uses. The model weights require a request-and-approval process from Google DeepMind, and commercial applications require a separate commercial license. Researchers who need to use AlphaFold 3 for drug discovery at a company, for instance, cannot legally download and run the weights without a commercial agreement. Open alternatives such as Chai-1 and Boltz-1 fill this gap but differ in training data, architectural choices, and performance on specific tasks.
Running AlphaFold 3 locally requires substantial GPU resources. The model is large, and generating multiple samples (which is recommended for reliability) multiplies the compute cost. For many academic groups without access to large GPU clusters, the AlphaFold Server remains the practical access point despite its rate limits.