AlphaProteo

AlphaProteo is Google DeepMind's generative protein-design system, announced September 5, 2024, that creates novel proteins engineered to bind tightly to a specified molecular target rather than predicting the structure of an already-known protein. It is available to academic and industry researchers through a limited Trusted Tester Program, with commercial drug-discovery applications developed at Isomorphic Labs. As of April 2026, it remains the most publicly documented computational binder-design system to achieve wet-lab-validated success on VEGF-A, a cancer-associated protein that had defeated earlier approaches.

At a glance

• Lab: Google DeepMind
• Released: September 5, 2024
• Modality: Biological (protein structure and sequence generation)
• Open weights: No (closed; Trusted Tester Program for research access)
• Context equivalent: Target protein structure input with optional hotspot residue specification
• Pricing: No public pricing; research access via application to the Trusted Tester Program; commercial use through Isomorphic Labs partnerships
• Distribution channels: AlphaProteo Trusted Tester Program (Colab notebooks); Isomorphic Labs drug-discovery pipeline

Origins

The AlphaProteo project grew directly from the research infrastructure and biological knowledge accumulated during the AlphaFold program. AlphaFold 3 addresses the forward problem in structural biology: given a protein sequence or complex of molecules, predict the three-dimensional structure. AlphaProteo addresses the inverse problem: given a target protein structure, generate a new protein that will bind to it with high affinity and specificity. The two capabilities are complementary rather than competing -- AlphaProteo was trained on more than 100 million predicted structures generated by AlphaFold, a dataset that would have been inaccessible before AlphaFold's own release.

Demis Hassabis has described protein design as the natural next step after structure prediction: once you can read the structural grammar of a protein, designing new proteins that obey that grammar becomes tractable. The preprint (arXiv 2409.08022) lists Vinicius Zambaldi as first author among 32 co-authors from Google DeepMind and the Francis Crick Institute. Crick researchers -- including Peter Cherepanov, Katie Bentley, and David LV Bauer -- handled wet-lab validation, synthesizing candidate binders and testing them in yeast surface display assays.

Before broadening the Trusted Tester Program, the team coordinated with the Nuclear Threat Initiative's AI Bio Forum and other biosecurity advisors to assess potential misuse risks -- a precaution that followed earlier scrutiny of AlphaFold's biosecurity implications.

Isomorphic Labs, the drug-discovery spinout founded by Hassabis in 2021, has been using AlphaProteo capabilities internally since launch. Isomorphic's partnerships with Eli Lilly and Novartis, valued at nearly $3 billion, have advanced AI-designed oncology candidates toward clinical trials, with the first phase-1 doses reported in 2025.

Capabilities

AlphaProteo takes a target protein structure as input, along with an optional specification of hotspot residues -- the binding site regions the researcher wants the new protein to contact. The system outputs both the structure and amino-acid sequence of a candidate binder. The generative model was trained on the Protein Data Bank plus AlphaFold's structural predictions, letting it learn the statistical patterns governing how proteins dock to one another across a vast range of folds and binding modes.

Binder candidates are generated in silico at scale -- typically hundreds per target, 50 to 140 amino acids long -- filtered computationally, and then a subset (47 to 172 per target in the published benchmarks) advances to wet-lab validation. A key practical claim is that AlphaProteo's binders reach usable affinity after a single round of screening without the iterative optimization rounds that traditionally take months.

The range of targets the system handles spans disease areas that have motivated decades of drug discovery. Tested targets include two viral proteins, BHRF1 (an Epstein-Barr virus anti-apoptotic protein) and SC2RBD (the SARS-CoV-2 spike protein receptor-binding domain), plus five proteins implicated in cancer, inflammation, and autoimmune disease: IL-7Ra (interleukin-7 receptor alpha, relevant to leukemia and autoimmunity), PD-L1 (an immune checkpoint implicated in tumor immune evasion), TrkA (a neurotrophin receptor linked to cancer), IL-17A (an inflammatory cytokine targeted by approved biologics), and VEGF-A (a promoter of tumor blood-vessel growth).

The system is described as closed-source in subsequent literature. Researchers gain access through the Trusted Tester Program, which provides Colab notebooks for preparing target inputs, but the model weights and inference infrastructure are not released publicly.

Benchmarks and standing

The preprint reports two main classes of comparison to prior methods: experimental success rates (the fraction of screened candidates that bind in the yeast display assay) and binding affinities (how tightly confirmed binders hold to their targets).

On success rates, AlphaProteo outperformed the next-best method by a factor of 5x on BHRF1, 8x on SC2RBD, and 700x on IL-17A. The 700-fold advantage on IL-17A is the most striking figure in the paper; the prior methods tested against it had near-zero success rates, while AlphaProteo generated a workable hit rate from the same screening budget. For the other targets (PD-L1, TrkA, IL-7Ra, VEGF-A), the paper reports that AlphaProteo had higher overall success rates than RFdiffusion-based designs tested in the same yeast display assay, but does not enumerate fold-differences for each.

On BHRF1 specifically, 88% of AlphaProteo candidates bound successfully in wet-lab validation -- a figure the team characterizes as unusually high for de novo binder design, where single-digit hit rates have historically been common.

On binding affinity, AlphaProteo binders bound their targets 3 to 300 times more tightly than the best designs from existing methods. The average binding improvement was reported as approximately 10-fold across the tested set.

VEGF-A is singled out as a landmark: it is described in the paper as the first target for which any computational method has generated successful de novo binders. VEGF-A had previously resisted de novo design approaches because of characteristics of its binding surface.

The one documented failure is TNFa (tumor necrosis factor alpha), a cytokine central to autoimmune diseases such as rheumatoid arthritis and a target of several approved biologics. Computational analysis identified TNFa as structurally difficult for de novo binder design, and AlphaProteo did not produce successful candidates against it. The team described this as an expected limitation to address in future iterations rather than an unexpected failure.

Since the September 2024 preprint, newer systems have reported higher hit rates on some of the same targets. By mid-2025, Latent-X (from a research group benchmarking on the same targets) reported a 52% hit rate on SC2RBD versus AlphaProteo's 12%, and 49% on PD-L1 versus AlphaProteo's 15%. PXDesign, a separate 2025 method, reported hit rates of 17% to 82% across six of seven diverse targets. These numbers suggest AlphaProteo's September 2024 benchmarks, while landmark at the time, have since been surpassed by successor methods in the academic literature.

Benchmark comparison in this field is complicated by differences in experimental protocols, screening depth, and target difficulty. The 2024 figures above remain significant as a calibration point; the April 2026 landscape for computational binder design has moved substantially since the initial release.

Access and pricing

AlphaProteo is not available as a public API and has no published per-query or subscription pricing.

Research access is through the AlphaProteo Trusted Tester Program. The GitHub repository for the program provides Colab notebooks that help researchers prepare and validate target protein inputs before submitting design requests. The program is described as a collaboration between Google DeepMind and participating research groups rather than a self-serve API; access requires application to the program and is granted selectively.

Commercial applications are pursued through Isomorphic Labs. Isomorphic uses protein-design capabilities -- including but not limited to AlphaProteo -- within its internal drug-discovery platform. Pharmaceutical companies gain access to these capabilities through research partnership agreements with Isomorphic, not through direct licensing of the model. The Novartis and Eli Lilly partnerships signed in early 2024 are examples of this model.

There is no indication as of April 2026 that Google DeepMind plans a general public release of AlphaProteo analogous to the AlphaFold Server or the open-source release of AlphaFold 3 model weights.

Comparison

AlphaProteo operates in a field where several distinct methodological approaches compete:

• RFdiffusion (David Baker lab, University of Washington). RFdiffusion is the most widely used prior-generation method for de novo protein binder design and serves as the primary baseline in the AlphaProteo preprint. It uses a diffusion model trained on protein backbone structures to generate new backbone geometries for a target binding site; sequences are then threaded onto the backbone using ProteinMPNN, a separate sequence design model from the same group. RFdiffusion is open-source and has been deployed by many academic labs. AlphaProteo's published benchmarks show higher success rates and binding affinities than RFdiffusion on the seven targets tested, but RFdiffusion's open availability has kept it in wide use for groups that cannot access the Trusted Tester Program.
• ProteinMPNN (David Baker lab). ProteinMPNN is not a binder-design system on its own but is typically used in tandem with RFdiffusion: RFdiffusion generates the backbone structure, ProteinMPNN designs the amino-acid sequence for that backbone. AlphaProteo is an end-to-end system that generates both structure and sequence jointly, removing the two-stage pipeline. The joint generation is part of why AlphaProteo's affinities are reported as stronger: sequence and structure are co-optimized rather than optimized sequentially.
• ESMFold and ESM-family models (Meta AI). ESMFold is a structure-prediction model rather than a generative binder-design tool, but ESM-based sequence priors underpin several academic binder-design methods -- a different technical lineage from AlphaProteo's AlphaFold-derived structural priors.
• Latent-X, PXDesign, SeedProteo (2025 successors). A cluster of methods published in 2025 report head-to-head benchmarks against AlphaProteo and in several cases surpass it on specific targets. These remain primarily academic tools as of April 2026, confirming that AlphaProteo's 2024 results have been extended -- though not yet matched commercially -- by the open research community.

Outlook

Open questions for the next 6 to 18 months:

• Broader program access. The Trusted Tester Program model limits who can use AlphaProteo in practice. Whether Google DeepMind moves toward a more open research access model -- comparable to the AlphaFold Server -- is the central access question. The AlphaFold Server precedent suggests that broader availability is plausible once biosecurity review criteria are met.
• TNFa and difficult targets. TNFa remains an unresolved challenge. Autoimmune diseases involving TNFa (rheumatoid arthritis, Crohn's disease, psoriasis) represent a large pharmaceutical market. Success against TNFa would significantly expand the practical scope of de novo binder design.
• Integration into Isomorphic Labs clinical pipeline. Isomorphic has advanced small-molecule AI-designed drugs into phase-1 trials. Whether protein binders designed with AlphaProteo follow a similar path -- and on what timeline -- is the key commercial question for the platform.
• Competition from open-source successors. The rapid progress of methods like Latent-X and PXDesign -- which are open-source and report higher hit rates on several benchmarks -- may reduce the closed AlphaProteo system's differentiation. Whether Google DeepMind broadens access or deepens the capability lead is the central strategic question.
• Connection to AlphaFold 4. A Nature article in early 2026 described a new Isomorphic Labs model as a major proprietary advance beyond AlphaFold 3. How that system relates to AlphaProteo's design capabilities is not yet public.

Sources

• AlphaProteo generates novel proteins for biology and health research -- Google DeepMind. Official announcement post, September 5, 2024.
• De novo design of high-affinity protein binders with AlphaProteo -- arXiv 2409.08022. Preprint, Zambaldi et al., 32 authors, September 12, 2024.
• AlphaProteo Trusted Tester Program -- GitHub. Colab notebooks for preparing design requests; access granted by application.
• Google DeepMind unveils AlphaProteo for AI drug design -- MobiHealthNews. Coverage of the September 2024 announcement.
• The Isomorphic Labs Drug Design Engine -- Isomorphic Labs. Isomorphic's description of its AI drug-discovery platform.
• Protein Optimization 102 -- Adaptyvbio. Third-party benchmarking context; cites AlphaProteo's 24.5% hit rate on IL-7Ra.
• Latent-X: An Atom-level Frontier Model for De Novo Protein Binder Design -- arXiv. 2025 successor method reporting comparative benchmarks against AlphaProteo on SC2RBD and PD-L1.
• 'An AlphaFold 4' -- Nature. Nature coverage of Isomorphic Labs' 2026 proprietary advance and its relationship to prior AlphaFold-family systems.