Chai-1
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,187 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,187 words
Add missing citations, update stale details, or suggest a clearer explanation.
Chai-1 is an open multimodal foundation model for biomolecular structure prediction released by the San Francisco startup Chai Discovery on 9 September 2024. It predicts atomic-resolution 3D structures for proteins, small-molecule ligands, DNA and RNA, multimeric complexes, covalent modifications, and antibody-antigen interfaces from sequence and chemical inputs, and it reaches accuracy comparable to or exceeding Google DeepMind's AlphaFold 3 on several public benchmarks, including a 77% success rate on the PoseBusters protein-ligand evaluation set.[^1][^2] A distinctive feature is its "MSA-free" inference mode, in which residue-level embeddings from a large protein language model substitute for multiple sequence alignments, allowing strong performance on single sequences alone.[^3] The original release made model weights and inference code available under a research-only license at github.com/chaidiscovery/chai-lab, with a subsequent relicensing under Apache 2.0 in late November 2024 permitting broad commercial use.[^4][^5] Chai Discovery followed Chai-1 with the Chai-2 antibody design system, unveiled in June 2025, which the company reports achieves around a 16% wet-lab hit rate on de novo antibody and nanobody design.[^6][^7]
| Property | Value |
|---|---|
| Type | Multimodal biomolecular structure prediction model |
| Developer | Chai Discovery |
| Release date | 9 September 2024[^1] |
| Technical report | bioRxiv 2024.10.10.615955 (posted 11 October 2024)[^3] |
| Repository | github.com/chaidiscovery/chai-lab[^2] |
| License (initial) | Non-commercial research use (code and weights)[^1] |
| License (since Nov 2024) | Apache 2.0 for "Chai-1(r)" code and weights[^4] |
| Default sampling | 5 trunk samples and 5 diffusion samples per recycle; confidence-ranked[^8] |
| Confidence metric | Interface predicted TM-score (ipTM) and related scores[^8] |
| Successor | Chai-2 (announced 30 June 2025)[^6] |
Chai Discovery was founded in early 2024 by Joshua Meier, Jack Dent, Matthew McPartlon, and Jacques Boitreaud, and incorporated in March 2024.[^9] CEO Joshua Meier had previously worked as a researcher at OpenAI (in 2018, during the GPT-1 and GPT-2 era), then at Meta's generative biology group where he co-developed the ESM1 transformer protein-language model, the precursor to the ESM-2 family whose embeddings later powered Chai-1's MSA-free mode. Most recently Meier had served as Chief AI Officer at the Vancouver, Washington biotech firm Absci, where he led the buildout of an AI-driven de novo antibody design program.[^10][^9] President Jack Dent is a Harvard computer-science classmate of Meier who previously led product and engineering efforts at Stripe, where he worked on Stripe Link and Stripe Capital.[^9] CTO Matthew McPartlon had worked with Meier on de novo antibody design at Absci and had earlier joined Proxima Bio for protein-protein interaction modeling, and Jacques Boitreaud studied RNA bioinformatics at McGill University before working as an AI scientist at the small-molecule company Aqemia.[^9]
According to TechCrunch's January 2026 reconstruction of the company's origin, the idea germinated during conversations between Meier, Dent, and OpenAI CEO Sam Altman as early as 2018, when Altman had proposed a proteomics company; Meier judged the technology premature at the time. After Meier spent intervening years at Meta and Absci and Dent built product at Stripe, the four founders eventually built Chai Discovery in 2024 while operating out of OpenAI's San Francisco Mission-district offices, with OpenAI's investment arm joining as a seed investor. Co-founder Dent told TechCrunch that "every line of code in our codebase is homegrown" and that the team had deliberately avoided fine-tuning off-the-shelf language models in favor of custom architectures purpose-built for molecular data.[^11]
Chai Discovery emerged from stealth on 9 September 2024, on the same day it published the Chai-1 model. Bloomberg reported the company had raised roughly $30 million in seed funding led by Thrive Capital, with participation from the OpenAI Startup Fund and Dimension Capital, at a roughly $150 million valuation.[^1][^9] Subsequent rounds expanded the investor base substantially: a $70 million Series A led by Menlo Ventures' Anthology Fund in August 2025 valued the company at $550 million and added former Pfizer Chief Scientific Officer Mikael Dolsten to the board, and a $130 million Series B in December 2025, co-led by General Catalyst and Oak HC/FT with participation from Glade Brook and Emerson Collective, raised Chai's valuation to $1.3 billion and total funding above $225 million. General Catalyst's Hemant Taneja and Oak HC/FT's Annie Lamont joined the board as part of that round.[^11][^12]
When Chai-1 appeared, the field had been transformed by Google DeepMind's AlphaFold 3, announced in May 2024, which generalized AlphaFold's protein-only modeling into a unified diffusion-based architecture that could place small molecules, nucleic acids, ions, and post-translational modifications, all within a single neural network. AlphaFold 3, however, was released only as a web server with restrictive query limits and initially without code or weights, leaving researchers without a self-hostable foundation model with comparable capabilities; the absence of openly distributed weights drew significant criticism from the open-science community.[^3][^12] Boltz-1 from the MIT Jameel Clinic and a handful of academic re-implementations followed in late 2024.[^13] Chai-1 was the first of these openly distributed alternatives to ship in functional form to researchers and drug-discovery teams, beating Boltz-1 to release by approximately two months.[^1][^12]
The competitive landscape mattered for the model's framing. Chai Discovery's launch materials positioned Chai-1 explicitly as an open counterpart to AlphaFold 3, citing matching or superior numbers on the same benchmarks DeepMind had used and offering a self-hosted code path. Independent observers, including ESM co-creator Sergey Ovchinnikov, noted on social media in late November 2024 that the move to Apache 2.0 turned Chai-1 into the most freely usable system in its class.[^4][^19]
The Chai-1 technical report, posted to bioRxiv on 11 October 2024 under the title "Chai-1: Decoding the molecular interactions of life," describes a neural network architecture that broadly follows the trunk-plus-diffusion-decoder design popularized by AlphaFold 3, with heavy use of pair-bias self-attention in the trunk.[^3] The model accepts an unusually wide range of optional input features: multiple sequence alignments, structural templates, residue-level embeddings from a protein language model, ligand SMILES, covalent-bond specifications, and experimental restraints such as cross-link mass-spectrometry contacts or epitope mappings derived from wet-lab data.[^3]
Inputs are converted to per-token features (one token per protein or nucleic-acid residue, plus per-atom tokens for ligands and chemical modifications). The trunk uses iterative pair and single representations updated through several rounds of self-attention and triangular updates, and the resulting features condition a diffusion module that denoises atomic coordinates over a sequence of steps to produce candidate 3D structures.[^3] A confidence head produces per-token plDDT-style scores and pair-wise PAE estimates, along with an interface predicted TM-score (ipTM) used to rank competing samples.[^8]
A central design choice in Chai-1 is the addition of an explicit input track carrying residue-level embeddings from a 3-billion-parameter protein language model in the ESM family (ESM-2), which can be used as a substitute for, or complement to, multiple sequence alignments.[^3] This allows Chai-1 to be run in a fully "MSA-free" or single-sequence mode in which no homology search is performed, dramatically reducing inference latency by eliminating the upstream MMseqs2 or JackHMMR step that dominates conventional AlphaFold pipelines.[^3][^14]
Performance in single-sequence mode remains strong. The technical report reports Cα LDDT of 0.849 on the CASP15 protein monomer set, compared with 0.801 for the 98-billion-parameter ESMFold-family model ESM3-98B, and a 69.8% DockQ success rate on protein-protein multimer prediction without MSAs, exceeding AlphaFold-Multimer 2.3's 67.7% with full MSAs.[^15][^3] On antibody-protein DockQ, Chai-1 in single-sequence mode reaches 47.9%, compared with 38.0% for AlphaFold-Multimer 2.3 with MSAs, and 52.9% when MSAs are also supplied to Chai-1.[^3]
Chai-1 is run at inference time as an ensemble. The default configuration uses 4 recycles, 5 distinct trunk samples (each driven by a different random seed in the trunk noise), and 5 diffusion samples per trunk, producing up to 25 candidate structures per query.[^8] These candidates are then ranked by the confidence head, with ligand-aware ipTM serving as the dominant ranking metric for protein-ligand tasks and a chain-pair ipTM for multimer interfaces.[^3][^8] The "confidence-ranked" Chai-1 variant reported in the paper takes the top sample under this ensemble-and-rank procedure, while a "single-shot" variant uses one trunk sample for direct comparison with single-pass baselines.[^3]
The model can be conditioned on optional experimental restraints, including known contacts between residues, epitope hints for antibody-antigen modelling, and explicit covalent-bond specifications between protein side chains and ligands. The technical report shows that providing four randomly sampled epitope residues for antibody-protein modelling more than doubles DockQ success across all quality cutoffs, and similar gains are reported for crosslink-derived contacts on protein-protein interfaces.[^3] This restraint interface mirrors the kind of partial wet-lab information that pharmaceutical structural-biology teams typically have on hand.
The technical report states that Chai-1 is trained on a combination of multiple sequence alignments and per-residue embeddings from a 3-billion-parameter protein language model, with structural targets drawn from public datasets including the Protein Data Bank.[^3] The report does not publish the precise number of model parameters or the full training-data composition. A community comparison by Boolean Biotech noted that the Chai-lab inference codebase is roughly 10,000 lines of Python, approximately half the size of the comparable Boltz-1 inference stack, and that the initial Chai-1 release shipped inference code only, without the training pipeline.[^14] The absence of training code distinguishes Chai-1 from Boltz-1, which was released with a complete training recipe, and is one of the most-noted limitations in independent commentary, since it constrains the community's ability to fine-tune the model on proprietary structural data or to retrain it with corrections.[^14]
The architecture's choice to feed protein language model embeddings on a dedicated input track, rather than only using them to construct MSA-like features, mirrors trends across the field. ESMFold (which uses only a protein language model and no MSA), Meta's ESM3 (which combines structural and sequence tokens), and OmegaFold all rely on the idea that learned embeddings of protein sequence contain enough evolutionary information to substitute for a homology search, at least for moderately conserved sequences. Chai-1's contribution is to fold this idea into a unified AlphaFold 3-style architecture that handles ligands, nucleic acids, and modifications, and to demonstrate that the resulting MSA-free system matches or exceeds the MSA-using AlphaFold-Multimer 2.3 on the relevant benchmarks.[^3]
The Chai-1 technical report and several independent reviews compare the model to AlphaFold 3, AlphaFold-Multimer 2.3, ESM-family folders, and RoseTTAFold All-Atom across structure-prediction benchmarks. Numbers below are drawn from the Chai-1 technical report version 1 unless otherwise noted.
| Benchmark | Metric | Chai-1 (best) | Comparator | Source |
|---|---|---|---|---|
| PoseBusters (protein-ligand) | success rate (RMSD < 2 Å + validity) | 77% | AlphaFold 3 ~76% | [^3][^16] |
| CASP15 monomers | Cα LDDT | 0.849 | ESM3-98B 0.801 | [^3] |
| Protein-protein multimer (no MSA) | DockQ success | 69.8% | AlphaFold-Multimer 2.3 (with MSAs) 67.7% | [^3] |
| Antibody-protein (no MSA) | DockQ success | 47.9% | AlphaFold-Multimer 2.3 (with MSAs) 38.0% | [^3] |
| Antibody-protein (with MSA) | DockQ success | 52.9% | AlphaFold-Multimer 2.3 38.0% | [^3] |
| Antibody-protein (no MSA, +4 epitope residues) | mean DockQ | 43.7% | unconstrained baseline | [^3] |
A subsequent independent evaluation published on Oxford's Bioinformatics in 2026 found that AlphaFold 3, AlphaFold 2-Multimer, Chai-1, and Boltz-1 produce broadly similar single-best predictions on a 2025 benchmark of antibody-antigen complexes, with AlphaFold 3 modestly ahead on certain interface classes; community reports also emphasized that benchmark gaps between these systems are sensitive to MSA settings, scoring choice, and sample budget.[^17][^14]
The initial 9 September 2024 release placed Chai-1 model weights and the inference Python package under a research-only license that permitted academic use but not redistribution or productized commercial use; Chai Discovery also operated a free web server that, per the company's terms, allowed commercial drug-discovery queries.[^1][^18][^4] This split arrangement drew criticism from some open-source observers who wanted self-hostable commercial use.
On 26 November 2024, the company posted from the @chaidiscovery account that "Chai-1 has always been available for commercial use via our server. Today, we're also making Chai-1(r) code and weights available under an Apache 2.0 license, which permits broad commercial use," and updated the chai-lab repository accordingly.[^4][^19] The relicensed model is sometimes referred to as "Chai-1r" in community writing.[^19] The chai-lab repository on GitHub continued to ship updates through 2025, with a v0.6.1 release in March 2025 adding template support and other refinements, and an Apache 2.0 LICENSE file applied to both code and weights.[^2]
Chai-1 is distributed as the chai_lab Python package (pip install chai_lab), with a Hugging Face mirror at huggingface.co/chaidiscovery/chai-1 hosting weights.[^2] Third-party servers, including BioLM, Neurosnap, Tamarind Bio, and Rowan, expose Chai-1 inference over hosted GPUs.[^20][^21] Oracle Cloud Infrastructure announced in 2025 that it powered parts of the Chai-2 antibody-design infrastructure for Chai Discovery.[^22]
The model attracted significant attention in both technical and trade press. Bloomberg's launch-day story, published 9 September 2024, emphasized the unusual speed with which a six-month-old startup had reached parity with established structure-prediction systems, and quoted Dimension Capital partner Zavain Dar that "in just a few months, Chai Discovery has brought Chai-1 to parity or supremacy over the existing" leaders.[^1] MarkTechPost called Chai-1 a "groundbreaking multi-modal foundation model" and highlighted its PoseBusters score and MSA-free capability.[^16] The AI Insider's coverage on 11 September 2024 characterized Chai-1 as an open alternative to AlphaFold 3 that could "transform" drug discovery, and singled out the model's ability to operate in single-sequence mode without significant accuracy loss as its most distinctive contribution.[^18] The Contrary Research company report described Chai-1 as one of the fastest paths from founding to state-of-the-art technical release in the AI biology space.[^9] TechCrunch's January 2026 retrospective described Chai Discovery, by then valued at $1.3 billion, as "one of the flashiest names in AI drug development" and credited the September 2024 Chai-1 release with establishing the company's technical credibility before its commercial push into antibody design.[^11]
Within structural biology, the response was more mixed. A 2026 community comparison by Boolean Biotech noted that Chai-1 and AlphaFold 3 "perform almost identically" across many metrics and that Chai-1's main practical advantages were a freely available code path and an MSA-free mode that meaningfully reduced wall-clock latency, while pointing out that the initial release omitted training code, limiting the ability of the community to fine-tune.[^14] A Data Helix / Medium head-to-head comparison published in early 2026 found that AlphaFold 3, Chai-1, and Boltz-1 produced overlapping but not identical predictions on a held-out test set, with Chai-1 winning some categories and AlphaFold 3 winning others, suggesting the choice between systems is increasingly a question of workflow and license rather than raw accuracy.[^13] Reviews on the PREreview platform raised concerns about benchmark choice, dataset overlap with training data, and the absence of training-code release, while still rating the technical contribution as substantial.[^23]
Adoption by infrastructure providers underlined the model's traction. Tamarind Bio, BioLM, Rowan Scientific, and Neurosnap all offer hosted Chai-1 inference, and Oracle Cloud Infrastructure publicly announced in 2025 that Chai Discovery had selected OCI to power Chai-2 antibody-design workloads, with Chai-1 continuing as the structural-scoring backbone of that platform.[^22][^19][^8][^21] Chai Discovery additionally announced partnerships with the UK's OpenBind consortium, and in January 2026 with Eli Lilly's TuneLab biologics group, with the latter explicitly citing Chai's generative design models as the technical anchor of the collaboration.[^9][^11]
On 30 June 2025, Chai Discovery announced Chai-2, a successor system that the company describes as a "multimodal generative model" for fully de novo antibody and nanobody design. According to the announcement, Chai-2 was prompted to generate up to 20 antibody or nanobody candidates per target across 52 diverse antigens, with at least one validated binder produced for 50% of those targets in a single round of wet-lab testing, with the entire AI-design-to-validation cycle completed in under two weeks.[^6] Crucially, the company emphasized that none of the 52 targets had a preexisting antibody or nanobody binder in the Protein Data Bank at training time, addressing concerns about memorization of known complexes.[^7]
The accompanying bioRxiv technical report, "Zero-shot antibody design in a 24-well plate" (posted 5 July 2025), reports a 15.5% average hit rate across all designed candidates (20.0% for VHH single-domain antibodies and 13.7% for scFv constructs), a figure the authors characterize as roughly two orders of magnitude above prior computational baselines that typically score below 0.1%.[^7][^24] The Chai-2 paper also reports a 68% success rate on miniprotein binders, often producing picomolar-affinity proteins. According to the report, 86% of designed antibodies showed zero or one developability flags under standard preclinical filters, and 24 of 28 evaluated target antigens yielded at least one drug-like candidate.[^7][^12]
Unlike Chai-1, Chai-2 has not been openly released. Chai Discovery has positioned it as a proprietary platform underpinning the company's commercial partnerships, and the company has not published weights or inference code for the antibody-design model.[^6] In January 2026, Chai Discovery announced a partnership with Eli Lilly's TuneLab biologics group to use Chai-2-class models in Lilly's discovery pipeline; Lilly's Aliza Apple, who heads TuneLab, said the partnership would combine "Chai's generative design models with Lilly's deep biologics expertise and proprietary data" to "push the frontier of how AI can design better molecules from the outset."[^11] (The user-supplied prompt referenced a "March 2025" Chai-2 announcement; the announcement was in fact made on 30 June 2025, with the technical report appearing on bioRxiv on 5 July 2025.)[^6][^7]
A follow-up December 2025 preprint by the Chai Discovery team, "Drug-like antibody design against challenging targets with atomic precision," extended the methodology to harder antigen classes and reported additional wet-lab validation data, including binders against historically difficult target classes such as G-protein-coupled receptors and ion channels.[^25] The combined effect of Chai-1 and Chai-2 has been to reposition Chai Discovery from a purely open-research entity into a platform company whose open structure model anchors a closed antibody-design service, a configuration that mirrors the relationship between Google DeepMind's AlphaFold 3 and Isomorphic Labs's internal drug-design pipelines.[^11][^12]
Within drug discovery, Chai-1 is used in several stages of the early-discovery pipeline:
The web server and the openly distributed chai-lab package have been adopted across both academic structural biology and pharmaceutical research teams, with the chai-lab GitHub repository accumulating roughly 1.9 thousand stars and Chai-1 being cited in several hundred subsequent papers as of early 2026.[^11] The model has also been incorporated into multi-tool benchmarking suites and tutorial articles aimed at computational chemists, including comparative reviews that put it head-to-head against AlphaFold 3 and Boltz-1 on identical input sets.[^13][^14]
Several limitations are noted by Chai Discovery, in independent benchmarks, and in community commentary:
| System | Release | Provider | Open weights | Notes |
|---|---|---|---|---|
| AlphaFold 3 | 8 May 2024 | Google DeepMind / Isomorphic Labs | No (web server; later limited code under non-commercial terms) | First unified protein-ligand-nucleic-acid model[^12] |
| Chai-1 | 9 Sep 2024 | Chai Discovery | Yes (Apache 2.0 since Nov 2024 for Chai-1(r)) | MSA-free with ESM-family embeddings; ensemble sampling[^3][^4] |
| Boltz-1 | Nov 2024 | MIT Jameel Clinic et al. | Yes (MIT license) | Academic re-implementation with full training code[^14] |
| RoseTTAFold All-Atom | 2024 | Baker lab | Yes | Earlier generation; protein-and-small-molecule unified folder[^16] |
| ESMFold (ESM3) | 2023-2024 | EvolutionaryScale | Partial | Language-model-only single-sequence folder[^3] |