EvolutionaryScale

EvolutionaryScale is an American artificial intelligence research company that built frontier generative models for biology, best known for ESM3, a multimodal generative protein language model that can reason over and generate protein sequence, structure, and function. It was founded in July 2023 by researchers led by Alexander Rives, together with Tom Sercu and Sal Candido, who had built the Evolutionary Scale Modeling (ESM) protein language models inside the Fundamental AI Research lab at Meta ^[4][5]. The company emerged from stealth on 25 June 2024 with ESM3 and a seed round of more than 142 million United States dollars ^[4][5]. EvolutionaryScale describes ESM3 as "the first generative model for biology that simultaneously reasons over the sequence, structure, and function of proteins" ^[2], and used it to design esmGFP, a novel fluorescent protein the company estimates is equivalent to "over 500 million years of natural evolution" ^[1][2].

The company was organized as a public benefit corporation and shipped the ESM3 family, the ESM Cambrian (ESM C) family, and a developer platform called Forge that exposed the models through an API, alongside distribution through Amazon Web Services and NVIDIA ^[2][3][4]. On 6 November 2025 the Chan Zuckerberg Initiative announced a strategic transaction under which the EvolutionaryScale team joined Biohub, the nonprofit medical research organization founded by Priscilla Chan and Mark Zuckerberg; financial terms were not disclosed, and Rives became Biohub's head of science ^[11][19][36]. The company's website now states that "EvolutionaryScale is now part of Biohub" and that current ESM models, APIs and related services are available at biohub.ai ^[26]. In May 2026 the former EvolutionaryScale team, working at Biohub, released ESMC, the ESMFold2 structure predictor, and a new ESM Atlas of 6.8 billion protein sequences ^[14][23][25].

EvolutionaryScale's strategy differed from that of Google DeepMind and Isomorphic Labs, whose AlphaFold 3 focused on structure and interaction prediction: EvolutionaryScale concentrated on generative protein design and on a scalable protein language model family that biologists could prompt with any combination of sequence, structure, and function ^[2][6]. ESM3 was published in the journal Science in January 2025 ^[1], and the company said it was trained with more compute than any other known model in biology ^[4].

What is EvolutionaryScale?

EvolutionaryScale is an AI research lab, based in New York and San Francisco at the time of its launch, that develops biological foundation models, the most prominent being the ESM3 generative protein language model ^[2][4][6]. It was organized as a public benefit corporation with a mission "to develop artificial intelligence to understand biology for the benefit of human health and society, through partnership with the scientific community, and open, safe, and responsible research" ^[2][4]. Rather than running its own drug pipeline, the company built models and offered access to them through its Forge platform and through partner clouds at Amazon and NVIDIA, and told TechCrunch it planned to make money through a combination of partnerships, usage fees, and revenue sharing ^{[4][5][9][10]}. The company was part of the wider AI for science movement applying frontier compute to the life sciences ^[5][6]. Since November 2025 its team has worked inside Biohub, and its models and API are now operated by Biohub ^[11][14][26].

How was EvolutionaryScale founded?

The scientific lineage of EvolutionaryScale runs back to the Evolutionary Scale Modeling, or ESM, project that Alexander Rives started and led inside Meta's Fundamental AI Research lab (FAIR), where Rives, Sercu, and Candido began developing AI models for proteins in 2019 ^[5][20]. Rives earned a bachelor's degree in philosophy and biology from Yale University and a doctorate in computer science from New York University ^[20]. The ESM group's 2021 paper in PNAS trained transformer protein language models of about 670 million parameters on 86 billion amino acids across 250 million protein sequences, and showed that information about protein structure and function emerged in the learned representations ^[17]. ESM2 followed with up to 15 billion parameters, and the team paired it with ESMFold, a structure predictor that the group used to build the ESM Metagenomic Atlas of more than 617 million predicted structures from metagenomic proteins ^[18].

According to the company, its founding team left Meta in April 2023 and started EvolutionaryScale, which was founded in July 2023 ^[4]. Fortune reported that in August 2023, as part of Mark Zuckerberg's "year of efficiency," Meta disbanded the research team of a dozen scientists that had trained the protein model, and that Rives spun up the startup with a core group of former Meta colleagues ^[6]. Rives took the role of chief executive and chief scientist of the new entity ^[5][6]. Rives told Fortune that the move was driven by scaling ambitions: "Meta is not a biotechnology company," he said, adding that "we were reaching the point where we really wanted to go to the next level of scaling these models. I think building a new company was really the right place to do that" ^[6]. Fortune placed the founders among a "Meta AI mafia" of former Meta researchers who had started companies ^[6].

The founding team included Tom Sercu and Sal Candido, both of whom had worked on the ESM models at Meta; Sercu had co-led the protein team at FAIR ^{[5][18][21][22]}. At its June 2024 launch the company employed roughly 20 people, according to TechCrunch ^[5].

How much funding has EvolutionaryScale raised?

EvolutionaryScale launched publicly on 25 June 2024, almost a year after its founding, with the simultaneous announcement of ESM3 and a seed financing of more than 142 million United States dollars ^[4][5][7]. The round was led by the technology investors Nat Friedman and Daniel Gross, together with Lux Capital ^[4][5][6]. Amazon and NVentures, the venture capital arm of NVIDIA, participated, along with angel investors ^[4][5][9]. No valuation for the round was disclosed by the company or reported by the outlets that covered it ^[4][5][6][7].

Rives told Fortune that biological foundation models "require a great deal of compute to build and train, similar to other frontier modeling efforts across AI," and that the size of the raise "really reflects the resources that we need to do that" ^[6]. The company said the money would be used to further expand the capabilities of its models ^[4]. TechCrunch noted that a pitch deck obtained by Forbes in August 2023 repeatedly emphasized that it could take a decade for generative AI models to help design therapies ^[5].

Funding history

Sources: ^{[4][5][11][13]}.

Who founded and led EvolutionaryScale?

The founding team carried over the working relationships of the ESM group at Meta ^[5][6]. Rives served as the public face of the company and the senior author on its scientific publications ^[1][6]. Sercu, as vice president of engineering, worked across engineering and AI research and led the company's organization building, hiring, and operations ^[9][21]. Candido, as chief technology officer, led or made foundational technical contributions to all of the company's flagship research efforts, according to his Biohub profile ^[22]. Thomas Hayes, Roshan Rao, and Halil Akin were the first three listed authors of the ESM3 paper ^[1].

Sources: ^{[5][6][9][19][20][21][22]}.

What is ESM3?

ESM3 is EvolutionaryScale's flagship model, which the company calls "the first generative model for biology that simultaneously reasons over the sequence, structure, and function of proteins" ^[2]. It was the scientific case for the entire company, resting on the bet that protein language models, which are transformers trained on sequences of amino acids, would keep improving with scale in the way that text language models had ^[2][5]. Earlier ESM work had shown that information about secondary and tertiary structure and mutational effects emerged in the representations of masked language models trained on protein sequences ^[17][18]. ESM3 was described by the company as an order of magnitude larger than ESM2 and natively multimodal and generative ^[2].

The core technical innovation of ESM3 was masked generative modeling over several tracks of discrete tokens. One track encoded the amino acid sequence of a protein. Another encoded a tokenized representation of local three dimensional structure, produced by a vector quantized variational autoencoder (VQ-VAE) that the team trained to compress atomic coordinates into a discrete vocabulary. Function was represented by keywords and InterPro annotations, and further tracks carried secondary structure and solvent accessibility ^[16]. All tracks were fused within a single latent space, and the model was trained to predict masked tokens, so that any track could be masked or revealed at inference time. This allowed a single network to perform structure prediction, sequence design conditioned on structure, function conditioned generation, and editing of arbitrary subsets of residues ^[2][16].

How was ESM3 trained?

ESM3 was trained at three scales: 1.4 billion, 7 billion, and 98 billion parameters, with the largest model using 216 transformer blocks ^[16]. The largest model was trained on 2.78 billion natural proteins drawn from sequence and structure databases including UniRef, MGnify, and the Joint Genome Institute, supplemented with predicted structures and with synthetic sequences generated by an inverse folding model, for a total of 3.15 billion protein sequences, 236 million protein structures, 539 million proteins with function annotations, and 771 billion unique tokens ^[16]. The 98B model was trained with 1.07 x 10^24 floating point operations, which the company summarized as: "We believe that ESM3 is the most compute ever applied to training a biological model" ^[2][16]. Training was performed on the Andromeda cluster, which uses NVIDIA H100 Tensor Core GPUs with NVIDIA Quantum-2 InfiniBand networking ^[9]. NVIDIA reported that the 98 billion parameter model used roughly 25 times more floating point operations and 60 times more data than ESM2 ^[9].

Sources: ^[2][9][16].

What is esmGFP and the 500 million years result?

To show that ESM3 could generate proteins with desired properties rather than only predict them, EvolutionaryScale asked the model to design new green fluorescent proteins that differed substantially in sequence from any natural GFP ^[1][2]. The team prompted the 7 billion parameter base model with a few residues from the core of natural GFP and used a chain of thought procedure over structure and sequence to generate candidates ^[2][16]. In a first experiment of 96 designs, several proteins fluoresced, including a dim one far from any natural protein (well B8); a second round of 96 designs, starting from B8, produced several proteins with brightness similar to natural GFPs, the brightest of which (well C10) the team named esmGFP ^[2]. esmGFP "differs by 96 mutations (out of 229 amino acids, 58% of the sequence is similar) to the closest fluorescent protein found in nature" ^[2]. Based on the rate at which natural GFPs diversify, the team estimated that "esmGFP represents an equivalent of over 500 million years of natural evolution" ^[2], a figure that became the title and central headline of the subsequent paper ^[1].

When was the ESM3 paper published?

The company released the ESM3 preprint on its website on 25 June 2024, and the preprint was posted to bioRxiv on 2 July 2024 ^[5][16]. Science received the manuscript on 24 July 2024 and accepted it on 7 January 2025. The paper, "Simulating 500 million years of evolution with a language model," was published online on 16 January 2025 and appeared in print in volume 387, issue 6736 (21 February 2025), pages 850 to 858 ^[1].

What models has EvolutionaryScale released?

EvolutionaryScale released two main model families during its independent life. The ESM3 family covered controllable generation across sequence, structure, and function. The ESM Cambrian family, announced on 4 December 2024, focused on representation learning rather than generation and was presented as a successor to ESM2 for embedding and downstream prediction tasks ^[3][28]. According to the company, the 300M ESM C model delivered performance similar to ESM2 650M, the 600M model rivaled the 3B ESM2 model, and the 6B model outperformed the best ESM2 models by a wide margin ^[3]. After the team moved to Biohub, the ESM C line was extended with ESMFold2 and new sparse autoencoders in May 2026 ^[14][23].

Sources: ^{[2][3][4][14][16][17][18][23][26][27][28]}.

The open models were released through the company's GitHub repository and published to PyPI as the package named esm ^[14]. At launch, the company made the 1.4B ESM3-open model available for non commercial use ^[4], and TechCrunch reported that the full 98 billion parameter model was available for non commercial use through Forge ^[5]. In 2025 the company's homepage still described ESM3-open as available "under a non-commercial license" ^[27]; by October 2026, the homepage and the ESM3 blog post described the ESM3 1.4B open model as released under the MIT license ^[2][26]. ESM C was launched for academic and commercial use under what the company called "a new license structure designed to promote openness" ^[28], and Biohub later released ESM C 6B, ESMFold2, and the ESM C sparse autoencoders under the MIT license ^[14]. The Hugging Face repositories for ESM3-open and ESM C are now published under Biohub's organization on Hugging Face ^[35].

What is the Forge platform?

Forge was EvolutionaryScale's cloud developer platform, through which it offered API access to the ESM3 and ESM C families ^[4][5]. The company opened the API as a closed beta on launch day in June 2024, and in December 2024 offered ESM C 6B on Forge for academic use ^[4][28]. After the move to Biohub, the API migrated from forge.evolutionaryscale.ai to biohub.ai, and the open source client library still refers to it in places as "Forge" ^[14]. Biohub also offers a batch API for larger workloads ^[14].

At launch the company also distributed ESM3 through two large cloud partners. AWS made the open version of ESM3 available on Amazon SageMaker and AWS HealthOmics, said the proprietary ESM3 family would follow on those services, and said Amazon Bedrock support would come later in 2024 ^[10]. EvolutionaryScale said that working with Amazon Web Services made the ESM3 family accessible to nine of the top ten global pharmaceutical companies, which already used AWS generative AI and health services ^[4]. NVIDIA said the full ESM3 family would be available to select customers as an NVIDIA NIM microservice, and that the small open version was coming to the NVIDIA BioNeMo platform ^[9].

Partnerships and distribution channels

EvolutionaryScale's stated commercial plan was to work with pharmaceutical companies to integrate ESM3 into their workflows and to share revenue with researchers for discoveries commercialized using ESM3 ^[5]. By contrast, Isomorphic Labs signed partnerships with Eli Lilly and Novartis in 2024, which TechCrunch reported could generate up to 3 billion dollars in milestone payments ^[31].

How does EvolutionaryScale compare to AlphaFold and other protein AI labs?

EvolutionaryScale operated among a small group of companies applying large models to biology. TechCrunch listed Isomorphic Labs, which already had contracts with large pharmaceutical companies, as well as Insitro, Recursion, and Inceptive as competitors at launch, while Fortune pointed to InstaDeep, Profluent, and Google DeepMind's AlphaFold ^[5][6]. Fortune compared EvolutionaryScale's plan, free open versions for academics plus a commercial version for pharmaceutical companies, with Google DeepMind's approach of offering AlphaFold to researchers while its spinout Isomorphic Labs works with pharmaceutical partners ^[6]. The relationship to the wider AlphaFold lineage is one of complementary problems: AlphaFold and the ESM team's own ESMFold predict structure from sequence, whereas ESM3 inverts and unifies that task by generating sequence, structure, and function together under one masked generative model ^[1][2]. An April 2026 IntuitionLabs report compared ESM3 with AlphaFold 3, Boltz-2, Chai-1, and RoseTTAFold ^[15].

Comparison with peer protein AI labs

Why is EvolutionaryScale a public benefit corporation?

EvolutionaryScale was organized as a public benefit corporation; its launch announcement used that term, while its blog called it "a public benefit company" ^[2][4]. The company described its mission as developing artificial intelligence to understand biology for the benefit of human health and society ^[2]. At the ESM3 launch it published a Responsible Development Framework, citing the 1975 Asilomar Conference as an inspiration ^[2]. The ESM3 preprint said the ESM3-open model was reviewed by a committee of technical experts who found that "the benefits of releasing the model greatly outweighed any potential risks" ^[16]. The company also said the scientific team behind the ESM models was among more than 160 signatories of a framework for responsible development of AI in biodesign ^[4]. AWS wrote that ESM3 "was architected to mitigate safety risks like generating hazardous proteins" ^[10].

After the move to Biohub, Biohub said it had established a safety team to assess the benefits and risks of its models before release, and that its freely accessible platform uses guardrails that detect and restrict keywords and sequences corresponding to controlled pathogens and toxins ^[14].

What happened to EvolutionaryScale in 2025?

On 6 November 2025 Mark Zuckerberg and Priscilla Chan announced that Biohub would launch a large initiative combining frontier AI with frontier biology, and that the EvolutionaryScale team would join Biohub ^[19]. Gunderson Dettmer, which represented EvolutionaryScale, described the deal as a strategic transaction with the Chan Zuckerberg Initiative ^[11]. Financial terms were not disclosed ^[36]. Rives took the role of head of science, leading "an integrated research strategy across experimental biology, data, and artificial intelligence," and kept his affiliations with the Broad Institute and MIT ^[19]. Biohub said the combined team would develop the datasets, laboratory technologies, and modeling innovations needed for the next generation of AI systems for biology, and that it would expand its compute capacity tenfold, to 10,000 GPUs by 2028 ^[19].

Endpoints News described the deal as an "apparent end" of the AI startup, with Biohub hiring the EvolutionaryScale team ^[13]. Axios reported that Zuckerberg and Chan were refocusing their philanthropy on the intersection of biology and AI ^[12], and GEN later wrote that the Chans would place the majority of their philanthropic effort in Biohub ^[23]. Zuckerberg said: "When we started, our goal was to help scientists cure or prevent all diseases this century. With advances in AI, we now believe this may be possible much sooner" ^[19].

Status of the models after the transaction

EvolutionaryScale's website now carries a notice that the company is part of Biohub and that current ESM models, APIs, and related services are available at biohub.ai under Biohub's terms of use ^[26]. The homepage states that all ESM3 models are available on the Biohub Platform ^[26]. The code repository moved from the evolutionaryscale organization on GitHub to Biohub/esm, and the API endpoint moved from forge.evolutionaryscale.ai to biohub.ai ^[14]. Open weights for ESM3-open and the ESM C models remain available on Hugging Face under the biohub organization ^[35].

What has the team released at Biohub?

On 27 May 2026 Biohub released a set of open models built on the ESM line: ESMC, a protein language model; ESMFold2, a structure prediction model built on the ESMC 6B model; and a new ESM Atlas ^[14][23][24]. GEN reported that the release came seven months after Biohub recruited the EvolutionaryScale team, and that Rives presented the work at the "AI in Biology" symposium at Cold Spring Harbor Laboratory that week ^[23]. The accompanying preprint, "Language Modeling Materializes a World Model of Protein Biology," was posted to bioRxiv on 4 June 2026, with Salvatore Candido as first author and Tom Sercu and Alexander Rives among the authors ^[25].

Biohub reported that ESMFold2 exceeded established methods for biomolecular complex prediction, including antibody and antigen interactions, and that a simple search procedure produced miniproteins and single chain antibodies with nanomolar binding affinities ^[25]. GEN reported that binders were designed against five targets in cancer and immunology (EGFR, PDGFR beta, PD-L1, CTLA-4, and CD45), with hit rates of 36 to 88 percent for mini-binders and 15 to 29 percent for antibody derived formats, and that ESMFold2 performed favorably against Chai-1, Boltz-1, and AlphaFold 3 in benchmarks ^[23]. The models were released under the MIT license for commercial and non commercial use ^[14][23].

Nature's news team wrote that the atlas eclipses the AlphaFold Database of predicted protein structures by more than 800 million entries ^[24]. Outside scientists quoted by Nature were broadly positive but cautious: computational biologist Gemma Atkinson of Lund University called the atlas "an extraordinary resource for biology," Martin Steinegger of Seoul National University questioned how well ESMFold2 predicts structures of proteins very different from known ones, and Sergey Ovchinnikov of MIT described the atlas as a supplement to the AlphaFold database rather than a replacement, while predicting that the open release would be widely used ^[24].

In August 2026 TIME named Rives to its TIME100 AI list, citing the ESM project at Meta and the ESM Atlas and ESMFold2 releases at Biohub ^[20][29]. In October 2026 Biohub, the U.S. Department of Energy, the National Institutes of Health, and other partners announced a combined 1.8 billion dollar commitment to generate AI-ready biological data, expanding Biohub's Virtual Biology Initiative; Rives was quoted in the announcement as Biohub's head of science ^[30].

Reception and impact

The June 2024 launch drew wide press coverage, much of it focused on the size of the seed round and on esmGFP ^[5][6][7][8]. Some coverage was skeptical: TechCrunch called the company's plan to scale into a general-purpose model for biotech "wildly ambitious" ^[5]. SiliconANGLE quoted Constellation Research analyst Holger Mueller, who said protein engineering had long been held back by computing limits, and noted that experts had warned that biological language models could potentially be misused to create bioweapons ^[7]. The Science publication in January 2025 gave the ESM3 work peer reviewed status ^[1].

After the transaction the models passed from a commercial startup to a nonprofit operator. By 2026 Biohub was releasing ESM models under the MIT license and offering them on a freely accessible platform ^[14][23]. Nature reported that ESMFold2 entered an increasingly crowded field of open and proprietary protein models ^[24].

Explain Like I'm 5

Proteins are tiny machines inside every living thing, and their shape decides what they do. EvolutionaryScale built a computer program, ESM3, that learned the patterns in billions of natural proteins the way a chatbot learns the patterns in sentences. Because it learned those patterns, it can invent brand new proteins that have never existed in nature. The company asked it to design a new glowing protein, and it made one so different from anything real that the team said it would have taken nature about 500 million years to come up with on its own.

See also

• List of neolabs
• Neolab
• Biohub
• Chan Zuckerberg Initiative
• ESM3
• ESMFold
• Enveda
• Formation Bio
• AlphaFold 3
• AlphaFold
• Protein language model
• Isomorphic Labs
• Meta AI
• AI for science
• AI drug discovery
• Nat Friedman
• Daniel Gross

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Hayes, T., Rao, R., Akin, H., et al. "Simulating 500 million years of evolution with a language model." Science 387(6736): 850-858. Published online 16 January 2025. science.org/...science.ads0018
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23EvolutionaryScale. "ESM3: Simulating 500 million years of evolution with a language model." Company blog, 25 June 2024. evolutionaryscale.ai/...esm3-release
3. ^1 ^2 ^3 ^4EvolutionaryScale. "ESM Cambrian: Revealing the mysteries of proteins with unsupervised learning." Company blog, 4 December 2024. evolutionaryscale.ai/...esm-cambrian
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22EvolutionaryScale press release. "EvolutionaryScale Launches with ESM3: A Milestone AI Model for Biology." Business Wire and Amazon press center, 25 June 2024. press.aboutamazon.com/...tone-ai-model-for-biology
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25TechCrunch. "EvolutionaryScale, backed by Amazon and Nvidia, raises 142 million USD for protein generating AI." 25 June 2024. techcrunch.com/...s-142m-for-protein-generating-ai
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16Fortune. "Meta axed his AI biology research team. Now this Meta AI mafia member's new startup has 142 million dollars to chase his protein generating dream." 25 June 2024. fortune.com/...-llm-biology-seed-round-142-million
7. ^1 ^2 ^3 ^4SiliconANGLE. "EvolutionaryScale raises 142 million USD in seed funding to accelerate AI powered protein discovery." 25 June 2024. siliconangle.com/...e-ai-powered-protein-discovery
8. ^Pharmaphorum. "EvolutionaryScale emerges with 142 million USD for AI drug discovery." 26 June 2024. pharmaphorum.com/...emerges-142m-ai-drug-discovery
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9NVIDIA blog. "EvolutionaryScale Debuts With ESM3 Generative AI Model for Protein Design." 25 June 2024. blogs.nvidia.com/...generative-ai-nim-bionemo-h100
10. ^1 ^2 ^3 ^4Wood, M. "Revolutionizing Generative Biology with AWS and EvolutionaryScale." AWS for Industries blog, 25 June 2024. aws.amazon.com/...y-with-aws-and-evolutionaryscale
11. ^1 ^2 ^3 ^4Gunderson Dettmer. "Gunderson Client EvolutionaryScale Team Joins Biohub in Strategic Transaction with Chan Zuckerberg Initiative." Client news, 6 November 2025. gunder.com/...tion-with-chan-zuckerberg-initiative
12. ^Axios. "Zuckerberg and Chan bet AI can cure all disease." 6 November 2025. axios.com/...zuckerberg-chan-biohub-ai-disease
13. ^1 ^2Endpoints News. "Zuckerberg backed Biohub hires EvolutionaryScale team in apparent end of AI startup." 6 November 2025. endpoints.news/...am-in-apparent-end-of-ai-startup
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16Biohub. ESM GitHub repository and README (formerly github.com/evolutionaryscale/esm). Accessed 27 October 2026. github.com/...esm
15. ^IntuitionLabs. "Comparing AI Biology Foundation Models: AlphaFold 3 and ESM3." 19 April 2026. intuitionlabs.ai/...y-foundation-models-comparison
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Hayes, T., Rao, R., Akin, H., et al. "Simulating 500 million years of evolution with a language model." bioRxiv preprint 2024.07.01.600583, posted 2 July 2024. biorxiv.org/...2024.07.01.600583v1
17. ^1 ^2 ^3Rives, A., Meier, J., Sercu, T., et al. "Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences." Proceedings of the National Academy of Sciences 118(15): e2016239118, 13 April 2021. pnas.org/...pnas.2016239118
18. ^1 ^2 ^3 ^4Lin, Z., Akin, H., Rao, R., et al. "Evolutionary-scale prediction of atomic-level protein structure with a language model." Science 379(6637): 1123-1130, 17 March 2023. science.org/...science.ade2574
19. ^1 ^2 ^3 ^4 ^5 ^6Biohub. "Biohub launches AI and biology initiative to cure disease." Press release, 6 November 2025. biohub.org/...ai-biology-cure-disease
20. ^1 ^2 ^3 ^4Biohub. "Alex Rives." Team profile. Accessed 27 October 2026. biohub.org/...alex-rives
21. ^1 ^2 ^3Biohub. "Tom Sercu." Team profile. Accessed 27 October 2026. biohub.org/...tom-sercu
22. ^1 ^2 ^3Biohub. "Salvatore Candido." Team profile. Accessed 27 October 2026. biohub.org/...salvatore-candido
23. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12GEN (Genetic Engineering and Biotechnology News). "Biohub Releases Protein Biology World Model to Address Disease." 27 May 2026. genengnews.com/...y-world-model-to-address-disease
24. ^1 ^2 ^3 ^4 ^5Naddaf, M. "Move over, AlphaFold: open-source model predicts shape of 1 billion proteins." Nature, 27 May 2026 (reproduced by Scientific American, 30 May 2026). nature.com/...d41586-026-01686-3 ; scientificamerican.com/...ds-on-alphafolds-efforts
25. ^1 ^2 ^3 ^4Candido, S., Hayes, T., Derry, A., et al. "Language Modeling Materializes a World Model of Protein Biology." bioRxiv preprint, posted 4 June 2026. biorxiv.org/...2026.06.03.729735v1
26. ^1 ^2 ^3 ^4 ^5 ^6EvolutionaryScale. Homepage, including the notice "EvolutionaryScale is now part of Biohub." Accessed 27 October 2026. evolutionaryscale.ai
27. ^1 ^2EvolutionaryScale. Homepage, archived 28 May 2025 by the Internet Archive. web.archive.org/...www.evolutionaryscale.ai
28. ^1 ^2 ^3 ^4 ^5EvolutionaryScale. "ESM Cambrian: Revealing the mysteries of proteins with unsupervised learning." Company blog as published December 2024, archived 12 December 2024 by the Internet Archive. web.archive.org/...esm-cambrian
29. ^TIME. "Alex Rives: The 100 Most Influential People in AI 2026." 27 August 2026. time.com/...alex-rives
30. ^Biohub. "International, cross-sector collaboration commits nearly $2 billion to build foundational data for AI models to predict and treat disease." Press release, 7 October 2026. biohub.org/...virtual-biology-initiative-expansion
31. ^1 ^2TechCrunch. "Alphabet's AI drug discovery platform Isomorphic Labs raises $600M from Thrive." 31 March 2025. techcrunch.com/...hic-labs-raises-600m-from-thrive
32. ^TechCrunch. "Cradle builds out its protein-design AI platform (and wet lab) with $73M in new funding." 26 November 2024. techcrunch.com/...-wet-lab-with-73m-in-new-funding
33. ^TechCrunch. "Profluent, spurred by Salesforce research and backed by Jeff Dean, uses AI to discover medicines." 25 March 2024. techcrunch.com/...an-uses-ai-to-discover-medicines
34. ^TechCrunch. "Generate Biomedicines raises $370M Series B with a focus on protein-based drugs." 18 November 2021. techcrunch.com/...h-a-focus-on-protein-based-drugs
35. ^1 ^2 ^3Hugging Face. "biohub" organization model listing (includes esm3-sm-open-v1, esmc-300m-2024-12, esmc-600m-2024-12, ESMC-6B, and ESMFold2). Accessed 27 October 2026. huggingface.co/biohub
36. ^1 ^2Associated Press. "Zuckerberg, Chan shift bulk of philanthropy to science, focusing on AI and biology to curb disease." 6 November 2025 (syndicated by WSLS). wsls.com/...sing-on-ai-and-biology-to-curb-disease