EvolutionaryScale

The models are the whole product. EvolutionaryScale, now part of Chan Zuckerberg Biohub, builds protein language models and distributes them: ESM3, which models sequence, structure and function together and generates proteins conditioned on any combination of the three; ESM Cambrian (ESMC) for representation learning; ESMFold2 for structure prediction of biomolecular complexes; and the ESM Atlas, a database of annotations for billions of proteins built with them. There is no pipeline, no laboratory franchise and no instrument line underneath: the model is what a customer reaches, through an API, a Python package or downloaded weights.

The oversight structure is published and one part is missing. What the models may be used for and what they may not is set out in binding terms: the acceptable use policy requires users to review and validate outputs, restricts use to scientific research, prohibits named categories of use with monitoring and suspension behind them, and says the services are not a substitute for professional advice in regulated fields. The model cards and the Limitations page state that outputs are computational predictions, not experimentally validated, and not intended for clinical or therapeutic use without further validation, and the ESM Atlas search is flagged as using an agent that may answer inaccurately. Nothing acts on its own: a design is a proposal until the user expresses and characterizes it.

The missing part is a threshold. Confidence measures are exposed (pLDDT, pTM) and described as imperfect, with no value at which the platform says an output should not be relied on, and no published failure rate to calibrate against.

The models are identified, the training data is characterized and the versions are published. ESM3 was published in Science in January 2025, and the company states the scale of the training run on its own site: 2.78 billion natural protein sequences, 771 billion unique tokens, 98 billion parameters in the largest model, and roughly 10^24 floating point operations of compute. Each model on the platform carries a model card with its architecture, supported modalities, training data, intended use and limitations, and a dated version identifier that also appears in the API call (esm3-open-2024-03), so a user can tell which model produced a result. The open model's weights and source are published on GitHub and Hugging Face, so outside researchers can run it and report where it fails, and the family extends to ESMC and ESMFold2 with their own pages.

One inconsistency remains in the published material: the ESM3 model card's overview gives training data as "3,000+ sequences and 700,000+ unique training tokens", which does not match the training scale the company publishes elsewhere or the paper.

A structurally short chain: the models are the company's own and one of them is obtainable. ESM3, ESMC and ESMFold2 are built in house and served from the company's own platform, with no outside model provider between a submitted sequence and an answer, and the open model's weights and source are published on GitHub and Hugging Face, so a user who does not want the chain at all can run it without one. The privacy policy names the categories of service provider that may touch platform data (email, database management and hosting, customer support, security) and states that access extends across affiliated entities.

The party list is not enumerated: no subprocessor list names the hosting provider or the individual services, and nothing states whether submitted sequences inform later model training.

Published research with a laboratory result behind it, and no therapeutic evidence. ESM3 was published in Science in January 2025, and its headline demonstration, esmGFP, is a fluorescent protein generated by the model and then expressed and characterized, far enough from natural fluorescent proteins that the company describes the gap in evolutionary terms; the design route is published as a tutorial a user can follow. The platform's scale is visible in the ESM Atlas, an open database of annotations for more than 6.8 billion proteins with 1.1 billion predicted structures.

What is absent is any medicine: no therapeutic program, no clinical candidate and no disclosed molecule designed with these models that reached development. Availability through an API and model hubs is distribution rather than measured adoption.

General privacy handling is published, and the questions AI raises are not answered. The Biohub privacy policy (last updated 2 June 2026) reaches the platform and is specific about what a user submits: queries, prompts and scientific information such as biological sequences, structures and related annotations are collected and associated with the user, generated outputs may be stored, and users are told not to include personal, sensitive, patient or other regulated data unless the service expressly asks for it. Retention is stated as a principle rather than a schedule: user submissions and usage records are kept as long as reasonably necessary to provide, secure, improve and evaluate the services, with actual periods varying.

What is not stated is whether submitted sequences are used to train or improve the models: the policy's product improvement clause covers research and evaluation of system performance in general terms and no categorical commitment excludes training on submissions, which for a customer's unpublished target sequence is the question that decides whether it can use the hosted API at all. Safety engineering is not addressed.

A scope position is stated plainly, with a service privacy document behind it. The Biohub privacy policy (last updated 2 June 2026) covers the platform and tells users not to include personal information, sensitive personal information, patient data or other regulated data in their submissions unless the service expressly requests it or Biohub has expressly permitted it. The acceptable use policy adds that users may not process or generate health, demographic, biometric or other sensitive personal data without the consent the law requires. Read together, those say the product is not built to receive protected health information and explain the scope, and the privacy policy is a substantive service document rather than a website notice.

No business associate status is stated and no agreement is published, which is consistent with the position, since there is no agreement to publish. Nothing published says what happens if patient derived sequence data or other regulated data reaches the platform, or whether Biohub will sign an agreement at all.

Nothing independent stands behind the stated controls. The privacy policy says reasonable administrative, technical and physical safeguards are in place and gives a security contact address; no SOC 2 report, ISO 27001 certificate, penetration testing cadence or trust center appears anywhere on the site, and the legal instruments that are published (terms of use, privacy policy, acceptable use policy, DMCA and ethics policies, model licenses) govern conduct rather than attest to controls.

The platform and its API share names with unrelated commercial products that do hold security certifications, and those certifications belong to those products, not to this vendor.

An intended purpose is published and the marketed capability matches it. Each model card carries an Intended Use section describing what the model is for (for ESM3, prompt driven generation of sequences and structures from motifs, partial coordinates, secondary structure constraints or function keywords) together with a Limitations and Risks section stating that it is not intended for clinical or therapeutic applications without further validation. The platform Limitations page (effective 27 May 2026) adds that outputs are computational predictions that establish no clinical, diagnostic or therapeutic conclusion without independent validation, and the acceptable use policy says the services are intended for scientific research and are not a substitute for professional advice in regulated fields. That is a positive statement of purpose and setting rather than a liability clause in terms of use, and nothing in the marketing claims a clinical use it disclaims.

No device claim is made and none appears to be required; a medicine designed with these models carries its own regulatory position, held by the organization that develops it.

A governance framework with published material a reviewer can evaluate the model against, short of results. Each model carries a model card naming its architecture, supported modalities, training data, intended use and its limitations and risks, and a separate platform Limitations page (effective 27 May 2026) states that outputs are computational predictions and do not establish functional, clinical or therapeutic conclusions without independent validation. That page also names where reliability varies: with the target, the sequence context, the training data available, and specifically for underrepresented lineages, short sequences and intrinsically disordered proteins. The acceptable use policy (27 May 2026) sets binding conditions on use, with stated monitoring and suspension for violations, and the company says its models are built and deployed under a responsible development framework, with the ESM3 paper recording outside expert input into that framework and into the decision to release the open model.

What is named is not measured: no results are published for the groups the Limitations page identifies, so the uneven performance is asserted rather than quantified, and no fairness or bias evaluation with a stated method appears.

Mechanisms let a user test an output, with nothing standing behind it. The open model's weights and source are published, so a user can run the system, probe where it fails and report it; the platform publishes confidence measures with a caution attached, the ESM3 model card noting that pLDDT and pTM are helpful but imperfect and that proposals may not be physically realizable; and every design is expressed and characterized in the user's own laboratory before it means anything.

No error rate for the models is published with a method and denominator, and no warranty, indemnity, service level or remediation commitment attaches to an output. The Limitations page states plainly that outputs are not experimentally validated, which is candor about the position rather than a commitment within it.

No clinical record integration exists: these are protein models reached through an API, and no clinical record system is involved. For a pharma or biotech buyer the record systems that matter are its research systems: electronic lab notebooks, laboratory information systems, research data platforms and cloud data environments.

What is published is a general integration surface rather than a named connection: a documented REST API with published endpoints, a developer console for keys and rate limits, an installable Python package, and model weights distributed through GitHub and Hugging Face, all of which a computational team can wire into its own stack. No electronic lab notebook, laboratory information system, research data platform or named customer deployment appears, which is integration available with no system named.

Options are stated, with residency and the processing path left open. Three routes are published: the hosted API on the Biohub Platform reached through an API key and the esm Python package, the open model's weights and source on GitHub and Hugging Face for a user to run on its own infrastructure, and the ESM Atlas and Fold tools in the browser. That range is real, and the residency answer differs by route: a user running the open weights keeps everything, while a user of the hosted models sends sequences to Biohub.

For the hosted path nothing is published about where the platform runs, in which country submitted sequences rest, whether processing crosses a border, or how one customer's submissions are isolated from another's.

A pricing basis is published without figures, and there is a route to start without a sales conversation. Access runs through the Biohub Platform: a user creates an API key in a developer console that manages keys, credits and rate limits, starts on free credits and applies for more, and installs the Python package to call the models. Credits and rate limits are the stated unit of charge. The open model goes further, with weights and source published on GitHub and Hugging Face under a stated license the company names as MIT on its own site, so that tier is free and its terms are readable before contact.

No figure is attached to a credit, no enterprise tier or rate card is published, and what a commercial user pays beyond the free allocation is not established.

Coverage is named broadly, with validation behind part of it. The models work on proteins as a class rather than on a disease area, and the company names the uses it is built for: designing enzymes that break down plastic, proteins that capture carbon, antibodies, and new medicines, with the ESM Atlas covering proteins across all of life. Validation sits behind one published demonstration, esmGFP, a fluorescent protein generated and then characterized in the laboratory and reported in Science, plus the structure prediction benchmarks the model pages describe.

The Limitations page names where coverage is thinner (underrepresented lineages, short sequences, intrinsically disordered proteins) without per group results, and no performance data is published for any particular application area. Breadth of applicability is not validated performance in a buyer's own problem.