PostEra

The model is the company. PostEra was founded on Alpha Lee's machine learning research at Cambridge, and every commercial proposition it makes rests on Proton's output rather than on chemists using Proton as an aid. Partners do not buy access to a tool and staff it themselves; they buy programs that Proton designs. Remove the models and there is no pipeline, no partnership and no product. This is the clearest case for the grade in the lane, and unlike several neighbours the vendor names the specific methods rather than resting on the word platform.

Human oversight is evident from the staffing rather than from any published policy. The team page lists a head of medicinal chemistry and several principal investigator medicinal chemists alongside the machine learning staff, which tells a reader that people decide what actually gets synthesised and tested. The active learning loop is described as selecting informative experiments, not as running them unattended.

What is missing is any statement of where the model's output is binding and where it is advisory, and how a chemist overrides a design the system ranks highly. The oversight is real but undocumented, which is the difference between this grade and a higher one.

The company states that its artificial intelligence is not a black box, and it is one of the few in this index that has earned the right to say so. The named methods are documented in the peer reviewed literature rather than asserted in marketing: molecular transformer reaction prediction, retrosynthetic route design, Bayesian uncertainty estimation and active learning.

The synthetic accessibility score is described publicly in enough detail to be reimplemented, Manifold exposes a versioned API that independent researchers have used and cited in their own papers, and a public help centre documents how the tools behave. What is withheld is the current picture: no architecture description, training data composition or model version history for Proton as it exists today. The publications establish the lineage, not the present system.

The methods are self originated and documented, which is more than most of this lane can show, and one external dependency is named openly: Manifold's retrosynthesis draws purchasable building blocks from the Mcule catalogue, with other catalogues supported. What is absent is any account of the current system's dependencies. No foundation model provenance, no compute or hosting provider, no third party licence, no open source component inventory is published for Proton as it runs today. The chemistry foundation models are described as the company's own, and on the published record there is no way to verify what sits underneath them.

The strongest evidence base graded in this lane, and it is checkable without taking the vendor's word for anything. Nine peer reviewed papers are published, including two written with Pfizer on live programs: a first in class target result in Science Advances and a kinase selectivity result in ACS Medicinal Chemistry Letters.

The COVID Moonshot effort PostEra helped launch in March 2020 was published in Science and produced a lead compound with demonstrated in vivo efficacy in preclinical infection models. Method papers in PNAS, ACS Central Science, Chemical Science and Chemical Communications cover bioactivity prediction on small noisy data, reaction prediction against human performance, uncertainty aware models and retrosynthetic route design.

The qualification a reader should carry: all of this is preclinical. No PostEra molecule has an approval, and the widely repeated figure of a first stage gate reached 40 percent faster than predicted comes from a company update rather than from a paper, so it belongs in the vendor claim column.

There is no protected health information at stake, so the patient facing half of this axis does not apply. The stewardship question that does apply is about partner data, and it is unanswered. The company states that Proton combines its proprietary data with chemistry foundation models, and it runs simultaneous multi year collaborations with competing pharmaceutical companies.

Nothing published states what a partner's chemistry data is permitted to train, how it is segregated between engagements, or what happens to it when a collaboration ends. Chemical structures and structure activity relationships are among the most closely held assets a drug company owns, which makes the silence more consequential here than the absence of a generic safety statement would be.

Graded neutrally because the obligation does not arise. Proton operates on chemical structures, synthetic routes and assay measurements. No protected health information enters the platform, no covered entity relationship exists, and no business associate agreement is required or published. This is the chemistry convention applied to a vendor it fits cleanly, unlike neighbours in this lane that work from patient derived samples and where the same axes apply in substance.

Two dedicated passes on 29 August 2026 across the company website navigation, the footer legal surface, the help centre and general search returned no certification, no trust centre, no security page and no attestation of any kind. The only legal document the site offers sits behind a footer link labelled Terms and Conditions, which serves a website privacy policy last updated 16 August 2022, assembled from a public template and covering website visitors rather than the platform.

Its security section is the standard acknowledgement that no method of internet transmission is fully secure. This is the weakest published security posture graded in this lane, and it sits under a company whose partners are two of the largest pharmaceutical companies in the world, which is the mismatch worth naming.

Nothing to clear and nothing adverse on the record. Preclinical discovery software is not a medical device, no PostEra product is FDA regulated, and no molecule from the pipeline or the partnerships has reached a stage that would create a regulatory standing to report. Two dedicated passes surfaced no warning letter, consent decree or enforcement action against the company as of 29 August 2026. The grade records an absence of applicable obligation rather than an absence of disclosure, which is a distinction this axis currently cannot express.

No governance framework, model card, bias statement or responsible use policy is published on any company surface as of 29 August 2026. The relevant failure mode here is not demographic bias but training distribution bias, where a model trained on known chemistry recommends only what resembles known chemistry, and on that specific question the company has published more than most.

Its Chemical Science work on Bayesian deep learning addresses models that report their own uncertainty, and its PNAS work addresses valid prediction from small and noisy datasets. That is method disclosure rather than governance, and it substitutes for governance only partially.

No terms of service exist on any public surface as of 29 August 2026. The footer link labelled Terms and Conditions resolves to a privacy policy, so the document a reader would go to for warranty, limitation of liability, indemnity or service level commitments does not exist. Manifold is licensed software with an open API and a documented help centre, and it is offered with no published terms governing its use at all. Whatever recourse a partner holds lives in a private collaboration agreement and is unavailable to anyone evaluating the vendor from outside.

No electronic health record surface exists and none would be appropriate; this product never touches a clinical system. Graded neutrally with the integration that does exist recorded in its place. Manifold is integrated into Optibrium's StarDrop, announced March 2023, giving StarDrop users retrosynthesis inside their existing design environment. The Manifold API accepts building block catalogues, with Mcule named as a supported source and customisation available. That is a real and checkable chemistry stack integration, which is the analogous capability for this lane.

Manifold is delivered as a hosted application with a public API, which is established from the API documentation and the StarDrop integration rather than from any deployment page. Beyond that nothing is published: no residency statement, no tenancy model, no on premises or private cloud option, no subprocessor list. The partnership model means the arrangements governing Pfizer and Amgen data almost certainly exist in contract, and equally certainly are not public. Graded neutrally under the chemistry convention, with the caveat that this is the third axis on this record where the convention is carrying weight the published evidence does not.

Unusually forthcoming about deal structure for a private company. Partners are named, the collaborations are dated, and the figures are itemised rather than gestured at: $260 million for the AI Lab announced January 2022, expansion to a stated $610 million in January 2025, $12 million upfront on that expansion, plus milestones and tiered royalties.

Amgen has an option to access some of the technology for its own in house programs once a threshold number of programs enter the collaboration, which the company disclosed rather than buried. What holds this below the top grade is everything outside the partnerships.

Funding totals reported by third party databases disagree by a factor of nearly four, ranging from roughly $24 million to $92 million, and the company publishes no funding page, investor list or pricing of its own to settle it.

Deliberately narrow and honest about it. The company works the chemistry stage of discovery only, and states plainly that biology and clinical development sit outside its current focus. Within chemistry the coverage is small molecule design, retrosynthesis, library enumeration and property optimisation, extended in 2025 to antibody drug conjugate payload optimisation. Therapeutic breadth spans oncology, antivirals and, in the internal pipeline, reproductive medicine. A buyer looking for target identification, translational biology or anything downstream of preclinical will not find it here, and the vendor does not pretend otherwise.