PostEra

The company describes itself as an artificial intelligence first biotechnology company and the description holds up against the band. Proton is the platform and it is the method: models design molecules against several competing property constraints at once, models find the synthetic route that makes each one and group molecules so they can be made in parallel from shared intermediates, and active learning models choose which experiments to run next. The make stage is the point, because chemical synthesis is the slowest part of medicinal chemistry and PostEra's distinctive claim is that its models solve for how a molecule gets made rather than only for what it should be.

Remove the models and nothing is left to sell. The company runs no software product since Manifold was withdrawn, employs a small team, and delivers candidates rather than services. The two fertility programs bought by EMD Serono in September 2026 were, on the company's own account, designed with Proton against targets that had resisted small molecule chemistry for decades.

The oversight structure is described with mechanisms at both the model level and the program level. At the model level, the platform is built as a closed design, make and test cycle in which every designed molecule is synthesized and assayed, active learning chooses the next experiments, and the models are trained to express what they do not know, which is the subject of a peer reviewed paper rather than a slogan. At the program level the published co-discovery process puts the partner at two of the three decision points: the partner selects the biological targets and defines the assay cascade and target product profile, PostEra drives discovery to a development candidate nomination, and both then agree who sponsors the asset forward.

One part is missing. Nothing published states a confidence threshold at which a prediction is not acted on, what happens when an assay contradicts a design, or which judgments inside the cycle require a chemist rather than allowing one. The company's own framing, that its benchmark is high quality chemical matter rather than topping a leaderboard, describes a philosophy rather than a control.

Six peer reviewed papers put the methods where they can be checked, and the approach itself is named and described stage by stage. Proton is the platform, structured as a closed design, make and test cycle: design combines proprietary data with chemistry foundation models to satisfy several competing property constraints at once, make ensures every designed molecule carries a reliable synthetic route and that molecules can be produced in parallel from shared intermediates, and test uses active learning to pick the most informative next experiments. The published papers give the underlying methods for reaction prediction, synthetic route design, model calibration and prediction from small and noisy data, with identifiers including doi 10.1126/science.abo7201, doi 10.1126/sciadv.ado4288 and doi 10.1021/acs.jmedchem.6c00923.

Two elements are missing. The chemistry foundation models are referred to as a category and never identified, leaving open what produces a design and whether any of it is licensed, and the proprietary data behind them is not characterized by size, source or coverage. No version identifier, release history or update practice is published for Proton or any component, so nothing ties the system running a 2026 program to the methods described in a 2023 paper.

A clear architecture with no model provider named. Proton's three stages are set out clearly, and the design stage is said to combine proprietary data with chemistry foundation models, but those models are never identified. Nothing says whether they are built in house, adapted from published work or licensed from a supplier.

No hosting or compute provider is named either, and there is no subprocessor list, no statement of what governs partner chemistry once it reaches the platform, and no disclosure of the contract research organizations or synthesis partners behind the make stage of a cycle whose whole premise is that molecules get made. The peer reviewed papers name methods rather than suppliers. For a company running discovery on two of the largest pharmaceutical companies' targets, the absence of any party list is the most conspicuous gap.

The published method record is strong, and the clinical record does not exist yet.

Six peer reviewed papers are listed with their identifiers, spanning the Journal of Medicinal Chemistry (doi 10.1021/acs.jmedchem.6c00923), ACS Medicinal Chemistry Letters (doi 10.1021/acsmedchemlett.5c00687), Science Advances (doi 10.1126/sciadv.ado4288), Science (doi 10.1126/science.abo7201, the open science COVID Moonshot project), ACS Central Science and PNAS. Their subjects are the platform itself: whether a model can outperform humans at predicting chemical reactions, how to train models that know what they do not know, how to design synthetic routes, how to make valid predictions from small and noisy data, and two accounts of real Pfizer partnered programs, one on kinase selectivity and one on a first in class target.

The commercial record is the second line and it is unusually concrete. On 9 September 2026 EMD Serono, the United States and Canada healthcare business of Merck KGaA, acquired two of PostEra's programs outright, a follicle stimulating hormone receptor agonist series for ovarian stimulation and a luteinizing hormone receptor agonist for male infertility and ovulation trigger. A large pharmaceutical company buying preclinical assets is a harder signal than a case study, and more than $1 billion in announced partnership ceilings across four multi year agreements with Pfizer and Amgen points the same way.

What is absent is a molecule in a person. Every program on the published pipeline is preclinical, including the two that were sold, whose safety and efficacy in humans the acquisition notice states have not been established. No trial, no clearance and no clinical timeline appears anywhere, so the platform's central claim, that it produces better candidates faster, has not yet met the test that decides it.

How information moves through this system is not described anywhere. The company's only privacy document sits inside the terms and conditions page, was last updated on 16 August 2022, and scopes itself to the Website, so it reaches visitors rather than anything the platform processes.

No patient data enters the platform: the inputs are chemical structures, synthetic routes and assay results, and there is no clinical program, no trial and no record feed. The question that does arise here is partner chemistry. PostEra runs discovery against Pfizer's and Amgen's targets and states that Proton combines proprietary data with chemistry foundation models, and nothing published says whether partner structures, assay data or results are retained, whether they train models that serve other programs, or how one collaboration's data is kept from another's. No retention window, training use position or de-identification practice is stated for any of it.

Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not published either.

No business associate status is stated anywhere, and the only privacy document reaches the website alone. It is carried inside the terms and conditions page, was last updated on 16 August 2022, and defines the Service it governs as the Website. Protected health information is not mentioned, the United States health privacy rule is not mentioned, and no statement says whether PostEra considers itself in or out of scope.

What is missing is a plain statement that the product does not process protected health information, an explanation of the scope, and a service level privacy document behind it. On the facts that statement would be easy to make, since this is a preclinical medicinal chemistry company with no clinical program and no patient data, but it has not been made. A separate company trading as Manifold at manifold.ai publishes a trust center carrying health privacy and other compliance claims; none of those attach to PostEra, and the shared name makes them easy to misattribute.

Nothing is published. There is no SOC 2 report, ISO/IEC 27001 certificate or HITRUST certification named anywhere, no trust center, no security page and no independent penetration testing on a stated cadence. The only document that touches the subject is a privacy notice inside the terms and conditions page, last updated on 16 August 2022, which covers the website.

One confusion is especially likely. A separate and unrelated company trading as Manifold at manifold.ai publishes a trust center carrying SOC 2 Type II, NIST 800-171, health privacy, state cloud program and European data protection compliance. PostEra's own retrosynthesis tool was also called Manifold, and is now withdrawn. None of those certifications attach to PostEra, and the name collision makes them easy to misassign. No public attestation covers the environment where a partner's chemistry and assay data sit.

As a preclinical medicinal chemistry company, PostEra makes no device claim, and the platform is scoped accordingly. Proton is not sold as a product, is not presented as a medical device, and carries no clearance. That is the appropriate position for what it is.

There is no regulatory milestone to record on either side of the pipeline. Every program in the published table sits at a discovery or preclinical stage, no investigational new drug clearance or trial start is named, and the September 2026 acquisition notice for the two fertility programs states directly that both are in preclinical development and that their safety and efficacy in humans have not been established. The published co-discovery model places the regulatory step outside PostEra by design: the partner and PostEra agree who sponsors an asset into investigational new drug enabling studies, and to date those positions belong to the partners.

Model behavior is evaluated in peer reviewed venues rather than asserted, and the papers ask the questions that matter most for governance. The published set includes work on whether a model can outperform humans at predicting chemical reactions, how to train models that know what they do not know, and how to make valid predictions from small and noisy data. Calibration and the boundary of a model's competence are the subject rather than a footnote, and the company states the position plainly on its own site: its AI is not a black box. The COVID Moonshot project, published in Science (doi 10.1126/science.abo7201), was run as open science with the data released, so outside groups could and did check the work.

What is not published is the breakdown. No results are given by target class, reaction class or chemical series, so where the models are weaker stays out of view, and no model card, no independent audit of model behavior and no governance framework document appears anywhere. The proprietary data that trains the design models is described as proprietary and is not characterized by size, source or coverage, which leaves the composition question that sits underneath any bias claim unanswered.

Mechanisms exist that let a prediction be challenged, and they are described rather than implied. PostEra states that its AI is not a black box, and publishes peer reviewed work on training models that know what they do not know, so uncertainty is treated as an output rather than hidden. Every designed molecule carries a synthetic route, which makes the chemistry behind a proposal inspectable. The co-discovery structure is published in three steps and puts a check on each one: the partner selects the targets and defines the assay cascade and target product profile, PostEra drives discovery to a development candidate, and the two then agree who sponsors the asset into investigational new drug enabling studies. Every molecule is made and assayed before anything advances.

Nothing stands behind the output. No accuracy figure is published for the design, retrosynthesis or property prediction models with a method and a denominator, so there is no error rate to hold the company to. No guarantee, indemnity or remediation obligation is published for a partner whose program follows a wrong prediction, and the risk is allocated inside collaboration agreements whose terms are not public.

No clinical record integration exists. There is no electronic health record touchpoint, no clinical workflow surface and no provider connection, because the platform serves PostEra's own medicinal chemistry programs and its partners' discovery programs rather than health systems.

No research system of record is named either, and this is a change. The company previously offered Manifold, a retrosynthesis tool licensed as software with a public application programming interface and an integration into Optibrium's StarDrop design environment. Manifold has been withdrawn: the site navigation and main pages no longer mention it, its own subdomain does not resolve, and its product address returns a not found page. PostEra now presents itself solely as an artificial intelligence first biotechnology company with no software product, so there is no application programming interface, no named research system integration and no interoperability claim left to verify.

The platform runs for PostEra and its collaborators rather than being deployed to a customer, so there is no tenancy model, hosting option or isolation commitment to evaluate. Location is implied rather than stated: the company's privacy document places it in Boston, Massachusetts and names Massachusetts, United States as the governing jurisdiction, and the work is described as running inside the company.

This was not always the shape of the offering. Manifold, the retrosynthesis component, was once licensed as software with a public application programming interface, which would have carried a hosting and residency question, and it is no longer offered. What remains is entirely internal, and nothing published names a cloud or compute provider, states where partner chemistry rests, or describes how data from one collaboration is separated from another's.

Pricing appears nowhere. There is no pricing page, no rate card and no unit of charge on any company surface, and the site navigation carries no product or pricing section at all, which is consistent with a company that sells collaborations rather than seats. Before contact, what is published is the structure of a deal, set out in three steps: the partner picks the targets and defines the assay cascade and target product profile, PostEra drives discovery to a development candidate, and the two agree who sponsors the asset onward, with the partnered pipeline carried on milestones and royalties.

The dollar figures in the public record are contract values rather than prices, and they measure the size of a collaboration: $260 million for the Pfizer AI Lab established in December 2021, a stated $610 million total on the January 2025 expansion that added an antibody drug conjugate collaboration, $12 million upfront on that expansion, and more than $1 billion in announced partnership ceilings across four multi year agreements with Pfizer and Amgen. Those are eligibility ceilings inclusive of milestones and royalties, not money received. One further transaction is published without a figure: EMD Serono acquired two fertility programs outright in September 2026, with no consideration disclosed.

Coverage is named at the level of disease, target and molecule. The published pipeline table gives the owner, disease, target, molecule and stage for each program: PMOS against AKR1C3 with PERA-5720, male infertility against the luteinizing hormone receptor with PERA-8103, and female infertility against the follicle stimulating hormone receptor, all wholly owned, alongside four preclinical partnered programs across obesity and oncology carried on milestones and royalties. The stated internal focus is women's health and fertility, and the platform focus is the chemistry stage of discovery rather than biology or clinical development, which the company says directly.

Validation reaches part of it. The two fertility programs were bought outright by EMD Serono in September 2026, and two Pfizer partnered programs, one on kinase selectivity and one on a first in class target, are written up in peer reviewed papers. What no published material establishes is breadth: four partnered programs are described only as obesity and oncology with no targets named, and every program in the table is preclinical, so where the platform has been shown to work is narrower than where it is offered.