Aqemia

The engine is the entire company. Aqemia has no laboratory of its own in the collaborations that generate its revenue, no clinical operation and no product other than designed molecules. Partners bring a target and Aqemia returns candidates, which means the model output is the deliverable rather than an input to one.

The claim goes further than most in this lane: the physics layer generates the data the generative model learns from, so the system is asserted to work on targets with no prior chemical data at all. Remove the algorithms and there is nothing left to sell.

The commercial boundary is stated more plainly than in most of this lane. The July 2026 Sanofi announcement says Aqemia designs molecules while Sanofi leads wet lab research, development and commercialisation, so nothing the model produces reaches an experiment without the partner deciding to run it. Servier's early pilot was structured the same way, as a blind test the partner scored. What is missing is the layer inside that boundary.

Nothing published describes how Aqemia's own chemists review, rank or reject what the generative model proposes, or whether any design advances without a person approving it. The oversight is evident at the contract level and undocumented at the bench level.

The method is named clearly and consistently, which puts Aqemia ahead of vendors that hide behind the word platform. Quantum and statistical mechanics algorithms compute binding free energy, that computation generates data, and the data trains a generative model that proposes molecules; the platform is called Qemi and earlier material refers to the same engine as the Launchpad. A public Science page describes the approach in general terms. But naming a method is not disclosing one.

No architecture, no model class, no training procedure, no compute scale and no version history is published, and the load bearing claim, that computed data substitutes for target specific experimental data, has never been demonstrated in a paper or an open benchmark. A reader can describe what Aqemia says it does and cannot check any part of it.

The company consistently describes both halves of the system as its own, the physics algorithms originating in a decade of academic research and the generative model built on top of them, with no external foundation model named or implied. Read straight, that is a claim of a very short supply chain, and it is the most that is offered. Nothing is published about compute providers, hosting, open source components or third party licences.

The only dependencies visible anywhere come from commercial technology profiling services rather than from the vendor, which list the GROMACS molecular dynamics package alongside Amazon and Google cloud infrastructure. Those are plausible for this kind of work and they are not the vendor's disclosure, so they are recorded here as third party observation rather than as anything Aqemia has confirmed.

The validation here is commercial rather than published, and one piece of it is better than anything else of its kind in this lane. Servier ran a blind test of the technology in early 2021 before entering a collaboration, which is an independent counterparty checking the claim under conditions the vendor did not control.

Servier then renewed in 2023 against a target it called undruggable, and in June 2024 both parties reported that a series of designed molecules had shown activity in vitro and in vivo. Sanofi has expanded twice, most recently in July 2026. Three candidates were reported in preclinical development, across more than a dozen internal programs.

What caps the grade: the blind test results were never published, no benchmark or accuracy figure exists on the public record, the targets and molecules are undisclosed, and two dedicated passes on 29 August 2026 across search, the company Science page and the news archive located no peer reviewed publication of any kind. A sophisticated buyer repeatedly paying is real evidence; it is not the same as a measurement anyone else can check.

No protected health information is at stake, so the patient facing half of the axis does not apply. The partner data question does apply, and Aqemia is in an unusual position on it. It runs simultaneous programs for Sanofi, Servier and Janssen, who compete, and it publishes nothing about how a partner's target information is segregated or what it is permitted to inform.

What partially answers the question is architectural rather than contractual: the company states its generative model learns from data its own physics engine computes rather than from partner assay measurements, which if accurate would materially limit what one partner's chemistry could teach a model serving another. That is a meaningful structural answer to a question most of this lane cannot answer at all, and it rests entirely on the vendor's description of its own system.

Graded neutrally because the obligation does not arise on two independent grounds. The platform operates on molecular structures, target information and computed binding energies, so no protected health information is present, and the company is French with its data processing governed by European regulation and the CNIL rather than by the United States framework this axis was written for. No business associate agreement is required and none is published.

Two dedicated passes on 29 August 2026 returned nothing. A targeted search on the company domain against the common certification terms surfaced only generic compliance content belonging to other organisations, and a direct review of every navigation section, the footer and the single legal document found no trust centre, no security page, no certification, no attestation and no penetration testing or incident disclosure statement. There is no security content on this website at all.

The company holds target information and molecular designs for at least three large pharmaceutical partners, which are among the most closely guarded assets those companies own, and it publishes nothing about how any of it is protected.

No product falls under a regulatory clearance pathway. Aqemia designs preclinical small molecules, which are not medical devices and require no marketing authorisation at this stage, and as a French company its primary regulator would be European rather than the FDA in any case. Two dedicated passes on 29 August 2026 found no warning letter, enforcement action or adverse regulatory standing. The grade records an absent obligation rather than an absent disclosure, a distinction this axis cannot currently express.

Nothing exists on any published surface as of 29 August 2026. The site carries five sections, covering the company, the science, the pipeline, careers and news, and none of them addresses model governance, evaluation practice, failure modes or responsible use. There is no model card, no validation methodology, no statement of how the company knows when a design is unreliable.

The relevant risk for a physics driven generator is that computed binding energies diverge from measured ones on target classes the method handles poorly, and the company has published neither the conditions under which its approach degrades nor any evidence that it has characterised them. Unlike neighbours in this lane that substitute peer reviewed method papers for a governance framework, Aqemia has published no papers either, so nothing stands in its place.

The page reached through the footer link labelled Terms of Use is a legal notice, not a contract. It asserts copyright, forbids linking to the site without prior written consent, and sets out how newsletter and contact data are handled. There are no terms of service, no warranty position, no limitation of liability, no indemnity and no service level commitment on any public surface as of 29 August 2026.

Less is at stake here than at vendors that license software to the public, since Aqemia sells nothing a stranger can buy, but the consequence for anyone evaluating from outside is the same: whatever recourse exists lives entirely inside private collaboration agreements and is invisible.

No electronic health record surface exists and none would be appropriate, so this is graded neutrally under the convention for discovery vendors. Worth recording in its place: unlike some neighbours in this lane, Aqemia publishes no integration into third party chemistry software either. There is no listed connector, no application programming interface and no partnership with a design suite vendor.

The platform runs inside the company and partners reach it through a collaboration agreement and joint teams rather than through software, which is a coherent model but leaves nothing for a buyer to integrate.

Aqemia publishes an actual residency statement, which is rare in this lane, and it comes with two problems. The legal notice states that personal data is processed within the European Union and not transferred to any entity outside it, places the site under French jurisdiction and the CNIL, and commits to holding non European subcontractors to clauses comparable to European Commission standard contractual clauses or to binding corporate rules.

The first problem is scope: this covers newsletter and contact form data from website visitors, not the target information and molecular designs a pharmaceutical partner entrusts to the platform, for which no residency, tenancy or hosting statement exists. The second is staleness. The document offers Privacy Shield as an example of an acceptable safeguard.

That framework was invalidated by the Court of Justice of the European Union in July 2020 and replaced in July 2023, so the commitment is being anchored to a mechanism that has not been valid for six years. The body of the notice also carries a 2025 copyright line under a 2026 footer, which suggests it is not being maintained.

Partners are named and the collaborations are dated: Servier from a 2021 pilot, Sanofi from 2020 antiviral work through the December 2023 agreement to the July 2026 expansion, and Janssen named in the chief executive's own account. Funding is disclosed with named investors, a €30 million Series A in 2022 led by Eurazeo with Bpifrance and Elaia, a €30 million extension led by Wendel Growth, and a $7.4 million France 2030 award in March 2025 with its purpose stated.

Three things hold this below the top grade. Individual payments are never broken out from the $140 million ceiling, including the original upfront. The same headline appears as $140 million in the English announcements and as €140 million when the chief executive describes it in a French interview, so the currency of the agreement is genuinely unsettled. And aggregator funding totals above $100 million do not reconcile with the rounds the company has itself disclosed.

Narrow by design and expanding along one axis only. The work is small molecule design at the discovery stage, with the partner or a spin out carrying anything downstream. Therapeutic breadth covers oncology and immuno oncology, extended into immunology, inflammation and central nervous system disease, and since the March 2025 France 2030 award into RNA targets and epitranscriptomics. More than a dozen internal programs are claimed. Nothing here reaches biologics, translational biology or clinical development, and the company does not suggest otherwise.