Numerion Labs

The models are the product. The company, now operating as Numerion Labs (atomwise.com forwards to numerionlabs.ai), describes an "AI chemistry superplatform" built from three machine learning components: COSMOS, a chemistry foundation model predicting drug function from structure; APEX, an enumerator for screening very large virtual libraries; and EXPO, optimization algorithms for turning binders into candidates.

Its programs in immune and inflammatory disease come out of that platform, as its earlier work did from AtomNet, the convolutional network the company ran as Atomwise. There is no instrument or assay business underneath.

The oversight is empirical and the selection rules are not published. In the AIMS study, compounds the model chose were synthesized and tested by more than 250 outside laboratories, so predictions were judged by wet lab results rather than by the company, and the platform page describes a discovery process that turns model selected binders into candidates through further optimization.

What is not published is how compounds are ranked and cut before synthesis, what confidence attaches to a prediction, or where chemists override the model on the current platform. Ask for the selection criteria and the review step between model output and synthesis.

The current platform's components are named and the method behind them is not described, while the previous model is in the literature. The platform page names COSMOS (a chemistry foundation model predicting drug function from structure), APEX (an enumerator for screening billions of virtual compounds, described in research published with NVIDIA in October 2025) and EXPO (optimization algorithms that adapt to a project without retraining).

AtomNet, the three dimensional convolutional network the company ran as Atomwise, was described in peer reviewed and open access papers. For COSMOS no architecture, training data, version or change practice is published. Ask how COSMOS differs from AtomNet, what it was trained on, and how model changes are communicated.

The platform is the company's own and the chain around it is not described. COSMOS, APEX and EXPO are presented as the company's proprietary components, and APEX research was run with NVIDIA, but no hosting or compute provider, subprocessor list or statement of who can access partner data is published; the privacy policy mentions service providers in other countries in general terms. The sensitive material is a partner's targets and compound data. Ask for the parties that process partner data and where.

Strong prospective evidence for the previous platform, and the current headline claim is unevidenced. The AIMS study (Scientific Reports, 2024, run as Atomwise) applied AtomNet prospectively to 318 targets nominated by more than 250 academic laboratories in 30 countries, with compounds tested by those laboratories, and found structurally novel hits for 235 targets, including targets with no known binders and no high quality structures. That is peer reviewed, prospective and multi site evidence of the kind the top band describes, for AtomNet.

The company now presents a successor platform, and its headline claims are about it: COSMOS as the "industry's most advanced predictive model", screening efficiency ten billion times higher and a chemistry domain 10,000 times larger than competitors. Those are published with no method, and the October 2025 APEX research with NVIDIA addresses screening speed, not hit rates. No candidate from either platform is shown in a clinical study on the company's pages. Ask for prospective hit rates for COSMOS and the stage of the lead programs.

General assurances that do not answer the questions AI raises. The platform works on protein structures and chemical libraries, with no patient data in the workflow, and the privacy policy asks people not to send health or genetic information.

What matters to a partner is its own targets and compound data, and nothing is published on how long that is kept, whether it trains the shared COSMOS model or project specific models (the platform page describes "project-bespoke model expertise without large datasets or re-training"), or how it is kept apart from other partners' work. The privacy policy covers personal information only and describes commercially reasonable security. Safety engineering is not addressed. Ask whether partner data trains shared models and how long it is retained.

The product does not handle protected health information, and the company says it does not want it. The platform works on protein structures and chemical libraries, and its privacy policy asks people not to provide sensitive personal information, naming health, biometrics and genetic characteristics, through the service or otherwise. Its counterparties are pharmaceutical companies and research laboratories rather than covered entities, so no business associate agreement arises. A partner's confidential targets and chemistry are governed by its collaboration agreement, not by this axis.

Nothing independent is published. The site carries no security page, trust center or certification, and the privacy policy says only that the company uses reasonable safeguards and cannot guarantee security. Descriptions of compute used for research are not controls, and certifications held by a cloud or hardware provider would belong to that provider. Partners share undisclosed targets, so the gap matters. Ask for any SOC 2 or ISO/IEC 27001 attestation in the company's own name.

No device claim is made and the platform is scoped as internal drug discovery. It finds and optimizes small molecules and makes no diagnostic or treatment recommendation, so no clearance applies. Regulatory standing later belongs to the drug candidates; the company's programs page shows no program detail, and no candidate from the platform is shown in a clinical study on its site.

Results published within the group where performance is most at risk. For a screening model the at risk group is targets unlike its training data, and the company's AIMS study (Scientific Reports, 2024, run as Atomwise) reported AtomNet's results on exactly those: targets with no known binders, targets without high quality crystal structures, and protein protein interaction and allosteric sites, with compounds chosen without manual cherry picking and tested by outside laboratories.

No comparison of performance across target classes is framed as a fairness or bias evaluation, and no governance framework is published. The current platform (COSMOS, APEX, EXPO) has no equivalent evaluation on the company's pages. Safety engineering is not addressed. Ask whether COSMOS has been evaluated on novel targets the way AtomNet was.

A published success and failure rate with its denominator, for the previous version of the platform. The company's own AIMS study (Scientific Reports, 2024, published as Atomwise with competing interests declared) applied AtomNet prospectively to 318 targets nominated by more than 250 academic laboratories in 30 countries and reports structurally novel hits for 235 of them, so 83 targets yielded nothing, with compounds synthesized and assayed by the outside laboratories.

That is a measure of the model's output with method and denominator stated, as a test result rather than performance on a partner's campaign. It describes AtomNet; the company now sells a successor platform (COSMOS, APEX, EXPO) with no equivalent published figure. No warranty, remediation or fee consequence for failed predictions is published. Ask for hit rates on targets like the partner's with the current platform.

No clinical record integration exists: the platform screens and designs small molecules against protein targets and has no provider workflow or patient data. For a pharma or biotech partner the record systems that matter are its research systems: electronic lab notebooks, laboratory information systems or compound registries, research data platforms and cloud data environments.

No connector to any of them is described for COSMOS, APEX or EXPO, no integration is claimed, and nothing says how targets, assay results and designs move between a partner's systems and the company's. That is no integration evidence on this axis, which for a discovery partnership records an absence rather than a defect. A partner should ask how data moves between its systems and the company's, in which direction, and in what format.

An internal platform with location implied rather than committed. The platform runs inside the company, and partners' targets would come to it; the privacy policy says the company is headquartered in the United States with service providers in other countries, and its APEX work was run with NVIDIA. No hosting provider, region or separation between partners' data is stated for the platform. Ask where partner data is stored and processed and how it is segregated.

No price is published, and the business model is discoverable. The company develops its own programs in immune and inflammatory disease and has worked through discovery collaborations; collaborations reported under the Atomwise name (Sanofi, Hansoh Pharma, Eli Lilly) carried headline potential values, but those are what particular partners negotiated, not a price a buyer can size a deal from, and the current site names no partners or terms. The only route in is a contact page. Ask whether the platform is offered to partners today, on what structure, and who owns molecules it designs.

Coverage is named, with validation behind the previous platform. The AIMS study (2024) validated AtomNet across every major therapeutic area and protein class, on 318 targets nominated by outside laboratories, including protein protein interaction and allosteric sites.

The company now describes its programs as focused on immune and inflammatory disease and its platform as covering "universal chemistry" for small molecules; the current platform's coverage has no published validation on the company's pages. Ask which target classes COSMOS has been tested on.