Atomwise

AtomNet is the product and the moat. It is a patented deep convolutional neural network that predicts small molecule bioactivity in three dimensions from structure activity data and protein structures, and the entire commercial proposition is that it substitutes for physical high throughput screening rather than assisting it. There is no instrument franchise, no assay business and no data brokerage underneath. The company's own published study frames the claim in exactly these terms, concluding that computational methods can substantially replace high throughput screening as the first step of small molecule discovery.

The oversight mechanism is empirical and unusually well evidenced. In the AIMS programme, model selected compounds were physically synthesized and assayed by more than 250 independent academic laboratories, so predictions were adjudicated by third party wet lab results rather than by the vendor. That is a stronger check than most peers can point to. Held at B because compound selection criteria, ranking thresholds and the rules for when a target is judged unsuitable for the method were not located in published form, so the decision layer between model output and experimental commitment remains undocumented.

The method is in the peer reviewed literature rather than in marketing copy. The core screening result was published in Scientific Reports in April 2024, and methodological work including AtomNet PoseRanker appeared in the Journal of Chemical Information and Modeling under an open access licence, with the architectural approach stated specifically as a three dimensional convolutional neural network operating on protein ligand structure. Author affiliations and competing financial interests are declared in the papers, with employees identified as such. A reviewer can read the method, examine the benchmarks and disagree in print.

The strongest platform level evidence in this category, and the reason is that it reports its own denominator. The AIMS study applied AtomNet prospectively to 318 targets through more than 250 academic laboratories across 30 countries and found structurally novel hits for 235 of them, which means 83 targets did not yield and the paper says so. It further stacked the deck against itself deliberately, reporting success on targets with no known binders, without high quality crystal structures and without manual cherry picking of compounds, and showing that selected molecules were novel scaffolds rather than modifications of known bioactives. Publishing the failure count alongside the success count is a candour benchmark this index has found rarely. Buyers should be equally clear about what this evidence is not: it is preclinical hit finding validation generated largely by third parties, not clinical validation, and no molecule from the platform with human efficacy data was located.

Not applicable in the provider sense and rated accordingly rather than penalized. The platform operates on protein structures and compound libraries with no patient data in the workflow. The confidentiality obligation that does exist covers partners' target and compound information and sits inside collaboration agreements.

Not applicable. Counterparties are pharmaceutical companies and academic laboratories engaged in preclinical discovery, not covered entities transferring protected health information.

No SOC 2, ISO 27001 or equivalent attestation was located and no trust center was found. Public materials describe the compute architecture, including NVIDIA GPU infrastructure on AWS with the WEKA data platform, but infrastructure description is not an attestation and should not be read as one during diligence.

The screening platform is not a regulated device and is not presented as one, which is correct. At asset level the disclosed lead is a novel allosteric TYK2 inhibitor discovered using AtomNet, with an IND filing described as planned as of 2024 reporting. No confirmed IND acceptance, registered trial or clinical data for a platform derived asset was located in this review, so regulatory standing is materially behind the peers in this category that have dosed patients.

The best governance disclosure in this category, though the company does not frame it as governance. The domain relevant bias question is whether a model only works where data already exists, and the AIMS study attacks that question directly by reporting performance on targets with no known binders and without high quality crystal structures, including protein protein interaction and allosteric sites that are historically underserved by computational screening. That is a published characterization of where the method does and does not hold. Held at B rather than A because there is no formal governance framework, no model card and no statement of intended and unsuitable use, so the disclosure is a byproduct of good science rather than a governance commitment.

Not applicable. This is a preclinical virtual screening platform with no provider workflow surface and no EHR touchpoint.

The commercial model is collaboration and screening engagement rather than deployed software, so tenancy and residency terms do not arise in the usual form and none were located. Partners submit targets and receive prioritized compounds, which means the practical data question is what happens to a partner's target information inside the vendor environment, and that is addressed by contract rather than by architecture.

No rate card exists and none would be meaningful given the engagement model, but headline deal structures are public and specific, with reported collaborations including Sanofi at approximately 1.2 billion dollars, Hansoh Pharma at up to approximately 1.5 billion dollars and Eli Lilly at up to approximately 560 million dollars in potential value. Per programme economics and the split between upfront, research funding and success milestones are not disclosed. Reporting also indicates the organization has become leaner while focusing on advancing AtomNet derived candidates, which is a material commercial fact for anyone assessing counterparty durability.

The broadest demonstrated coverage in this category, and demonstrated is the operative word. The AIMS study reports finding novel hits across every major therapeutic area and protein class, tested across 318 targets nominated by more than 250 independent laboratories in 30 countries rather than selected by the vendor, and explicitly includes historically difficult classes such as protein protein interactions and allosteric sites. Modality is confined to small molecules, which is the real limit here rather than therapeutic area.