Numerion Labs
Numerion Labs, formerly known as Atomwise, is a San Francisco AI drug discovery company developing small molecules for immune and inflammatory disease, led by chief executive Steve Worland. Its platform combines COSMOS, a structure based chemistry foundation model trained to predict molecular binding and function; APEX, an enumerator that in research published with NVIDIA in October 2025 retrieved the top one million compounds from a 10 billion compound library in under 30 seconds on a single GPU; and EXPO, optimization algorithms for turning novel binders into drug candidates.
As Atomwise it built AtomNet, a three dimensional convolutional network for structure based virtual screening, and published the AIMS study (Scientific Reports, 2024): across 318 prospective projects, 296 of them run with academic labs and screening centers in 30 countries, AtomNet found at least one bioactive compound in 73% of the academic projects. Collaborations under the Atomwise name included Sanofi, Hansoh Pharma and Eli Lilly.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
Discovery runs on what the company calls 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.
Models choose compounds and laboratory assays decide what advances. In the AIMS study, AtomNet ranked compounds by predicted binding probability, and the workflow clustered the top ranked molecules and algorithmically picked the highest scoring member of each cluster, with no manual cherry picking. The picks were then screened in the laboratory: by the academic collaborators in the AIMS program, and, for internal programs, at contract research organizations after synthesis at Enamine. On the current platform, EXPO optimization algorithms carry novel binders toward drug candidates, and the chief executive describes APEX as letting chemist and computer work together in real time.
How the current platform's picks are ranked and cut before synthesis, and where chemists override them, is not published.
COSMOS is a structure based, generative pretrained foundation model trained to predict molecular binding and function, and APEX uses it to prioritize compounds. APEX pairs deep learning surrogates with GPU accelerated enumeration over structured chemical spaces; its paper, coauthored by NVIDIA subject matter experts, is on arXiv with code on GitHub. EXPO is a set of expert optimization algorithms, deployed per project without large datasets or retraining. AtomNet, the three dimensional convolutional network the company ran as Atomwise, is described in peer reviewed and open access papers, the 2024 AIMS study among them.
For COSMOS, the architecture, training data and version history are not published.
COSMOS, APEX and EXPO are the company's own components, and NVIDIA subject matter experts coauthored the APEX research. The company's privacy policy names the kinds of service providers that handle personal information for it: customer support, hosting, analytics, email delivery, marketing and database management.
No hosting or compute provider, subprocessor list or statement of who can access partner data is published for the platform.
AtomNet, the model the company ran as Atomwise, was tested prospectively in the AIMS study (Scientific Reports, 2024), which reports successes and failures from 318 projects: 22 internal programs and 296 collaborations with 482 academic labs and screening centers at 257 institutions in 30 countries. At least one bioactive compound turned up in 215 of the 296 academic projects, a 73% success rate the authors compare with about 50% for high throughput screening, and 91% of the internal programs produced hits that reconfirmed in dose response testing. Hits came for targets with no known binders and for targets without high quality crystal structures.
For the current platform, the APEX research with NVIDIA reports benchmark retrieval of the top one million compounds from a 10 billion compound library in under 30 seconds. The company's headline figures, a chemistry operating domain 10,000 times larger than competitors and screening efficiency 10 billion fold higher, come without the comparisons behind them.
Inputs are protein targets and chemical libraries and outputs are ranked, optimized small molecules, so patient records play no part in the work. The sensitive material is a partner's own targets, compound data and assay results, the project data EXPO adapts its models to.
How long that data is kept, whether it trains COSMOS, and how one partner's work is kept apart from another's is not published.
The platform works on protein structures and chemical libraries, and the company's 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, and a partner's confidential targets and chemistry are governed by its collaboration agreement.
The company makes no statement on business associate status.
The company's privacy policy describes organizational, technical and physical safeguards for the personal information it collects, and adds that security risk is inherent in internet technologies, so it cannot guarantee security. Partners hand the platform confidential targets and chemistry.
No security certification, trust center or dedicated security page is published.
Discovery is the platform's whole job: it finds and optimizes small molecules and makes no diagnostic or treatment recommendation, so no device clearance applies to it. Regulatory exposure comes later, through the small molecule candidates its immune and inflammatory disease programs produce, which follow the drug pathway.
The company shows no platform candidate in a clinical study.
The AIMS study put AtomNet to work where a screening model is most likely to fail and published the results: for 207 of the 296 academic targets, the training data held no active molecule for that target or any closely related protein, and hits came for 53 of 72 protein protein interaction sites and 46 of 58 allosteric sites. Hit rates are broken out by protein and enzyme class, and the authors report that no class fell outside the model's reach.
No equivalent evaluation or governance framework is published for COSMOS, APEX or EXPO.
Failures are counted alongside successes in the company's own AIMS study, published as Atomwise with the authors' financial competing interest declared: of 296 academic projects, 215 produced at least one bioactive compound, and all 318 prospective experiments are reported. The academic compounds were bought off the shelf, purity checked and screened by the collaborating laboratories, so the result measures the model's picks rather than outcomes on a paying partner's campaign. The figure describes AtomNet, not the successor platform built on COSMOS, APEX and EXPO.
No warranty, remediation or fee consequence for failed predictions is published.
Work on the platform moves through protein structures, chemical libraries and assay results rather than clinical records, and partners bring their targets to the company directly.
No connector to a partner's lab notebooks, laboratory information systems, compound registries or research data platforms is described.
The company runs the platform in house for its own programs and for partners, whose targets come to it. Its office is at 250 Sutter Street in San Francisco, the privacy policy places the company in the United States with service providers in other countries, and the APEX benchmarks ran on a single NVIDIA GPU.
No hosting provider, region or separation between partners' data is stated for the platform.
The company develops its own small molecule programs in immune and inflammatory disease and has worked through discovery collaborations. As Atomwise it signed collaborations with Sanofi, Hansoh Pharma and Eli Lilly, whose headline values reflected negotiated deals, and gave academic laboratories screening access through the AIMS program.
No price, rate card or current partnership terms are published.
The company designs small molecule drugs, and its own programs target immune and inflammatory diseases. The APEX research reports high scoring, drug like compounds across kinases, GPCRs, proteases and nuclear receptors, and the platform is pitched as a universal chemistry foundation model. AtomNet's AIMS study spanned every major therapeutic area and protein class, with oncology at 35% of academic projects, infectious diseases at 27% and neurology at 9%.
No program target or development stage is published.
Alternatives to Numerion Labs
Other products that do the same kind of work as Numerion Labs, and what each one does.
Stanford spinout, now trading as Genesis Molecular AI, building GEMS, a small molecule discovery platform its own chemists and its partners' programs run on.
Iktos is a Paris generative AI and robotics company for small molecule drug discovery, founded in 2016.
Valo Health uses AI on large scale human data to find drug targets and design small molecules, through its Opal Computational Platform.
Computational chemistry software company (Nasdaq: SDGR) whose physics based molecular simulation platform is licensed to pharmaceutical, biotechnology, industrial and academic organizations, alongside a proprietary therapeutics pipeline built on the same platform.
Variational AI develops Enki, a generative model for designing small molecules against a research team's target product profile.
Superluminal Medicines chases the same target class as Nabla Bio and Antiverse and attacks it from the opposite direction.
Compared With
Each comparison carries a written verdict, the buyer conditions that favor each vendor, and a graded side by side. Pairs that cross a category boundary are grouped separately, and their verdicts state where the boundary sits rather than manufacturing a head to head.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Not published
|
Discovery collaboration and screening engagement with milestone and royalty economics. Academic access has been provided through a collaborative research program. | — | Not published. | Vendor Published |
The Atomwise era collaborations with Sanofi, Hansoh Pharma and Eli Lilly carried headline potential values, and the company now presents itself as developing its own programs in immune and inflammatory disease. Neither the split of those deals between upfront payments, research funding and milestones nor any current partnership terms have been disclosed.