Drug discovery · AI chemistry
Numerion Labs
Formerly Atomwise. Rebuilt under a new name and a new chief executive, with more published platform validation than most of its peers and nothing yet in patients.
Thirty seconds
On October 29, 2025, a San Francisco drug discovery company put out a release with a benchmark and a new name in it.
The benchmark: a search protocol called APEX had retrieved the best one million compounds from a library of more than 10 billion, in under 30 seconds, on a single graphics card. Not a sample of the library. Every compound in it.
The name was Numerion Labs. The release handled the rest in one clause. This was the company “formerly known as Atomwise”.
Atomwise was among the first to argue that deep learning could replace the physical screening deck at the front of drug discovery, and it raised money on that argument for a decade. The rename followed a 30 percent staff cut, the exit of its founding chief executive, and a round roughly a third the size of the one before it. The technology came through all of that. Whether a medicine does is still open.
The problem it picked
Drug discovery begins by finding a small molecule that binds a protein. High throughput screening pushes physical compounds through robotic assays. It works, and it is bounded by what sits in the plate.
Make on demand catalogues broke that bound. Libraries such as Enamine REAL list billions of compounds that have not been synthesised but can be, in weeks. Scoring all of them is the hard part. The company’s own preprint puts the limit bluntly: a routine campaign “may be limited to scoring fewer than 0.1% of the available compounds, leaving potentially many high scoring compounds undiscovered.”
AtomNet, the original model, predicted binding from three dimensional structure. APEX is an engineering answer: train a cheap surrogate model, factorise it across the reactions and building blocks that define the library, cache those pieces once, then stream the catalogue through a priority queue on the GPU. The heavy cost moves to a one time step, and the search becomes close to free.
Three founders, none of them still there
Atomwise came out of Toronto and through Y Combinator’s Winter 2015 batch. Abraham Heifets, a computer scientist with a doctorate from the University of Toronto and a stint at IBM Research, ran it as chief executive. Izhar Wallach co-founded it on the algorithm side and is still a director on the latest SEC filing. Alex Levy, the third co-founder, left in 2018 and now runs an AI weather forecasting company.
Heifets sold scale. In 2018 he told TechCrunch that the vision was “to become one of the most prolific and diverse life science research groups in the world, working at a scale that is truly unprecedented.” Two years later, announcing a $123 million round, he pointed at a different target: “We are also going to focus on what is unique about us: drugging the undruggable.”
Endpoints News reported in October 2024 that “Atomwise’s founding CEO Abraham Heifets has left the company after a dozen years”. The Y Combinator directory still lists him under former founders, now described as “Founder and CEO of A Company in Stealth Mode.”
The decade, in order
March 2018: a $45 million Series A led by Monsanto Growth Ventures, DCVC and B Capital Group. June 2019: Eli Lilly agrees to pay up to $1 million per target for as many as 10 targets, with up to $550 million in later milestones. August 2020: a $123 million Series B co-led by B Capital Group and Sanabil Investments. August 2022: Sanofi pays $20 million upfront for up to five targets, against more than $1 billion in possible milestones.
Then the arrow turned. In December 2022, four months after the Sanofi deal, the company cut 30 percent of its staff, which Endpoints News reported the following month. In September 2023 it nominated its first development candidate, an orally bioavailable allosteric TYK2 inhibitor, appointed a chief medical officer to carry it into trials, and said it was “targeting IND submission in 2H 2024”.
April 2024 brought the platform’s best evidence: a Scientific Reports paper covering 318 separate projects. Six months later Endpoints reported the founder’s departure. In February 2025 the board named Steve Worland, who had led Anadys Pharmaceuticals through its sale to Roche and co-founded eFFECTOR Therapeutics, and disclosed a Series C. In October 2025 the name changed, and in March 2026 a fresh offering appeared in EDGAR under the new one.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Platform validation, peer reviewed | Scientific Reports, April 2, 2024. A prospective campaign across 318 projects. The authors conclude that computational methods “can substantially replace HTS as the first step of small-molecule drug discovery.” GEN reported novel hits for 235 of 318 targets, a 74% project success rate against a literature estimate of 50% for physical screening. | Public record |
| APEX search protocol | arXiv preprint posted October 2025, co-authored with NVIDIA staff, code on GitHub. Benchmarks use two libraries the authors built themselves. Not peer reviewed as of September 22, 2026. | Company-stated |
| Lead candidate, preclinical data | ATMW-DC presented at ACR Convergence on November 17, 2024 and in the EULAR 2024 abstracts. TYK2-JH2 IC50 of 12 pM, at least 350 fold selectivity over JAK1, JAK2 and JAK3, and 65% to 69% inhibition of ear swelling in an IL-23 driven mouse model. | Public record |
| Clinical entry | The September 2023 nomination release targeted an IND submission in the second half of 2024. No IND confirmation and no first in human study has been published in the two years since. | Claim outruns record |
| Registered trials | No trial found on the public registry under either company name as of September 22, 2026. | Not found |
| Federal research awards | None found under either name. The 25 NSF records returned by a name search all belong to unrelated companies whose names contain the word Labs. | Not found |
| Partnered validation | Eli Lilly, June 2019, up to $1 million per target for as many as 10 targets and up to $550 million in milestones. Sanofi, August 2022, $20 million upfront for up to five targets with more than $1 billion in milestones. No new pharmaceutical partnership since. | Independent |
| Capital raised | Four Form D filings under CIK 0001736153 report $238,580,439 sold between 2018 and 2026. The 2026 filing is the first under the Numerion Labs name. | Public record |
Read plainly: this platform has been tested in public across more targets than almost any other in its category, and the 318 target paper is the strongest single artefact an AI screening company has put into the literature. What the record does not contain is a patient. Three years after the first candidate was nominated and two years past the IND target date, the published evidence for that molecule stops at potency numbers in mice.
What to watch
- Any IND filing or first in human study for ATMW-DC. The stated target was the second half of 2024.
- Whether the $35 million offering opened in March 2026 closes. $15,205,055 of it was unsold at the filing date.
- Peer review for APEX, which is still a preprint with company and NVIDIA authors.
- A new pharmaceutical partnership. The last one announced was Sanofi, in August 2022.
- Whether the public pipeline page ever names a program. It carried none on September 22, 2026.
In their words
“to become one of the most prolific and diverse life science research groups in the world, working at a scale that is truly unprecedented”
Abraham Heifets, co-founder and CEO, TechCrunch, 2018 · Independent
“We chose to invest based on the impressive results we saw from Atomwise in our own hands.”
Kiersten Stead, Monsanto Growth Ventures, quoted by TechCrunch, 2018 · Investor
“We believe the best path to discovering first-in-class and best-in-class medicines is to explore unique and untapped chemical space, which is what our AI enables. If you want to differentiate, it helps to start different.”
Abraham Heifets, CEO and co-founder, company release, 2023 · Company release
“Generally the predictive power of virtual screening platforms has been extremely limited.”
Gregory Bowman, professor, University of Pennsylvania, GEN, 2024 · Independent
“I first wrote about Atomwise back in 2015, when I found the hype around the company a bit hard to take”
Derek Lowe, In the Pipeline, Science, 2024 · Independent
“Atomwise spent most of its first 12 years talking up the potential of AI in making molecules.”
Andrew Dunn, Endpoints News, 2025 · Independent
“With APEX, we’ve demonstrated that it is now possible to virtually evaluate billions of molecules in seconds, enabling a real-time marriage of creativity between chemist and computer.”
Steve Worland, CEO, company release, 2025 · Company release
Related companies
Sources
- Public recordForm D, Atomwise Inc., CIK 0001736153
- Public recordForm D, Atomwise Inc., CIK 0001736153
- Public recordForm D, Atomwise Inc., CIK 0001736153
- Public recordForm D, Numerion Labs, Inc., CIK 0001736153
- Public recordAI is a viable alternative to high throughput screening: a 318-target study
- Public recordATMW-DC allosteric TYK2 inhibitor abstract
- PreprintAPEX search protocol, with NVIDIA co-authors
- IndependentDeep learning startup Atomwise raises $45M
- IndependentAtomwise Lands $123M To Advance AI Drug Discovery
- IndependentSanofi signs $1.2B pact with Atomwise
- IndependentAtomwise laid off 30% of staffers amid strategic shift
- IndependentFounding CEO leaves Atomwise
- IndependentAtomwise taps biotech vet as CEO to lead pared-down company
- IndependentAI Does Not Make It Easy
- IndependentAtomwise AI platform as an alternative to high throughput screening
- CompanyNumerion Labs publishes APEX research with NVIDIA
- CompanyAtomwise nominates first AI-driven development candidate
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
