Life sciences software · AI molecular simulation
Angstrom AI
A University of Cambridge spinout building generative AI molecular simulations meant to substitute for wet lab chemistry in early drug development. Its physics constrained models now back a published crystal structure method built with AstraZeneca, even as basic facts like who runs the company and how much it has raised are not settled in the public record.
Ten thousand times more information, one frame at a time
Angstrom AI's website shows a plot of supercooled water, held below freezing at negative 40 degrees Celsius, a condition that breaks most molecular dynamics software because the molecules barely move. Its Y Combinator launch post says the generative model packs about 10,000 times more information per step than a conventional simulator. A companion video shows 64 water molecules first simulated by a physics model that took a week of compute on 8 A100 GPUs, then reproduced far faster by a diffusion model trained to imitate it.
The pitch is not that the AI is smarter than physics. It is that the AI has learned to imitate a slow physics simulation quickly enough to make the imitation useful.
The gap it is aimed at
Before a drug reaches a patient, chemists need to know how it behaves in the body: whether it dissolves, binds its target, or holds a stable crystal form. The company names three existing methods and why each falls short. Wet lab experiments are accurate but slow and expensive. Machine learning predictions, it cites AlphaFold, are fast but only as good as the lab data that trained them. Molecular dynamics simulates the physics directly, but more accurate models cost more compute, making it compute bottlenecked rather than data bottlenecked.
Angstrom AI's answer is to keep the compute heavy physics model, MACE, and add a second generative model to produce similar output far faster. “We use diffusion models to accelerate MACE simulations, making them computationally affordable,” the team wrote in its 2024 launch post. “Our models generate states consistent with physics, but the transitions between states are non-physical and significantly faster.”
Whether that shortcut holds up on real drug molecules, not just water and methane, is what the company's own papers are starting to test.
Two PhDs, two professors, one physics model
The company describes itself as a team of two PhDs and two professors from Cambridge. Javier Antoran and Laurence Midgley did their doctorates there in generative AI and probabilistic modeling. Javier, per the launch post, scaled probabilistic AI methods 1000x during his PhD and turned down a Big Tech research scientist role to start the company. Laurence developed FAB, a method for training generative models from physical equations without training data, and worked at InstaDeep before its 2023 BioNTech acquisition.
Their supervisor Jose Miguel Hernandez-Lobato, is a Professor of Machine Learning at Cambridge with close to 20 years in the field and, per the company's YC profile, citations above 15,000. The fourth co-founder, Gabor Csanyi, built MACE, the model the company is built around. Combined, the company says the two professors' research has been cited more than 40,000 times.
Company records disagree on who runs the business day to day. Data platform Preqin and research firm Tracxn both name Javier Antoran as chief executive. A separate directory, Entangled Future, names Laurence Midgley as co-founder and CEO. UK corporate filings do not use the title CEO at all, so neither claim can be checked against that record.
A UK entity, an SF address, and two departures
Angstrom Ai Ltd was incorporated in the UK on April 12, 2024, registered on Great Portland Street in London, per Companies House. Gabor Csanyi joined as a director on June 17, 2024. The company's website and Y Combinator listing both give San Francisco as its base, and a street address there, 2325 Third Street, appears in the author affiliations of its two published papers.
Companies House filings also show two board departures. Laurence Midgley's directorship ended May 6, 2025. Javier Antoran's ended June 19, 2026, about three months before this page was written. Neither filing states a reason, and no independent reporting found explains either exit. It is possible operations moved to a US entity and the UK board was pared down for that reason; the record does not say.
The company says it has produced the first physically accurate generative AI simulation of multiple interacting molecules, and published water solubility results within the error range of wet lab experiments. It also says it started a $150,000 pilot with an unnamed pharma company on solubility estimation. A separate, confirmed collaboration surfaced in 2026: a preprint on crystal structure prediction lists co-authors from AstraZeneca's own CSP team, named directly in a company LinkedIn post.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Solvation free energy paper | J Am Chem Soc, Jan 27, 2026. Csanyi and researcher J Harry Moore list an Angstrom AI, San Francisco affiliation. Subchemical accuracy for solvation free energies. | Public record |
| MACE-OFF force field paper | J Am Chem Soc, May 19, 2025. Csanyi and Moore again carry the affiliation. The transferable force field the simulations build on. | Public record |
| CSP-MACE-A preprint with AstraZeneca | arXiv, May 27, 2026. Midgley, Antoran and Csanyi plus five named AstraZeneca scientists. Close to DFT accuracy across 47 test compounds, far faster. | Public record |
| AstraZeneca collaboration, as described | A LinkedIn post names five AstraZeneca CSP scientists, matching the paper's authors. AstraZeneca has issued no public statement of its own. | Company-stated |
| $150,000 pharma pilot | Described in the company's YC profile as a solubility project. Partner not named, no independent confirmation found. | Company-stated |
| Water solubility results | Company says it published water solubility results within wet lab error range. No standalone paper found beyond the solvation paper above. | Not independently found |
| Chief executive | Preqin and Tracxn name Antoran as CEO. Entangled Future names Midgley. Companies House shows both left the UK board and uses no CEO title. | Differs across sources |
| Funding raised | No SEC Form D found. Preqin cites Pioneer Fund, Sep 2024, no amount. Tracxn cites $500,000, Jul 2024. GetLatka reports zero. | Differs across sources |
| Team size | YC lists 5. Tracxn counts 4 as of June 2026. GetLatka estimates 6. | Differs across sources |
Read plainly: the science has a paper trail, three papers in two years, one with a named pharma co-author. The business side, who leads it, how it is funded, how many people work there, is reported inconsistently, and none of it is backed by a filing this research could find.
What to watch
- Whether Angstrom AI publishes results on drug-like molecules beyond water, methane and the AstraZeneca crystal set, the harder test of the wet lab replacement claim.
- Whether AstraZeneca or another pharma partner confirms a commercial relationship in its own materials, rather than only in Angstrom AI's posts and papers.
- Whether a Form D appears under the company's name or a variant of it, settling the conflicting seed funding reports.
- Whether the UK entity is wound down now that two of its four founding directors have left its board.
In their words
“We use diffusion models to accelerate MACE simulations, making them computationally affordable. Our models generate states consistent with physics, but the transitions between states are non-physical and significantly faster.”
Angstrom AI, Y Combinator launch post, 2024 · Company
“Our AI introduces about 10,000 times more information per step compared to traditional simulations.”
Angstrom AI, Y Combinator launch post, 2024 · Company
“I'm the co-founder of Ångström AI which substitutes wet lab experiments with physically accurate GenAI molecular simulations for clients in Pharma and BioTech. Before founding Ångström AI, I was a research engineer at InstaDeep (acquired by BioNTech 2023), and pursued a PhD at the University of Cambridge in generative AI models for molecular systems. I love coding and surfing, reach out if you are in the Bay Area and want to catch some waves.”
Co-founder and CTO, Y Combinator company profile, 2024 · Company
“The highlight of this project has been working closely with AstraZeneca's CSP team (Sten Nilsson Lill, Emma Eriksson, Felix Faber, Lars Tornberg, Anders Broo) for the past several months to understand how they do CSP and how Angstrom AI can build something useful for them!”
Angstrom AI, company LinkedIn post, 2026 · Company
“By running multiple orders of magnitude faster than DFT, CSP-MACE-Å enables energy and free energy evaluation of far more candidate structures, providing greater confidence when derisking solid forms.”
Midgley et al., arXiv preprint 2605.28905, May 27, 2026 · Public record
Related companies
Sources
- Public recordCompanies House overview, ANGSTROM AI LTD, company 15640060
- Public recordCompanies House filing history, ANGSTROM AI LTD
- Public recordComputing Solvation Free Energies of Small Molecules with Experimental Accuracy
- Public recordMACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules
- Public recordDFT Accuracy on Crystal Structure Prediction with ML Interatomic Potentials, arXiv 2605.28905
- Public recordEDGAR company search, companies named Angstrom, Form D filings
- IndependentAI Insider Profile: Angstrom AI Aims to Revolutionize Drug Discovery
- IndependentFrom Wet Labs to AI: How Angstrom AI is Transforming Pharma
- DirectoryAngstrom AI company profile
- Data aggregatorAngstrom AI, Inc. asset profile
- Data aggregatorAngstrom AI company and funding profile
- Data aggregatorAngstrom AI revenue and team estimate
- CompanyAngstrom AI company profile and founder bios
- CompanyLaunch YC: Angstrom AI, Gen AI molecular simulations
- CompanyAngstrom AI website, Our Science
- CompanyAngstrom AI company page, AstraZeneca CSP-MACE-A post
- AcceleratorAngstrom AI portfolio profile
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
