Boltz trains BoltzMol-1 ADMET models on Excelra's curated GOSTAR data

Boltz PBC, the Cambridge, Massachusetts AI lab behind the Boltz structure prediction and design models, is training the ADMET models in its BoltzMol-1 system on curated data from Excelra's GOSTAR database, the two companies announced on 10 November 2026¹. Excelra is based in Hyderabad, India¹.

What the partnership covers

The GOSTAR data feeding BoltzMol-1 consists of curated measurements of three properties: kinetic solubility, lipophilicity and Caco-2 permeability¹. Boltz uses the resulting ADMET models as an early screen that discards compounds with poor properties before a chemist buys or makes them¹. GOSTAR itself covers more than 11 million small molecules and their biological properties, is ISO 9001-certified, and is sold as an application, through APIs or as a downloadable machine learning dataset¹.

Boltz cites results from a recent technical report. In prospective tests on ten difficult targets, most of which were missing from the affinity training data, the company confirmed binders or active compounds for six of them while testing only 28 to 96 compounds per target¹. The release sets this against conventional high-throughput screens of thousands to millions of compounds, where hit rates are often below 0.15%¹. Boltz also says its GOSTAR-trained ADMET models matched or beat the leading open-source ADMET package on strictly out-of-distribution chemistry, with the biggest improvements in solubility and permeability¹.

Gabriele Corso, Boltz co-founder and chief executive, said the dataset's scale "helps us build models that generalize to chemistry far outside typical training data"¹. Excelra chief executive Anandbir Brar said curated solubility data at this scale "is exactly what public sources lack"¹.

These are the companies' own claims in a press release. The technical report has not been described as peer reviewed, the release does not name the open-source comparator, and no financial terms were given¹.

What it means for the market

The deal points to where competition in AI chemistry is moving. Structure and binding models are increasingly available, including open ones such as Boltz's own family, so the differentiator becomes the quality of property data used to predict whether a hit can become a drug¹. Brar framed the goal as models that "hold up on prospective data, not just retrospective benchmarks", which is the test discovery teams care about when deciding which compounds to synthesise¹.

For medicinal chemistry and discovery leaders, the practical claim is cost: 28 to 96 compounds per target instead of large screening campaigns, with an ADMET filter applied before money is spent on synthesis¹. If that holds up in wider use, it reduces the wet-lab spend needed to reach a hit and shifts budget towards data licences and model access.

For curated data vendors, the partnership is a template. Commercial datasets that public sources cannot match are becoming training inputs for AI labs, not only lookup tools for chemists¹. Pharma companies weighing in-house ADMET models against vendor platforms will want to see the out-of-distribution results published in full before relying on them.