Boltz PBC has partnered with Excelra to power its AI drug discovery models with GOSTAR™ data, the companies announced. Curated GOSTAR™ measurements of kinetic solubility, lipophilicity and Caco-2 permeability train the ADMET models in BoltzMol-1, which act as an early triage layer, filtering out unfavorable compounds before purchase or synthesis.

Prospective Testing Results

In a recent technical report, Boltz prospectively tested ten challenging targets, most absent from its affinity training data, and confirmed binders or active compounds for six, using just 28–96 compounds per target. The company contrasted this with conventional high-throughput screens, which use thousands to millions of compounds and where hit rates are often below 0.15%.

Boltz said its GOSTAR™-trained ADMET models matched or exceeded the leading open-source ADMET package on strictly out-of-distribution chemistry, with the largest gains in solubility and permeability.

Executive Comment

Gabriele Corso, Co-founder and CEO of Boltz PBC, said the GOSTAR™ dataset's scale helps the company build models that generalize to chemistry far outside typical training data. He described it as an important input across Boltz's modeling work that lets the company save a significant amount of time.

Anandbir Brar, CEO of Excelra, said the company is proud to support Boltz as they push the frontier of AI-driven drug discovery. He said curated and solubility data at this scale is exactly what public sources lack, and what GOSTAR™ was built to provide: a foundation for models that hold up on prospective data, not just retrospective benchmarks.

About the Companies

Excelra describes itself as a data science and AI partner to life science organizations. Its GOSTAR™ database offers a curated, ISO 9001-certified view of over 11 million small molecules and their biological properties, available as an application, via APIs, or as a downloadable dataset for machine learning.

Boltz PBC describes itself as an applied AI research lab and public benefit corporation behind the Boltz family of open and proprietary models for protein structure prediction, protein design, and small-molecule hit discovery.