Excelra data will train Boltz AI models for drug discovery

Curated data for early compound triage

The partnership will provide Boltz with curated measurements from Excelra’s GOSTAR database covering kinetic solubility, lipophilicity and Caco-2 permeability. These measurements will be used to train the ADMET models in BoltzMol-1.

According to the companies, the models act as an early triage layer for drug discovery. Their role is to identify compounds with unfavorable properties before researchers commit to purchasing or synthesizing them, potentially helping teams focus experimental work on more promising candidates.

Excelra describes GOSTAR as an ISO 9001-certified database containing a curated overview of more than 11 million small molecules and their biological properties. The data are available through an application, APIs or as a downloadable dataset for machine learning.

Boltz reports results across 10 difficult targets

Boltz said a recent technical report prospectively tested 10 challenging target structures. Most of those targets were largely absent from its affinity training data, and the company confirmed binders or active compounds for six of them.

The work used between 28 and 96 compounds per target. That compares with the thousands to millions of compounds used in conventional high-throughput screening, where hit rates are often below 0.15%, according to the source.

Boltz also reported that the GOSTAR-trained ADMET models matched or exceeded a leading open-source ADMET package on purely out-of-distribution chemistry. The largest improvements were reported for solubility and permeability.

Why the partnership matters for drug discovery

The companies are positioning the partnership around a gap in publicly available data. Excelra said processed data, including solubility data at this scale, is missing from public sources and that GOSTAR was developed to provide a foundation for models tested with prospective data rather than only retrospective benchmarks.

For Boltz, the data will support models intended to generalize to chemical processes beyond typical training data. Gabriele Corso, co-founder and chief executive of Boltz PBC, said the GOSTAR dataset had become an important input for the company’s modeling work and helped it save substantial time, according to the announcement.

Boltz PBC is an applied AI research laboratory and public benefit corporation behind open and proprietary models for protein structure prediction, protein design and small-molecule drug discovery. Excelra is a data science and AI partner for life-science organizations. The announcement does not disclose commercial terms or a launch date for the partnership.