Axiom Bio Releases Human Liver Toxicity Dataset with 130,000 Compounds to Advance AI-based Toxicity Prediction
San Francisco-based startup Axiom Bio, previously backed by $15 million in seed funding from investors including Amplify Partners, Dimension Capital, and Zetta Ventures, has released a publicly accessible data exploration tool featuring a primary human liver toxicity dataset. This dataset includes detailed liver toxicity profiles covering more than 130,000 distinct molecules.
See also: Axiom Bio Launches with $15M to Replace Animal Testing with AI for Drug Toxicity Prediction
The release builds upon Axiom’s proprietary wet-lab protocols and its automation platform, which allows high-throughput, multiplexed screening with primary human hepatocytes, a challenging and costly cell type traditionally difficult to scale. According to co-founder Brandon White, the key innovation lies in simultaneously achieving high-throughput capacity and preserving physiological relevance. The team conducted over 120 separate batches using a consistent donor pool.
Axiom's web tool for the dataset exploration
Unlike conventional assays that typically rely on single endpoints, Axiom’s assays measure 10–20 toxicity-related cellular phenotypes simultaneously from each well, including apoptosis, necrosis, mitochondrial fission, endoplasmic reticulum stress, stress granule formation, and microtubule stability. The ability to capture this extent biological information per compound across thousands of conditions per experiment is specifically designed to improve AI-based predictions of drug-induced liver injury (DILI).
In total, the newly released dataset, "X-Axiom", encompasses:
- 130,000+ unique small molecules, covering 1,200+ unique biological targets and 50,000+ chemical scaffolds.
- Specialized compound classes, including 3,300+ macrocyclic compounds and 9,500+ PROTACs and molecular glues.
- Over 700 clinical-stage molecules referenced from FDA’s DILIrank dataset, linking experimental data to clinically observed outcomes.
- High-content imaging data featuring 394+ million individually labeled cells and 9+ billion mitochondria, annotated through more than 7,000 human labeling hours.
This release follows Axiom’s earlier announcement at the Society of Toxicology (SOT) conference in March 2025, where it shared pilot validation results (Ewald et al. 2025) showing predictive sensitivity of around 70–75% and specificity approaching 90% in identifying liver-toxic compounds, reportedly matching or surpassing existing laboratory-based liver assays at a significantly reduced cost.
The dataset aligns with recent industry shifts toward adopting human-based, AI-supported methods over traditional animal testing, particularly following FDA initiatives aimed at transitioning away from animal models. Axiom plans future expansions of its predictive platform to cover other organ systems, including brain, heart, and immunogenicity.
Explore the dataset: axi.om/dataset
Topics: AI & Digital
