toxra.ai is a toxicological risk assessment platform that consolidates the entire workflow—from chemical search and data retrieval to in-silico prediction and report generation—into a single, provenance-tracked environment. It combines an expert rule-based system (Pandecta) with a statistical model (Precedenta) to provide transparent, defensible mutagenicity predictions. The platform explicitly names data gaps and recommends specific OECD, EPA, or ICH studies to close them.
Funding
Funding not disclosed
Founders
Product
Problem
Toxicological risk assessments require assembling data from disparate sources—IRIS, PubChem, ECHA, ATSDR, IARC, and PubMed—and reconciling conflicting values, missing endpoints, and outdated spreadsheets. This manual stitching is time-consuming, error-prone, and leaves assessors without a clear audit trail. Additionally, for most chemicals, the required experimental study was never performed, and regulatory pressure is moving away from generating new animal data toward new approach methodologies, squeezing assessors from both sides.
Solution
toxra.ai provides an assessment instrument, not a search engine, that brings every step of a toxicological risk assessment into one chemical context, from first identifier to signed report. The platform retrieves data with provenance attached to every value, names data gaps with the specific OECD, EPA, or ICH study that would close them, and offers in-silico NAMs (QSAR) for prediction. It pairs two complementary prediction engines—Pandecta, an expert rule-based system, and Precedenta, a statistical model—that reason differently against the same molecule and report their agreement or disagreement, making the reasoning visible and defensible in a regulatory submission.
Target Audience
Primary users are toxicologists, regulatory affairs professionals, and risk assessors in pharmaceutical, chemical, and agrochemical industries who need to produce defensible, review-ready assessments under ICH M7 and other regulatory frameworks.
Features
- Provenance at the moment of retrieval: every number traces back to its source and version, enabling defensible values
- Gap naming with remedy: missing endpoints are explicitly flagged with the specific OECD, EPA, or ICH study that would close them
- Pandecta: a clean-room rule-based alert system with 78 active alerts, 0.844 sensitivity against Hansen 2009, and silence reported as silence (never as safe)
- Precedenta: a statistical engine trained on curated experimental data, with ROC-AUC 0.897 on internal hold-out, 0.849 sensitivity on the external EPA CompTox-genetox Ames panel, and a sensitivity-biased threshold holding false negatives to 7.9%
- Conformal prediction: abstains on uncertain structures, routing them to expert review rather than classifying them badly, with 0.89–0.91 specificity on confident calls
- Neighbour display: analogues that drove the result arrive with measured outcomes and structural similarity, allowing the reasoning to be read like legal precedent