ToxIndex generates a 47-page chemical safety report in three hours - work that usually takes six months

Only about one in ten chemicals in commercial use has been properly assessed for safety. One in a hundred gets a comprehensive evaluation. The testing costs close to $20 billion annually, and the backlog keeps growing. Meanwhile, AI-driven drug discovery is generating thousands of novel molecules that each need safety evaluation, and only 9,000 toxicologists serve all of North America.

That arithmetic is the backdrop for ToxIndex, an agentic AI platform developed by Insilica Inc. that launched publicly ahead of the Society of Toxicology's 65th Annual Meeting in San Diego. The platform's pitch: it can produce comprehensive, source-traceable toxicological risk assessments in hours rather than months.

How the platform assembles a risk assessment

ToxIndex works by deploying AI agents that programmatically access and orchestrate three tiers of safety evidence. The first tier draws on over 600 containerized open-source toxicology models, including a proprietary transformer model trained on 254 million human chemical activity measurements. The second indexes existing laboratory data from a knowledge graph containing over 60 billion data triples, built with NSF funding and incorporating ToxCast, Tox21, ChEMBL, PubChem, and 1,200 additional curated sources. The third tier covers physicochemical and ADME profiling through REACH dossiers, EPA databases, and computational tools.

Every claim in a generated report carries full provenance - traceable to the specific database, table, row, and column of origin. Each claim also receives a Klimisch reliability score, the standard rating system used by regulatory agencies to assess data quality.

In a proof-of-concept demonstration, the platform generated a 47-page risk assessment with 944 individually sourced claims for dodecanedioic acid (DDDA) in under three hours. That type of assessment conventionally requires three to six months of expert manual work.

The regulatory context driving adoption

The timing reflects converging regulatory pressures. The FDA Modernization Act 2.0 eliminated mandatory animal testing requirements and explicitly endorsed computational alternatives. The EPA has committed to ending mammalian testing by 2035. Europe's REACH compliance burden - estimated at 9.5 billion euros - has created massive review backlogs, and the EU's cosmetics animal testing ban has made computational safety assessment a strategic necessity for one of the world's largest consumer industries.

Meanwhile, 88% of new chemical reviews already exceed statutory deadlines, a bottleneck that computational approaches could help relieve if they meet regulatory standards for documentation and traceability.

Academic roots and institutional backing

ToxIndex was created by Thomas Luechtefeld, who completed his doctoral work in computational toxicology at the Johns Hopkins Center for Alternatives to Animal Testing (CAAT) under Thomas Hartung. Hartung - who spent seven years heading the European Commission's validation body for alternative testing methods (ECVAM) and advises the EPA, FDA, EFSA, and OECD - has publicly endorsed the platform.

The two also collaborate within the European ONTOX consortium, a 17.2 million euro EU initiative developing non-animal toxicity prediction methods. Plans are underway to share the platform's underlying concepts on arXiv, code on GitHub, and the transformer model with NIH and OECD partners.

Important qualifications

Several caveats deserve attention. The proof-of-concept used DDDA, a relatively well-characterized chemical with substantial existing data. Performance on novel molecules with sparse safety data - precisely the compounds that most urgently need assessment - may differ. The platform's value depends heavily on the quality and coverage of its underlying data sources, and toxicological databases are known for gaps, inconsistencies, and variable quality.

Regulatory acceptance is not guaranteed by technical capability. Agencies may require extensive validation before accepting AI-generated assessments as substitutes for traditional expert review, particularly for high-stakes decisions involving chemicals with significant human exposure. The platform's traceability features address this concern but do not eliminate it.

Insilica is a commercial entity, and NeuroSense funded the platform's development. The endorsements, while from credible scientists, come from collaborators with financial and institutional ties to the project. Independent validation by regulatory agencies and external researchers will be necessary to establish the platform's reliability at scale.

That said, the gap between safety assessment capacity and demand is real, growing, and poorly served by current methods. A tool that can systematically integrate the scattered landscape of toxicological data, models, and methods - with auditable provenance - addresses a genuine structural problem in chemical safety science.