An AI platform just produced a 47-page toxicology report in three hours - a task that usually takes months

Chemical safety testing costs nearly $20 billion per year globally. For that investment, roughly one in ten commercially used chemicals has been assessed. One in a hundred gets a comprehensive evaluation. The bottleneck isn't scientific knowledge - it's the sheer labor of pulling together data from hundreds of sources, running it through dozens of models, and documenting every claim with traceable provenance. A single risk assessment can take three to six months of expert work.

ToxIndex, an agentic AI platform built by Insilica Inc., claims to compress that timeline to hours. In a proof-of-concept demonstration, the system generated a 47-page, 944-claim toxicological risk assessment for dodecanedioic acid (DDDA) in under three hours. Every claim carries a reliability score and traces back to its specific source database, table, row, and column of origin.

The 2007 vision, finally assembled

The National Research Council's 2007 report Toxicity Testing in the 21st Century laid out a roadmap: replace traditional animal-based safety testing with integrated computational models, in vitro assays, and pathway-based approaches. In the nearly two decades since, the scientific community built most of the pieces - validated quantitative structure-activity relationship (QSAR) models, high-throughput screening libraries like ToxCast and Tox21, adverse outcome pathway frameworks, and curated regulatory databases.

What was missing, according to Thomas Hartung, director of the Center for Alternatives to Animal Testing (CAAT) at the Johns Hopkins Bloomberg School of Public Health, was the engineering to make it all work together at scale. Hartung, who spent seven years heading the European Centre for the Validation of Alternative Methods (ECVAM) at the European Commission and advises the EPA, FDA, EFSA, and OECD, has endorsed ToxIndex as the platform that bridges that gap.

Three tiers of evidence, one auditable workflow

The platform integrates three categories of New Approach Methodologies (NAMs) into a single pipeline. In silico predictions draw on over 600 containerized open-source toxicology models plus a proprietary transformer model trained on 254 million human chemical activity measurements. In vitro evidence comes from a knowledge graph of more than 60 billion data triples - co-developed with NICEATM under National Science Foundation Proto-OKN funding - that encompasses ToxCast, Tox21, ChEMBL, PubChem, and 1,200 additional curated sources. Physicochemical and ADME (absorption, distribution, metabolism, excretion) profiling rounds out the evidence base with data from REACH dossiers, EPA databases, and tools like the OECD QSAR Toolbox.

The system's AI agents programmatically access and orchestrate these resources, routing data through the appropriate models and assembling the output into a structured report. When evidence gaps remain, the platform can route procurement requests to testing laboratories. It updates automatically as new literature is indexed by a document crawler that continuously processes regulatory and scientific publications from agencies worldwide.

Regulatory pressure as catalyst

The timing reflects converging mandates. The FDA Modernization Act 2.0 eliminated the requirement for mandatory animal testing and explicitly endorsed computational alternatives. The EPA committed to eliminating mammalian testing by 2035. Europe's 9.5-billion-euro REACH compliance burden has created massive backlogs, and the EU animal testing ban for cosmetics has made AI-first safety assessment a strategic necessity for one of the world's largest consumer products industries.

Meanwhile, AI-driven drug discovery is generating thousands of novel molecules that each require safety evaluation. Only about 9,000 toxicologists serve all of North America, and 88% of new chemical reviews already exceed statutory deadlines. The workforce simply cannot scale to meet demand through traditional methods.

Trust and traceability

In toxicology, accuracy is a matter of human safety. Hartung has emphasized that 95% accuracy is insufficient when lives are at stake, and that trust must be earned through institutional relationships, transparent methodology, and rigorous validation. The platform's full-provenance design - where every claim links to its originating data - is built to meet or exceed current REACH and TSCA standards for regulatory documentation.

ToxIndex developer Thomas Luechtefeld, who completed his doctoral work in computational toxicology under Hartung at CAAT, also collaborates within the European ONTOX consortium, a 17.2-million-euro EU Horizon 2020 initiative developing non-animal toxicity prediction methods. The underlying transformer model and concepts are being shared on arXiv and GitHub, with discussions underway with NIH and OECD for broader public access.

What this doesn't tell us

A proof-of-concept on a single compound is not validation at scale. The 47-page DDDA report demonstrates speed and traceability, but the question of accuracy across diverse chemical classes - including novel molecules with limited existing data - remains to be answered through broader deployment and independent verification. Regulatory agencies will need to assess whether AI-generated reports meet their standards in practice, not just in principle.

The platform is also a commercial product from a private company, and Hartung serves as a scientific advisor to Insilica - a relationship that should be weighed when evaluating his endorsement, however well-credentialed he is.

Insilica will present and demonstrate ToxIndex at the Society of Toxicology 65th Annual Meeting in San Diego, March 22-25, 2026.