Lunai Bioworks Tests AI Chemical Risk Screening Model

In a press release, Lunai Bioworks shared results from a BioSymetrics case study of structure based screening for chemical outputs generated by large language models and other AI systems. The study evaluated 9,667 compound entries using public NIH Tox21 data for acetylcholinesterase activity.

The selected model achieved a 0.88 AUROC on a held out test set, with 33% precision at 70% recall. It produced about 5.1 times enrichment compared with the 6.4% prevalence of active entries in the test set. Training, validation, and test data were separated by chemical scaffold.

The screening method analyzes molecular structure independently of the system that generated a candidate compound. It could help prioritize compounds for further computational analysis, expert review, or biological testing.

The study measured assay defined acetylcholinesterase activity. It did not establish clinical neurotoxicity, actual chemical hazards, prevention of misuse, or performance against all chemical threat classes. Lunai said prospective validation is still required and clarified that OpenAI and Anthropic did not participate in or endorse the study.