A multi-pathway mechanism-informed read-across framework for screening developmental and reproductive toxicity priorities among data-poor PFAS.

Abstract

Per- and polyfluoroalkyl substances (PFAS) form a large chemical universe, yet comparable developmental and reproductive toxicity points of departure are available for only a small fraction of chemicals. We developed and evaluated a multi-pathway mechanism-informed read-across approach for screening data-poor PFAS while separating prediction from regulatory hazard classification. An EPA CompTox EPAPFASINV inventory was standardized with RDKit, restricted to 69 PFAS in four strict acid classes, linked to ToxCast bioactivity across seven nuclear-receptor and thyroid-related pathway groups, and evaluated against a narrow matrix of eight PFAS with rat oral developmental or reproduction/developmental endpoints. Three GenRA-style models were compared: structure similarity, structure plus chemical-domain gating, and a mechanism-informed model. Validation was preliminary because endpoint counts were small and model coverage differed. On the common target-endpoint subset predicted by all models, the mechanism-informed model had similar aggregate error to the structural baselines (MAE 1.41 log10 units) but lower coverage, indicating a coverage-performance tradeoff, not broad superiority. In the hierarchical screening application, 44 of 61 data-poor PFAS received predictions; 14 were classified as high screening priority, 19 as moderate, 11 as lower, and 17 as insufficient evidence. Most ether-containing replacement PFAS lacked same-class in vivo source endpoints. These results support use of the approach for prioritizing testing needs and documenting analogue rationales; they do not establish a validated regulatory point-of-departure model.