Chemical-space-aware routing for improved PFAS toxicity prediction: large-scale pretraining, conditional fine-tuning, and leakage control

Abstract Predicting the toxicity of per- and polyfluoroalkyl substances (PFAS)—persistent environmental pollutants with bioactivity data available for only a small fraction of inventoried compounds-remains challenging because PFAS-specific training sets are small. Here, we quantify how much domain-specific data is actually required and introduce chemical-space-aware routing, in which a structure-based rule automatically directs PFAS-like molecules to a dedicated prediction panel within an open ADMET platform. We trained directed message-passing neural networks (Chemprop) on a 26-bioassay PFAS panel (n = 1,012) derived from the Cheng–Ng dataset, comparing training from scratch, large-scale multi-task pretraining on ~ 53,000 non-PFAS compounds from the same assay collection, and PFAS fine-tuning, under random and scaffold (perfluoroalkyl chain-length) splitting regimes with three random seeds and a Morgan-fingerprint random-forest baseline. Canonical-structure overlap analysis revealed that the PFAS panel is a strict subset of the pretraining collection (1,012/1,012 compounds), so pretraining without exclusion fabricates an apparent fine-tuning gain of up to + 0.36 AUROC. After decontamination, scratch-trained models showed a large generalization gap (random test AUROC 0.751 vs. scaffold 0.607 ± 0.044) and were outperformed by the random-forest baseline on the scaffold split (0.769); a pretrained model that had never encountered any PFAS achieved 0.798, nearly closing the gap, while fine-tuning added gains only under random splitting (0.864 vs. 0.804) and not on novel chain lengths (0.812 ± 0.032 vs. 0.798). The routing rule detected 82 of 86 held-out PFAS (95.3% sensitivity) and flagged 41 of 368 approved drugs (88.9% specificity), notably, 14 of the 41 flagged drugs are perfluorocarbons or fluorinated inhalational anesthetics chemically akin to PFAS, revealing a previously undocumented overlap between the approved-drug and PFAS chemical spaces. Category-level profiles placed CYP-mediated metabolism as the dominant axis for PFOA and PFOS, both of which appear unremarkable on conventional drug-oriented ADMET summaries. The code, decontaminated data, and model weights are freely available at https://github.com/zhanting-yang/PFAS-GNN (DOI: 10.5281/zenodo.22869943).