{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,12]],"date-time":"2026-10-12T12:49:19Z","timestamp":1791809359160,"version":"4.3.4"},"reference-count":43,"publisher":"American Chemical Society (ACS)","issue":"19","license":[{"start":{"date-parts":[[2026,9,24]],"date-time":"2026-09-24T00:00:00Z","timestamp":1790208000000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"published-print":{"date-parts":[[2026,10,12]]},"abstract":"<jats:title>Abstract<\/jats:title>\n                  <jats:p>Per- and polyfluoroalkyl substances (PFAS) comprise more than 21,000 chemicals, yet fewer than 1.3% have any experimental bioactivity measurement, making structure-based hazard prediction the gating problem for prioritizing the untested majority. Existing PFAS models are single-task and rarely evaluated under leakage-aware cross-validation. We present the first multitask chemical foundation-model benchmark for PFAS bioactivity, fine-tuning off-the-shelf MolFormer-XL with 156 ToxCast classification heads under Tanimoto-Butina cluster-out cross-validation, which is appropriate because 89% of PFAS have empty Murcko scaffolds. The model reaches macro AUROC 0.764, beating a Morgan-fingerprint XGBoost baseline by +0.18 and winning on 155 of 156 end points. That baseline is, however, partly degenerate: binary Morgan fingerprints map a quarter of the compounds onto identical vectors, and sweeping eight 1D\/2D\/3D representations on the same folds raises the best classical model to 0.688, so the margin properly attributable to the foundation model is +0.076. A matched factorial ablation shows that backbone fine-tuning and multitask learning are synergistic: multitask learning helps only once the backbone is fine-tuned (+0.08). This interaction is backbone- and task-dependent: it reproduces on Tox21 for two backbones but is absent for ChemBERTa-2 on PFAS, and label-free encoder diagnostics show this is not because ChemBERTa-2 represents PFAS poorly, so static representation quality does not predict postfine-tuning performance. PFAS-specific domain-adaptation pretraining gives no detectable improvement (\u0394 = \u2013 0.01 to \u2013 0.03). Mondrian-by-class conformal calibration improves rare-active-class coverage from 0.36 to 0.78, and an applicability-domain analysis shows only 26% of untested PFAS lie close enough to the training set for the model to be applied without caveat. The representation also transfers to soil-water partitioning (Kd) regression, where it holds at Spearman \u03c1 = 0.75 even on PFAS chemistries absent from training while the fingerprint baselines collapse.<\/jats:p>","DOI":"10.1021\/acs.jcim.6c02202","type":"journal-article","created":{"date-parts":[[2026,9,24]],"date-time":"2026-09-24T22:02:20Z","timestamp":1790287340000},"page":"12518-12533","source":"Crossref","is-referenced-by-count":0,"title":["A Chemical Foundation\nModel Benchmark for PFAS Bioactivity\nPrediction: Multi-Task Fine-Tuning Beats Fingerprint Baselines, PFAS\nDomain Pretraining Does Not Help"],"prefix":"10.1021","volume":"66","author":[{"ORCID":"https:\/\/orcid.org\/0009-0003-9341-0526","authenticated-orcid":true,"given":"Yue","family":"Zhu","sequence":"first","affiliation":[{"name":"Georgia Institute of Technology , , 225 North Avenue NW , , ,","place":["Atlanta, Georgia, United States, 30332"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Qingyang","family":"Liu","sequence":"additional","affiliation":[{"name":"Northeastern University , , 360 Huntington Avenue , , ,","place":["Boston, Massachusetts, United States, 02115"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"316","published-online":{"date-parts":[[2026,9,24]]},"reference":[{"key":"2026101208080073800_cit1","unstructured":"U.S. Environmental Protection Agency.\n          \n          PFASSTRUCT\nV6: Curated list of per- and polyfluoroalkyl substances; 2024. https:\/\/comptox.epa.gov\/dashboard\/chemical-lists\/PFASSTRUCT."},{"key":"2026101208080073800_cit2","doi-asserted-by":"publisher","first-page":"13970","DOI":"10.1021\/acs.est.9b04833","article-title":"Using Machine\nLearning to Classify Bioactivity for\n3486 Per- and Polyfluoroalkyl Substances (PFASs) from the OECD List","volume":"53","author":"Cheng","year":"2019","journal-title":"Environ. Sci. Technol."},{"key":"2026101208080073800_cit3","doi-asserted-by":"publisher","first-page":"2887","DOI":"10.1021\/jm9602928","article-title":"The Properties of Known Drugs. 1. Molecular Frameworks","volume":"39","author":"Bemis","year":"1996","journal-title":"J. Med. Chem."},{"key":"2026101208080073800_cit4","doi-asserted-by":"publisher","first-page":"1256","DOI":"10.1038\/s42256-022-00580-7","article-title":"Large-scale chemical language representations capture molecular structure\nand properties","volume":"4","author":"Ross","year":"2022","journal-title":"Nat. Mach. Intell."},{"key":"2026101208080073800_cit5","article-title":"Massively Multitask\nNetworks for Drug\nDiscovery","author":"Ramsundar","year":"2015","journal-title":"arXiv"},{"key":"2026101208080073800_cit6","doi-asserted-by":"publisher","first-page":"80","DOI":"10.3389\/fenvs.2015.00080","article-title":"DeepTox: Toxicity\nPrediction using Deep Learning","volume":"3","author":"Mayr","year":"2016","journal-title":"Front. Environ.\nSci."},{"key":"2026101208080073800_cit7","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2209.01712","article-title":"ChemBERTa-2: Towards Chemical Foundation\nModels","author":"Ahmad","year":"2022","journal-title":"arXiv"},{"key":"2026101208080073800_cit8","doi-asserted-by":"publisher","first-page":"7678","DOI":"10.1021\/acs.est.4c13284","article-title":"Modeling PFAS Sorption in Soils Using Machine Learning","volume":"59","author":"Fabregat-Palau","year":"2025","journal-title":"Environ. Sci. Technol."},{"key":"2026101208080073800_cit9","doi-asserted-by":"publisher","first-page":"488","DOI":"10.1021\/acs.est.3c06561","article-title":"Elucidating Key Characteristics of PFAS Binding to\nHuman Peroxisome Proliferator-Activated Receptor Alpha: An Explainable\nMachine Learning Approach","volume":"58","author":"Maeda","year":"2024","journal-title":"Environ. Sci. Technol."},{"key":"2026101208080073800_cit10","doi-asserted-by":"publisher","first-page":"590","DOI":"10.3390\/toxics13070590","article-title":"New QSAR Models\nto Predict Human\nTransthyretin Disruption by Per- and Polyfluoroalkyl Substances (PFAS):\nDevelopment and Application","volume":"13","author":"Evangelista","year":"2025","journal-title":"Toxics"},{"key":"2026101208080073800_cit11","doi-asserted-by":"publisher","first-page":"835","DOI":"10.1021\/acs.chemrestox.5c00424","article-title":"New Docking, Molecular\nDynamics, and QSAR Models to\nPredict Disruption of Human and Rat Transthyretin Function by Per-\nand Polyfluoroalkyl Substances (PFAS)","volume":"39","author":"Almeida","year":"2026","journal-title":"Chem.\nRes. Toxicol."},{"key":"2026101208080073800_cit12","article-title":"Uncovering\nHepatotoxicity of Per- and Polyfluoroalkyl Substances via Liver-Type\nFatty Acid Binding Protein Binding: A Graph Convolutional Network\nApproach","author":"Bondarenko","year":"2024","journal-title":"arXiv"},{"key":"2026101208080073800_cit13","doi-asserted-by":"publisher","first-page":"98","DOI":"10.3390\/toxics11020098","article-title":"A Machine Learning Model to Estimate Toxicokinetic\nHalf-Lives of Per- and Polyfluoro-Alkyl Substances (PFAS) in Multiple\nSpecies","volume":"11","author":"Dawson","year":"2023","journal-title":"Toxics"},{"key":"2026101208080073800_cit14","doi-asserted-by":"publisher","first-page":"5793","DOI":"10.1021\/acs.jcim.1c01204","article-title":"Uncertainty-Informed\nDeep Transfer Learning of Perfluoroalkyl and Polyfluoroalkyl Substance\nToxicity","volume":"61","author":"Feinstein","year":"2021","journal-title":"J. Chem. Inf. Model."},{"key":"2026101208080073800_cit15","doi-asserted-by":"publisher","first-page":"1017","DOI":"10.1021\/acs.estlett.2c00530","article-title":"Harnessing\nSemi-Supervised Machine\nLearning to Automatically Predict Bioactivities of Per- and Polyfluoroalkyl\nSubstances (PFASs)","volume":"10","author":"Kwon","year":"2023","journal-title":"Environ. Sci. Technol. Lett."},{"key":"2026101208080073800_cit16","doi-asserted-by":"publisher","first-page":"748","DOI":"10.1126\/science.ado6638","article-title":"Predictions\nof groundwater PFAS occurrence at drinking water supply depths in\nthe United States","volume":"386","author":"Tokranov","year":"2024","journal-title":"Science"},{"key":"2026101208080073800_cit17","doi-asserted-by":"publisher","first-page":"969","DOI":"10.1021\/acsestwater.3c00134","article-title":"Prediction\nof 35 Target Per- and\nPolyfluoroalkyl Substances (PFASs) in California Groundwater Using\nMultilabel Semisupervised Machine Learning","volume":"4","author":"Dong","year":"2024","journal-title":"ACS ES T Water"},{"key":"2026101208080073800_cit18","doi-asserted-by":"publisher","first-page":"109504","DOI":"10.1016\/j.envint.2025.109504","article-title":"Unveiling PFAS hazard\nin European surface waters using\nan interpretable machine-learning model","volume":"199","author":"Zhao","year":"2025","journal-title":"Environ.\nInt."},{"issue":"1","key":"2026101208080073800_cit19","doi-asserted-by":"publisher","first-page":"10918","DOI":"10.1038\/s41467-024-55320-9","article-title":"Elucidating governing factors of\nPFAS removal by polyamide membranes using machine learning and molecular\nsimulations","volume":"15","author":"Jeong","year":"2024","journal-title":"Nat. Commun."},{"key":"2026101208080073800_cit20","doi-asserted-by":"publisher","first-page":"176575","DOI":"10.1016\/j.scitotenv.2024.176575","article-title":"Elucidating\nper- and polyfluoroalkyl substances (PFASs) soil-water partitioning\nbehavior through explainable machine learning models","volume":"954","author":"Xie","year":"2024","journal-title":"Sci. Total Environ."},{"key":"2026101208080073800_cit21","doi-asserted-by":"publisher","first-page":"624","DOI":"10.1021\/acs.estlett.9b00476","article-title":"A Machine Learning\nApproach for Predicting Defluorination of Per- and Polyfluoroalkyl\nSubstances (PFAS) for Their Efficient Treatment and Removal","volume":"6","author":"Raza","year":"2019","journal-title":"Environ. Sci. Technol. Lett."},{"key":"2026101208080073800_cit22","doi-asserted-by":"publisher","first-page":"30743","DOI":"10.1021\/acsomega.1c04778","article-title":"Exploring\nChemical Reaction Space\nwith Reaction Difference Fingerprints and Parametric t-SNE","volume":"6","author":"Andronov","year":"2021","journal-title":"ACS Omega"},{"key":"2026101208080073800_cit23","doi-asserted-by":"publisher","first-page":"513","DOI":"10.1039\/C7SC02664A","article-title":"MoleculeNet: A Benchmark\nfor Molecular Machine Learning","volume":"9","author":"Wu","year":"2018","journal-title":"Chem. Sci."},{"key":"2026101208080073800_cit24","doi-asserted-by":"crossref","first-page":"8342","DOI":"10.18653\/v1\/2020.acl-main.740","article-title":"Don\u2019t Stop\nPretraining: Adapt Language Models to Domains and Tasks","volume-title":"Proceedings of the 58th Annual Meeting of the Association\nfor Computational Linguistics (ACL)","author":"Gururangan","year":"2020"},{"key":"2026101208080073800_cit25","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2207.06010","article-title":"Does GNN Pretraining Help Molecular\nRepresentation?","author":"Sun","year":"2022","journal-title":"arXiv"},{"key":"2026101208080073800_cit26","doi-asserted-by":"publisher","first-page":"722","DOI":"10.1021\/acs.chemrestox.6c00143","article-title":"Multitask Pretraining\nFramework for Improving Predictivity of Machine Learning Chemical\nBioactivity Models for Low-Data Endpoints","volume":"39","author":"Wichrowski","year":"2026","journal-title":"Chem.\nRes. Toxicol."},{"key":"2026101208080073800_cit27","first-page":"1263","article-title":"Neural Message\nPassing for Quantum Chemistry","volume-title":"Proceedings\nof the 34th International Conference on Machine Learning","author":"Gilmer","year":"2017"},{"key":"2026101208080073800_cit28","doi-asserted-by":"publisher","first-page":"176095","DOI":"10.1016\/j.scitotenv.2024.176095","article-title":"Molecular designing of potential\nenvironmentally friendly PFAS based\non deep learning and generative models","volume":"953","author":"Yang","year":"2024","journal-title":"Sci.\nTotal Environ."},{"key":"2026101208080073800_cit29","doi-asserted-by":"publisher","first-page":"9","DOI":"10.1021\/acs.jcim.3c01250","article-title":"Chemprop:\nA Machine Learning Package for Chemical Property Prediction","volume":"64","author":"Heid","year":"2024","journal-title":"J. Chem. Inf. Model."},{"key":"2026101208080073800_cit30","doi-asserted-by":"publisher","first-page":"109253","DOI":"10.1016\/j.compchemeng.2025.109253","article-title":"Exploring\nData Augmentation: Multi-Task Methods for Molecular Property Prediction","volume":"201","author":"Javaid","year":"2025","journal-title":"Comput. Chem. Eng."},{"key":"2026101208080073800_cit31","article-title":"On-line compression\nmodeling I: conformal prediction","volume-title":"Algorithmic\nLearning in a Random World","author":"Vovk","year":"2005"},{"key":"2026101208080073800_cit32","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2107.07511","article-title":"A\nGentle Introduction to Conformal Prediction and Distribution-Free\nUncertainty Quantification","author":"Angelopoulos","year":"2022","journal-title":"arXiv"},{"key":"2026101208080073800_cit33","volume-title":"Mondrian\nConfidence Machine","author":"Vovk","year":"2003"},{"key":"2026101208080073800_cit34","doi-asserted-by":"publisher","first-page":"1596","DOI":"10.1021\/ci5001168","article-title":"Introduction of Conformal\nPrediction in Predictive Modeling: A Transparent and Flexible Alternative\nto Applicability Domain Determination","volume":"54","author":"Norinder","year":"2014","journal-title":"J. Chem.\nInf. Model."},{"key":"2026101208080073800_cit35","doi-asserted-by":"publisher","first-page":"3370","DOI":"10.1021\/acs.jcim.9b00237","article-title":"Analyzing\nLearned Molecular Representations for Property\nPrediction","volume":"59","author":"Yang","year":"2019","journal-title":"J. Chem. Inf. Model."},{"issue":"8","key":"2026101208080073800_cit36","doi-asserted-by":"publisher","first-page":"913","DOI":"10.1111\/ecog.02881","article-title":"others Cross-validation\nstrategies for data with temporal, spatial, hierarchical, or phylogenetic\nstructure","volume":"40","author":"Roberts","year":"2017","journal-title":"Ecography"},{"key":"2026101208080073800_cit37","doi-asserted-by":"publisher","first-page":"747","DOI":"10.1021\/ci9803381","article-title":"Unsupervised\nData Base Clustering Based on Daylight\u2019s Fingerprint and Tanimoto\nSimilarity: A Fast and Automated Way to Cluster Small and Large Data\nSets","volume":"39","author":"Butina","year":"1999","journal-title":"J. Chem. Inf. Comput. Sci."},{"key":"2026101208080073800_cit38","doi-asserted-by":"publisher","first-page":"17671","DOI":"10.1021\/acs.est.3c00026","article-title":"Machine\nLearning in Environmental\nResearch: Common Pitfalls and Best Practices","volume":"57","author":"Zhu","year":"2023","journal-title":"Environ. Sci. Technol."},{"key":"2026101208080073800_cit39","doi-asserted-by":"publisher","first-page":"1225","DOI":"10.1021\/acs.chemrestox.6b00135","article-title":"ToxCast Chemical Landscape: Paving the Road to 21st Century Toxicology","volume":"29","author":"Richard","year":"2016","journal-title":"Chem. Res. Toxicol."},{"key":"2026101208080073800_cit40","unstructured":"U.S. Environmental Protection Agency.\n          \n          ToxCast &\nTox21 Data Spreadsheet (invitroDB v4.3); 2024. https:\/\/www.epa.gov\/comptox-tools\/exploring-toxcast-data."},{"key":"2026101208080073800_cit41","doi-asserted-by":"crossref","first-page":"785","DOI":"10.1145\/2939672.2939785","article-title":"XGBoost: A Scalable\nTree Boosting System","volume-title":"Proceedings of the\n22nd ACM SIGKDD International Conference on Knowledge Discovery and\nData Mining","author":"Chen","year":"2016"},{"key":"2026101208080073800_cit42","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.1905.12265","article-title":"Strategies for Pre-training Graph Neural Networks","author":"Hu","year":"2020","journal-title":"arXiv"},{"key":"2026101208080073800_cit43","unstructured":"MAPIE Contributors\nMAPIE: Model Agnostic Prediction Interval Estimator; 2024. https:\/\/github.com\/scikit-learn-contrib\/MAPIE."}],"container-title":["Journal of Chemical Information\nand Modeling"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/pubs.acs.org\/doi\/pdf\/10.1021\/acs.jcim.6c02202","content-type":"application\/pdf","content-version":"vor","intended-application":"unspecified"},{"URL":"https:\/\/pubs.acs.org\/jcisd8\/article-pdf\/66\/19\/12518\/68644918\/acs.jcim.6c02202.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"syndication"},{"URL":"https:\/\/pubs.acs.org\/jcisd8\/article-pdf\/66\/19\/12518\/68644918\/acs.jcim.6c02202.pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,10,12]],"date-time":"2026-10-12T12:26:54Z","timestamp":1791808014000},"score":1,"resource":{"primary":{"URL":"https:\/\/pubs.acs.org\/jcisd8\/article\/66\/19\/12518\/5438022\/A-Chemical-Foundation-Model-Benchmark-for-PFAS"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,9,24]]},"references-count":43,"journal-issue":{"issue":"19","published-online":{"date-parts":[[2026,9,24]]},"published-print":{"date-parts":[[2026,10,12]]}},"URL":"https:\/\/doi.org\/10.1021\/acs.jcim.6c02202","relation":{},"ISSN":["1549-9596","1549-960X"],"issn-type":[{"value":"1549-9596","type":"print"},{"value":"1549-960X","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,9,24]]}}}