High-throughput prediction of PFAS binding affinities with human liver-fatty acid binding protein using machine learning and QSAR model

The binding affinity with human liver-fatty acid binding protein (K_{d FABP}) is a key parameter to characterize the accumulation potential of per- and polyfluoroalkyl substances (PFAS) in animal liver. However, the vast number of PFAS, combined with the limited available commercial standards, makes it a major challenge to measure their K_{d FABP}. In this study, the K_{d FABP} of 44 PFAS standards, and 72 PFAS extracted and semi-quantified in environmental samples by suspect screening analysis were measured using ultrafiltration methods. Several machine learning regression algorithms were developed to predict the K_{d FABP}, and extreme gradient boosting regression exhibited the best performance. Of note, intrinsic molecular descriptors, such as AATS0d, AATS8pe, VR1_A, ATSC1d, and AATSC2Z were found to be the primary factors to affect the binding affinities. The optimized model was then applied to predict the K_{d FABP} values of 9117 PFAS listed by U.S. EPA. DTXSID40896722, which features perfluorinated branches connected through sulfonyl linkages, exhibits the lowest K_{d FABP} value. Additionally, by combining the predicted K_{d FABP} of 76,216 artificial-intelligence-generated PFAS, it was found that chemical fragments containing carbon-fluorine and ether moieties are conducive for the binding. This study holds significant importance in de novo design of environmentally friendly PFAS.