QSAR Model Development to predict PPARα activation by PFAS using human in vitro data

German Federal Institute for Risk Assessment [email protected] | bfr.bund.de/en Follow us QSAR Model Development to predict PPARαactivation by PFAS using human in vitro data Wiebke Alker1, Periklis Tsiros2, Thorsten Buhrke1, Haralambos Sarimveis2, Albert Braeuning1 Introduction Results Funding: This project has received funding from the H2020 programm under Grant Agreement No. 101037509 Unpublished data, ©2024, Wiebke Alker, German Federal Institute for Risk Assessment, All rights reserved. 1German Federal Institute for Risk Assessment, 2National Technical University of Athens Perand polyfluoroalkyl substances (PFAS) are man-made chemicals, being used in almost all industry branches and many consumer products. The use of many legacy PFAS is being reduced due to their high persistence and hazardous effects. Therefore, novel PFAS are increasingly used for numerous industrial applications, whereas at the same time there is no or only few data on their toxicity. Due to the large number of PFAS, it is impossible to perform hazard characterization on all PFAS based on experimental data. Here in silico models are a helpful tool to obtain hazard insights and to prioritize experiments accordingly. The nuclear receptor PPARαplays an essential role in lipid metabolism and is being activated by fatty acids. However, it has also been shown to be activated by many PFAS. Contact: [email protected] Tested PFAS •Sulfonic PFAS have a lower potential to induce PPARαthan carbocylic PFAS •PFECA demonstrate the highest potential in PPAR-α induction Experiments on PPARαinduction with 33 PFAS QSAR Model Development Model Screening of 10,000 PFAS congeners Abbrevations: BMD: Benchmarkdose, HFPO-TA: 2,3,3,3-Tetrafluoro-2-(1,1,2,3,3,3-hexafluoro-2-(perfluoropropoxy)propoxy)propanoic acid, PFAS: Perand polyfluoroalky substances, PFCA: perfluoroalkyl carboxylic acids, PFECA: perfluoroalkyl (poly-)ether carboxylic acids, PFESA: perfluoroalkyl (poly-)ether sulfonic acids, PFOA: Perfluorooctanoic acid, PFO2HpA: Perfluoro-3,6-dioxaheptanoic acid, PFSA: perfluoroalkyl sulfonic acids, Figure 2: HEK293T cells were cotransfected with plasmids pGAL4- (UAS)5-TK-LUC, pGAL4-hPPARαLBD and pcDNA3-Rluc for 5h. Cells were then incubated with PFAS for 24h before luciferase activity was measured. Values were normalized to Renilla reniformis luciferase activity and compared to untreated cells. Subsequently Benchmark dose (BMD) modelling was performed with a Benchmark response (BMR) of 50%. Conclusions •The developed QSAR model can be used to prioritize in vitro experiments with PFAS on PPARαinduction •GA selection was successful in dimensionality reduction and important feature recognition •A C-Atom bonded to both a hydroxyl and an ether group is predicted to increase the potential of a PFAS to induce PPARα, resulting in a low BMD •In total 33 PFAS •Legacy & common: PFCA, PFSA •Poor-data: PFECA, PFESA (linear/ branched, mono-/ polyether) •Genetic algorithm (GA) selection produced a model with an R² score of 0.86 in 5-fold cross-validation •Seven descriptors were selected, five representing a mixture of topological, electronic, and geometric molecular properties and two fingerprints encoding the presence of carbon atoms near an ether group. •1048 PFAS ( ca. 10%) of the screened PFAS are within the applicability domain •Reflecting patterns observed in the training data, the most potent PFAS are predicted to be those that have the fingerprint “Ether-Bit” ( C-Atom bonded to a hydroxyl and an ether group in molecular structure) and exhibit high ATSC5dv values Figure 4: 10,000 compounds from the STRUCTV4 EPA list were screened to predict their BMD for a PPARαinduction at a BMR of 50%. Two applicabilitys domain methods, i.e. bounding box, and leverage, were combined to create a more robust applicability domain definition. Only compounds that were considered “in domain” by both applicability domain definitions were considered in the applicability domain. Here, predicted BMD values are plotted against the most influential feature ATSC5dv. Figure 3: A QSAR model was developed using the results of the BMD analysis and computational descriptors from python packages ‘Mordred’ and ‘RDKit’. Due to the limited number of congeners, all were included in model development. A simple linear model was selected to avoid overfitting and increase interpretability. Feature selection was performed using GA selection, optimizing for 5-fold cross validation R2. PFOA Nafion by product 1 Figure 1: Examples of experimentally tested PFAS, highlighted is the fingerprint „Ether-Bit“, a C-Atom bonded to a hydroxyl and an ether group. It‘s presence in the molecular structure is predicted to result in a low BMD. HFPO-TA PFO2HpA yes no Fingerprint „Ether-Bit“ in molecular structure