In Silico Prediction of PFAS Retention in Humans: Integrating QSAR Models for Albumin Binding and Half-Life Estimation

In Silico Prediction of PFAS Retention in Humans: Integrating QSAR Models for Albumin Binding and Half-Life Estimation Vasileios Minadakis 1Periklis Tsiros 1Paraskevi Papakyriakopoulou 1 Haralambos Sarimveis 1 1National Technical University of Athens, School of Chemical engineering, 9 Heroon Polytechniou St, 15780, Athens, Greece P25-12 Introduction Perand Polyfluoroalkyl substances (PFAS) are synthetic chemicals that pose environmental concerns due to their long half-lives and strong protein binding properties that have been reported. A decisive role is played by their binding to human serum albumin (HSA), which is the most abundant protein in human blood. This study focuses on developing in silico tools that predict both the binding affinity of PFAS to HSA and their half-lives in humans. We propose a linear computational pipeline consisting of two QSAR models: the first predicts PFAS-albumin association constants, while the second uses these predictions along with additional molecular features to estimate half-lives, enabling identification of structural characteristics that influence retention time in organisms. Datasets Binding Affinities (log KA) data The data for the albumin binding model were retrieved from the study of Chen et al. [1], where the authors experimentally evaluated the binding affinities of 58 PFAS congeners through equilibrium dialysis, which is considered as gold standard. Specifically, they evaluated dissociation constant KD, which is related to association constant by KD= 1/KA.The estimated values of KAwere transformed into log KAto reduce the range and the distances of the values. The digital representation of the congeners in the datasets was accomplished through their canonical SMILES, followed by the estimation of RDKit descriptors and Extended Connectivity Fingerprints (ECFP). RDKit descriptors provide a wide range of molecular features that quantify physicochemical and structural properties of the molecules. ECFP generate a binary vector that encompasses information about substructures of the molecules. The final number of descriptors was more than two thousand. The large number of features and the existing overlap between the two sets of descriptors indicated the need for feature elimination. Consequently, the features of the dataset were filtered by variance and those with variance equal to 0 were removed. Next, we eliminated the highly correlated features using a threshold value of 0.95. Finally, random splitting was applied to the dataset, to retain 20% as test data for model evaluation. Half-life (t1/2) data The half-life dataset contains data from 16 studies, including a total of 30 PFAS congeners. Some of these studies overlapped, meaning that they provide the half-lives of common PFAS congeners. This introduced variability (aleatoric uncertainty) in the half-life values. The first step was to identify and remove the outliers per PFAS congener by applying the Interquantile Range (IQR) method and then To address this, the mean and standard deviation of half-life was computed for each PFAS, and new samples were drawn from these distributions. Additionally, RDKit descriptors and ECFP were calculated, as in the case of binding affinities. By applying the same feature elimination strategy, we managed to reduce considerably the number of features. The final half-life dataset was enriched with the predictions of the Albumin binding affinity QSAR model. Finally, the PFAS congeners were split in train and test datasets by applying the Kennard-Stone algorithm on the most Principal Components of the generated physicochemical descriptors, ensuring feature space similarity between them, which strengthens the model’s validity by ensuring evaluation on unseen congeners. Methodology PFAS-Albumin Binding Affinity QSAR - Random Forest PFAS–HSA binding affinities were modeled using a Random Forest regression with iterative forward feature selection. Starting from an empty set, features were added one by one based on 5-fold crossvalidation R2. The procedure stopped when no further improvement was observed. Model performance was then evaluated on unseen PFAS congeners. Half-life QSAR - Heteroscedastic Gaussian Process Regressor This modeling task was carried out by applying an alternative version of the Gaussian Process Regressor (GPR), that extends for data with heteroscedastic noise. This model is called Heteroscedastic Gaussian Process Regressor (HGPR) and was introduced by Ozbayram et al.[2] to extend the basic concept of the Gaussian process to datasets with heteroscedastic aleatoric uncertainty. This extension suggests modifying the structure of the kernel by exploiting polynomial regression for noise modeling. Generally, in the case of data with homescedastic noise, the posterior predictive distribution of f∗as well as ¯ f∗ and Cov(f∗), which are the model’s predictions for mean and covariance matrix for unseen data, for given Xand ytrain data and X∗unseen data, are estimated as: f∗|X,y,X∗∼ N(¯ f∗, Cov(f∗)) ¯ f∗=K(X∗,X)>(K(X,X) + diag(σ2(X)))−1y Cov(f∗) = K(X∗,X∗)−K(X∗,X)>(K(X,X) + diag(σ2(X)))−1K(X∗,X) + diag(σ2(X∗)) (1) where X∈RN×pand X∗∈RM×p, with Nand Mbeing the number of rows in train and test datasets respectively and pthe number of features, while K(X,X)∈RN×N,K(X∗,X∗)∈RM×M and K(X∗,X)∈RM×Nare the (cross-) covariance matrices estimated by a kernel function k(·,·) and σ2is the noise that occurs from the aleatoric uncertainty of the data. The Automatic Relevance Determination Radial Basis Function (ARD-RBF) was used as the kernel function. The contribution of Ozbayram et al. [2] is the modification of the following term included in Cov(f∗), by extending it with a term for polynomial regression of heteroscedastic noise: σ2(X) =  kal exp   d X i=1 p X j=1 θijxi j   2 (2) where ddenotes the polynomial degree, pis the number of features, θij are the polynomial coefficients and σ(X)2=k2 al, where kal is optimized during the training process. The previously described feature selection strategy was applied in this case study as well. Results Table 1 summarizes model performance, showing consistent cross-validation and test scores, confirming generalizability. Figure 1 visualizes predictive ability on unseen data, while figure 2 presents selected features and SHAP values, highlighting the strong link between LogKAand half-lives (Figure 2b). Finally, figure 3 visualises the chemical fragments of each ECFP descriptor included in each model. R2LogKAQSAR Half-life QSAR 5-fold CV 0.645 0.715 Test 0.641 0.679 Table 1. Performance metrics for PFAS models Predictions on Test Datasets (a) LogKa QSAR predictions on test dataset (b) Half-life QSAR predictions on test dataset Figure 1. Predicted against observed values for both models on test sets. Selected Features - SHAP Values (a) SHAP values for LogKa QSAR (b) SHAP values for Half-life QSAR Figure 2. Comparison of SHAP values for LogKa and Halflife Visualisation of Selected ECFP descriptors (a) ECFP descriptors of LogKa QSAR. (b) ECFP descriptors of Half-life QSAR. Figure 3. Visualisation of ECFP descriptors selected for each model. Jaqpot: Model Deployment QSAR models are hosted on Jaqpot, an open-source platform for online model deployment. Users can run the models directly through the web interface without coding: the binding affinity model is available at https://app.jaqpot.org/dashboard/models/2028, while the half-life model is available at https://app.jaqpot.org/dashboard/models/2179. Acknowledgments This work has been financially supported by the SCENARIOS project (Grant Agreement 101037509) which has been funded by the European Commission under the Horizon 2020 Programme. References [1] R. Chen et al. “Deriving Membrane–Water and Protein–Water Partition Coefficients from In Vitro Experiments for Perand Polyfluoroalkyl Substances (PFAS)”. In: Environmental Science & Technology 59 (2025) [2] O. Ozbayram et al. “Heteroscedastic Gaussian Process Regression for material structure–property relationship modeling”. In: Computer Methods in Applied Mechanics and Engineering 431 (2024) www.eurotox2025.com EUROTOX Congress 2025 - Athens [email protected]