Evaluating Molecular Representations for Predicting Cyclodextrin-PFAS Binding Energy with Machine Learning

Abstract

Per- and polyfluoroalkyl substances (PFAS) persist in water systems and resist conventional removal methods such as activated carbon, which shows reduced efficiency with short-chain PFAS and in the presence of dissolved organic matter. Cyclodextrin-based polymers (CDPs) have emerged as sustainable alternatives, with competitive and selective PFAS adsorption capabilities. These polymers consist of glucose-based cyclodextrin (CD) units that can form host–guest inclusion complexes with PFAS pollutants. However, these binding interactions are not fully understood or quantified. We conducted an evaluation of machine learning approaches to model these host–guest interactions, providing insights into predictive capabilities for later CDP design. This study systematically compares molecular representations (Mordred, ECFP, ChemBERTa, UniMol2, etc.) across several machine learning architectures to predict CD-PFAS binding energies. First, we generated molecular embeddings of 3459 experimental host–guest pairs in the OpenCycloDB data set and 63 external CD-PFAS pairs. We then compared these embeddings via AlignedUMAP visualizations and nearest neighbor analyses. Next, we trained and evaluated predictive models using these embeddings on the OpenCycloDB data set, exploring the effectiveness of transfer learning and finetuning techniques. We finally tested model generalizability on two external experimental CD-PFAS binding data sets. All embeddings captured relevant chemical features, where UniMol2 differed most from other methods in embedding space analysis. Predictive models performed variably based on embedding choice and architecture, with the best-performing combination achieving moderate accuracy on the OpenCycloDB data set. Embeddings pretrained on large molecular data sets and finetuning the ChemBERTa embeddings both showed predictive improvements. However, external validation revealed limited generalizability to CD-PFAS complexes, highlighting domain shift challenges. Notably, leave-one-out cross-validation on the external PFAS-specific data indicated that training on in-domain data improved predictive performance at the cost of generalizability. This work demonstrates that molecular representation choice is critical for small-data host–guest binding prediction. However, domain shift between general CD data and specialized CD–PFAS applications remains a fundamental challenge, for which transfer learning and finetuning may offer potential solutions for future data-driven pipelines for CDP design and sustainable PFAS removal.