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

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Materials Science

Background:

• Per- and polyfluoroalkyl substances (PFAS) are persistent pollutants resistant to conventional water treatment.
• Cyclodextrin-based polymers (CDPs) offer a sustainable alternative for PFAS adsorption via host-guest complexation.
• Understanding and quantifying cyclodextrin-PFAS binding interactions is crucial for optimizing CDP design.

Purpose of the Study:

• To evaluate machine learning (ML) approaches for modeling host-guest interactions between cyclodextrins (CDs) and PFAS.
• To compare various molecular representations and ML architectures for predicting CD-PFAS binding energies.
• To assess the generalizability of ML models and the impact of transfer learning and finetuning.

Main Methods:

• Generated molecular embeddings for experimental host-guest pairs using representations like Mordred, ECFP, ChemBERTa, and UniMol2.
• Compared embeddings using AlignedUMAP visualizations and nearest neighbor analyses.
• Trained and evaluated ML models on the OpenCycloDB dataset, incorporating transfer learning and finetuning, and tested on external CD-PFAS datasets.

Main Results:

• All molecular embeddings captured relevant chemical features, with UniMol2 showing distinct embedding space characteristics.
• ML model performance varied by embedding and architecture, achieving moderate accuracy on the primary dataset.
• Pretraining embeddings and finetuning ChemBERTa embeddings improved predictive performance.
• External validation revealed limited generalizability due to domain shift challenges between general CD data and specific CD-PFAS applications.

Conclusions:

• Molecular representation is critical for accurate host-guest binding prediction in small-data scenarios.
• Domain shift presents a significant challenge for applying ML models to specialized CDP-PFAS applications.
• Transfer learning and finetuning show potential for addressing domain shift in data-driven CDP design for sustainable PFAS removal.