Interpretable convolutional neural networks for sequence-based classification and discovery of plastic-degrading...

The rapid accumulation of plastic waste has emerged as a critical environmental threat, driving the need for scalable and effective biodegradation solutions. Major plastic types, including biopolyesters, aliphatic polyesters, and aromatic polyesters, are widely used in industrial and consumer products and exhibit distinct chemical structures and degradability profiles, posing challenges for systematic enzyme classification. Hydrolytic plastic-degrading enzymes (PDEs) offer a promising solution, yet their functional classification remains limited by insufficient annotations and enzymatic diversity. In this study, we present an explainable deep learning framework, plastic-degrading enzyme prediction via interpretable CNN (PEPIC), to classify enzymes across nine plastic substrate types directly from protein sequences. Using a curated data set of experimentally validated plastic-degrading enzymes (181 sequences) and an expanded homologous data set generated via sequence similarity search (~5,900 sequences), we benchmarked PEPIC against state-of-the-art approaches and evaluated both predictive performance and interpretability. PEPIC demonstrated statistically significant improvements in F1-score compared to state-of-the-art methods. PEPIC calculated contribution scores at the amino acid level, indicating how individual residues influence the predictions. The model interpretation revealed that regions with high contribution scores aligned with key catalytic residues. Homology-based structural modeling demonstrated that residues with high contribution scores mapped to known catalytic and substrate-binding regions of plastic-degrading enzymes and reflected structural differences across plastic classes, supporting the biological relevance of PEPIC's predictions. PEPIC identified an uncurated enzyme as a potential PET-degrading candidate. This work provides an interpretable framework for plastic-degrading enzyme discovery, which can accelerate biotechnological solutions for plastic waste management and support data-driven strategies toward sustainable environmental remediation.IMPORTANCEWe propose an explainable deep learning-based approach, named PEPIC, that can effectively classify enzymes across nine plastic substrate categories relevant to hydrolytic PDE activity and provide trustworthy predictions by identifying active and binding sites that align with prior biological knowledge. PEPIC is the first study that demonstrated the high potential of deep learning-based approaches for plastic-degrading enzymes prediction using large data sets. PEPIC not only significantly improved predictive performance compared to the current state-of-the-art models but also provided the trustworthiness of the prediction. PEPIC was thoroughly assessed by intensive and comprehensive experimental settings, and PEPIC enhances the model interpretation for trustworthy predictions and potential new biological knowledge discovery. This work offers scientific advances in accelerating the discovery of plastic-degrading enzymes, contributing to sustainable plastic waste management using a novel AI technique.