Biology subjects

Dumontet, L.

Publications and source records attributed to Dumontet, L..

3 recordsLinked to original sources

bioRxiv2025

Plastic-hydrolytic enzyme classification using explainable deep learning

The rapid accumulation of plastic waste has emerged as a critical environmental threat, driving the need for scalable and effective biodegradation solutions. 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, PEPIC, to classify nine types of PDEs directly from protein sequences. Using a curated dataset of experimentally validated enzymes and an expanded homologous dataset, we built an explainable deep learning model based on convolutional neural networks (PEPIC) for plastic-degrading enzyme prediction. We benchmarked PEPICs performance against state-of-the-art approaches. First, PEPIC demonstrated statistically significant improvements in predictive performance compared to state-of-the-art methods. Second, PEPIC calculates contribution scores for each amino acid in the protein sequence, indicating their influence on the predictions. The model interpretation revealed that regions highlighted by high contribution scores matched conserved catalytic triads and substrate-binding clefts across PET-, PCL-, and PLA-degrading enzymes. Furthermore, structural modeling confirmed the trustworthiness of PEPICs predictions. Finally, PEPIC predicted an uncurated enzyme as a PET-degrading enzyme, which was biologically validated to hydrolyze bis(2-hydroxyethyl) terephthalate (BHET). These findings demonstrate that PEPIC provides accurate and trustworthy predictions of PDEs, facilitating the discovery of novel enzymes and supporting the development of sustainable plastic biodegradation technologies.

bioinformatics↗

bioRxiv2025

HIT-EC: Trustworthy prediction of enzyme commission numbers using a hierarchical interpretable transformer

Accurate and trustworthy prediction of Enzyme Commission (EC) numbers is critical for understanding enzyme functions and their roles in biological processes. Despite the success of recently proposed deep learning-based models, there remain limitations, such as low performance in underrepresented EC numbers, lack of learning strategy with incomplete annotations, and limited interpretability. To address these challenges, we propose a novel hierarchical interpretable transformer model, HIT-EC, for trustworthy EC number prediction. HIT-EC employs a four-level transformer architecture that aligns with the hierarchical structure of EC numbers, and leverages both local and global dependencies within protein sequences for this multi-label classification task. We also propose a novel learning strategy to handle incomplete EC numbers. HIT-EC, as an evidential deep learning model, produces trustworthy predictions by providing domain-specific evidence through a biologically meaningful interpretation scheme. The predictive performance of HIT-EC was assessed by multiple experiments: a cross-validation with a large dataset, a validation with external data, and a species-based performance evaluation. HIT-EC showed statistically significant improvement in predictive performance when compared to the current state-of-the-art benchmark models. HIT-ECs robust interpretability was further validated by identifying well-known conserved motifs and functional regions in the CYP106A2 enzyme family. HIT-EC would be a robust, interpretable, and reliable solution for EC number prediction, with significant implications for enzymology, drug discovery, and metabolic engineering. The open-source code is publicly available at: https://github.com/datax-lab/HIT-EC.

bioinformatics↗

bioRxiv2025

Interpretable Kolmogorov-Arnold Networks for Enzyme Commission Number Prediction

Accurate prediction of enzyme commission (EC) numbers remains a significant challenge in bioinformatics, limiting our ability to fully understand enzyme functions and their roles in biological processes. This paper presents the integration and evaluation of the next paradigm of deep learning architecture, named Kolmogorov-Arnold network (KAN), in state-of-the-art models for predicting EC numbers. KAN modules are incorporated into current state-of-the-art models to assess their impact on predictive performance. Additionally, we introduce a novel interpretation method, specifically designed for KANs, to identify relevant input features for a given prediction, addressing a current limitation in KANs. Our evaluation demonstrates that the integration of KANs significantly enhances predictive performance compared to the state-of-the-art deep learning models, with up to a 4.35% increase in micro-averaged F1 score and a 4.1% increase in macro-averaged F1 score. Moreover, our novel interpretation method not only enhances the predictions trustworthiness but also facilitates the discovery of motif sites within enzyme sequences. This innovative approach provides deeper insights into enzyme functionality and highlights potential new targets for research. The results underscore KANs effectiveness in improving enzymatic classification and advancing our understanding of enzyme structures and functions. The open-source code is publicly available at: https://github.com/datax-lab/kan_ecnumber.

bioinformatics↗