EnzyDiff: Sequence-based classification of mutation-induced enzyme activity direction using latent diffusion denoising

Abstract

Objective Prediction of mutation-induced enzyme activity changes is important for protein engineering, drug response analysis, and precision medicine. Existing methods often depend on structural information or complex external pipelines, limiting reproducibility and practical use. Methods We developed EnzyDiff, a lightweight diffusion-based framework that encodes mutation-induced activity perturbations from enzyme sequences before and after mutation, together with mutation identity and position. The model was evaluated on the P450 deep mutational scanning dataset using five-fold cross-validation, with additional analyses of robustness, representation visualization, and case-based occlusion sensitivity. After preprocessing, 3,289 labeled mutation entries were retained, including 240 Increased samples and 3,049 Decreased samples. Results In five-fold cross-validation, EnzyDiff achieved an accuracy of 0.933, F1-score of 0.497, AUROC of 0.803, and PR-AUC of 0.495 on the P450 dataset. It remained stable under reversed mutation settings, showed clear class separation in learned representations, and highlighted mutation-relevant positions in interpretability analyses. However, the moderate F1-score also indicates that minority-class prediction remains challenging under strong class imbalance. Conclusions EnzyDiff is a lightweight, reproducible, and interpretable framework for enzyme activity direction classification. It offers a practical sequence-based solution for functional mutation assessment and downstream biological analysis.