Automated enzyme classification is hindered by high-dimensional feature spaces and extreme class imbalance. In this study, we introduce EC-Design, a machine learning framework that utilizes principal component analysis (PCA) and Fisher Score-based feature selection on 134,153 validated sequences. Challenging the perceived superiority of ensemble methods, k-nearest neighbors (k-NN) achieved a top accuracy of 74.59% and a macro-F1 of 0.6859, significantly outperforming six comparative machine learning methods, including random forest, ensemble bagging, linear discriminant analysis, logistic regression, support vector machine, and multilayer perceptron. The model demonstrated robust generalization (74.37 ± 0.49%) and excellent discriminative power (mean AUC = 0.937). Performance analysis revealed a distinct trade-off: majority classes (EC1-EC3) exhibited high precision (>0.83), while minority classes (EC4-EC7) achieved high recall (>0.83). Feature importance analysis revealed that dipeptide patterns containing asparagine and glycine serve as key discriminative features, which aligns with established catalytic motifs and provides biologically interpretable insights into enzyme function. The EC-Design framework establishes instance-based learning as an efficient and accurate approach for large-scale enzyme annotation, offering a transparent alternative to complex ensemble models.
Xia et al. (2026) studied this question.