Enzyme Classification Dataset with 134,153 Validated Sequences

134,153 validated enzyme sequences form the basis for a machine learning framework for enzyme function prediction. The dataset was contributed by author Huanghui Xia and was last updated in April 2026. It supports the EC-Design framework, which applies PCA and Fisher Score-based feature selection to these sequences.

Use Cases

• Predict enzyme commission (EC) numbers from dimensionality-reduced sequence features using k-NN or other classifiers.
• Analyze feature importance for biological interpretation, focusing on discriminative dipeptide patterns like those containing asparagine and glycine.
• Investigate class imbalance trade-offs by modeling high-precision majority classes (EC1-EC3) versus high-recall minority classes (EC4-EC7).
• Benchmark model generalization and discriminative power using reported metrics like macro-F1 (0.6859) and mean AUC (0.937).

Strengths

• Large-scale dataset containing 134,153 validated enzyme sequences.
• Framework achieved a reported top accuracy of 74.59% and robust generalization (74.37 ± 0.49%).
• Feature analysis provides biologically interpretable insights, linking dipeptide patterns to catalytic motifs.

Limitations

• Extreme class imbalance is noted between majority (EC1-EC3) and minority (EC4-EC7) enzyme classes.
• The dataset's specific feature columns and raw sequence data are not described in the provided input.
• License is CC-BY-NC-4.0, prohibiting commercial use.

Provenance

Source

figshare, author Huanghui Xia.

Collection Method

Machine learning framework (EC-Design) applied to validated enzyme sequences, utilizing PCA and Fisher Score-based feature selection.

Time Range

null

Freshness

Last updated April 2026.

Geography

null