Prediction of enzyme function using an interpretable optimized ensemble learning framework.

Area of Science:

• Bioinformatics
• Computational Biology
• Enzyme Catalysis

Background:

• Accurate enzyme function prediction is crucial for biological research, especially for novel sequences.
• Existing machine learning tools face challenges in feature extraction, interpretability, and generalization.

Purpose of the Study:

• To develop an interpretable machine learning method for accurate enzyme function prediction.
• To address limitations of current ML tools in feature extraction and interpretability.

Main Methods:

• Constructed a dedicated enzyme function dataset.
• Developed SOLVE (Soft-Voting Optimized Learning for Versatile Enzymes), an ensemble ML model using random forest, LightGBM, and decision trees.
• Employed tokenized subsequences from primary protein sequences and Shapley analyses for interpretability.

Main Results:

• SOLVE achieved high prediction accuracy, distinguishing enzymes from non-enzymes and assigning Enzyme Commission (EC) numbers.
• The method effectively mitigated class imbalance using focal loss penalty.
• Identified functional motifs at catalytic and allosteric sites, enhancing interpretability.

Conclusions:

• SOLVE offers a streamlined, accurate, and interpretable approach for high-throughput enzyme function prediction using only primary sequence data.
• Outperforms existing tools on independent datasets, showing promise for biological research and therapeutic drug design.