Optimizing Model Performance and Interpretability: Application to Biological Data Classification.

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Achieving high performance accuracy and interpretability simultaneously is a significant challenge in biological data classification.
• Transcriptomic data classification requires methods that are both precise and understandable.

Purpose of the Study:

• To develop a novel framework for transcriptomic-data-based classification that optimizes both performance accuracy and result interpretability.
• To select features, models, and a meta-voting classifier that enhance classification outcomes and biological insight.

Main Methods:

• A four-step feature selection process including metabolic pathway identification, principal component analysis, minimal gene set selection, and adversarial sample filtering.
• Utilizing adversarial samples for optimal classification model selection and constructing a meta-voting classifier.

Main Results:

• The framework achieved comparable prediction performance to full-gene models in binary classification (F1-score differences: -5% to 5%).
• In ternary classification, the framework showed improved performance (F1-score differences: -2% to 12%) while maintaining excellent interpretability.
• Selected feature genes provided clear biological insights.

Conclusions:

• The developed framework effectively integrates feature selection, adversarial sample handling, and model optimization for biological data classification.
• This approach offers a valuable tool for computational biology, balancing predictive accuracy with high interpretability.