PROTEAN: An Explainable AI Pipeline for Protein Classification in Support of Precision Medicine

Area of Science:

• Biochemistry and Bioinformatics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Soluble N-ethyl-maleimide-sensitive factor Attachment Protein Receptors (SNARE) proteins are crucial for intracellular trafficking and disease.
• Differentiating SNARE from NONSNARE proteins is difficult due to sequence and structural similarities.
• Traditional AI models lack transparency, limiting their use in sensitive biomedical fields.

Purpose of the Study:

• To develop PROTEAN, a methodology combining machine learning and explainable AI (XAI) for accurate and interpretable protein classification.
• To address the challenge of distinguishing SNARE from NONSNARE proteins.
• To enhance the trustworthiness of AI in biomedical applications.

Main Methods:

• A three-phase pipeline: data preprocessing on a balanced dataset (D128), training and evaluation of classifiers (SVMs, NNs), and interpretation using SHAP and LIME XAI models.
• Utilized a balanced dataset of 128 SNARE and NONSNARE protein sequences.
• Employed SHAP and LIME for model interpretation to identify key protein descriptors.

Main Results:

• The Medium Gaussian Support Vector Machine (SVM) classifier achieved 92.1% accuracy, 94.8% sensitivity, and 89.5% specificity on a 458-protein test set.
• SHAP and LIME provided consistent, biologically relevant explanations for classification decisions.
• Identified amino acid composition and sequence order as critical features for protein classification.

Conclusions:

• PROTEAN demonstrates that integrating interpretable AI with balanced datasets improves protein classification performance and transparency.
• The findings are vital for developing trustworthy AI systems in biomedicine.
• Offers new tools for disease mechanism analysis, biomarker discovery, and personalized medicine.