Protein homology detection with biologically inspired features and interpretable statistical models

Area of Science:

• Computational biology
• Bioinformatics
• Protein science

Background:

• Computational protein classification methods like string kernels and Fisher-SVM are successful but lack biological interpretability.
• Understanding protein superfamily relationships is crucial for biological research.

Purpose of the Study:

• To develop a biologically interpretable computational method for protein superfamily detection.
• To identify a small subset of critical features (positions and residues) in protein sequences.

Main Methods:

• Development of a biologically motivated feature set.
• Application of a sparse classifier utilizing the feature set.
• Evaluation on a benchmark dataset for protein superfamily detection.

Main Results:

• The proposed method achieves performance comparable to state-of-the-art techniques.
• The identified sparse critical features align with known biological findings.
• The models provide interpretable insights into protein sequence determinants.

Conclusions:

• Biologically motivated features and sparse classification offer an interpretable alternative for protein superfamily detection.
• This approach enhances understanding of protein sequence-function relationships.
• The method demonstrates the potential for discovering novel biological insights through computational analysis.