A theoretical analysis of PhysioChem-K-mer features for protein classification using controlled synthetic benchmarks

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Science

Background:

• Standard k-mer methods represent amino acids as categorical tokens, neglecting their inherent physicochemical properties.
• Integrating physicochemical properties into bioinformatics tasks is underexplored as a systematic alternative to k-mer counting.

Purpose of the Study:

• To introduce PhysioChem-K-mer, a framework for transforming protein sequences into physicochemical property-based feature spaces.
• To test the hypothesis that property-based representations capture functional constraints more effectively than traditional amino acid-based methods.

Main Methods:

• Developed the PhysioChem-K-mer framework to generate feature spaces from protein sequences based on physicochemical properties.
• Created a controlled benchmark of 1500 synthetic sequences across 10 protein families, excluding evolutionary patterns.
• Evaluated the framework on real UniProt/Swiss-Prot data (11,620 sequences, 10 families) for generalizability.

Main Results:

• Hydropathy-based PhysioChem-K-mer achieved 81.33% accuracy on synthetic data, a 44.33% absolute gain over standard 3-mer methods (37.00%).
• On real data, PhysioChem-Hydropathy achieved 64.63% accuracy, a 47.68% absolute gain over the 3-mer baseline (16.95%).
• The framework reduced features by 73.9% and training time by 81.6% on real data.

Conclusions:

• Physicochemical properties offer a vital source of information for protein classification, outperforming traditional k-mer methods.
• PhysioChem-K-mer provides an interpretable and computationally efficient alternative for protein sequence analysis.
• The framework demonstrates practical generalizability across both synthetic and real-world biological data.