Amino Acid Encoding Methods for Protein Sequences: A Comprehensive Review and Assessment

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in structural biology

Background:

• Amino acid encoding is a foundational step for machine learning-based protein structure and function prediction.
• Existing encoding methods have not received sufficient attention or comprehensive review.
• Effective encoding is vital for both residue-level and sequence-level protein property predictions.

Purpose of the Study:

• To systematically classify and comprehensively review various amino acid encoding methods.
• To assess the performance of different encoding strategies on benchmark tasks.
• To identify promising encoding approaches for future protein prediction research.

Main Methods:

• Categorization of encoding methods into five groups: binary, physicochemical, evolution-based, structure-based, and machine-learning.
• Selection and comparison of 16 representative encoding methods.
• Evaluation using large-scale benchmark datasets for protein secondary structure prediction and protein fold recognition.

Main Results:

• Evolution-based position-dependent scoring matrix (PSSM) encoding demonstrated superior performance.
• Structure-based and machine-learning encoding methods, particularly neural network-based distributed representations, show significant potential.
• Comparative analysis highlights the strengths and weaknesses of different encoding categories.

Conclusions:

• Amino acid encoding significantly impacts the success of protein prediction models.
• PSSM is currently the top-performing encoding method for these tasks.
• Emerging machine learning and structure-based methods offer exciting avenues for future advancements in protein bioinformatics.