Entropy and Variability: A Second Opinion by Deep Learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Analysis of amino acid distributions in protein alignments is crucial for human genetics, protein engineering, and drug design.
• Common measures include variability, average hydrophobicity, and Shannon entropy.
• Entropy-variability analysis reduces residue distributions to two key numbers: Shannon entropy and variability.

Purpose of the Study:

• To apply a deep learning method for analyzing multiple sequence alignments of all human proteins.
• To identify unsupervised features that best describe variability patterns in protein sequences.

Main Methods:

• Utilized a deep learning, unsupervised feature extraction approach.
• Trained an auto-encoder neural network on 27,835 human protein multiple sequence alignments.
• Extracted two key features representing variability patterns.

Main Results:

• The two unsupervised learned features closely resemble traditional entropy and variability measures.
• These features effectively capture the most informative aspects of residue distributions in protein alignment columns.
• Demonstrated the power of deep learning in reducing dimensionality while preserving key information.

Conclusions:

• Unsupervised deep learning features provide a robust representation of protein sequence variability.
• The findings validate entropy and variability as fundamental descriptors of protein sequence evolution.
• This approach offers new insights into analyzing large-scale protein sequence data.