Automatic discovery of cross-family sequence features associated with protein function

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Predicting protein function from amino acid sequences is crucial for uncharacterized protein families and comparative genomics.
• Current machine learning methods predict predefined functional categories or subcellular locations, potentially missing underlying biological relationships.
• Human-designated functional classes may not perfectly align with biological reality, limiting discovery.

Purpose of the Study:

• To develop a self-supervised data mining approach for discovering protein function directly from sequence data.
• To identify relationships between sequence features and functional annotations without requiring prior biological assumptions.
• To explore novel sequence-to-function correlations beyond established categories.

Main Methods:

• Employed a self-supervised data mining technique utilizing genetic programming.
• Co-evolved amino acid-based regular expressions and keyword-based logical expressions.
• Trained on protein sequences and their associated UniProt/Swiss-Prot annotations.

Main Results:

• Successfully identified relationships between sequence features and functional annotations.
• Detected strong correlations with cellular compartment targeting and broad functional roles like inhibition, biosynthesis, transcription, and defense.
• Found distinct sequence motifs for specific functions, e.g., polyglutamine repeats linked to transcription more than nuclear location.

Conclusions:

• Developed a novel approach for knowledge discovery in annotated sequence data.
• The technique effectively identifies functionally important sequence features without expert knowledge.
• Revealed unexpected links between biological processes, such as ubiquitination and transcription, by analyzing protein function from a sequence perspective.