Exploring Large Protein Sequence Space through Homology- and Representation-based Hierarchical Clustering

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Exploring protein sequence space is crucial for understanding protein function and relationships.
• Traditional sequence similarity networks face limitations in scalability and hierarchical homology visualization.
• Current methods struggle with analyzing very large protein sequence datasets.

Purpose of the Study:

• To present an innovative sequence analysis pipeline addressing limitations of traditional methods.
• To enable scalable exploration of large protein sequence datasets.
• To enhance understanding of protein sequence-function relationships.

Main Methods:

• Developed a hierarchical visualization approach for homology.
• Utilized protein language model embeddings as an alternative homology metric to BLAST.
• Employed HMMs or vector representations for unbiased representative sequence sampling.
• Applied the pipeline to FMN/F420-binding split barrel and nuclear transport factor 2-like superfamilies.

Main Results:

• Hierarchical visualization captures full homology ranges across protein superfamilies.
• Protein language model embeddings provide comparable results to BLAST for identifying isofunctional families.
• Unbiased sequence sampling improves phylogenetic analysis.
• The pipeline is scalable to approximately 445,000 sequences on desktop computers.

Conclusions:

• The developed pipeline offers a scalable solution for exploring large protein sequence datasets.
• Innovations in visualization and homology metrics enhance protein sequence-function analysis.
• Publicly available code (ProteinClusterTools) facilitates broader research applications.