Clustering FunFams using sequence embeddings improves EC purity.

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Science

Background:

• Protein functional families aid in understanding protein roles and annotation transfer.
• Current Functional Families (FunFams) within CATH superfamilies show inconsistencies, with 7% of EC-annotated FunFams having mixed functions.
• Ensuring functional purity in protein clusters is crucial for accurate biological interpretation.

Purpose of the Study:

• To develop and validate a computational approach for creating more functionally pure protein sub-families.
• To enhance the reliability of protein function annotation through improved clustering methods.
• To leverage advanced language models for protein sequence representation and functional classification.

Main Methods:

• Utilized protein language model embeddings (ProtBERT) optimized for CATH superfamily discrimination (PB-Tucker).
• Employed DBSCAN clustering on embedding distances to group Functional Families (FunFams) and identify outliers.
• Applied the method to existing FunFams and families based on sequence similarity, assessing purity with EC and binding annotations.

Main Results:

• The proposed embedding-based clustering method doubled the number of pure clusters per FunFam compared to random clustering.
• The approach demonstrated success not only on FunFams but also on families derived from sequence similarity.
• Observed improved functional consistency for both EC and binding annotations, suggesting broader applicability.

Conclusions:

• The novel clustering approach significantly enhances the functional purity of protein families.
• This method improves the reliability of transferring annotations within protein groups.
• The approach is broadly applicable to various protein groupings and functional annotations.