CATHe2: Enhanced CATH superfamily detection using ProstT5 and structural alphabets.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in genomics

Background:

• The CATH database classifies protein domain structures and evolutionary relationships.
• AlphaFold has accelerated protein structure prediction, necessitating automated classification methods.
• The previous CATHe classifier used protein Language Model (pLM) embeddings for superfamily prediction.

Purpose of the Study:

• To develop an improved automated classifier, CATHe2, for CATH superfamilies.
• To enhance protein domain classification accuracy and efficiency.
• To leverage recent advancements in protein Language Models and incorporate 3D structural data.

Main Methods:

• Utilized updated pLM embeddings (ProstT5) and 3Di sequence embeddings representing 3D structure.
• Implemented a fine-tuned feed-forward neural network (FNN) classifier architecture.
• Trained and evaluated models on a dataset of remote protein homologues with a 20% sequence identity threshold.

Main Results:

• The best CATHe2 model achieved 92.2% accuracy and 82.3% F1 score on the largest dataset (∼1700 superfamilies).
• This represents a substantial improvement over the previous CATHe version (85.6% accuracy, 72.4% F1 score).
• Even a simplified CATHe2 version using only amino acid sequences showed significant performance gains.

Conclusions:

• CATHe2 offers a more accurate and robust method for automated CATH superfamily classification.
• The integration of advanced pLMs and 3D structural information is crucial for improving protein classification.
• This advancement aids in understanding protein evolution and function in the era of large-scale structure prediction.