CATHe2: Enhanced CATH superfamily detection using ProstT5 and structural alphabets

TL;DR

CATHe2 is an improved automated system for classifying protein domain superfamilies using updated language models and structural data.

Contribution

CATHe2 introduces ProstT5 and structural alphabet embeddings to enhance CATH superfamily classification accuracy and F1 score.

Findings

CATHe2 achieves 92.2% accuracy and 82.3% F1 score, a significant improvement over CATHe.

Using ProstT5 and 3Di embeddings boosts performance by 9.9% in F1 score and 6.6% in accuracy.

A simplified version using only AA sequences still improves F1 score by 6.7% and accuracy by 6.6%.

Abstract

The CATH database is a free publicly available online resource that provides annotations about the evolutionary and structural relationships of protein domains. Due to the flux of protein structures coming mainly from the recent breakthrough of AlphaFold and therefore the non-feasibility of manual intervention, the CATH team recently developed an automatic CATH superfamily (SF) classifier called CATHe, which uses a feed-forward neural network (FNN) classifier with protein Language Model (pLM) embeddings as input. Using the same dataset of remote homologues (with a 20% sequence identity threshold), this paper presents CATHe2, which improves on CATHe by switching the old pLM ProtT5 for one of the most recent versions called ProstT5, and by incorporating domain 3D information into the classifier through Structural Alphabet representation, specifically, 3Di sequence embeddings. Finally, CATHe2 implements a new version of the FNN classifier architecture, fine-tuned to perform at the CATH superfamily prediction task. The best CATHe2 model reaches an accuracy of 92.2% ± 0.7% with an F1 score of 82.3% ± 1.3%, which constitutes an improvement of 9.9% on the F1 score and 6.6% on the accuracy, from the previous CATHe version (85.6% ± 0.4% accuracy and 72.4% ± 0.7% F1 score) on its largest dataset (∼1700 superfamilies). This model uses ProstT5 amino acid (AA) sequence and 3Di sequence embeddings as input to the classifier, but a simplified version requiring only AA sequences, already improves CATHe’s F1 score by 6.7% ± 1.3% and accuracy by 6.6% ± 0.7% on its largest dataset.

Figures