SelfAT-Fold

SelfAT-Fold predicts protein folds by applying Residue-Based and Motif-Based Self-Attention Networks together with Support Vector Machines to improve fold recognition for tertiary structure analysis.

Key Features:

• Residue-Based Self-Attention Network (RSAN): captures global associations among residues along protein sequences.
• Motif-Based Self-Attention Network (MSAN): captures global associations among structural motifs along protein sequences.
• Fold-specific attention features: derives fold-specific attention features from RSAN and MSAN for fold characterization.
• Integration with Support Vector Machines (SVMs): combines fold-specific attention features with SVM classifiers for protein fold prediction using the LE benchmark dataset independent from training data.
• Benchmark performance: demonstrated performance exceeding 34 state-of-the-art predictors and stable results on the independent SCOP_TEST dataset.
• Interpretability via attention: attention mechanisms provide interpretable signals for analyzing protein fold characteristics.
• MSAN architecture components: includes a motif convolution layer, a self-attention layer, and a fully connected layer.
• NLP-inspired self-attention mechanism: employs a self-attention mechanism inspired by natural language processing to measure global residue and motif associations.

Scientific Applications:

• Protein structure prediction: improves recognition of protein folds to support tertiary structure analysis.
• Structural biology studies: enables analysis of fold characteristics and relationships among residues and motifs.
• Feature extraction for downstream analyses: provides fold-specific attention features for use in further computational analyses.
• Drug discovery and design support: supplies structural insights that can inform drug discovery and design processes.

Methodology:

SelfAT-Fold implements two self-attention networks (RSAN and MSAN); MSAN comprises a motif convolution layer, a self-attention layer, and a fully connected layer; the self-attention mechanism is inspired by NLP to measure global associations among residues and motifs; fold-specific attention features are combined with SVM classifiers and evaluated on the LE benchmark dataset and the independent SCOP_TEST dataset.