HN-PPISP

HN-PPISP predicts protein-protein interaction (PPI) sites from raw protein sequences, secondary structures, and Position Specific Scoring Matrices (PSSM) using a hybrid neural network that integrates a Multi-layer Perceptron Mixer (MLP-Mixer), attention modules, TextCNN, and Bi-LSTM to model local and global sequence dependencies.

Key Features:

• Hybrid Network Architecture: Combines an MLP-Mixer module with a two-stage multi-branch module inspired by Transformers, TextCNN, and Bi-LSTM to integrate local and global feature extraction.
• Local Feature Extraction (MLP-Mixer): Uses the MLP-Mixer module to extract local sequence features without relying on convolutional operations.
• Global Feature Capture (Two-stage Multi-branch): Stage one employs improved attention modules to assign attention scores to raw sequences, secondary structures, and PSSM; stage two aggregates information via parallel branches using TextCNN, Bi-LSTM, and various activation functions.
• Attention Mechanism: Assigns distinct attention scores to different feature types (raw sequences, secondary structures, PSSM) to appropriately weight input channels.
• Long-Distance Dependency Modeling: Incorporates a Bi-LSTM branch to model long-distance sequence dependencies that are challenging for CNN-based methods.
• Benchmark Performance: Demonstrated superior performance on real-world public datasets, outperforming seven notable baselines.

Scientific Applications:

• PPI Site Prediction: Predicts residue-level protein-protein interaction sites from sequence and derived features.
• Drug Discovery Support: Identifies putative interaction interface residues that can inform target and inhibitor design.
• Molecular and Systems Biology: Provides interaction-site annotations useful for studying biochemical mechanisms and interaction networks.

Methodology:

Stage one uses improved hybrid attention modules to extract features from raw protein sequences, secondary structures, and PSSM; stage two encodes and aggregates these features via a multi-branch network that includes parallel TextCNN and Bi-LSTM branches, various activation functions, and an MLP-Mixer module for local feature encoding.