ProtAttn-QuadNet: An attention-based deep learning framework for protein-protein interaction prediction using...

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions but experimentally identifying them is challenging.
• Protein Language Models (PLMs) have advanced sequence-based PPI prediction by encoding protein sequence information.
• Existing methods often struggle with capturing complex, reciprocal dependencies between interacting protein pairs.

Purpose of the Study:

• To introduce ProtAttn-QuadNet, an attention-based deep learning framework for accurate PPI prediction.
• To leverage ProtBERT embeddings and a novel quad-stream attention mechanism for modeling protein pair dependencies.
• To provide a reliable computational tool for large-scale PPI prediction.

Main Methods:

• Developed an attention-based deep learning framework, ProtAttn-QuadNet.
• Utilized ProtBERT embeddings for deep contextual encoding of protein sequences.
• Employed a quad-stream attention mechanism with multi-level self- and cross-attention layers to model protein pair relationships.

Main Results:

• ProtAttn-QuadNet achieved high accuracy (97.16% on balanced data, 99.19% on oversampled data) and AUC-ROC (99.00%, 99.76%) on an independent UniProt dataset.
• The model outperformed several state-of-the-art PPI prediction methods.
• Statistical validation confirmed the model's predictive significance and reliability.

Conclusions:

• ProtAttn-QuadNet effectively models reciprocal dependencies in protein pairs using an attention-based deep learning approach.
• The framework offers a significant advancement in computational prediction of protein-protein interactions.
• This method provides a powerful and accurate tool for large-scale biological research.