Inferred global dense residue transition graphs from primary structure sequences enable protein interaction...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Accurate prediction of protein-protein interactions (PPIs) is vital for understanding cellular mechanisms and drug discovery.
• Current in-silico methods often rely on computationally intensive approaches like Protein Language Models (PLMs) or Graph Neural Networks (GNNs) on 3D structures.

Purpose of the Study:

• To investigate computationally less intensive alternatives for predicting protein-protein interactions (PPIs).
• To introduce and evaluate a novel two-stage graph representation learning framework for PPI prediction via link prediction.

Main Methods:

• Developed ProtGram to model protein primary structure as a hierarchy of globally inferred n-gram graphs with residue transition probabilities as edge weights.
• Proposed DirectGCN, a custom directed graph convolutional neural network with specialized convolutional layers and a learnable gating mechanism.
• Applied DirectGCN to ProtGram graphs to learn residue and protein-level embeddings using an attention mechanism for PPI prediction.

Main Results:

• DirectGCN demonstrated comparable performance to established methods on general node classification benchmarks, excelling on complex, directed, and dense heterophilic graphs.
• The full ProtGram-DirectGCN framework achieved robust predictive power for PPI prediction, even when trained on limited data.

Conclusions:

• A globally inferred, directed graph-based representation of sequence transitions presents a computationally distinct and potent alternative to resource-intensive PLMs for PPI prediction.
• The ProtGram-DirectGCN framework shows promise for various bioinformatics tasks beyond PPI prediction.