Inferred global dense residue transition graphs from primary structure sequences enable protein interaction prediction via directed graph convolutional neural networks

Islam Akef Ebeid^1*Haoteng Tang²Pengfei Gu²

• ¹Texas Woman's University, Denton, United States
• ²The University of Texas Rio Grande Valley, Brownsville, United States

Accurate prediction of protein-protein interactions (PPIs) is crucial to understanding cellular functions and advancing drug discovery. While some in-silico methods leverage direct sequence embeddings from Protein Language Models (PLMs), others apply Graph Neural Networks (GNNs) to topological features from PPI networks or 3D protein structures. Here, we introduce a novel approach ProtGram, that models protein primary structure through a hierarchy of globally inferred n-gram graphs. For each n-gram level, residue transition probabilities, aggregated from a large sequence corpus, define the edge weights of an underlying directed graph of paired residues. We propose a custom directed graph convolutional neural network, DirectGCN, featuring a unique convolutional layer that processes information through separate path-specific (incoming, outgoing, undirected) and shared transformations. These components are then combined via a learnable gating mechanism scoring each aggregated path. The efficacy of DirectGCN is established on standard node classification benchmarks, where its performance is on par with established methods on general benchmark datasets yet highlights its specialization for complex, directed, limited and dense heterophilic graph structures. Subsequently, DirectGCN is applied to the hierarchy of n-gram graphs ProtGram to learn residue-level embeddings, which are then pooled via an attention mechanism to generate protein-level embeddings. We name our two-stage graph representation learning framework ProtGram-DirectGCN. The main focus of this study is to investigate less intensive alternative models to PLMs for the downstream task of PPI prediction via link prediction. Our method achieves robust predictive power, suggesting that our globally inferred directed graph-based representation of global sequence transitions offers a potent and computationally distinct alternative to resource-intensive PLMs on the task of PPI prediction. Future work includes testing ProtGram-DirectGCN on a wide range of bioinformatics tasks.