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

TL;DR

This work introduces ProtGram-DirectGCN, a graph-based framework that predicts protein-protein interactions by learning from globally inferred, directed n-gram graphs derived from primary protein sequences. ProtGram constructs hierarchical G_n graphs where edges encode residue transition frequencies, and DirectGCN processes these directed, weighted graphs with path-specific aggregation and learnable gating to produce residue- and protein-level embeddings. The approach achieves strong PPI predictive power, notably in low-data settings, while remaining computationally lighter than large PLMs; ablations show the importance of gating and context level. The work provides biologically meaningful attention maps that highlight motifs associated with interactions and offers a scalable, structure-light alternative that can complement structure-based methods in proteome-wide analyses.

Abstract

Introduction Accurate prediction of protein-protein interactions (PPIs) is crucial for understanding cellular functions and advancing drug development. Existing in-silico methods use direct sequence embeddings from Protein Language Models (PLMs). Others use Graph Neural Networks (GNNs) for 3D protein structures. This study explores less computationally intensive alternatives. We introduce a novel framework for downstream PPI prediction through link prediction. Methods We introduce a two-stage graph representation learning framework, ProtGram-DirectGCN. First, we developed ProtGram. This approach models a protein's primary structure as a hierarchy of globally inferred n-gram graphs. In these graphs, residue transition probabilities define edge weights. Each edge connects a pair of residues in a directed graph. The probabilities are aggregated from a large corpus of sequences. Second, we propose DirectGCN, a custom directed graph convolutional neural network. This model features a unique convolutional layer. It processes information through separate path-specific transformations: incoming, outgoing, and undirected. A shared transformation is also applied. These paths are combined via a learnable gating mechanism. We apply DirectGCN to ProtGram graphs to learn residue-level embeddings. These embeddings are pooled via attention to generate protein-level embeddings for prediction. Results We first established the efficacy of DirectGCN on standard node classification benchmarks. Its performance matches established methods on general datasets. The model excels at complex, directed graphs with dense, heterophilic structures. When applied to PPI prediction, the full ProtGram-DirectGCN framework delivers robust predictive power. This strong performance holds even with limited training data.

Figures (6)

• Figure 1: An example of dummy sequences separated by a space indicating multiple proteins. The figure shows how the transitions between the residues or characters are conceptualized as a directed, weighted, dense graph where the weights are the transition frequencies calculated as counts or probabilities. In addition, the figure shows how we split the directed adjacency into an $A_{in}$ and an $A_{out}$. Later, we describe how to overcome the non-hermitian nature of these two matrices to make them suitable for graph convolutional neural networks.
• Figure 2: $ProtGram-DirectGCN$ full pipeline.
• Figure 5: The plot displays the Receiver Operating Characteristic (ROC) curves comparing the performance of protein embeddings generated by the proposed $ProtGram-DirectGCN$ method against the state-of-the-art ProtT5 and ESM language models, in addition to Word2Vec. The evaluation is for the downstream task of Protein-Protein Interaction (PPI) link prediction, with this specific chart illustrating the results from the average of a 5-fold cross-validation. All models perform significantly better than random chance (dashed line). This visualization confirms that while the proposed graph-based method captures a strong predictive signal for protein interactions, both ProtT5 and ESM models serves as a higher-performing benchmark in this experiment.
• Figure 6: The figure illustrates the models' performance metrics reflecting the results in table \ref{['tab:ppi_performance_results']}
• Figure 7: Ablation results for varying gating modes versus different n-gram levels when training $ProtGram-DirectGCN$