Enhancing PPI explainability with geometric deep learning
Summary
Understanding protein–protein interactions (PPIs) is essential for elucidating biological processes and advancing drug discovery. Traditional experimental techniques, while precise, are expensive and difficult to scale. Computational methods, particularly those using deep learning, have emerged as powerful alternatives. Among them, Graph Neural Networks (GNNs) are especially suited to model biomolecular structures, as they can capture spatial and relational features from 3D protein conformations. However, many GNN-based approaches suffer from poor interpretability, limiting their adoption in biomedical contexts where model transparency is essential.
This thesis presents a deep learning framework combining GNNs and attention mechanisms for multiclass classification of PPIs based on structural interface types. The model builds upon the Struct2Graph architecture and adapts it to handle 31 PPI classes from the PINDER dataset, each consisting of 100 holo complexes. Proteins are represented as graphs with residues as nodes and edges defined by proximity. These graphs are encoded using two GCNs with shared weights, followed by a mutual attention layer that emphasizes important residues during interaction prediction. The resulting representations are used to classify the interaction type via a fully connected neural network with a softmax output.
The model was trained using 5-fold cross-validation and evaluated on an independent test set. It achieved consistently high accuracy and macro F1-score across all folds (0.97), demonstrating strong predictive performance. To assess robustness, additional experiments were conducted using only interface residues, defined by varying distance thresholds (from 10 Å to 4 Å). As the threshold decreased, model performance declined (down to F1 = 0.78 at 4 Å), suggesting that while interface information is essential, broader structural context improves classification.
A central goal of this work was interpretability. The mutual attention mechanism was used to identify and visualize the residues most influential in each prediction. Attention scores were displayed as 2D heatmaps and projected onto 3D structures using PyMOL. Additionally, attention values were aggregated by Cluster ID after aligning receptor and ligand sequences with multiple sequence alignment (MSA). These visualizations revealed attention concentration at interaction interfaces, though some attention was also assigned to distal regions, possibly reflecting structural shortcuts learned by the model.
Several limitations were identified: (1) the interpretability analysis relied on a limited number of examples; (2) attention maps were generated only for one fold, and (3) attention mechanisms may be biased by graph topology. These factors highlight the need for complementary interpretability techniques, such as gradient-based saliency methods, to validate and extend the current findings.
In summary, this thesis introduces a framework that combines high predictive accuracy with biologically meaningful interpretability for PPI classification. It represents a step toward more transparent AI applications in structural biology, with promising implications for drug discovery and protein engineering.