EGRET: edge aggregated graph attention networks and transfer learning improve protein-protein interaction site prediction

Sazan Mahbub, Md. Shamsuzzoha Bayzid

TL;DR

This paper introduces EGRET, a novel deep learning model that significantly improves the prediction of protein-protein interaction (PPI) sites on isolated proteins. By combining graph attention networks with edge aggregation and leveraging transfer learning from pre-trained language models, EGRET achieves state-of-the-art performance and offers better interpretability than previous methods.

What Problem Does This Paper Address?

Identifying the specific sites on proteins that mediate interactions with other proteins (PPI sites) is crucial for understanding biological processes, disease mechanisms, and for drug discovery. Traditional experimental methods for this are slow and costly. While computational methods have advanced, accurately predicting these sites, especially on isolated proteins where partner information is unknown, remains a challenge. Existing methods often struggle to effectively capture complex structural information, particularly long-range interactions between residues that are not physically close in the protein sequence.

Methods

EGRET is a deep learning architecture built upon Graph Neural Networks (GNNs), specifically a variant of Graph Attention Networks (GATs). Here's a breakdown of its key components:

  • Graph Representation: Proteins are represented as graphs where each amino acid residue is a node. Edges connect residues to their k-nearest neighbors in 3D space.
  • Local Feature Extractor (λ): A 1D convolutional neural network is used to capture local sequential information around each residue, reducing the dimensionality of initial features.
  • Edge Aggregated Graph Attention Layer (γ): This is the core innovation. Unlike standard GATs, EGRET's attention mechanism explicitly incorporates edge features (like distance and relative orientation between residues) alongside node features. This allows the model to better leverage the 3D structural information and long-range interactions.
  • Transfer Learning with ProtBERT: The initial node features are derived from ProtBERT, a pre-trained transformer-based language model for proteins. This allows EGRET to benefit from knowledge learned from vast amounts of protein sequence data, significantly improving performance and reducing training time.
  • Node-Level Classification: A final layer predicts the probability of each residue being an interaction site.

Datasets

The study evaluated EGRET on several benchmark datasets:

  • DeepPPISP Dataset: A combined dataset from Dset_186, Dset-72, and PDBset_164, widely used for PPI site prediction.
  • MaSIF Dataset: A substantially larger dataset compiled from various sources, including PRISM, ZDock, PDBBind, and SAbDab.
  • Dockground Dataset: A benchmark for evaluating protein-protein docking methods, used to assess EGRET's performance on known complexes.

Key Insights and Ideas

  • Edge Aggregation is Crucial: Explicitly incorporating edge features into the attention mechanism significantly enhances the model's ability to capture structural nuances and long-range interactions, leading to improved PPI site prediction.
  • Transfer Learning Boosts Performance: Using pre-trained embeddings from models like ProtBERT provides a powerful starting point, enabling EGRET to achieve high accuracy with less training data and time compared to methods relying solely on traditional features like PSSM.
  • Long-Range Interactions Matter: EGRET's architecture, particularly its edge aggregation, is better equipped to handle long-range interactions between residues, which are critical for accurate PPI site prediction, especially as protein length increases.
  • Interpretability Potential: While deep learning models are often black boxes, the attention scores in EGRET show patterns that correlate with biological insights, suggesting potential for understanding why certain residues are predicted as interaction sites.

Key Results

  • State-of-the-Art Performance: EGRET outperformed existing methods on six out of seven evaluation metrics on the DeepPPISP benchmark dataset, including key metrics like F1-score, AUROC, AUPRC, and MCC.
  • Significant Improvements: EGRET achieved 4.8%, 12.5%, and 14.4% higher F1-score, AUPRC, and MCC, respectively, compared to the previous best method (DELPHI), with statistically significant improvements.
  • Faster Feature Generation: Generating ProtBERT-based features for EGRET was dramatically faster (minutes) than generating traditional DeepPPISP features (hours), especially for longer proteins.
  • Edge Aggregation Benefits: EGRET showed performance improvements over a similar GAT model without edge aggregation (GAT-PPI), highlighting the positive impact of incorporating edge information.
  • High Recall on Dockground: EGRET demonstrated high recall (often >0.9) on the Dockground dataset, indicating its ability to identify a substantial portion of true interaction sites even when evaluated against more recent complex data.

Limitations and Future Work

  • Interpretability Needs Further Study: While attention scores show promising correlations, further research is needed to fully understand their relationship to true interaction sites and to design attention mechanisms that more closely align with biological predictions.
  • Exploring Other Pre-trained Models: The study used ProtBERT, but future work could investigate other pre-trained language models from ProtTrans.
  • Expanding to Other Binding Sites: EGRET's architecture could be adapted for predicting other types of binding sites, such as DNA or ligand binding sites.
  • Pairwise PPI Site Prediction: The current work focuses on isolated proteins. Future research could extend EGRET to pairwise PPI site prediction, which involves considering partner protein information.

Summary updated

Abstract

Protein–protein interactions (PPIs) are central to most biological processes. However, reliable identification of PPI sites using conventional experimental methods is slow and expensive. Therefore, great efforts are being put into computational methods to identify PPI sites.
BibTeX
@article{mahbub2022egret,
  author = {Sazan Mahbub and Md. Shamsuzzoha Bayzid},
  title = {{EGRET: edge aggregated graph attention networks and transfer learning improve protein-protein interaction site prediction}},
  journal = {Briefings in Bioinformatics},
  year = {2022},
  url = {https://openreview.net/forum?id=EriFLCrU9if},
}

Loading related content...

Figure