Rapid Prediction of Lipid Interaction Sites on Pleckstrin Homology Domains Using Deep Graph Neural Networks and Molecular Dynamics Simulations

Kyle I. P. Le Huray, Frank Sobott, He Wang, Antreas C. KallibioRxiv 2023 · DOI

Abstract

Interactions between membrane proteins and specific lipid molecules play a major role in cellular biology, but characterizing these interactions can be challenging due to the complexity and physicochemical properties of membranes. Molecular dynamics (MD) simulations allow researchers to predict protein-lipid interaction sites and generate testable models. MD simulations are however computationally expensive and require specialist expertise. In this study, we demonstrate that graph neural networks trained on coarse-grained MD simulation data can predict phosphoinositide lipid interaction sites on Pleckstrin Homology (PH) domain structures, a large family of membrane binding domains. The predictions are comparable to the results of simulations and require only seconds to compute. Comparison with experimental data shows that the model can predict known phosphoinositide interaction sites and can be used to form hypotheses for PH domains for which there is no experimental data. This model is a next generation tool for predicting protein-lipid interactions of PH domains and offers a basis for further development of models applicable to other membrane protein classes.

Introduction

Peripheral membrane proteins (PMPs) serve as critical "molecular switches" within the cell, transiently attaching to membranes to regulate signaling, metabolism, and transport. Identifying the specific residues where these proteins bind to lipids, such as phosphoinositides, is essential for both basic biological understanding and drug discovery.

Traditionally, researchers have faced a significant bottleneck in this area. Experimental methods like NMR or X-ray crystallography often fail to preserve lipid-protein complexes during processing. While computational Molecular Dynamics (MD) simulations can bridge this gap, they are prohibitively expensive; a single protein study can consume tens of thousands of CPU hours. This research introduces a lightweight deep learning alternative that provides MD-level insights with a speedup of over six orders of magnitude.

Methodology

The proposed model utilizes a Graph Attention Network (GAT-v2) architecture designed to capture the complex physicochemical rules of membrane binding.

Graph Representation and Features

Proteins are represented as fully-connected graphs where each amino acid residue acts as a node. This dense connectivity allows the model to learn cooperative influences between distant residues. Each node is defined by a 25-dimensional feature vector encompassing amino acid identity, secondary structure class, solvent accessible surface area, and local charge within 8 Å. Edges encode geometric relationships, specifically the Euclidean distance and unit vectors between alpha-carbons.

Overview of the feature processing pipeline and the multi-head graph attention network architecture.

Model Architecture

The network consists of two Multi-Head Graph Attention layers followed by a Node Distributed Dense Layer. The attention mechanism dynamically weights the importance of neighboring residues, effectively learning the underlying biophysics of lipid interaction. Unlike binary classifiers, the model predicts a continuous normalized contact frequency, preserving the nuanced data generated by Coarse-Grained Molecular Dynamics (CGMD).

Results

Performance and Speed

The model was validated using a dataset of 100 PH domains with ground truth derived from high-throughput CGMD simulations. While the original MD dataset required approximately 46,000 supercomputing hours, the GNN generates predictions in 0.3 to 1 second on a standard laptop.

Across extensive k-fold cross-validation, the model achieved a Mean Squared Error (MSE) of approximately 0.024 and a sensitivity of ~0.74 at a high-confidence threshold (0.8 normalized contact frequency). This demonstrates that the model is highly effective at identifying critical lipid interaction "hotspots."

Comparison of GNN predictions (blue) against MD simulation ground truth (red) for various PH domains, showing high alignment in contact frequency patterns.

Generalization and Validation

The model demonstrated remarkable generalization, successfully predicting binding sites for yeast (Slm1) and plant (VAN3-binding protein) proteins despite being trained primarily on mammalian data. Furthermore, the model is compatible with AlphaFold-predicted structures, significantly expanding its utility for proteins that lack experimental crystal structures. External validation against NMR chemical shift perturbation data for the Arap3 PH1 domain showed perfect alignment with experimental findings.

Discussion and Conclusion

This study proves that a relatively simple GNN can encapsulate the complex, collective behavior of protein-lipid interactions that previously required massive computational resources. By learning fundamental biophysical principles rather than simple sequence patterns, the model serves as a "physics in a box" solution for PMP research.

However, limitations remain. The model currently focuses on isolated PH domains and is dependent on the quality of the underlying CGMD training data, which may oversimplify certain electrostatic or stereochemical nuances. Future work will involve expanding the training set to include a wider variety of lipid types and other peripheral membrane protein families, moving toward a universal predictor for membrane-protein interactions.

Summary updated

Abstract

Abstract Interactions between membrane proteins and specific lipid molecules play a major role in cellular biology, but characterizing these interactions can be challenging due to the complexity and physicochemical properties of membranes. Molecular dynamics (MD) simulations allow researchers to predict protein-lipid interaction sites and generate testable models. MD simulations are however computationally expensive and require specialist expertise. In this study, we demonstrate that graph neural networks trained on coarse-grained MD simulation data can predict phosphoinositide lipid interaction sites on Pleckstrin Homology (PH) domain structures, a large family of membrane binding domains. The predictions are comparable to the results of simulations and require only seconds to compute. Comparison with experimental data shows that the model can predict known phosphoinositide interaction sites and can be used to form hypotheses for PH domains for which there is no experimental data. This model is a next generation tool for predicting protein-lipid interactions of PH domains and offers a basis for further development of models applicable to other membrane protein classes.
BibTeX
@misc{huray2023rapid,
  author = {Kyle I. P. Le Huray and Frank Sobott and He Wang and Antreas C. Kalli},
  title = {{Rapid Prediction of Lipid Interaction Sites on Pleckstrin Homology Domains Using Deep Graph Neural Networks and Molecular Dynamics Simulations}},
  year = {2023},
  doi = {10.1101/2023.12.22.573003},
  howpublished = {bioRxiv},
}
Figure