Abstract
Metalloproteins play indispensable roles in various biological processes ranging from reaction catalysis to free radical scavenging, and they are also pertinent to numerous pathologies including cancer, HIV infection, neurodegeneration, and inflammation. Discovery of high-affinity ligands for metalloproteins powers the treatment of these pathologies. Extensive efforts have been made to develop in silico approaches, such as molecular docking and machine learning (ML)-based models, for fast identification of ligands binding to heterogeneous proteins, but few of them have exclusively concentrated on metalloproteins. In this study, we first compiled the largest metalloprotein-ligand complex dataset containing 3079 high-quality structures, and systematically evaluated the scoring and docking powers of three competitive docking tools (PLANTS, AutoDock Vina and Glide SP) for metalloproteins. Then, a structure-based deep graph model called MetalProGNet was developed to predict metalloprotein-ligand interactions. In the model, the coordination interactions between metal ions and protein atoms and the interactions between metal ions and ligand atoms were explicitly modelled through graph convolution. The binding features were then predicted by the informative molecular binding vector learned from a noncovalent atom-atom interaction network. The evaluation on the internal metalloprotein test set, the independent ChEMBL dataset towards 22 different metalloproteins and the virtual screening dataset indicated that MetalProGNet outperformed various baselines. Finally, a noncovalent atom-atom interaction masking technique was employed to interpret MetalProGNet, and the learned knowledge accords with our understanding of physics.
Introduction
Metalloproteins are biological heavyweights, making up nearly half of the human proteome and playing critical roles in oxygen transport, signal transduction, and enzyme catalysis. Because they are implicated in diseases like cancer and neurodegeneration, they are prime therapeutic targets. However, computational drug discovery has long faced a "metal problem."
Most traditional molecular docking programs and machine learning models treat metal ions (such as Zinc, Magnesium, or Iron) as standard atoms. This "coarse handling" ignores the intricate coordination geometries and electronic interactions that define metal-ligand binding. Consequently, while general tools perform adequately for standard proteins, their accuracy often drops significantly when faced with the unique physics of a metallic binding site.
The MetalProGNet Architecture
To address these challenges, researchers developed MetalProGNet, a structure-based deep graph model that represents the protein-ligand complex using three distinct graphs: a Protein Graph, a Ligand Graph, and a Protein-Ligand Interaction Graph.
Key Innovations:
- Explicit Coordination Modeling: Unlike previous models, MetalProGNet explicitly encodes coordination interactions between metal ions and protein/ligand atoms as specific edge types in the graph.
- Atom–Atom Interaction Networks (AAINs): The model uses graph convolutions to learn "molecular binding vectors" from noncovalent interactions.
- Atomic Environment Vectors (AEVs): To ensure the model understands 3D space, it incorporates AEVs—features describing the radial and angular local environment of each atom, ensuring the model is rotationally and translationally invariant.
Data and Performance
The study utilized a newly curated set of 3,079 high-quality structures from PDBbind 2020, currently the largest dataset of its kind. MetalProGNet was benchmarked against traditional docking tools (PLANTS, AutoDock Vina) and general-purpose machine learning baselines.
- Superior Accuracy: On the internal test set, MetalProGNet achieved a Pearson correlation coefficient ($R_p$) of 0.703, significantly higher than RosENet (0.600) and NNScore2.0 (0.629).
- Virtual Screening Power: In the DEKOIS2.0 benchmark, MetalProGNet achieved a ROC_AUC of 0.942 for the mmp2 target, whereas traditional baselines often hovered near 0.50 (random guessing).
- Generalization: On an independent ChEMBL dataset of ~25,000 active ligands, the model maintained a competitive edge across 22 different metalloprotein families.
Interpretability: Physics Emerges from Data
A standout feature of MetalProGNet is its interpretability. By employing a noncovalent atom-atom interaction masking technique, the researchers determined which interactions the model prioritized. They found that the model naturally learned physics-based principles—such as the Lennard-Jones potential and hydrogen bonding distances—without being explicitly programmed with these rules. This confirms that the model's predictions are grounded in the actual physical chemistry of the binding site.
Conclusion
MetalProGNet demonstrates that treating metal ions as unique entities rather than "regular" atoms is essential for accurate protein-ligand modeling. By filling the gap in scoring power where traditional docking tools fail, MetalProGNet provides a robust new framework for the discovery of high-affinity ligands for metalloprotein targets.
Summary updated