🧬 biology

EquiPocket: an E(3)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction

The paper introduces EquiPocket, an E(3)-equivariant Geometric Graph Neural Network that overcomes the limitations of existing voxel-based CNN methods by effectively modeling irregular protein structures, rotation invariance, and surface geometry to achieve superior ligand binding site prediction.

Original authors: Yang Zhang, Zhewei Wei, Ye Yuan, Chongxuan Li, Wenbing Huang

Published 2026-08-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yang Zhang, Zhewei Wei, Ye Yuan, Chongxuan Li, Wenbing Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the microscopic world of living cells, life depends on a constant, precise conversation between molecules. Large proteins act as receptors, waiting for small chemical messengers called ligands to arrive and trigger a response. This interaction does not happen randomly across the entire surface of the protein; it occurs in specific, often hidden, hollows or pockets. Finding these pockets is the first critical step in designing new medicines, as drugs must fit into these spaces to work. For decades, scientists have tried to map these locations using computer models, but the proteins themselves are messy, irregular shapes that resist simple digital representation. Traditional methods often tried to force these complex 3D structures into neat, blocky grids, much like trying to fit a jagged rock into a square box. This approach frequently missed the finer details of the protein's surface or failed when the protein was rotated or changed size, leading to inaccurate predictions about where a drug might bind.

A team of researchers has now proposed a new way to solve this puzzle, moving away from rigid grids and toward a more fluid, geometric approach. They developed a system called EquiPocket, which treats a protein not as a collection of pixels in a 3D image, but as a network of connected points. In this model, every atom is a node, and the bonds between them are the connections. The system is designed to understand the shape of the protein no matter how it is turned or shifted in space, a property that mimics the way physical laws work in the real world. By focusing on the surface of the protein—the very edge where a drug would touch—the system gathers detailed information about the local geometry, such as the angles and distances between atoms, rather than just looking at a coarse, blocky approximation.

The researchers built this system with three distinct parts that work together to understand the protein. First, the system examines the immediate neighborhood of each atom on the surface, using a virtual probe to measure the shape of the surrounding space. This allows it to detect the subtle curves and crevices that define a binding pocket. Second, it looks at the protein as a whole, analyzing the chemical types of the atoms and how they are connected to understand the broader structure. Finally, it passes information between the surface atoms, allowing the system to learn how the local shapes combine to form the larger binding site. To handle the fact that proteins vary wildly in size, from small molecules to massive structures, the team added a special layer that adjusts its focus based on how crowded the atoms are in a given area. This ensures the system works equally well on a small protein with a few hundred atoms and a giant one with thousands.

When tested against existing methods on several standard datasets, this new approach showed clear advantages. It was able to predict the location of binding sites more accurately than previous techniques, including those based on deep learning that use 3D grids. The old grid-based methods often struggled with large proteins, sometimes failing to find a binding site at all because the protein was too big to fit into their fixed-size digital boxes. The new system, by contrast, adapted naturally to proteins of any size. It also proved more reliable when the protein was rotated, a common issue for older models that would give different answers depending on the protein's orientation. The researchers found that combining the local surface details with the global chemical structure was essential; removing either part caused the system's performance to drop significantly.

The study also revealed that the system could learn to predict not just where a binding site is, but also the direction in which a drug molecule would approach it. This extra layer of geometric understanding helped the model perform even better, particularly on smaller proteins. While the system requires more computing power than the simplest geometric tools, it remains faster and more efficient than the heavy 3D grid methods, striking a balance between speed and precision. The work suggests that by respecting the natural, irregular geometry of proteins and treating them as interconnected networks rather than static images, scientists can build more reliable tools for drug discovery. The researchers have made their code available, allowing others to test and build upon this new way of seeing the molecular world.

Technical Summary: EquiPocket

Problem Statement

Predicting ligand binding sites on target proteins is a fundamental step in drug discovery, facilitating downstream tasks such as protein-ligand docking and structure-based molecular generation. While deep learning methods, particularly Convolutional Neural Networks (CNNs), have achieved state-of-the-art performance by treating proteins as 3D voxelized images, they suffer from four critical limitations:

Irregular Structure Representation: CNNs rely on fixed-size voxel grids (e.g., 70 Å × 70 Å × 70 Å), which fail to capture the irregular shapes of proteins and discard atoms outside the fixed boundary, leading to inaccurate modeling of large proteins.
Rotation Sensitivity: Voxelization depends on fixed coordinate bases; rotating a protein alters the voxel grid, changing the prediction despite the physical invariance of binding sites. Data augmentation or local grids mitigate this but do not guarantee theoretical rotation invariance.
Insufficient Surface Characterization: Binding pockets are defined by surface geometry. CNN-based methods encode surface atoms within voxels surrounded by empty space, providing only coarse geometric information that fails to capture the intricate local interactions and geometry of surface atoms.
Protein Size Shift: Existing models struggle to generalize across datasets with varying protein sizes (from hundreds to tens of thousands of atoms), often failing on larger proteins due to fixed receptive fields or grid constraints.

Methodology

The authors propose EquiPocket, an $E(3)$ -equivariant Geometric Graph Neural Network (GNN) designed to address the aforementioned issues without voxelization. The framework treats atoms as nodes and interactions as edges, naturally handling irregular structures and ensuring invariance to Euclidean transformations (rotations and translations).

The architecture comprises three primary modules:

1. Local Geometric Modeling Module

To address the lack of detailed surface characterization, this module extracts local geometric information for each surface atom using Surface Probes.

Surface Probes: Using the MSMS algorithm, a probe of a specific radius is moved along the protein to generate a Solvent Accessible Surface (SAS) and Solvent Excluded Surface (SES). These probe coordinates form a set $S$ .
Feature Extraction: For each surface atom, the surrounding probes are analyzed to compute relative position vectors and angles (e.g., distances between probes, angles relative to the atom and the probe center).
Aggregation: These geometric descriptors are aggregated via pooling (mean and max) and processed through Multi-Layer Perceptrons (MLPs) to create an $E(3)$ -invariant geometric embedding $g_i$ for each surface atom.

2. Global Structure Modeling Module

This module captures the chemical and spatial context of the entire protein, acknowledging that global structure influences pocket formation.

Chemical Graph Modeling: A Graph Attention Network (GAT) processes chemical features (atom types, electron counts, bonds) to update chemical embeddings.
Spatial Graph Modeling: An Equivariant Graph Neural Network (EGNN) processes the 3D coordinates and updated chemical features to capture spatial interactions. The output is an invariant feature vector $c''_i$ for each atom, which encodes both chemical and global spatial information.

3. Surface Message Passing Module

This module refines the features of surface atoms by performing equivariant message passing on the surface graph $G_S$ .

Surface-EGNN: It utilizes a custom equivariant layer that updates both invariant features ( $h$ ) and equivariant coordinate matrices ( $X$ ). The message passing incorporates relative distances and angles between nodes and their neighbors, ensuring the output remains equivariant to input transformations.
Dense Attention Output Layer: To address the protein size shift issue, a novel output layer is introduced. It calculates the spatial density of neighbors for each atom across different distance hops. A learnable attention mechanism (using Sigmoid) weights the hidden features from different network layers based on this density distribution. This allows the model to adaptively balance the receptive field scope for proteins of varying sizes, preventing over-smoothing in small proteins and insufficient context in large ones.

Optimization

The model is trained with two objectives:

Binding Site Prediction ( $L_b$ ): A Dice loss is used to predict the probability of a surface atom being part of a binding site (defined as being within 4 Å of a ligand atom).
Relative Direction Prediction ( $L_d$ ): A cosine loss is applied to predict the relative direction from a surface atom to its nearest ligand atom, enhancing the extraction of local geometric features.

Key Contributions

First $E(3)$ -Equivariant GNN for Binding Sites: EquiPocket is the first framework to apply $E(3)$ -equivariant GNNs to ligand binding site prediction, eliminating the need for voxelization and ensuring rotation/translation invariance.
Three-Module Architecture: The integration of local geometric modeling (via surface probes), global structure modeling (chemical + spatial), and equivariant surface message passing provides a comprehensive representation of protein pockets.
Dense Attention Output Layer: A novel mechanism to handle variable protein sizes by adaptively weighting layer contributions based on local spatial density, addressing the generalization gap in existing methods.
Superior Performance: Extensive experiments demonstrate that EquiPocket outperforms state-of-the-art CNN-based, geometric-based, and other graph-based methods across multiple benchmarks.

Experimental Results

The framework was evaluated on standard benchmarks: COACH420, HOLO4K, and PDBbind2020, using metrics such as Distance to Center of Center (DCC), Distance to Center of Atom (DCA), and Failure Rate.

Performance: EquiPocket achieved the highest success rates, with a DCC of 0.423 and DCA of 0.656 on COACH420, significantly outperforming CNN-based methods like DeepSurf (DCC 0.386) and geometric methods like Fpocket (DCC 0.228).
Ablation Studies:
- Removing the local geometric module or global structure module individually caused a 10–20% drop in performance, confirming the necessity of both.
- Removing the Surface Message Passing module (EquiPocket-LG) resulted in a significant performance drop, highlighting the importance of surface-level interaction modeling.
- Removing the Dense Attention layer showed a marked performance decrease for proteins with fewer than 3,000 atoms, validating its role in handling size shifts.
- Removing the Direction Loss also degraded performance, particularly for smaller proteins, confirming the utility of predicting relative ligand directions.
Efficiency: EquiPocket is faster than 3D-CNN methods (e.g., DeepSurf) while maintaining higher accuracy, as it avoids the computational cost of 3D grid generation and convolution.

Significance

The paper claims that EquiPocket represents a significant advancement by moving away from the rigid constraints of voxel-based CNNs. By leveraging $E(3)$ -equivariance and explicit geometric modeling of surface probes, the framework naturally handles irregular protein structures and rotation invariance. The introduction of the Dense Attention layer specifically addresses the challenge of protein size variability, a critical issue in practical drug discovery scenarios where protein sizes vary widely. The authors assert that their framework achieves superior prediction accuracy and computational efficiency compared to existing state-of-the-art methods, providing a robust tool for downstream tasks like molecular generation and docking.