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

arXiv:2302.12177 · q-bio.BM, cs.LG · Submitted 2026-08-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "EquiPocket: an E(3)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction".

Jane: The paper was written by Yang Zhang, Zhewei Wei, Ye Yuan, Chongxuan Li and Wenbing Huang from Gaoling School of Artificial Intelligence, Renmin University of China and Beijing Institute of Technology.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, folks! Today we're diving into a paper that's got a mouthful of a title — "EquiPocket: an E(three)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction." Jane, I'm going to need you to break that down for me before my brain melts.

Jane: Happy to, Tom! So the title is basically saying three things. First, "EquiPocket" is their new method. Second, "E(three)-equivariant" means the model is smart about rotations and translations — if you spin the protein around, the prediction stays the same. And third, it's a graph neural network, which is a way of learning from data that's structured like a network of connected points.

Tom: And "ligand binding site prediction" — that's the actual job, right? Finding where a small molecule would attach to a protein?

Jane: Exactly. Think of it like finding the docking station on a protein. If you're designing a drug, you need to know where on the protein that drug is going to latch on. The paper is from researchers at Renmin University and Beijing Institute of Technology, and they're tackling this problem in a fresh way.

Tom: So what's wrong with the old ways? I mean, people have been doing this for years.

Jane: The old ways mostly used three dee convolutional neural networks. They'd chop the protein into little cubes, like voxels, and run a three dee CNN over that grid. But that has real problems. Proteins aren't cubes — they're irregular blobs. And if you rotate the protein, the grid changes, so the prediction can change too, which makes no sense physically.

Tom: So the paper is saying, "let's stop pretending proteins are Lego bricks and treat them like the messy three dee objects they are."

Jane: Precisely. They use a graph where each atom is a node, and edges connect atoms that are close in space. That's much more natural. And because the model is equivariant, it doesn't care if you rotate the whole protein — the answer is the same.

Tom: That sounds like a solid upgrade. But I'm guessing there's more to it than just swapping the architecture?

Jane: Oh, definitely. The paper identifies four issues with the old CNN approach, and they address all of them. We'll get into the details in a bit, but the headline is that they beat the state-of-the-art methods on several benchmarks.

Tom: I love a good benchmark beatdown. So who are we talking about here — the authors?

Jane: Yang Zhang, Zhewei Wei, Ye Yuan, Chongxuan Li, and Wenbing Huang. They're from Renmin University and Beijing Institute of Technology. And they've made the code public, which is always a nice touch.

Tom: Public code, new architecture, better results. This is shaping up to be a fun episode. So what's the big idea behind the method itself?

Jane: The big idea is that they don't just look at the atoms — they also look at the protein's surface. They generate something called surface probes, which are points that trace the outer boundary of the protein. Then they use those probes to understand the geometry of the surface, which is where binding sites actually live.

Tom: So it's not just about the atoms themselves, but the shape of the protein's skin?

Jane: You got it. And that's what makes this different from just slapping a graph neural network on a protein and hoping for the best. The surface geometry is crucial, and the old methods didn't really capture it well.

Tom: Alright, I'm hooked. Let's dig into the actual method and see how they pulled this off.

Summary: Tom: So we're back with "EquiPocket: an E(three)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction." Jane, you mentioned the surface probes — how does that actually work in practice?

Jane: So they use a tool called MSMS to generate the solvent-accessible surface of the protein. Imagine rolling a ball over the protein's surface — that ball traces out a shape, and the points where the ball touches are the surface probes. Each probe is associated with the nearest protein atom.

Tom: So you get a cloud of points hugging the protein's surface. Then what?

Jane: For each atom on the surface, they look at the probes around it and compute geometric features — distances between probes, angles, that sort of thing. This gives each atom a local geometric fingerprint. That's their first module.

Tom: And that's the local part. What about the global structure?

Jane: The second module processes the whole protein graph. They use a GAT, which is a graph attention network, to handle chemical bonds, and then an EGNN — an equivariant graph neural network — to handle the spatial structure. This gives them a global understanding of the protein's layout.

Tom: So you've got local geometry from the surface and global structure from the whole protein. Then what?

Jane: Then comes the third module, which is the surface message passing. They take the surface atoms and run equivariant message passing over them — this lets atoms talk to their neighbors on the surface. It's like the atoms are sharing information about their local geometry with each other.

Tom: And this is where the E(three)-equivariance really kicks in?

Jane: Right. The model updates both invariant features — things like atom type — and equivariant features — things like coordinates. The clever part is that the message function uses distances and angles, which are invariant to rotation and translation, so the whole thing stays equivariant.

Tom: I see. So the model can output not just a probability for each atom, but also a predicted direction toward the ligand. That's pretty neat.

Jane: Exactly. They actually add a second training task: predicting the relative direction from each surface atom to the nearest ligand atom. That helps the model learn better geometric features.

Tom: And then there's the dense attention output layer, right? What's that about?

Jane: That's their answer to the protein size problem. Proteins vary hugely in size — some have a few hundred atoms, some have tens of thousands. If you have a fixed number of message-passing layers, small proteins get over-smoothed and large proteins don't get enough information flow.

Tom: So they made the model adaptive?

Jane: Yes. For each atom, they compute how many neighbors are within different distance ranges. That tells them about the local density. Then they use that to weight the features from different layers — so a small protein might rely more on early layers, while a large protein uses later layers more.

Tom: That's a smart way to handle the size shift. And the results?

Jane: They tested on three benchmarks — COACH420, HOLO4K, and PDBbind2020. Their full model, EquiPocket, beats all the baselines, including DeepSurf and P2rank, on both DCC and DCA metrics. DCC is the distance to the true binding site center, and DCA is the distance to the nearest ligand atom.

Tom: So it's not just a small improvement — they're clearly ahead?

Jane: On COACH420, their DCC success rate is zero point four two three, while DeepSurf gets zero point three five four. On HOLO4K, it's zero point three three seven versus zero point two seven seven. And on PDBbind, it's zero point five four five versus zero point four nine two. That's a solid jump across the board.

Tom: Wow. And the failure rate — the percentage of proteins where they can't find any binding site — is much lower too, right?

Jane: Yes, their failure rate is zero point zero five one on COACH420, compared to zero point zero seven five for DeepSurf. So they're not just more accurate — they're more reliable.

Tom: I'm impressed. But I'm curious — what does this mean for actually using it in drug discovery?

Improvements: Tom: So we're still on "EquiPocket: an E(three)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction." Jane, you mentioned the improvements over CNN-based methods. Let's dig into what they actually fixed.

Jane: Sure. The paper calls out four specific issues with the old CNN approach. First, CNNs need voxelization — chopping the protein into a fixed-size grid. That's bad for irregular proteins, and it discards atoms that fall outside the grid. EquiPocket doesn't need a grid at all.

Tom: So no more losing atoms at the edges. What's the second issue?

Jane: Rotation sensitivity. If you rotate a protein, the voxel grid changes, so the CNN's prediction changes. That's physically wrong — rotating a protein shouldn't change where the binding site is. EquiPocket is E(three)-equivariant, so it's invariant to rotations and translations by design.

Tom: That's a fundamental advantage. And the third issue?

Jane: Surface characterization. The old methods didn't really capture the protein surface geometry well. EquiPocket explicitly uses surface probes to model the local geometry of each surface atom, which is where binding sites actually form.

Tom: And the fourth issue?

Jane: Protein size shift. Different proteins have wildly different sizes, and the old methods didn't adapt. EquiPocket's dense attention output layer adjusts the receptive field based on the local density around each atom.

Tom: So they've addressed all four issues. But I'm curious — what about the practical side? How does this actually run?

Jane: Good question. They report that EquiPocket takes about thirty-seven seconds per one hundred proteins for inference. That's faster than DeepSurf, which takes six hundred forty-one seconds, and comparable to Kalasanty at eighty-six seconds. So it's not just more accurate — it's also more efficient.

Tom: That's a huge difference. six hundred forty-one seconds versus thirty-seven seconds — that's like an order of magnitude faster.

Jane: Exactly. And they also did ablation studies to show each module matters. If you remove the local geometric modeling, performance drops. If you remove the global structure modeling, it drops too. And the full model with the surface message passing is the best.

Tom: So every piece is pulling its weight. What about the dense attention — did they show it actually helps?

Jane: They did. They compared EquiPocket with and without the dense attention layer. For proteins with fewer than four thousand atoms, the attention layer gives a noticeable boost. For larger proteins, it doesn't hurt. So it's a win-win.

Tom: And the direction loss — the second training task?

Jane: That also helps, especially for smaller proteins. Without it, the DCC drops by about ten percent for proteins under three thousand atoms. So the direction prediction isn't just a gimmick — it genuinely improves the learned geometry.

Tom: So the improvements are real and measurable. But I'm wondering — what does this mean for the broader field? Is this just a better tool, or does it change how people think about the problem?

Jane: I think it changes the conversation. For a long time, people treated binding site prediction as a three dee image segmentation problem. This paper shows that treating it as a geometric graph problem is not only more principled but also more practical. It's faster, more accurate, and more robust.

Tom: And it opens the door for other equivariant GNNs to be applied to similar problems in drug discovery.

Jane: Exactly. The same architecture could be adapted for protein-protein interaction sites, or even for predicting other functional regions on proteins. The surface probe idea is pretty general.

Tom: So what's the catch? Is there anything they didn't address?

Jane: Well, they trained on scPDB, which is a specific dataset. The test sets are different, but they're all protein-ligand complexes. It would be interesting to see how it performs on truly novel protein folds or on proteins with no known binding sites. But that's future work.

Tom: Fair enough. Let's wrap this up and talk about the big picture.

Conclusion: Tom: Alright, we're wrapping up our discussion of "EquiPocket: an E(three)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction." Jane, give us the final takeaway.

Jane: The takeaway is that EquiPocket shows you can do better than three dee CNNs for binding site prediction by using an equivariant graph neural network that explicitly models the protein surface. It's faster, more accurate, and more robust to protein size variations.

Tom: And it's not just a theoretical exercise — they've got public code and strong results on three benchmarks.

Jane: Right. The code is on GitHub, so anyone can try it. And the results are consistent across COACH420, HOLO4K, and PDBbind2020. That's a solid evidence base.

Tom: I want to bring in Lu, our senior researcher, to get a bigger picture take. Lu, what excites you most about this?

Lu: Thanks, Tom. What excites me is that this is a clean example of how geometric deep learning can outperform brute-force approaches. The old CNN methods were essentially treating proteins as images, which is a lossy representation. EquiPocket respects the actual geometry of the system, and that pays off.

Tom: And Meng, from the engineering side — does this look practical to you?

Meng: Yeah, I'm impressed by the efficiency. thirty-seven seconds per one hundred proteins is actually usable in a high-throughput screening pipeline. And the fact that it doesn't need a fixed grid means you can feed it proteins of any size without preprocessing headaches.

Tom: Lalam, what about the bigger cultural or scientific impact?

Lalam: I think this is a step toward more reliable in-silico drug discovery. If we can predict binding sites accurately and quickly, we can screen more compounds, explore more protein targets, and potentially reduce the number of failed experiments. That has real implications for how fast we can develop new medicines.

Tom: So it's not just an academic improvement — it could actually speed up drug development.

Lalam: Exactly. And the equivariance property means the model is robust to how the protein is oriented in space, which is a practical concern when you're dealing with experimental structures that might be in different coordinate frames.

Jane: And that's what makes this paper exciting — it's not just a better number on a benchmark. It's a better way of thinking about the problem.

Tom: Well said, Jane. So we're saying goodbye to EquiPocket, but I have a feeling we'll be seeing more work building on this. Thanks for joining us, everyone.

Jane: And remember, the code is out there — go try it on your favorite protein. Until next time!

Tom: See you on the next paper!

Yang Zhang, Zhewei Wei, Ye Yuan, Chongxuan Li, Wenbing Huang

Gaoling School of Artificial Intelligence, Renmin University of China · Beijing Institute of Technology

q-bio.BM, cs.LG

Submitted: 2026-08-18

Updated: 2026-08-19

Comments: Accepted to ICML 2024 (Oral)

Code: https://github.com/fengyuewuya/EquiPocket

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 66/100

The gist: EquiPocket is an E(3)-equivariant Graph Neural Network (GNN) for ligand binding site prediction, proposed to address four critical issues in existing CNN-based methods: 1) defective in representing

Key concepts

E(3)-equivariant
This means the model is smart about rotations and translations. If you rotate a protein in space, the prediction for the binding site remains the same because it is invariant to these changes. This makes it physically sensible.
Graph Neural Network (GNN)
A GNN is a way of learning from data structured like a network of connected points. In this paper, each atom in the protein is a node, and edges connect atoms that are close in space, allowing the model to learn from the protein's structure.
Surface Probes
These are points generated by rolling a ball over a protein's surface. Each probe is associated with the nearest protein atom, and they are used to compute geometric features that give each surface atom a local geometric fingerprint.
Dense Attention Output Layer
This layer makes the model adaptive to different protein sizes. It computes how many neighbors are within various distance ranges to determine local density, which then weights features from different message-passing layers based on that density.

Terminology

Summary

EquiPocket is an E(3)-equivariant Graph Neural Network (GNN) for ligand binding site prediction, proposed to address four critical issues in existing CNN-based methods: 1) defective in representing irregular protein structures due to voxelization; 2) sensitive to rotations; 3) insufficient to characterize the protein surface; 4) unaware of protein size shift. The framework comprises three modules: the first one to extract local geometric information for each surface atom, the second one to model both the chemical and spatial structure of protein, and the last one to capture the geometry of the surface via equivariant message passing over the surface atoms. The paper also proposes a dense attention output layer to alleviate the effect incurred by variable protein size. Extensive experiments on several representative benchmarks (COACH420, HOLO4K, PDBbind2020) demonstrate the superiority of our framework to the state-of-the-art methods. The key contributions are: 1) being the first to apply an E(3)-equivariant GNN for ligand binding site prediction, which is free of the voxelization process, able to model irregular protein structures by nature, and insensitive to any Euclidean transformation; 2) the three-module design, with the first and last modules tackling the surface characterization issue; 3) the dense attention output layer to adaptively balance the scope of the receptive field for each atom based on the density distribution of the neighbor atoms; 4) extensive experiments showing superiority over SOTA methods. The model is trained with a Dice loss for binding site prediction and a cosine loss for relative direction prediction, with the final loss being L = Lb + Ld. Ablation studies show that the local geometric modeling module, global structure modeling module, and surface message passing module each contribute significantly, with the full EquiPocket achieving DCC improvements of approximately 20% over the version without the surface message passing module. Hyperparameter analysis covers probe radius (default 1.5), cutoff θ (default 6), and depth of surface-egnn (default 4), balancing performance and GPU memory.

Improvements for AI systems

Based on the paper, here are specific improvements I can make to AI systems, along with what the improved system can do:

  • Replace CNN-based voxelization approaches with an E(3)-equivariant GNN that directly processes atoms as graph nodes with 3D coordinates.

  • Implement a dual-channel equivariant feature representation: one channel for atom coordinates, one for surface center coordinates.

  • Use invariant message passing that computes relative distances and angles between coordinate channels.

What the improved system can do: Process irregular protein structures without discretization artifacts, maintain rotation/translation/reflection invariance by construction, and avoid information loss from fixed-size voxel grids.

  • Integrate solvent-accessible surface (SAS) computation using MSMS to generate surface probe points.

  • For each surface atom, aggregate local geometric descriptors from surrounding probes: distances to neighboring probes, angles to surface center, and relative position to protein atom.

  • Combine both pooled MLP-transformed features and MLP-transformed pooled features for richer representation.

  • Compute per-atom spatial density vectors by measuring the proportion of protein atoms within increasing distance thresholds (0 to L hops).

  • Generate attention weights via Sigmoid-activated MLP on these density vectors.

  • Concatenate layer-wise hidden features weighted by these attention scores, rather than using only the final layer.

  • Add a secondary task: predict the relative direction from each surface atom to its nearest ligand atom.

  • Use the equivariant coordinate output from the network to compute predicted direction vectors.

  • Combine Dice loss for binding site classification with cosine loss for direction prediction.

  • Implement a three-stage architecture: (1) local geometric module for surface atoms, (2) global structure module combining chemical GNN (GAT) and spatial EGNN, (3) surface message passing module with equivariant updates.

  • Use the global module to encode whole-protein context, then selectively propagate information only over surface atoms.

  • Initialize each surface atom's equivariant matrix with both its 3D position and the center of its surrounding surface probes.

  • Use a specialized invariant function that computes distances and angles between the four points (two per atom) for edge messages.

  • Update coordinates via mean-aggregated relative position vectors weighted by learned functions.

The improved AI system can:

  • Predict ligand binding sites on proteins of any size (from 500 to 10,000+ atoms) with consistent accuracy

  • Maintain performance under arbitrary 3D rotations/translations without data augmentation

  • Detect both shallow and deep pockets by combining surface geometry with global protein context

  • Operate in resource-constrained environments (single A100 GPU) while outperforming methods requiring 40x more parameters

  • Provide both binding site probabilities and directional information for downstream docking tasks

Abstract

Predicting the binding sites of target proteins plays a fundamental role in drug discovery. Most existing deep-learning methods consider a protein as a 3D image by spatially clustering its atoms into voxels and then feed the voxelized protein into a 3D CNN for prediction. However, the CNN-based methods encounter several critical issues: 1) defective in representing irregular protein structures; 2) sensitive to rotations; 3) insufficient to characterize the protein surface; 4) unaware of protein size shift. To address the above issues, this work proposes EquiPocket, an E(3)-equivariant Graph Neural Network (GNN) for binding site prediction, which comprises three modules: the first one to extract local geometric information for each surface atom, the second one to model both the chemical and spatial structure of protein and the last one to capture the geometry of the surface via equivariant message passing over the surface atoms. We further propose a dense attention output layer to alleviate the effect incurred by variable protein size. Extensive experiments on several representative benchmarks demonstrate the superiority of our framework to the state-of-the-art methods.

Sources

Related papers