GDEGAN: Gaussian Dynamic Equivariant Graph Attention Network for Ligand Binding Site Prediction
Animesh, Plaban Kumar Bhowmick, Pralay Mitra
TL;DR
GDEGAN introduces a novel attention mechanism for equivariant GNNs that adaptively captures local residue variations, significantly improving binding site prediction accuracy in proteins for drug discovery.
Contribution
The paper presents GDEGAN, a new Gaussian dynamic attention mechanism that enhances equivariant GNNs by modeling local residue variations for better binding site prediction.
Findings
Achieves 37-66% improvement in DCC success rate
Achieves 7-19% improvement in DCA success rate
Outperforms existing methods on multiple datasets
Abstract
Accurate prediction of binding sites of a given protein, to which ligands can bind, is a critical step in structure-based computational drug discovery. Recently, Equivariant Graph Neural Networks (GNNs) have emerged as a powerful paradigm for binding site identification methods due to the large-scale availability of 3D structures of proteins via protein databases and AlphaFold predictions. The state-of-the-art equivariant GNN methods implement dot product attention, disregarding the variation in the chemical and geometric properties of the neighboring residues. To capture this variation, we propose GDEGAN (Gaussian Dynamic Equivariant Graph Attention Network), which replaces dot-product attention with adaptive kernels that recognize binding sites. The proposed attention mechanism captures variation in neighboring residues using statistics of their characteristic local feature…
Peer Reviews
Decision·Submitted to ICLR 2026
Strength 1：Uses a local Gaussian kernel with adaptive bandwidth from neighborhood statistics and a learnable temperature, yielding context-aware attention suited to heterogeneous protein surfaces. Strength 2：Provides formal analysis showing the proposed attention preserves SE(3) equivariance under the chosen feature representation, giving a clear geometric justification. Strength 3 ：Demonstrates consistent improvements over strong baselines across standard pocket benchmarks, supported by ablatio
Weakness 1： The paper claims to capture geometric structure and handle variation among neighboring residues, but the evidence is mostly indirect. It should demonstrate which previously hard geometric challenges are now addressed, with targeted analyses rather than only aggregate metrics and visuals. Weakness 2：Comparisons with prior graph-attention variants are incomplete, especially kernelized attention methods. A deeper analysis against these baselines is needed to substantiate the claimed con
The paper is well organized in terms of the limitations of the current binding site prediction models. The adoption of GotenNet and modifying the vanilla attention with the proposed approach to make the attention more dynamic and aware of an atom's local neighborhood is a an interesting approach. The key contribution is the idea of computing a neighboring atom’s features from Gaussian distribution defined by the target atom’s local neighborhood in the model. The results in Table 1 show that GDEG
From the results in Table 1 it shows the proposed method beats GotenNet on all three datasets except on the failure rate. However, from the ablation study shows (in Table 2) that the main boost comes from the ESM embeddings for both methods. As the authors show that the proposed approach is beneficial as structural heterogeneity increases. Since protein ligand binding site is determined by the chemical fingerprint it would be interesting to see if the method relied no only on the C-alpha atoms
The novelty lies in the successful adaptation and application of a probabilistic, variance-aware attention mechanism to the domain of 3D equivariant graph representations for a critical bioinformatics task. While building upon a strong backbone (GotenNet), the proposed attention module is a distinct and impactful innovation. It provides a more physically grounded inductive bias by assuming that variance in learned features is a meaningful signal, a departure from standard similarity-based dot-pr
**Key Flaw:** The most critical weakness is the ambiguity in the formulation of the core Gaussian attention mechanism. Specifically, the dimensionality of the neighborhood statistics μ_i and (σ_i)^2 in Equations 5 and 6 is unclear in the context of Equation 7. Since h_j is a high-dimensional feature vector, μ_i and (σ_i)^2 should also be vectors (element-wise mean and variance). However, Equation 7 uses (σ_i)^2 as if it were a scalar value for modulating the attention kernel's bandwidth. This la
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsComputational Drug Discovery Methods · Protein Structure and Dynamics · Machine Learning in Bioinformatics