SigmaDock: Untwisting Molecular Docking With Fragment-Based SE(3) Diffusion
Alvaro Prat, Leo Zhang, Charlotte M. Deane, Yee Whye Teh, Garrett M. Morris
TL;DR
SigmaDock introduces a fragment-based SE(3) diffusion model for molecular docking, significantly improving pose prediction accuracy and generalization over existing deep learning and physics-based methods.
Contribution
It presents a novel fragmentation scheme and a diffusion model that reassembles ligand fragments, achieving state-of-the-art docking success rates and surpassing classical methods.
Findings
Achieves Top-1 success rates above 79.9% on PoseBusters
Outperforms recent deep learning approaches in accuracy
Surpasses classical physics-based docking methods in reliability
Abstract
Determining the binding pose of a ligand to a protein, known as molecular docking, is a fundamental task in drug discovery. Generative approaches promise faster, improved, and more diverse pose sampling than physics-based methods, but are often hindered by chemically implausible outputs, poor generalisability, and high computational cost. To address these challenges, we introduce a novel fragmentation scheme, leveraging inductive biases from structural chemistry, to decompose ligands into rigid-body fragments. Building on this decomposition, we present SigmaDock, an SE(3) Riemannian diffusion model that generates poses by learning to reassemble these rigid bodies within the binding pocket. By operating at the level of fragments in SE(3), SigmaDock exploits well-established geometric priors while avoiding overly complex diffusion processes and unstable training dynamics. Experimentally,…
Peer Reviews
Decision·ICLR 2026 Poster
The motivation for fragment based docking is clear and compelling. The mathematical derivations are clear and straightforward albeit appear mostly in the appendices, not as part of the main paper.
There are a number of unsubstantiated and inconsistent claims, starting with the abstract. The presentation needs to be revised; any claims must accompany clear empirical or theoretical support. E.g., in section 2.2.2, the authors say that "... breaking the product structure [leads] to ill-conditioned and often degenerate dynamics during training and sampling". What's the evidence for ill-conditioned and degenerate dynamics? All claims of this kind require either supporting evidence or be remove
* The paper does a good job of motivating the problem formulation. The analysis of the right space to diffuse over for molecular docking is thoughtful, with theoretical results supporting the conceptual arguments. * The technical exposition and proofs in the appendix are clear and nicely done. * The preservation of performance at low sequence similarity to the training set is very nice to see.
**Significance** * The topic is somewhat stale, with wide consensus on Euclidean diffusion and co-folding for molecular docking. While contributions that challenge the consensus are welcome, they should provide a compelling value proposition rather than retread problem formulations that are no longer of primary interest. * The historical interest in docking to rigid receptors largely stems from computational considerations and works focusing only on this task should not be encouraged, as holo s
1. The paper is well-written, clearly motivated, and theoretically sound. It provides a strong argument for the limitations of torsional models and convincingly presents the fragment-based approach as a superior alternative. 2. The proposed method is novel and principled. The combination of the FR3D fragmentation scheme, soft geometric constraints, and an SE(3) diffusion process introduces strong and chemically-aware inductive biases into the model. 3. The empirical results are highly impr
1. The paper argues that its fragmentation scheme (FR3D) helps manage the degrees of freedom, but the analysis could be more thorough. 2. The definition of the protein binding pocket is a critical input for any docking model, yet it is not discussed in the main paper. Typically, this involves selecting atoms within a certain radius of the ground truth ligand, making this radius a key hyperparameter. The sensitivity of SIGMADOCK's performance to this pocket definition is not analyzed. An abl
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 Materials Science