SigmaDock: Untwisting Molecular Docking With Fragment-Based SE(3) Diffusion

AI summary

Overview

Research area: Computational drug discovery, specifically molecular docking, combining geometric deep learning with Riemannian diffusion models on the SE(3) manifold.

Technical level: Advanced. The paper assumes familiarity with Lie groups (SE(3), SO(3)), Riemannian score-based diffusion, equivariant graph neural networks, and structural chemistry concepts such as torsions, dihedrals and holonomic constraints.

Scope: The paper proposes a fragment-based SE(3) diffusion model, SigmaDock, for rigid-receptor re-docking, and evaluates whether decomposing ligands into rigid bodies yields more reliable and generalisable poses than torsional diffusion or classical physics-based docking.

What This Paper Is About

Molecular docking predicts how a small molecule (ligand) sits inside a protein's binding pocket. Recent deep learning methods for this task are often reported as beating classical docking software, but independent work has shown those gains shrink or disappear once chemical plausibility of the generated poses is checked. The authors argue that existing deep learning docking models are either too data- and compute-hungry (co-folding models such as AlphaFold3) or poorly conditioned (torsional diffusion models), and they set out to fix the torsional approach by diffusing over rigid-body fragments instead of torsion angles.

Key Contributions

1. A fragmentation scheme (FR3D) for SE(3) diffusion. Ligands are decomposed into rigid-body fragments by breaking rotatable bonds, then adjacent fragments are recursively and stochastically merged until an irreducible set of size m is reached. This reduces the learnable degrees of freedom relative to the naive torsion-free decomposition of m̂ = (k + 1) fragments, and provides a stream for data augmentation. The number of fragments is bounded by 1 ≤ m ≤ k + 1.

2. Soft triangulation geometric constraints. For each torsional bond connecting adjacent fragments, triangles (A, B, C) and (B, C, D) are defined using neighbouring atoms. The authors prove (Lemma 1) that conditioning on cross-fragment distances ||A − C|| and ||B − D|| uniquely determines the relevant bond lengths and bond angles without restricting changes in dihedral angles. The distance mismatch from the reference RDKit conformer is fed as an edge feature, vanishing as t → 0 so that only dihedrals and rigid-body roto-translations remain free.

3. An SO(3)-equivariant architecture tailored to fragment geometry. SigmaDock augments EquiformerV2 with virtual nodes and edges to build a hierarchical topology, role-specific node and edge featurisations, and messages that decay smoothly to zero near a cutoff. An equivariant prediction head based on Newton-Euler equations from rigid-body mechanics resolves the non-uniqueness of local fragment coordinate frames; Theorem 2 states that training and sampling are invariant to the choice of local coordinate axes, and that the score model is SO(3)-equivariant.

4. An argument for why torsional models underperform. Theorem 1 states that standard molecular topologies make torsional updates a nonlinear mapping to Cartesian coordinates, inducing a highly entangled, non-product measure, whereas disjoint rigid fragments yield a factorised product of Haar measures on SE(3)^m. This is offered as the explanation for the discrepancy between the promise and the empirical results of torsional models.

Main Findings

• State-of-the-art PoseBusters performance: SigmaDock reaches Top-1 success rates (RMSD < 2 Å and PB-valid) above 79.9% on the PoseBusters set, compared to 12.7–32.8% reported by recent deep learning approaches.

• First deep learning method to beat physics-based docking under this split: The authors state that SigmaDock is the first deep learning approach to surpass classical physics-based docking under the PB train-test split.

• Generalisability to unseen proteins: The model demonstrates consistent generalisation to unseen proteins (Figure 4), with PoseBusters(v2) serving as a temporal-split validation set of 308 protein-ligand complexes with unseen protein sequences realised from 2021 onwards.

• AF3-level accuracy at lower cost: With a fraction of the training data, training and sampling time, and lower test-train leakage, the authors report reaching AlphaFold3-level performance and substantially outperforming previous generative methods on the re-docking task.

• Validity of the fragmentation assumption: Aligning ligand conformations from the conformational manifold to the bound manifold via joint roto-translational and torsional registration gives RMSDs substantially below both standard reported docking error rates and the 2 Å success threshold. An example is shown for ligand BFL in PDB 1Q4G with an optimised alignment RMSD of 0.11Å, which the authors use to justify treating bond length and bond angle variation as negligible for docking.

• Fragment reduction in practice: FR3D empirically reduces the number of fragments to approximately m ≈ 2/3 m̂.

• No separate confidence model needed: Because of SigmaDock's reliability in generating chemically plausible samples, the authors propose ranking N_seeds samples with a cheap heuristic based on pseudo binding energy and physicochemical checks (bond angles, bond lengths, internal energy) instead of a separately trained confidence model.

Methodology in Plain English

Rather than treating a ligand as a cloud of atoms or as a chain of rotatable bonds, the authors cut it into rigid chunks connected by rotatable bonds. Each chunk keeps a fixed internal shape, so the only things the model has to predict are where each chunk sits (position) and how it is rotated (orientation) — an element of SE(3) per chunk.

The generative model is a diffusion model defined on that product space. A forward process gradually adds noise by translating and rotating each fragment independently until the fragments are scattered according to a known stationary distribution: Gaussian noise for positions and the uniform distribution on SO(3)^m for orientations. A neural network is trained by score matching to reverse this process, and sampling runs the reverse SDE to reassemble the fragments into a predicted bound pose.

To keep the problem well posed, the authors merge fragments where possible (FR3D), keep the retained bond lengths and angles by adding dummy atoms at either side of broken bonds, and prune dummy atoms that become over-constrained once their bond is merged. They also add cross-fragment distance constraints that implicitly pin down bond angles without freezing dihedrals. The network itself is an SO(3)-equivariant graph neural network built on EquiformerV2, with a prediction head derived from rigid-body mechanics so the model's output does not depend on how each fragment's local coordinate frame was arbitrarily oriented.

The experimental setting is deliberately conservative: the receptor is held fixed in its holo conformation and the binding pocket is known, i.e. rigid-receptor re-docking. Training uses PDBBind(v2020), a curated set of 19,443 crystallographically obtained protein-ligand complexes, and the authors restrict themselves to this dataset for fair comparison. Validation uses PoseBusters(v2) (308 complexes, unseen sequences from 2021 onwards) and the Astex dataset of 85 diverse, highly curated complexes. The binding pocket is defined as all residues with any atom within a stochastic cutoff around any ligand atom, where the cutoff is d_0 + N(0, σ_r) with d_0 defaulting to 5Å and σ_r to 1Å during training.

Why This Matters

Impact on research. The paper reframes the deep-learning-versus-physics debate in docking by showing that the weakness of torsional diffusion models may be a modelling artefact rather than an inherent limit of learned docking, and by demonstrating gains under conditions that include chemical plausibility checks (the PB-valid criterion) rather than RMSD alone. It also provides a formal argument (Theorem 1) for why the torsional parametrisation induces entangled measures, which is a claim other geometric deep learning work can test and build on.

Real-world applications.

• High-throughput virtual screening, where millions of protein-ligand pairs must be scored and faster inference is the bottleneck.
• Lead optimisation, where rigid-receptor docking is standard industrial practice and pose reliability directly affects the decision to synthesise a compound.
• Structure-based drug design campaigns that need chemically valid poses without training their own large co-folding model.
• Benchmarking and validation tooling, since the paper argues for evaluations that combine RMSD with physicochemical validity rather than RMSD alone.

Industry relevance. The authors explicitly frame their choice of a fixed holo receptor with a known pocket as reflecting industrial HTVS and lead-optimisation practice, where rigid-receptor docking is tractable at scale. Their emphasis on reduced training data, reduced compute, and no separate confidence model is aimed at making the method usable by groups that cannot train co-folding-scale models. The claim of surpassing classical physics-based docking under the PoseBusters train-test split is the headline result for practitioners already relying on such software.

Future Directions

• Flexible docking. The authors note that the fragment SE(3)^m formulation extends naturally to flexible docking by treating selected protein side chains as additional fragments, and leave this to future work.
• Adaptation to co-folding. The same formulation is presented as adaptable to co-folding, where the protein is not assumed rigid, which the authors also leave as future work.
• Data augmentation via FR3D. Because FR3D performs a stochastic search over merge actions rather than biasing a single fragmentation order, the authors describe it as a promising stream for data augmentation; how much this helps in practice is left open.
• Better characterisation of the degrees-of-freedom trade-off. The paper gives bounds (k + 6 ≤ DoF ≤ 6m̂ under naive fragmentation, narrowing as FR3D reduces m) and points to Appendix D.4 for a more extensive analysis, leaving the practical calibration of the fragmentation level as an open question.

Target Audience

Researchers and practitioners in computational chemistry, structural bioinformatics and geometric deep learning who are already familiar with diffusion models and equivariant architectures. It will be most useful to those working on molecular docking benchmarks, those building generative models on Lie groups such as SE(3), and industrial scientists evaluating whether a learned docking method is reliable enough to replace or supplement physics-based docking in virtual screening pipelines. Readers without a background in Riemannian geometry or structural chemistry will find the method sections demanding, though the motivation and experimental results are accessible.

Authors’ 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, we show SigmaDock achieves state-of-the-art performance, reaching Top-1 success rates (RMSD&lt;2 &amp; PB-valid) above 79.9% on the PoseBusters set, compared to 12.7-30.8% reported by recent deep learning approaches, whilst demonstrating consistent generalisation to unseen proteins. SigmaDock is the first deep learning approach to surpass classical physics-based docking under the PB train-test split, marking a significant leap forward in the reliability and feasibility of deep learning for molecular modelling.