{"id":14046,"date":"2026-03-25T14:19:02","date_gmt":"2026-03-25T14:19:02","guid":{"rendered":"https:\/\/www.blopig.com\/blog\/?p=14046"},"modified":"2026-03-29T13:19:40","modified_gmt":"2026-03-29T12:19:40","slug":"sigmadock-untwisting-molecular-docking-with-fragment-based-se3-diffusion","status":"publish","type":"post","link":"https:\/\/www.blopig.com\/blog\/2026\/03\/sigmadock-untwisting-molecular-docking-with-fragment-based-se3-diffusion\/","title":{"rendered":"SigmaDock: untwisting molecular docking with fragment-based SE(3) diffusion"},"content":{"rendered":"\n

Alvaro Prat, Leo Zhang, Charlotte Deane, Yee Whye Teh, & Garrett M. Morris
International Conference On Learning Representations (ICLR 2026)<\/a><\/em><\/p>\n\n\n\n

Molecular docking sits at the heart of structure-based drug discovery. If we can reliably predict how a small molecule binds in a protein pocket, we can prioritize compounds faster, reason about interactions more clearly, and build better pipelines for hit discovery and lead optimization. But in practice, docking is still a difficult problem: classical methods are often robust but imperfect, while recent deep learning approaches have sometimes looked promising on headline metrics without consistently producing chemically plausible poses.<\/p>\n\n\n\n

SigmaDock was built to address exactly that gap. Instead of treating docking as a problem of directly diffusing on torsion angles or unconstrained atomic coordinates, SigmaDock represents ligands as collections of rigid fragments and learns how to reassemble them inside the binding pocket using diffusion on SE<\/mtext>(<\/mo>3<\/mn>)<\/mo><\/mrow>\\text{SE}(3)<\/annotation><\/semantics><\/math>. In plain English: rather than trying to \u201cwiggle\u201d every flexible degree of freedom in a tangled way, SigmaDock breaks the ligand into chemically meaningful rigid pieces and learns where those pieces should go, and how they should reorient, to recover a valid bound pose.<\/p>\n\n\n
\n
\"\"<\/a>
Figure 1:<\/strong> Illustration of SigmaDock using PDB 1V4S and ligand MRK. We create an initial conformation of a query ligand where we define our m<\/mi>m<\/annotation><\/semantics><\/math> rigid body fragments (colour coded). The corresponding forward diffusion process operates in SE<\/mtext>(<\/mo>3<\/mn>)<\/mo>m<\/mi><\/msup><\/mrow>\\text{SE}(3)^m<\/annotation><\/semantics><\/math> via independent roto-translations.<\/figcaption><\/figure>\n<\/div>\n\n\n\n\n\n\n

Why fragment-based diffusion?<\/strong><\/h2>\n\n\n\n

A central observation behind SigmaDock is that torsional models are elegant in theory, but difficult in practice. Small changes in torsion angles can create large, non-local changes in Cartesian coordinates, which makes the learning problem highly entangled. SigmaDock avoids that by moving to fragment space: each rigid fragment is handled as a body in SE<\/mtext>(<\/mo>3<\/mn>)<\/mo><\/mrow>\\text{SE}(3)<\/annotation><\/semantics><\/math>, and the model learns rigid-body translations and rotations rather than directly predicting torsional updates. The result is a cleaner geometric parameterization and a more stable generative problem.<\/p>\n\n\n\n

This is paired with a fragmentation strategy called \u201cFR3D<\/strong>\u201d, which does more than just cut every rotatable bond. A na\u00efve split would often create too many fragments and inflate the degrees of freedom. FR3D recursively merges adjacent fragments to reach an irreducible representation, reducing the number of independent rigid bodies while preserving the chemistry that matters. On top of that, SigmaDock adds soft triangulation constraints<\/strong> so the model keeps bond lengths and bond angles consistent across fragment boundaries \u2014 without locking the dihedral degrees of freedom it still needs to explore.<\/p>\n\n\n

\n
\"\"<\/a>
Figure 2<\/strong>: (a) Illustration of a dihedral \u03d5ABCD<\/sub> across torsional bond BC, defined as the angle between planes ABC and BCD, across two adjacent benzene rings in ligand BFL; (b) Bound (red) and aligned (green) poses for BFL in PDB 1Q4G with an optimised alignment RMSD of 0.11 \u00c5; C: Conformational ensembles generated from \u03c0Mc<\/sub>\u00a0for ligands SKF, CEL, and IH5 respectively. Notably, the most significant structural changes are derived from torsions across the rotatable bonds.<\/figcaption><\/figure>\n<\/div>\n\n
\n
\"\"<\/a>
Figure 3<\/strong>: Illustrative example of how FR3D reduces the number of fragments (colour coded) required to represent rigid bodies on ligand TNK into irreducible form. (a) Defining fragments by breaking all torsional bonds (ribbons); (b) FR3D recursively attempts to reduce the k<\/mi>k<\/annotation><\/semantics><\/math> torsional
bonds and removes over-constrained dummies in the process (denoted by the coloured rings), which otherwise define a dihedral across the merged fragment; (c) Over-constrained dummies removed and triangulation edges displayed under a different stochastic reduction (equiprobable to the fragmentation shown in b).<\/figcaption><\/figure>\n<\/div>\n\n\n

What makes SigmaDock different architecturally?<\/strong><\/h2>\n\n\n\n

At the modelling level, SigmaDock uses an SE(3)<\/strong> Riemannian diffusion process<\/strong> over fragment translations and rotations, together with an SO(3)-equivariant architecture<\/strong> built on top of EquiformerV2-style reasoning for protein\u2013ligand geometry. The model also augments the graph with virtual nodes and edges to improve long-range information flow, and uses pseudo-force predictions to make the score parameterization invariant to arbitrary choices of local fragment coordinate axes.<\/p>\n\n\n\n

That sounds abstract, but the practical point is simple: SigmaDock was designed so that the geometry, chemistry, and inductive biases all line up with the actual structure of the problem. The paper\u2019s main message is not just that a bigger model can do better docking, but that a better geometric formulation can make the learning task fundamentally easier and more reliable.<\/p>\n\n\n\n

How well does it work?<\/strong><\/h2>\n\n\n\n

In the paper, SigmaDock is evaluated on the standard re-docking setup with a fixed receptor and known pocket. On the PoseBusters split, the reported Top-1 success rate with both RMSD < 2 \u00c5 and PB-validity is 79.9%<\/strong>, and the paper presents SigmaDock as the first deep-learning docking method to surpass classical docking under the intended PoseBusters train\/test split<\/strong>. The paper also reports strong generalization to unseen proteins and near-perfect Top-1 performance above 90% on the Astex Diverse Set.<\/p>\n\n\n\n

Just as importantly, the paper emphasizes that SigmaDock achieves these results without relying on a separate confidence model<\/strong> to filter poor generations. Instead, it uses lightweight post hoc<\/em> ranking based on pseudo-binding energy and physicochemical checks. The paper also reports that performance is competitive with AF3-level docking performance while using a much smaller training set and substantially faster sampling.<\/p>\n\n\n

\n
\"\"<\/a>
Figure 4<\/strong>: Performance benchmarks. Left: Comparative performance of SigmaDock\u00a0on the PoseBusters v2 and Astex Diverse Sets against other methods. Extracted from\u00a0Abramson et al.\u00a0(2024);\u00a0Buttenschoen et al. (2024). (*) Denotes classical docking; (**) Are not open-source. Right: Performance breakdown across sequence similarity splits in the PB set.<\/figcaption><\/figure>\n<\/div>\n\n\n

What\u2019s in the codebase?<\/strong><\/h2>\n\n\n\n

The GitHub repo is the official implementation for training, sampling, and evaluation<\/strong>. It includes:<\/p>\n\n\n\n