Template-Guided 3D Molecular Pose Generation via Flow Matching and Differentiable Optimization

TL;DR

The paper tackles template-guided 3D pose generation for small molecules in protein pockets by combining a flow-matching based alignment model with differentiable coordinate optimization. It develops FM-MA, which conditions conformer generation on a 3D template ligand via a velocity-field predictor and harmonic prior, followed by a differentiable pose optimization that jointly optimizes shape, pharmacophore, pocket complementarity, and internal energy terms. A new AlignDockBench benchmark of 369 template–query pairs is introduced, and experiments show that the proposed method (FMA and FMA-PO/+) outperforms traditional docking tools and open-access alignment baselines, especially in low-template-similarity and highly flexible ligands. The work highlights the value of incorporating known ligand geometries into generative modeling to improve pose prediction, with implications for data augmentation and downstream predictive tasks, and outlines future directions toward receptor-aware conditioning and flexible receptor scenarios, while noting runtime considerations for large-scale applications.

Abstract

Predicting the 3D conformation of small molecules within protein binding sites is a key challenge in drug design. When a crystallized reference ligand (template) is available, it provides geometric priors that can guide 3D pose prediction. We present a two-stage method for ligand conformation generation guided by such templates. In the first stage, we introduce a molecular alignment approach based on flow-matching to generate 3D coordinates for the ligand, using the template structure as a reference. In the second stage, a differentiable pose optimization procedure refines this conformation based on shape and pharmacophore similarities, internal energy, and, optionally, the protein binding pocket. We introduce a new benchmark of ligand pairs co-crystallized with the same target to evaluate our approach and show that it outperforms standard docking tools and open-access alignment methods, especially in cases involving low similarity to the template or high ligand flexibility.

Figures (16)

• Figure 1: Overview of the pipeline. The method comprises two stages: (1) 3D of the ligand to the template; and (2) Pose Optimization, including shape and pharmacophores scores, energy minimization, and optional refinement based on protein pocket information.
• Figure 2: Evolution of the query ligand's conformation (in blue) during the FM-MA denoising process. The template ligand is shown in pink. Starting from a noisy initial conformation ($t=0$), the model progressively refines the 3D coordinates to align the query ligand with the template pose ($t=1$).
• Figure 3: Performance of molecular alignment and docking methods on AlignDockBench in a cross-docking scenario. Bar plots (left y-axis) indicate the percentage of molecules with RMSD below $2$Å, while the line plot (right y-axis) shows the number of molecules in each bin. Each figure stratifies the results based on a different structural or chemical property of the query ligand.
• Figure S1: Illustration of FMA model pipeline.
• Figure S2: Molecular graph $\mathcal{G}$ with query and template ligand. Blue nodes represent atoms, while orange nodes denote functional groups. The graph incorporates multiple edge types: purple edges for covalent bonds, green edges for dense interconnections, and orange edges linking functional groups to their constituent atoms.