Research map / 3D vision / Face and shape
Template-Guided 3D Molecular Pose Generation via Flow Matching and Differentiable Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FM-MAPO, a template-guided ligand pose generator that uses a flow-matching alignment model followed by differentiable shape/pharmacophore/pocket/internal-energy optimization, and introduces the AlignDockBench benchmark.
Problem
Predicting 3D ligand conformations/poses in protein binding sites when a co-crystallized template ligand is available, especially for query ligands with low similarity to the template or high flexibility.
Approach
A flow-matching model conditioned on a 3D template ligand (graph with query and template atoms plus BRICS fragment nodes, harmonic prior) generates query poses aligned to the template; these are then refined by directly optimizing all atomic coordinates with differentiable Gaussian shape and pharmacophore Tanimoto scores, GAFF2 internal energy, and an optional pocket term.
Key results
- FM-MAPO+ achieves mean RMSD 1.62 Å and 77.78% of molecules below 2 Å on AlignDockBench cross-docking, versus 2.54 Å / 54.2% for LS-align and 3.39 Å / 47.15% for Vina.
- FM-MAPO (without the all-sample refinement) reaches 1.86 Å and 69.38% at 3.66 s per molecule.
- FM-MAPO and FM-MAPO+ outperform all baselines across Tanimoto similarity, Murcko scaffold similarity, atom-count and rotatable-bond strata.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add pose optimization to FMA (ablation) | FMA-MAPO · AlignDockBench cross-docking | mean RMSD (Å): 1.97 ± 1.36 -> 1.86 ± 1.42 | Helped |
| Refine all sampled poses before selection instead of refining only the top-ranked pose (ablation) | FMA-MAPO · AlignDockBench cross-docking | mean RMSD (Å): 1.86 ± 1.42 -> 1.62 ± 1.33 | Helped |
| Use FMA-PO+ instead of LS-align | AlignDockBench cross-docking · AlignDockBench | mean RMSD (Å): 2.54 ± 2.00 -> 1.62 ± 1.33 | Helped |
| Refine all sampled poses before selection (ablation) | FMA-MAPO · AlignDockBench cross-docking | average runtime (s): 3.66 -> 27.96 | Hurt |
Surprising findings
- ROSHAMBO, a GPU shape/pharmacophore alignment tool, has the lowest percentage below 2 Å (30.35%) among the reported methods, below even rDock docking.
- Vina and rDock are not template-aware, yet FM-MAPO+ is compared against them under a cross-docking setting with the template pocket.
Limitations
- Computationally more intensive than non-deep-learning methods, per the authors.
- Assumes rigid receptor; flexible receptor / induced-fit scenarios are not addressed.
- Benchmark pairs are filtered to STS > 0.5 and pocket RMSD < 1.2 Å, which may limit generality.
Experimental setup
Datasets: PDB same-pocket ligand pairs ChEMBL DUD-E DEKOIS PoseBusters
Benchmarks: AlignDockBench PoseBusters
Metrics: mean RMSD % molecules with RMSD < 2Å average runtime per molecule
Baselines: AutoDock Vina rDock FitDock LS-align ROSHAMBO
Builds on: flow matching DiffDock GAFF2 ROCS
Closest papers by meaning
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.88)
- Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design (ICML 2024 · similarity 0.88)
- Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge (ICML 2024 · similarity 0.86)
- 3D Molecule Generation from Rigid Motifs via $\mathrm{SE}(3)$ Flows (NeurIPS 2026 · similarity 0.86)
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · similarity 0.86)
- ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.