FMA-PO

FMA-PO — Flow Molecular Alignment with Pose Optimization — predicts the three-dimensional conformation a small molecule adopts inside a protein binding site, using an already-solved co-crystal structure of a related ligand as a geometric template. This is the situation most medicinal chemistry projects are in: once one compound has been crystallized against the target, every subsequent analogue can borrow that pose as a prior. Classical ligand-based superposition tools exploit exactly this, but they compare shapes and pharmacophores without knowing anything about the receptor, and they degrade sharply when the new molecule diverges chemically from the template.

The method splits the problem in two. A flow-matching generative model, FMA, produces 3D coordinates for a query molecule from its 2D graph, conditioned on the template ligand's structure. A second, non-learned stage then refines those coordinates by differentiable optimization against shape overlap, pharmacophore similarity, internal energy, and — optionally — complementarity to the binding pocket. Only the first stage is trained; it is fitted once on a broad corpus and applied to arbitrary new protein-ligand pairs at inference without per-target retraining.

FMA-PO was developed by Noémie Bergues, Arthur Carré and colleagues at Iktos, with academic collaborators at Institut Pasteur and Université Paris Cité, and was accepted at NeurIPS 2025. The work also contributes AlignDockBench, a public benchmark of template-query ligand pairs co-crystallized against the same target.

#Key Features

• Template conditioning: The query and template ligands are encoded in a single joint graph, so attention lets the reference conformation inform every atom of the generated pose without requiring an explicit atom-to-atom mapping between the two molecules.
• Harmonic prior: Initial coordinates are drawn from a distribution built on the molecular graph Laplacian rather than from Gaussian noise, so bonded atoms begin near each other and the flow spends its capacity on placement, not on rediscovering connectivity.
• Coordinate-level refinement: The optimization stage updates all atomic positions directly rather than rigid-body transforms or torsion angles, allowing fine-grained correction of both placement and internal geometry.
• Optional receptor awareness: Pocket complementarity enters only as a refinement term, so the same trained model serves both ligand-based and structure-aware workflows.
• Robustness to dissimilar templates: Accuracy holds up in the low-Tanimoto and high-flexibility bins of AlignDockBench, where classical alignment tools fall off most steeply.

#Technical Details

FMA is a compact 3.5M-parameter graph neural network. Nodes represent heavy atoms of both ligands plus BRICS-derived functional-group nodes; edges encode covalent bonds, atom-to-fragment membership, and radial-basis-encoded distances between fragments within and across the two molecules. Features pass through a multi-head attention encoder with edge bias, then a time-conditioned vector field network predicts the flow-matching velocity, integrated with Euler steps at sampling time. Training used 301,348 co-binding ligand pairs covering 111,678 unique molecules from the PDB, augmented with ChEMBL compounds following the BindingNet protocol; any training molecule with Morgan-fingerprint Tanimoto above 0.5 to a benchmark ligand was removed to prevent leakage.

AlignDockBench comprises 369 query complexes mapped to 61 diverse templates spanning kinases, proteases, nuclear receptors, and GPCRs, with query pockets required to align to the template below 1.2 Å backbone RMSD. On this benchmark FMA alone reaches 1.97 Å mean RMSD with 64.8% of poses under 2 Å in 0.83 s per molecule; adding refinement to the top-ranked pose (FMA-PO) gives 1.86 Å and 69.4%; refining all sampled poses before selection (FMA-PO+) gives 1.62 Å and 77.8% at 28 s per molecule. The strongest baselines are LS-align at 2.54 Å / 54.4%, FitDock at 2.93 Å, ROSHAMBO at 2.87 Å / 30.4%, AutoDock Vina at 3.39 Å / 47.3%, and rDock at 4.52 Å / 34.4%. FMA and its variants are the only methods that produce a pose for all 369 queries.

#Applications

The natural home for FMA-PO is hit-to-lead and lead optimization, where a co-crystal structure already exists and chemists need reliable 3D poses for many analogues — for free-energy calculations, pharmacophore modeling, or inspection of a proposed modification. It suits 3D ligand-based virtual screening equally well, with a known active's conformation seeding the search across a compound library. AlignDockBench fills a gap of its own: existing cross-docking benchmarks either omit reference ligands entirely or restrict themselves to chemically near-identical pairs.

#Impact

FMA-PO shows that a small conditional generative model paired with a physics-informed refinement stage can beat established docking software on template-guided pose prediction while running one to two orders of magnitude faster in its single-pose configuration. Splitting learning from scoring is deliberate: the flow model handles the combinatorially hard placement problem, while interpretable, tunable terms enforce chemical validity. Two caveats bound its use. The trained FMA weights and inference code have not been released, so only AlignDockBench — on Zenodo under CC BY-NC-ND 4.0 — is reproducible from the paper. And the method requires a co-crystallized template, leaving genuinely novel targets to conventional docking.