Flow-matching 3D molecule generator conditioned on conformational ensembles, composing shape, pharmacophore and pocket conditions at inference.
No providers recorded yet. Browse all providers
An agonist and an antagonist can occupy the same receptor pocket; what separates them is which receptor conformation the ligand stabilises. A dual-target binder has the mirror-image requirement — one molecule that adopts a bioactive pose in two unrelated pockets. Neither is visible to a generative model that places one conformer in one site, because both are properties of the distribution of shapes a molecule can access rather than of any single pose.
EnsCondFlow reframes molecular design as an optimisation over both the modes and the properties of a molecule's conformational ensemble. Modes are the shapes the ensemble should visit — a shape profile, a pharmacophore profile or a protein pocket; properties are aggregate scalars over the whole distribution, such as mean polar surface area or mean pairwise RMSD between conformers. The framework conditioning on both axes at once is called ensemble-conditioned guidance: each condition is encoded once and run through the generator for its own endpoint prediction, and those predictions are combined at every integration step with weights that may be positive (target this mode) or negative (avoid it). Because composition happens at sampling time rather than in the loss, one checkpoint covers combinations the training data never contained.
The work comes from the Molecular AI group in Discovery Sciences at
AstraZeneca in Gothenburg with the Department of Computer Science
and Engineering at Chalmers University of Technology, posted to arXiv
in September 2026 under the paper title "Ensemble-Conditioned Molecular Design"; the code package
is enscondflow. It sits beside the same group's FlexiFlow, which samples
several conformers per molecule, and trains on the SPINDR corpus from the FLOWR line
(FLOWR.root).
The network holds roughly 58.6M learnable parameters: a shared 8-layer profile encoder for shape and pharmacophore points (2.5M), a pocket encoder of the same size, a property embedder (0.08M), and a 16-layer generator of hidden dimension 384 (53.5M) interleaving cross-attending transformer layers with graph transformer layers under adaptive layer normalisation. Generation is continuous and categorical flow matching, predicting coordinates, atom types and bonds jointly. Training was one run over GEOM Drugs (300K molecules with CREST conformer ensembles) and SPINDR (35K PDB- derived protein-ligand complexes, filtered to 28K by QED), 200 epochs in bf16 on a single A100 in about two days, with conditions independently masked. On the multi-mode shape benchmark (1K held- out scaffolds, 496 targeting and 535 avoidance pairs), molecules shift clearly toward the target shape against a size-matched virtual screening baseline, with connected validity above 0.97 and uniqueness above 0.99 in all but one run. Across 179 SPINDR test systems, raising the PSA target from 80 to 140 Ų moves the achieved mean from 111 to 163 Ų while interaction recovery holds near its 0.954 baseline.
In a dual-target study against acetylcholinesterase and monoamine oxidase B, conditioning on donepezil's and safinamide's pharmacophores together produced 23% of molecules docking within 1 kcal/mol of both reference drugs, against 9% and 7% for either reference alone. In active-state- selective design on the A2A adenosine receptor, adding the inactive-state pocket as an avoided condition raised the mean docking preference for the active state from +0.47 to +1.26 kcal/mol while generated chemistry stayed distant from NECA. Off-target selectivity and scaffold hopping use the same machinery.
Ensemble-conditioned guidance makes the conformational distribution, rather than a single docked pose, the object a generative model is steered by — and does it at test time, so a new design task is a new set of conditions rather than a new training run. The stated limits: property control is directional rather than calibrated, compressed toward the unconditioned baseline at the extremes of a sweep; sampling cost grows linearly with the number of mode conditions; invariantly encoded conditions pull more weakly than the one defining the reference frame; and docking is the binding oracle in both case studies, with no wet-lab follow-up. Code is MIT-licensed and the checkpoint and datasets are on Zenodo; the preprint is not yet peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.