Structure-based drug design model that docks, links, grows and designs 3D ligands with one pocket-conditioned generator, wet-lab validated.
No providers recorded yet. Browse all providers
A structure-based design campaign is not one question but several. Where does this fragment sit in the pocket? What linker joins these two fragments? What replaces the R-group pointing into the hydrophobic patch? Each has had its own specialized model, featurization and checkpoint. LDDM (Large Drug Discovery Model) collapses the set into one pocket-conditioned generator by never fixing, during training, which parts of the molecule are known.
The training task is fragment-level masked modelling. Every training ligand is cut into retrosynthetically meaningful pieces with BRICS, and each piece is assigned one of three roles at random: designed (graph and coordinates both noised), docked (graph given, coordinates noised), or context (left untouched to condition the rest). A model trained across that space of masking patterns then performs a task at inference by choosing what to mask — a fully masked pocket is de novo design, a known graph with unknown coordinates is docking, a gap between two context fragments is linker design, and constraining one atom's position is covalent docking.
LDDM was developed at EPFL in the groups of Bruno Correia, Philippe Schwaller and Michael Bronstein, with experimental collaborators at UCSF, ETH Zurich and iBET, and posted to bioRxiv in September 2026. It succeeds DrugFlow from the same lab, whose network and generative machinery it reuses under the new objective.
The backbone is the E(3)-equivariant heterogeneous graph neural network introduced with DrugFlow, over ligand atoms and pocket residues with a 10 A interaction cutoff, retaining self-conditioning, the confidence head and virtual nodes; the input layer adds three binary flags marking whether each bond type, atom type and coordinate is known. Coordinates are generated by conditional flow matching and discrete atom and bond types by Markov bridges, with coordinate sampling optionally embedded in an SDE. Training used a curated set of 836,259 protein-ligand complexes covering 454,662 unique chemical structures, pooled from CrossDocked, BindingNet v2 and Gnina-docked BigBind records, filtered by PoseBusters and GenBench3D validity and deduplicated against the test set at 0.5 Tanimoto.
On PoseBusters local docking, LDDM reaches top-1 success under 2 A RMSD in 79.5% of valid cases, above Boltz-2 run with the same pocket conditioning, and it leads the Scarpino covalent docking benchmark out of the box. Around 65% of its linkers pass all quality filters, roughly 22% above DiffLinker, and over 60% of de novo designs pass the full PoseBusters and GenBench3D battery, about double DrugFlow. Over 90% of samples fall below 0.5 Tanimoto similarity to training molecules, and programmable generation yields more than four times as many molecules with improved hydrogen-bonding profiles as REINVENT.
The model suits hit finding and early optimization where designs must be ordered rather than admired: R-group exploration around a scaffold, linker design in fragment campaigns, pose prediction for covalent warheads, and de novo generation restricted to a make-on-demand catalog. The peptide case study extends it to non-natural amino acid substitution, and the confidence head ranks candidates before synthesis.
What distinguishes the work is the scale of prospective validation. Redesigning both flanks of a VHL-CDO1 molecular glue, 67% of LDDM's modifications retained nanomolar ternary-complex affinity, and a non-natural amino acid substitution improved a cathepsin S peptide inhibitor from 70.4 nM to 19.9 nM in one round. De novo designs gave a 14.7 uM PGK1 binder confirmed by NMR, a BRD4 ligand in the acetyl-lysine pocket, and a SARS-CoV-2 Nsp3 Mac1 inhibitor at 46.2 uM whose crystal structure matches the designed pose within 2.35 A. The authors are candid about the ceiling: affinities stay micromolar, a KRAS series gave signals they could not separate from artifacts, and a Pin1 design crystallized in an unpredicted binding mode. Code is MIT-licensed and two checkpoints are archived, the BindingNet-free one under MIT and the main one under CC BY-NC 4.0; Enamine REAL cannot be redistributed, so the repository ships a smaller open building-block library.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.