2026-09-19
LDDM unifies 3D docking and design via fragment masking. PoseBusters success is 79.5%; one wet-lab cycle produced synthesizable, confirmed ligands on five targets.
Virtual screening can outrun a wet lab, but it still searches libraries that already exist or can be bought. Drug-like chemical space is at least a billion times larger, so exhaustive search does not scale. Generative models can build molecules from scratch. 2D SMILES and graph models never see pocket geometry directly. 3D generators do, and then often emit compounds nobody can make. Without synthesis, there is no experimental loop, and the method cannot improve on real binding data.
LDDM (Large Drug Discovery Model), from EPFL's LPDI group, puts local docking, covalent docking, fragment linking, fragment growing, and de novo design into one pocket-conditioned generator. Sampling can be restricted to on-demand chemical space such as Enamine REAL. The practical bet is a single design cycle that yields compounds you can order and assay.
The network is the geometric heterogeneous GNN from the same group's DrugFlow. Continuous atom coordinates are generated with equivariant flow matching; discrete atom and bond types with Markov bridges. Three extra binary flags mark whether a bond type, atom type, or coordinate is known. Training data is not the PDB's fewer than 50,000 non-redundant complexes. It is 836,259 pocket-ligand pairs assembled from CrossDocked, BindingNet v2, and BigBind, then filtered for geometry and chemistry.
The training objective is fragment-level masking, the molecular analogue of filling in missing tokens. Molecules are split with BRICS into retrosynthetically meaningful pieces. Each fragment is randomly assigned one of three roles:
At inference, the same checkpoint becomes a different tool by changing how much is given. Full graph yields docking. Pinning the covalent atom yields covalent docking. Two fragments yield a linker. Nothing given yields de novo design. Per-atom uncertainty, inherited from DrugFlow, ranks poses: 100 conformations per target, keep one.
Programmable design turns that conditioning into an iterative loop. Sample full molecules in the pocket, drop them on global filters such as QED, then score fragments on hydrogen bonds, geometry, and unsatisfied donors or acceptors. Accepted fragments land in a fragment tree; a UCB rule borrowed from Monte Carlo tree search allocates the next sampling budget. The synthesizable variant skips BRICS cuts. It matches generated molecules against a building-block library, then enumerates reachable products with reaction templates. Every step still emits a complete molecule and prunes afterward. Combinatorial docking grows incomplete R-groups and gets more expensive as catalogs grow. Here a larger virtual space makes a matching substructure easier to hit, so sampling can get cheaper.
In silico, top-1 success for docking is RMSD below 2 Å:
| Task | LDDM | Baseline |
| PoseBusters local docking | 79.5% | Boltz-2 67.7%, Vina 58.8% |
| Covalent docking | 80.1% | ICM-Pro 62.6% |
| Linker validity | 65% | about 22 points above DiffLinker |
| De novo molecule validity | 60.6% | FLOWR 57.8%, predecessor DrugFlow 25.9% |
Uncertainty tracks docking RMSD with Spearman ρ = 0.59. About 90% of samples sit below 0.5 maximum Tanimoto similarity to the training set. On local, structure-dependent goals such as adding hydrogen bonds versus the reference ligand, programmable generation produces more than four times as many accepted molecules as the structure-implicit RL baseline REINVENT. Across seven targets and a 4,000-sample budget each, only the synthesizable strategy consistently returns Enamine REAL members that pass every filter. One-shot sampling often returns almost none.
Five wet-lab campaigns, each a single design round:
| Target | Task | Made | Outcome |
| VHL-CDO1 molecular glue | Hold the hydroxyproline anchor, redesign both flanks | 9 | 67% kept nanomolar ternary affinity |
| Cathepsin S | Non-canonical amino acids on a covalent peptide inhibitor | 11 | IC50 70.4 nM to 19.9 nM |
| PGK1 | One-shot de novo, then buy from REAL | 6 | 4 hits; best Z856 KD 14.7 µM (CHR-6494 control 30 µM) |
| BRD4 | Synthesizable de novo | 8 | 3 hits; Z967 about 40 to 60 µM |
| SARS-CoV-2 Nsp3 Mac1 | Synthesizable de novo plus fragment growing | 57+31 | 6 compounds above 15% competition at 200 µM; best Z8035 IC50 46.2 µM |
Two Mac1 crystal structures: ligand RMSD 2.35 Å for Z8035 and 1.43 Å for Z8066. CACHE #3 screened 1,739 candidates and reported best novel hits around 18 µM; this paper assayed 57 de novo designs. NMR on PGK1 shows a Glu-319 shift consistent with the C-terminal nucleotide pocket that LDDM targeted.
For structure-based design, this is a single model that docks, links fragments, builds de novo, and can stay inside a catalog you can actually order from. The wet-lab bar is a small one-cycle list, not a thousand-compound lottery. Four of six purchasable PGK1 molecules bound. Mac1 crystals match the designed poses. Validity jumped from DrugFlow's 25.9% to 60.6%.
These are still hits, not drug candidates. The best BRD4 binder sits at 40 to 60 µM, three orders of magnitude behind (+)-JQ1 at 30 nM. The useful claim is the loop: generate, constrain to make-on-demand space, assay, and in the Mac1 case confirm the pose, all in one round.
The paper reports its misses. Three of five de novo KRAS ligands gave GCI signal, but the response amplitude is too high to rule out aggregation or nonspecific binding. Covalent Pin1 designs bound by GCI; the crystal pose of one compound disagrees with the prediction, which the authors blame on affinity too weak to beat a crystal contact with a neighboring subunit. Even the successes are mostly micromolar, small, and sparsely interacting. The authors themselves say better pose scoring and property predictors are needed before this reaches drug-like affinity at the hit stage.
The docking comparison is not fully fair. Pocket definitions change task difficulty; LDDM uses an 8 Å residue cutoff. AlphaFold 3 reaches similar docking numbers without a pocket or even a holo structure, though the supplement shows stronger memorization. Restricting generation to Enamine REAL trades away the open chemical space that is supposed to be the point of generative models versus screening. BRD4 and PGK1 hits have GCI and NMR, not crystals. Mac1's 46.2 µM is weaker than CACHE #3's roughly 18 µM. Only one of eleven peptide variants beat the parent sequence. Training complexes are synthetic; filters are not a substitute for experimental structures.