ED2Mol

ED2Mol is a structure-based generative model for small-molecule drug design that conditions on electron density rather than on pocket geometry alone. Most pocket-aware generators place atoms or fragments by reasoning about neighbouring protein atoms; ED2Mol instead first predicts what the ligand's electron density in that pocket ought to look like, then builds a molecule that fills it. The idea is borrowed from crystallography, where tools such as LigandFit place a rigid core into an experimental density map and iteratively attach the rest of the ligand to the remaining density.

The motivation is a practical failure mode of generative drug design: models routinely produce molecules with excellent docking scores but implausible physics — poses that clash with the receptor, bond geometries that no chemist would draw, or scaffolds that collapse on re-docking. Conditioning on a density field gives the generator both the external shape complementarity of the pocket and an internal, physically grounded target for where atoms should sit.

ED2Mol was developed by the Medicinal Chemistry and Bioinformatics Center at Shanghai Jiao Tong University School of Medicine with academic and industry collaborators, first posted as a preprint in December 2024 and published in Nature Machine Intelligence in 2025. Code and trained weights are released under an MIT licence.

#Key Features

• Electron-density imputation: A variational autoencoder predicts the ligand electron density a pocket would accommodate, portraying a space of "pseudo-ligands" before any atoms are placed.
• Stratified core placement: A peak-density search algorithm seats a rigid core fragment into the strongest region of the imputed density, mirroring how crystallographers fit a ligand into an experimental map.
• Equivariant fragment extension: An equivariant graph neural network selects the hydrogen to grow from and predicts the torsion angle of the new bond, attaching fragments one at a time into unoccupied density.
• Fragment-based lead optimization: Supplying a known active as a pre-located core skips the search step, letting the same model grow or replace fragments around an existing scaffold.
• Allosteric-pocket generalization: The fragment protocol produces compounds suited to the small, superficial pockets typical of allosteric sites, not only conserved orthosteric ones.

#Technical Details

The density-imputation model was trained on 49,987 high-quality pocket–ligand pairs filtered from roughly 780,000 PDB ligand-binding pockets, split 7:1.5:1.5 with test proteins held below 70% sequence identity to training and validation proteins, and with any pocket sharing more than 70% structural similarity to a benchmark pocket excluded. The fragment-extension networks were trained on drug-like ZINC molecules decomposed along rotatable bonds into growth trajectories, giving a growth-point dataset of 2,986,278 samples from 94,519 molecules and a torsion-angle dataset of 17,978,368 samples from 4,497,302 molecules, with torsions discretized into 36 bins.

Evaluation covered DUD-E (94 orthosteric pockets) and ASB-E (112 allosteric pockets curated from ASBench), generating 1,000 molecules per pocket against Pocket2Mol, GraphBP, ResGen, FLAG, and TargetDiff. ED2Mol reached ligand-efficiency success rates of 67.3% on DUD-E and 68.9% on ASB-E; 97.3 ± 0.4% of its poses passed the 18 PoseBusters validity criteria, against 63.1 ± 0.4% for the next-best model; and its combined "reliable success rate" — efficient, PoseBusters-valid, and stable under re-docking — was 34.4 ± 0.5%, roughly double TargetDiff at 17.1 ± 1.0% and Pocket2Mol at 15.0 ± 2.0%.

#Applications

The workflow points a frozen model at a user-supplied receptor PDB and pocket centroid, so it slots into hit finding and lead elaboration without per-target training. The authors ran four real drug design campaigns: de novo generation against FGFR3 and CDC42, and lead optimization by fragment replacement on the RNA helicase Brr2 and fragment growing on PPARγ. Wet-lab assays confirmed FGFR3 orthosteric inhibitors, CDC42 allosteric inhibitors, and GCK allosteric activators, and an X-ray co-crystal structure validated a directly generated binding mode.

#Impact

ED2Mol makes a concrete case that a physically meaningful intermediate representation — electron density — buys reliability that pure geometry-conditioned generation has struggled to deliver, and its ASB-E benchmark gives the field a harder allosteric test than the orthosteric-heavy DUD-E set. The released weights and MIT-licensed code make the method directly reusable, though the paper itself carries a more restrictive CC BY-ND preprint licence and the release does not report a parameter count.