The generative model LDDM designed binding molecules for five proteins, and X-ray crystallography confirmed its 3D prediction for one target

The generative model LDDM designed binding molecules for five proteins, and X-ray crystallography confirmed its 3D prediction for one target

On September 18, a preprint from an EPFL team appeared on bioRxiv describing LDDM, a generative model that designs drug molecules to fit the structure of a target protein. The authors synthesized and tested the model's proposed compounds against five different proteins: binding was confirmed in all five cases, and for one target X-ray crystallography showed that the molecule's actual pose in the binding pocket closely matched the prediction.

Computational models have long proposed molecules tailored to specific proteins, but they almost always hit the same wall: the designed molecule is difficult to synthesize, and the cycle of designing, testing, and refining the model never starts. AdaptiveFlow tackles this by screening billions of ready-made compounds; LDDM takes a different route by embedding synthesizability into the generation process itself. The molecule grows fragment by fragment, and at each step the model accepts only variants built from commercially available chemical building blocks. Training followed a technique borrowed from language models: fragments of known molecules were masked, and the model learned to reconstruct them from the shape of the binding pocket. The training set comprised 836,000 examples, 16 times the total number of resolved protein-ligand structures in the world's principal structural database.

Testing began with modification of an existing compound. A molecular glue binds two proteins at once and marks one for destruction: it engages VHL, which tags proteins for degradation, thereby directing the cell to destroy CDO1. LDDM redesigned both flanks of the glue at the CDO1 interface, and 67% of the synthesized variants retained nanomolar binding affinity. For a peptide inhibitor of cathepsin S, an enzyme involved in cancer, the model (never trained on peptides) improved binding 3.5-fold by substituting a single amino acid, from 70.4 to 19.9 nM.

Next came fully de novo design, without an existing template. For PGK1, an energy metabolism enzyme linked to cancer, Parkinson's disease, and ALS, the best LDDM molecule bound more tightly (14.7 μM) than the control compound (30 μM), as confirmed by nuclear magnetic resonance. BRD4 is another cancer target: its leading inhibitor acts for only 0.9 hours, and alternatives carry a risk of severe thrombocytopenia. For this target, LDDM designed a molecule with an affinity of 40 to 60 μM, and NMR confirmed that its binding mode matched the prediction.

The most stringent test involved Mac1, a SARS-CoV-2 enzyme the virus uses to suppress the host cell's immune defense. An open competition had already been held for this target: 23 teams screened 1,739 molecules and achieved a best result of approximately 18 μM. LDDM tested only 57 molecules and found a weaker but chemically distinct hit at 46.2 μM. X-ray crystallography, however, confirmed the molecule's pose in the binding pocket to within 1.4 to 2.3 angstroms, roughly the width of one or two atoms.

Success was not universal: for KRAS the signal could not be distinguished from artifact, and for Pin1 the crystal structure revealed a pose different from the one the model had predicted. The authors note that most hits so far fall short of the potency required for a drug candidate.

"This limitation may seem to undercut the main advantage of generative modeling over screening, but we consider it a pragmatic compromise necessary for experimental validation of designed compounds," the preprint's authors write.

The lead author, Ilya Igashov, is a MIPT graduate and a doctoral student at EPFL supervised by Bruno Correia and Oxford professor Michael Bronstein. LDDM continues a four-year research trajectory from the same laboratory, progressing from fragment-based molecular assembly through fully de novo design to predicting whether a compound can be synthesized.