Review Claim by Claim

 Long Explanation

LDDM: A Unified, Synthesizable 3D Generative Model — Critical Evidence Review

What is established (reported). LDDM trains one equivariant flow-matching + Markov-bridge model with fragment-level masked modeling (BRICS fragmentation) so that a single checkpoint natively supports docking, covalent docking, fragment linking/growing, and de novo design. On the PoseBusters test set (308 targets, PDB entries since 2021), LDDM achieved 79.5% top-1 docking success at 2 Å RMSD — with the authors candidly noting pocket-definition differences make redocking comparisons imperfect — and it beat specialized covalent-docking baselines on the Scarpino benchmark (207 complexes, 54 targets) PoseBusters and covalent docking benchmarks. Training used 836,259 curated complexes (CrossDocked, BindingNet v2, BigBind; ~454,662 unique structures), and synthetic-data training improved validity (PoseBusters pass 71.3%→79.5%, 3D validity 53.5%→63.6%) versus PDB-only training. Over 60% of de novo designs passed strict PoseBusters + GenBench3D validity; ~65% of linkers passed all filters, exceeding DiffLinker by ~22%.

Wet-lab validation (reported, direct evidence). Five case studies produced confirmed ligands: 67% of nine LDDM-introduced VHL-CDO1 molecular-glue modifications retained nanomolar ternary affinity by SPR; one non-natural peptide improved cathepsin S inhibition 3.5-fold (70.4→19.9 nM) despite LDDM never training on peptides; PGK1 Z856 gave Kd 14.7 µM (GCI + HSQC NMR site confirmation), beating the 30 µM positive control; BRD4 Z967 bound at 40–60 µM with NMR shifts matching the predicted acetyl-lysine pocket; and Mac1 de novo design Z8035 inhibited at IC50 46.2 µM with a crystallographic ligand RMSD of 2.35 Å — from only 57 de novo candidates vs 1,592 chemically novel compounds screened in CACHE #3 (best ~18 µM) Five experimental case studies with biophysical and structural validation.

Honest failures (reported). KRAS GCI signals had abnormally high amplitudes precluding artifact exclusion (Z755, 85.7 µM, questionable); a Pin1 covalent design crystallized in a pose different from prediction due to crystal-packing contacts. Many validated hits are far below drug-candidate affinity, which the authors attribute to small ligand size and limited interactions — an author interpretation, not a proven mechanism.

Blind spots (BGPT inference). All validation came from the model's own laboratories with no blinded, identical-pipeline comparison against competing generative tools (e.g., DiffSBDD, FLOWR) or exhaustive screening of the same virtual spaces, so reported hit rates cannot be attributed to LDDM specifically versus the Enamine REAL prefiltering and heavy post-hoc filter stacks (GNINA re-docking, AlphaFold3/Boltz co-folding, gradient-boosted pose scorers). Only binding was measured — no functional, selectivity, or cellular data. In silico generative design work broadly remains at hit-identification maturity; comparable efforts like reaction-guided fragment growth frameworks also show promise but lack this paper's experimental depth MOZAIC reaction-guided fragment growing framework.

What would change this conclusion: independent replication of hit rates on fresh targets without Enamine-REAL restriction; prospective pose-accuracy stats on many crystal structures (only 3 structures solved here, one deviating 5.07 Å); and functional selectivity data. Code is available at github.com/lpdi-epfl/lddm, supporting partial reproducibility LDDM source code repository.