Arne Schneuing

A Unified 3D Generative Model for Synthesizable Structure-Based Drug Design 1. The paper introduces LDDM (Large Drug Discovery Model), a unified 3D generative framework that jointly samples atom types, covalent bonds, and 3D coordinates conditioned on a protein pocket—so the same model can do de novo design, fragment growing/linking, (non)covalent docking, and even non-canonical peptide side-chain docking. 2. A key methodological idea is fragment-based masked modeling (inspired by masked language modeling): molecules are BRICS-fragmented, and during training different fragments are masked in different “roles” (design: graph+coords missing; docking: coords missing but graph known; context: fully known). This trains one model that naturally supports many structure-based tasks by changing what is masked/conditioned. 3. The generative core combines equivariant flow matching for continuous 3D coordinates with Markov bridge models for discrete atom/bond types, plus per-atom uncertainty estimates. That uncertainty becomes a practical scoring signal: it correlates with pose RMSD (Spearman ρ = 0.59) and can highlight which substructures are unreliable. 4. Training data is scaled via a curated synthetic complex dataset (836,259 protein–ligand complexes; 454,662 unique molecules) assembled from CrossDocked, BindingNet v2, and BigBind (BigBind ligands docked with Gnina), then heavily filtered with PoseBusters + GenBench3D and outlier filters. The model’s samples cover the training chemical space while staying novel: ~90% of generated molecules have <0.5 max Tanimoto similarity to the training set. 5. Benchmarking shows “unified but competitive” performance: on PoseBusters local docking, LDDM reaches 79.5% top-1 success at 2 Å RMSD (valid cases), and on a covalent docking benchmark it achieves 80.1% top-1 success at 2 Å—outperforming specialized covalent docking baselines in the cited comparison. 6. Beyond one-shot generation, the paper proposes programmable generation: an iterative fragment-tree search that alternates global filtering (e.g., QED thresholds) with local fragment-level filtering (pose plausibility, hydrogen-bond satisfaction, target interactions), allocating sampling budget with a UCB-style rule (Monte Carlo tree search flavor). This improves sampling efficiency for local, structure-dependent objectives versus SMILES RL (REINVENT), especially for hydrogen-bonding criteria. 7. A second practical innovation is synthesizable generation: programmable generation constrained by building blocks + reaction templates (demonstrated with Enamine REAL). Instead of generating “anything” then hoping it’s makeable, LDDM generates full molecules but only accepts expansions that match precomputed reaction products; partial matches use an MCS-anchored hybrid docking step. In comparisons with equal sampling budgets, this approach yields far more candidates that are both filter-passing and present in REAL space. 8. Experimental validation spans five targets and multiple modalities, emphasizing “few syntheses, high hit rates” rather than massive screening. In a VHL–CDO1 molecular glue campaign, LDDM redesigned two CDO1-contacting flanks while keeping the VHL anchor fixed; across both flank campaigns, 67% of synthesized LDDM modifications retained nanomolar ternary-complex affinity by SPR. 9. LDDM also optimized a covalent non-natural peptide inhibitor of cathepsin S (CTSS): docking non-canonical side chains (54 building blocks; 270 single mutants; 200 conformers each) led to one improved design (J2) with 3.5× better potency, improving IC50 from 70.4 nM to 19.9 nM in FRET assays—despite LDDM not being trained on peptides. 10. De novo hit-finding case studies show micromolar binders with orthogonal validation and pose accuracy: for PGK1, 4/6 purchasable one-shot designs bound (best KD 14.7 µM) with NMR CSPs supporting binding at the targeted nucleotide pocket; for BRD4, synthesizable designs produced binders (e.g., Z967 ~40–60 µM) with NMR CSPs consistent with the acetyl-lysine pocket; for SARS-CoV-2 Nsp3 Mac1, 6/88 synthesized candidates showed competition in HTRF (best IC50 46.2 µM), and X-ray structures for two designs showed ligand RMSDs of 2.35 Å and 1.43 Å vs designed poses. 💻Code: github.com/lpdi-epfl/lddm. 📜Paper: biorxiv.org/content/10.64898… #ComputationalDrugDiscovery #GenerativeAI #StructureBasedDrugDesign #DiffusionModels #FlowMatching #MolecularDocking #FragmentBasedDesign #SyntheticAccessibility #DrugDesign #ChemicalBiology

A Unified 3D Generative Model for Synthesizable Structure-Based Drug Design 1. The paper introduces LDDM (Large Drug Discovery Model), a unified 3D generative framework that jointly samples atom types, covalent bonds, and 3D coordinates conditioned on a protein pocket—so the same model can do de novo design, fragment growing/linking, (non)covalent docking, and even non-canonical peptide side-chain docking. 2. A key methodological idea is fragment-based masked modeling (inspired by masked language modeling): molecules are BRICS-fragmented, and during training different fragments are masked in different “roles” (design: graph+coords missing; docking: coords missing but graph known; context: fully known). This trains one model that naturally supports many structure-based tasks by changing what is masked/conditioned. 3. The generative core combines equivariant flow matching for continuous 3D coordinates with Markov bridge models for discrete atom/bond types, plus per-atom uncertainty estimates. That uncertainty becomes a practical scoring signal: it correlates with pose RMSD (Spearman ρ = 0.59) and can highlight which substructures are unreliable. 4. Training data is scaled via a curated synthetic complex dataset (836,259 protein–ligand complexes; 454,662 unique molecules) assembled from CrossDocked, BindingNet v2, and BigBind (BigBind ligands docked with Gnina), then heavily filtered with PoseBusters + GenBench3D and outlier filters. The model’s samples cover the training chemical space while staying novel: ~90% of generated molecules have <0.5 max Tanimoto similarity to the training set. 5. Benchmarking shows “unified but competitive” performance: on PoseBusters local docking, LDDM reaches 79.5% top-1 success at 2 Å RMSD (valid cases), and on a covalent docking benchmark it achieves 80.1% top-1 success at 2 Å—outperforming specialized covalent docking baselines in the cited comparison. 6. Beyond one-shot generation, the paper proposes programmable generation: an iterative fragment-tree search that alternates global filtering (e.g., QED thresholds) with local fragment-level filtering (pose plausibility, hydrogen-bond satisfaction, target interactions), allocating sampling budget with a UCB-style rule (Monte Carlo tree search flavor). This improves sampling efficiency for local, structure-dependent objectives versus SMILES RL (REINVENT), especially for hydrogen-bonding criteria. 7. A second practical innovation is synthesizable generation: programmable generation constrained by building blocks + reaction templates (demonstrated with Enamine REAL). Instead of generating “anything” then hoping it’s makeable, LDDM generates full molecules but only accepts expansions that match precomputed reaction products; partial matches use an MCS-anchored hybrid docking step. In comparisons with equal sampling budgets, this approach yields far more candidates that are both filter-passing and present in REAL space. 8. Experimental validation spans five targets and multiple modalities, emphasizing “few syntheses, high hit rates” rather than massive screening. In a VHL–CDO1 molecular glue campaign, LDDM redesigned two CDO1-contacting flanks while keeping the VHL anchor fixed; across both flank campaigns, 67% of synthesized LDDM modifications retained nanomolar ternary-complex affinity by SPR. 9. LDDM also optimized a covalent non-natural peptide inhibitor of cathepsin S (CTSS): docking non-canonical side chains (54 building blocks; 270 single mutants; 200 conformers each) led to one improved design (J2) with 3.5× better potency, improving IC50 from 70.4 nM to 19.9 nM in FRET assays—despite LDDM not being trained on peptides. 10. De novo hit-finding case studies show micromolar binders with orthogonal validation and pose accuracy: for PGK1, 4/6 purchasable one-shot designs bound (best KD 14.7 µM) with NMR CSPs supporting binding at the targeted nucleotide pocket; for BRD4, synthesizable designs produced binders (e.g., Z967 ~40–60 µM) with NMR CSPs consistent with the acetyl-lysine pocket; for SARS-CoV-2 Nsp3 Mac1, 6/88 synthesized candidates showed competition in HTRF (best IC50 46.2 µM), and X-ray structures for two designs showed ligand RMSDs of 2.35 Å and 1.43 Å vs designed poses. 💻Code: github.com/lpdi-epfl/lddm. 📜Paper: biorxiv.org/content/10.64898… #ComputationalDrugDiscovery #GenerativeAI #StructureBasedDrugDesign #DiffusionModels #FlowMatching #MolecularDocking #FragmentBasedDesign #SyntheticAccessibility #DrugDesign #ChemicalBiology