Submission Details

Design Rationale:

CHEMICAL PROFILE Molecules designed by Benjamin P. Brown Graduate student in the laboratory of Jens Meiler, Ph.D. Email: benjamin.p.brown@vanderbilt.edu Notes on the chemical profile: Design protocol: Fragments co-crystalized with COVID-19 main protease (released by the Diamond Xchem group) served as starting scaffolds for combinatorial chemistry with an in-development Meiler Lab algorithm called BCL::LinkFragments. After filtering ~800,000 fragment combinations for physchem properties, geometry (i.e. pocket complementarity), and predicted activity, we performed focused library design (BCL::FocusedLibraryDesign) on a subset of the best new molecules. The variation of this algorithm utilized for this study incorporated a conventional supervised feed-forward deep neural network (DNN) as a pose-dependent protein-ligand interface scorer. Fragments were perturbed in a Monte Carlo - Metropolis fashion using alchemical transformations, and refined at each step to minimize clashes, optimize pose orientation, and filter out unstable/non-drug-like modifications. Molecules are optimized for clash resolution and interaction score. The best molecules undergo a final short run of BCL::FocusedLibraryDesign, and minor manual modifications are selectively made to intentionally increase probing of the structure-activity relationship (SAR). Finally, compounds are re-docked with RosettaLigand. The top-scoring complex is taken to be the final pose. Note that in multiple instances there is more than 1 well-populated binding pose. Here, we simply took the best scoring pose. Also note that the pose optimizing the RosettaLigand score is not the same pose that optimizes the DNN score. Nomenclature: XLogP, XLogS, and XdG_Hyd are LogP, LogS, and dg_hydration (kcal/mol) predictions made using a multi-tasking deep neural network trained in the BCL. LipinksiDruglike returns 1 if there are less than 2 Lipinksi violations. RS_Score_Raw and RSCONVOL_Score_Raw are two novel (in preparation) scoring functions developed in the BCL that use a DNN to prediction the binding affinity of small molecules to receptors (in units of -log(Kd)) using either pose-dependent or pose-dependent+pose-independent protein-ligand hybrid descriptors. RS_AD and RSCONVOL_AD are applicability domain (AD) metrics of the DNN scores; values less_equal 0.90 are generally considered good and mean we trust the corresponding activity prediction reasonably well. RosettaLigandInterfaceScore is the interface_delta_X (protein-ligand interaction score) from Rosetta v3.11 of the best pose in Talaris2014 Rosetta energy units (REU). MoleculeComplexity is a metric by Ertl. et al., 2009, Journal of Cheminformatics, and generally less than 2.5 indicates the molecule is reasonably synthesizable.

Other Notes:

If you try anything from this cluster, I recommend c1(CN2CCN(C(=O)C)CC2)cc(c(c2cc(ncn2)C)c(F)c1)OCC and Clc1c(c(cc(CN2CCN(C(=O)C)CC2)c1)OCC)c1cc(ncn1)C Again, activity predictions, physchem calculations, and docking score vs. rmsd plots and poses available upon request. Select examples from each of the clusters we submit (BEN-VAN) will be analyzed with MD simulations. Please reach out if you have any questions or would like specific molecules of ours to be simulated. We typically use Amber18 ff14sb +gaff2 forcefields with TIP3P water and HMassRepart 4 fs timestep. Binding free energies are completed in triplicated with MM-PBSA and the interaction entropy (IE). For similar compounds we can do GPU-TI numerically integrated with 12-point Gaussian quadrature.

Inspired By:

AAR-POS-d2a4d1df-18

MED-COV-4280ac29-15

AAR-POS-d2a4d1df-24