cPLA2 inhibitors

Cytosolic phospholipase A2 (GIVA cPLA2) is involved in a number of inflammatory diseases
Synthetic inhibitors of GIVA cPLA2 attracted great interest as potential antiinflammotory drugs

Desing Strategies:
* The ChEMBL database contains more than 340 records with the relevant data on biological activity – opportunity for ligand based virtual screening
* Structure-based design using crystal structure of human cytosolic phospholipase A2


Schematic representation of the proposed workflow:
Key steps:
1. Removal of compounds with undesirable properties:
* Substructure filters for removal of PAINS
* Filters using non MedChem-friendly SMARTS
* Removal of compounds with undesirably high lipophilicity (cLogP predicted by Molsoft software)
2. Virtual screening guided by consensus of ligand based methods and docking
3. Structural diversity picking: Min-Max algorithm

Ligand based strategy – Summary

1. Filtering stock collection by lipophilicity values (cLogP) predicted by ICM models
2. Classification machine learning models as an initial virtual screening filter:
* The ChEMBL database containing 340 records* with the relevant data on biological activity of known cPLA2 inhibitors was used as a source of training data. Decoys in the training set were generated by DUD-E methodology; decoys in the test set were collected from the ChEMBL data  
* Two independent ML models: Random forest built on count-based Morgan fingerprints and XGBoost model built on ECFP4 fingerprints
* Scores from the above mentioned models were combined
3. Diversity picking of selected compounds to enhance chemical space coverage
4. LB selection from stock collection results in ~100K diverse compounds ready for SB studies

Structural information:
The PDB database contains only one structure (uniprot ID P47712, PDB code: 1CJY) with the desired domains (CAP/Lid).
PDB structure of Lid region (code: 1CJY) includes unresolved loops and side chains. These regions were modeled using FREAD approach. To enhance sampling of these regions additional molecular dynamics studies were carried out.

Key steps:
* Simulation of 10 ns molecular dynamics with OpenMM engine in explicit solvent to improve sampling of the Lid/CAP region
* Binding site detection using ICM algorithm. This binding site appears to be stable during simulation
* Clustering the MD trajectory by the RMSD values of the CAP region to perform docking studies: the three most populated clusters were selected
* Molecular docking and scoring were carried out by Flare (Cresset)
* Electrostatic complementary calculations (Flare, Cresset) as an additional structure-based approach

Structure based strategy – Summary

* 10 ns molecular dynamics to sample binding site conformations
* Consensus docking to the three most populated clusters generated by MD
* Electrostatic complementary calculations
* SB selection based on docking and EC calculations yields ~17K compounds with the best scores
* Diversity picking of selected compounds yields 5.6K compounds
* Distributions of basic physicochemical properties and representative structures are shown in the following slides