Computational Drug Discovery

Construct protein-protein interaction (PPI) networks from STRING and BioGRID databases, integrate gene ontology (GO) annotations, KEGG and Reactome pathway data, and disease-gene associations from GeneCards and OMIM. Candidate targets are scored by network centrality (degree, betweenness, closeness) and tissue/disease expression from GTEx and GEO datasets. Output is a prioritised target list with evidence rationale.

Retrieve experimental structures from RCSB PDB; where no experimental structure exists, generate homology models using MODELLER or AlphaFold2 predictions as templates. Structures are energy-minimised, protonation states assigned at physiological pH (PDB2PQR/PROPKA), and binding sites identified using fpocket or DoGSiteScorer. Co-crystallised ligands are removed and grid boxes defined around each site for docking.

Apply Lipinski Rule-of-Five and Veber filters to reduce the screening library. Generate pharmacophore models from known active compounds (if available) and use them for ligand-based pre-screening via Pharmer or RDKit. Prepare 3D conformers (Omega/Open Babel) and assign partial charges. Libraries include ChEMBL subsets, ZINC20, or client-provided SMILES sets, processed to PDBQT format for docking.

Ensemble docking run against multiple receptor conformations to account for induced fit. AutoDock Vina primary docking with AutoDock 4 re-scoring on top poses. Binding modes are inspected for key interactions (H-bonds, hydrophobic contacts, pi-stacking) using PyMol and LigPlot+. Hits are filtered by binding energy threshold and interaction quality before ADMET evaluation.

GROMACS MD simulations of top-ranked protein-ligand complexes using CHARMM36 or AMBER ff19SB force fields with GAFF2 ligand parameters. Systems solvated in TIP3P water with explicit 0.15 M NaCl. Production runs of 100-500 ns following energy minimisation and NPT equilibration. Analyses: RMSD, RMSF, radius of gyration, H-bond occupancy, and MM-PBSA/MM-GBSA binding free energy estimation.

Multi-endpoint ADMET prediction: aqueous solubility, Caco-2 permeability, P-gp substrate likelihood, blood-brain barrier penetration, CYP450 inhibition (1A2, 2C9, 2C19, 2D6, 3A4), hERG cardiotoxicity, AMES mutagenicity, and LD50 acute toxicity. Compounds are scored against ICH M7 mutagenic impurity thresholds and flagged for known PAINS scaffolds (Pan-assay interference compounds).

Composite scoring integrating docking energy, MM-PBSA free energy, ADMET scores, and structural novelty. Where target bioactivity data is available (ChEMBL), a graph neural network or XGBoost model predicts pIC50 for final ranking. Morgan fingerprint similarity analysis ensures chemical diversity across the final candidate list. Cluster representatives are selected to maximise scaffold coverage.

A ranked candidate list (typically 5-20 compounds) with full supporting data. Deliverables include: SDF file of hit compounds with 3D poses, per-compound interaction tables (H-bond donors/acceptors, key residues), ADMET property spreadsheet, MD trajectory summaries with RMSD plots, and a written methodology report with references. All files are formatted for direct handoff to synthetic chemistry or in-vivo teams.