Target: Glucokinase Regulatory Protein (GCKR) — allosteric site
Indication: Type 2 Diabetes Mellitus
Primary Lead: pgen_00050 (phenylsulfonyl-piperazine-fluorophenyl)
Pipeline Duration: 5 days (2026-04-11 ~ 04-15)
Total Compute Cost: ~$44
An end-to-end computational drug discovery pipeline that starts from disease-level omics data and ends with a lead compound ready for synthesis. Built and executed entirely with cloud compute (GCP + AWS).
Phase 0 Multi-omics 7,068 genes ranked → GCKR selected
Phase 1 Target validation 18 crystal structures, AMG allosteric site
Phase 2 Virtual screening 49,948 compounds docked → 23 PLIP-validated hits
Phase 3 MD + MM-GBSA 10ns simulations, binding free energy (ΔG)
Phase 4 ADMET 17 endpoint predictions per compound
Phase 5 DDI 7 T2DM comedication interaction checks
Phase 6 PK simulation Dosing, Cmax/IC50 estimation
Phase 7 IND report Final compound selection + recommendations
| Stage | Compounds | Method |
|---|---|---|
| Library | 49,948 | Synthetic drug-like, Lipinski + PAINS filtered |
| Docking | 49,948 → 200 | AutoDock Vina + composite scoring |
| PLIP interaction | 200 → 23 | ≥60% AMG pharmacophore reproduction |
| Ensemble docking | 23 → 10 | 5 crystal structures, consistency ranking |
| MD + MM-GBSA | 10 → 7 | 10ns GROMACS, gmx_MMPBSA binding ΔG |
| ADMET/DDI/PK | 3 evaluated | Go/no-go decision |
| Compound | ΔG (kcal/mol) | ADMET | DDI | Role |
|---|---|---|---|---|
| pgen_00050 | -10.64 | GO | Clean | Primary lead |
| enum_00184 | -24.32 | NO-GO (Caco-2) | Clean | Backup (needs optimization) |
| Compound | ΔG ± SD (kcal/mol) | Interpretation |
|---|---|---|
| enum_00184 | -24.32 ± 6.35 | Strong binding, poor absorption |
| enum_00554 | -16.72 ± 2.56 | Moderate, CYP inhibition issue |
| pgen_00050 | -10.64 ± 2.28 | Moderate, clean ADMET profile |
| enum_00555 | +3.79 ± 2.87 | Non-binder (eliminated) |
drugdiscovery/
├── config.yaml # Target and binding site configuration
├── pipeline_log.yaml # Full execution log (all phases)
├── CLAUDE.md # Execution rules and constraints
│
├── phase0_omics/ # Phase 0: Target discovery
│ ├── 01_geo_search.py # GEO dataset retrieval
│ ├── 02_deg_analysis.R # limma DEG analysis
│ ├── 03_integrate_degs.py # Cross-dataset integration
│ ├── 04_gwas_genes.py # GWAS catalog mining
│ ├── 05_ppi_network.py # STRING PPI network
│ ├── 06_druggability_filter.py # DGIdb druggability
│ ├── 07_pathway_enrichment.R # GO/KEGG enrichment
│ └── 08_target_selection.py # Final target ranking
│
├── phase1_target/ # Phase 1: Target validation
│ └── run_phase1.py # UniProt → PDB → ChEMBL profiling
│
├── phase2_screening/ # Phase 2: Virtual screening
│ ├── 00_prepare_receptor.py # 4MQU → PDBQT receptor
│ ├── 01_extract_controls.py # AMG co-crystal ligand extraction
│ ├── 01b_score_activity_correlation.py
│ ├── 02_prepare_library.py # Drug-like library generation
│ ├── 03_docking.py # AutoDock Vina batch docking
│ ├── 04_scaffold_enum.py # AMG scaffold enumeration
│ ├── 05_hit_selection.py # Composite scoring + selection
│ └── run_unidock.py # Uni-Dock GPU docking
│
├── data/results/ # All output data
│ ├── phase0/ # DEGs, GWAS, PPI, druggability
│ ├── phase1/ # Target profiles, PDB structures
│ ├── phase2/ # Docking scores, PLIP, ensemble results
│ ├── phase4_admet.csv # ADMET predictions
│ ├── phase5_ddi.csv # DDI risk assessment
│ └── phase6_pk.csv # PK simulation results
│
├── obsidian-vault/ # Documentation (Obsidian notes)
│ ├── 00 Pipeline Dashboard.md
│ ├── Phase 0 - Target Discovery.md
│ ├── Phase 1 - Target Validation.md
│ ├── Phase 2 - Hit Discovery.md
│ ├── Phase 3-6 - Lead Optimization.md
│ └── Phase 7 - IND Report.md # ← Final report
│
└── shared/
└── log_utils.py # YAML pipeline log utilities
- Transcriptomics: GEO datasets (GSE41762 pancreatic islets, GSE25724 liver) → limma differential expression (logFC>0.5, adj.P<0.05)
- GWAS: EBI GWAS Catalog T2DM loci (P < 5×10⁻⁸)
- PPI: STRING network (confidence > 0.7), hub gene identification
- Druggability: DGIdb GraphQL query, structural druggability assessment
- Docking: AutoDock Vina (exhaustiveness=4), Uni-Dock (GPU)
- Scoring: Rank-normalized composite (docking 35%, pharmacophore 30%, SA 20%, QED 15%)
- Validation: PLIP protein-ligand interaction profiling against AMG co-crystal reference
- Ensemble: 5 GCKR crystal structures (4MQU, 4LY9, 4PX2, 4MRO, 4MSU)
- Engine: GROMACS 2024.5 (CUDA, NVIDIA L4 GPU)
- Force field: AMBER99SB-ILDN (protein) + GAFF2 (ligand via acpype)
- Protocol: EM → NVT 100ps → NPT 100ps → Production 10ns
- Binding energy: gmx_MMPBSA (GB, igb=5, 0.15M salt, last 5ns)
- ADMET: 17 endpoints (Caco-2, HIA, CYP inhibition, hERG, AMES, DILI, etc.)
- DDI: CYP metabolism overlap with 7 T2DM comedications
- PK: One-compartment model, QD/BID dosing simulation
| Phase | Platform | Instance | GPU |
|---|---|---|---|
| 0-2 | GCP | e2-standard-8 (Spot) × 5 | — |
| 2 (Uni-Dock) | GCP | g2-standard-4 | L4 |
| 3 (MD) | GCP + AWS | g2-standard-8 + g6.xlarge × 4 | L4 |
| 4-6 | AWS | t3.medium | — |
Storage: AWS S3 (s3://dd-pipeline-drugdiscovery/, 3.7GB) + GCP GCS (backup)
- Synthesis: pgen_00050 custom synthesis (Enamine REAL)
- In vitro binding: SPR or ITC for experimental Kd
- SAR expansion: 10-20 analogs around pgen_00050 core
- Hit-to-Lead: Optimize IC50 from µM to nM range
If you use this pipeline or find it helpful:
Xavier Yu (2026). Computational Drug Discovery Pipeline: GCKR Allosteric Modulator for T2DM.
GitHub: https://github.com/brianyu43/drugdiscovery
MIT