AI-Accelerated CASP4 Inhibitor Screening
An AI-based virtual screening pipeline that identified repositionable DrugBank compounds as candidate CASP4 (Caspase-4) inhibitors for Alzheimer's disease, combining ChemBERTa-based ML ranking, molecular docking, MD simulation, and MM/PBSA free-energy analysis.
Manuscript status: accepted; citation to be added upon publication. Code: github.com/mubashirhassangcul/Caspase4-inhibitor-screening Data: SidraBhatti/caspase4-inhibitor-screening-data
The full 5-stage pipeline: (1) data curation & preprocessing, (2) molecular
structure retrieval & descriptor calculation (ChemBERTa + RDKit), (3)
machine learning — Random Forest classification and regression, (4)
molecular docking & binding-site analysis (AutoDock Vina, PrankWeb), (5) MD
simulation & MM/PBSA analysis (GROMACS).
Role & Attribution
Sidra Bhatti led the AI-based prediction pipeline: ChemBERTa embedding generation, hybrid feature engineering, and development of the Random Forest classifier and regressor used to prioritize candidates. Molecular docking, MD simulations, and MM/PBSA analysis were performed by collaborators (see the GitHub repository for the full author list).
Pipeline
DrugBank Drug-Lib compounds
↓
Structure-based virtual screening (VSTH/Tianhe-2 vs. CASP4, PDB 6NRY) — ~1,739 initial hits
↓
ChEMBL (fetch experimental pIC50 labels; active if pIC50 ≥ 8.2)
↓
Feature extraction: ChemBERTa SMILES embeddings (768-dim) + RDKit physicochemical descriptors (6)
↓
Random Forest Classifier (active vs. inactive; 500 trees, 774-dim hybrid features)
↓
Random Forest Regressor (pIC50 prediction on actives)
↓
Top-10 candidate selection → ADMET/PK-PD filtering
↓
Binding-site ID (PrankWeb) + AutoDock Vina docking vs. CASP4
↓
Top-5 candidates + donepezil reference → 100-ns MD in triplicate (GROMACS, CHARMM36)
↓
MM/PBSA free-energy ranking → benchmarking vs. donepezil
Method
Feature engineering. Each screened compound's SMILES string is encoded with a pretrained ChemBERTa transformer (768-dim embedding), concatenated with 6 RDKit physicochemical descriptors, giving a 774-dim hybrid feature vector per compound.
Random Forest classifier. 500 trees, trained on the 774-dim hybrid features to distinguish active vs. inactive compounds (active defined as ChEMBL pIC50 ≥ 8.2).
Random Forest regressor. Trained on the experimental pIC50 values of active compounds only, used to rank candidates by predicted potency.
Results
| Metric | Value |
|---|---|
| Classifier accuracy | ~95% |
| Classifier ROC-AUC | 0.73 |
| Enrichment (top 10%) | ~3.5× |
AI-Ranked Shortlist
Top-10 compounds by RF classifier/regressor score, before ADMET/PK-PD filtering:
| DrugBank ID | Predicted pIC50 | Predicted active prob. | Mol. Weight (g/mol) | Log P | HBD | HBA | Lipinski |
|---|---|---|---|---|---|---|---|
| DB08897 | 8.481 | 0.724 | 484.66 | 4.67 | 1 | 6 | Yes |
| DB09477 | 8.513 | 0.722 | 348.40 | 1.13 | 3 | 4 | Yes |
| DB06202 | 8.446 | 0.720 | 413.56 | 5.73 | 1 | 3 | No |
| DB08882 | 8.299 | 0.716 | 472.55 | 1.15 | 1 | 10 | Yes |
| DB05316 | 8.707 | 0.704 | 427.56 | 4.67 | 1 | 3 | Yes |
| DB00519 | 8.917 | 0.700 | 430.55 | 2.77 | 2 | 5 | Yes |
| DB01068 | 8.595 | 0.700 | 315.72 | 3.04 | 1 | 4 | Yes |
| DB00722 | 8.490 | 0.682 | 405.50 | 1.24 | 4 | 5 | Yes |
| DB00234 | 8.675 | 0.678 | 313.40 | 3.19 | 1 | 4 | Yes |
| DB13867 | 9.509 | 0.676 | 444.52 | 3.47 | 2 | 5 | Yes |
Final Candidates (after ADMET/PK-PD filtering, went to docking)
This set differs from the raw AI-ranked shortlist above — 7 compounds overlap, 3 were filtered out and replaced after ADMET/PK-PD screening:
| DrugBank ID | Compound | Vina Docking Score (kcal/mol) |
|---|---|---|
| DB00439 | Cerivastatin | −6.7 |
| DB00519 | Trandolapril | −5.4 |
| DB01068 | Clonazepam | −6.6 |
| DB01544 | Flunitrazepam | −9.1 |
| DB05316 | Pimavanserin | −7.7 |
| DB06202 | Lasofoxifene | −7.5 |
| DB06203 | Alogliptin | −6.1 |
| DB08882 | Linagliptin | −6.6 |
| DB08897 | Aclidinium | −6.3 |
| DB09477 | Enalaprilat | −7.3 |
Structures:
 DB00439 Cerivastatin |
 DB00519 Trandolapril |
 DB01068 Clonazepam |
 DB01544 Flunitrazepam |
 DB05316 Pimavanserin |
 DB06202 Lasofoxifene |
 DB06203 Alogliptin |
 DB08882 Linagliptin |
 DB08897 Aclidinium |
 DB09477 Enalaprilat |
MD / MM-PBSA Validation
The top 5 candidates (DB00519, DB01068, DB06202, DB08882, DB05316) plus the donepezil reference underwent 100-ns MD simulation in triplicate (GROMACS/CHARMM36):
Backbone RMSD — structural stability over the trajectory.
Per-residue RMSF — local flexibility.
Radius of gyration — overall compactness.
Solvent-accessible surface area.
Protein-ligand center-of-mass distance.
Total system energy.
Binding free energies: DB05316 (−21.4 kcal/mol) and DB00519 (−20.9 kcal/mol) exceeded even the reference compound donepezil, identifying them as the most promising CASP4 inhibitors.
Data availability
DrugBank raw data (drugbank.xml) and structure/SMILES columns are
excluded per DrugBank's terms of use — see the companion
Dataset repo
for the processed screening/feature tables that are safe to share.
Citation
Manuscript accepted; BibTeX will be added here once publication details are final.