← Back to Home
🧬 T2D Multi-Target Drug Discovery
AI-Driven Virtual Screening for Type 2 Diabetes · Cornell University
Overview
We screen 12,301 real ChEMBL bioassay records (IC50 measurements) across 6 T2D drug targets using a multi-task neural network. The pipeline predicts activity and potency simultaneously, then generates novel drug candidates via a conditional VAE. All evaluation uses scaffold split — the model is tested on entirely unseen chemical scaffolds, reflecting realistic drug discovery conditions.
📦 ChEMBL IC50 Data
→
🧬 Multi-Task Predictor (V100)
→
🔬 Conditional VAE Generator
→
💊 ADMET + Drug Scoring
Data: Real ChEMBL Bioassays
| Target | Mechanism | Records | Active (IC50 < 1μM) |
|---|
| DPP-4 | Incretin degradation inhibitor | 4,827 | 3,909 (81%) |
| SGLT2 | Renal glucose reabsorption blocker | 2,098 | 1,945 (93%) |
| PPARγ | Insulin sensitizer | 1,954 | 1,255 (64%) |
| GLP-1R | Incretin receptor agonist | 1,321 | 1,056 (80%) |
| α-Glucosidase | Carbohydrate digestion inhibitor | 350 | 45 (13%) |
| AMPK | Cellular energy sensor activator | 4,962 | 4,129 (83%) |
⚠️ Known issues with this data: 78.6% of records are "active" — reflecting publication bias (negative results rarely published). IC50 values aggregate different labs, assay formats, and conditions. Activity threshold (1μM) is standard but arbitrary.
Model Performance (Scaffold Split)
Three architectures tested on identical data/split. All use multi-task learning (binary activity + pIC50 regression). v4 shows that deep learning on molecular graphs significantly outperforms hand-crafted fingerprints.
| Model | Params | ROC-AUC | PRC-AUC | pIC50 R | Time |
|---|
| FP+Desc (v3) | 791K | 0.7276 | 0.5330 | 0.4412 | 49s |
| GAT (v4) | 307K | 0.8910 | 0.9604 | 0.7255 | 235s |
| Transformer (v4) | 3.3M | 0.7749 | 0.8997 | 0.5505 | 533s |
| Ensemble (v4) | — | 0.8872 | 0.9608 | — | 786s |
Key findings: Graph Attention Network (GAT) achieves 0.891 AUC with only 307K parameters — 22% improvement over Morgan FP baseline (0.728). The GAT directly learns molecular graph structure (atoms = nodes, bonds = edges) via 3-layer attention mechanism. SMILES Transformer (character-level) captures complementary sequential patterns. Ensemble (75% GAT + 25% Transformer) marginally improves robustness. pIC50 correlation jumps from 0.44 → 0.73, indicating GAT captures finer SAR signal.
What This Means
- GNN > FP: Learning graph representations beats hand-crafted fingerprints for scaffold generalization.
- Still honest metrics: 0.89 AUC on scaffold split (unseen scaffolds) is strong but realistic. Random split would give >0.95 (data leakage).
- EF@1% = 1.3: Top 1% enrichment remains weak (barely better than random). Ranking performance is the bottleneck, not discrimination.
- Ready for production? No. Needs experimental validation (synthesize top candidates, measure IC50 in vitro).
Top Candidates
Generated by Conditional VAE, scored by composite metric (QED, synthetic accessibility, PAINS, hERG risk). 120 candidates across 6 targets; top 5 shown.
| # | Target | Score | QED | SA | SMILES |
|---|
| 1 | DPP4 | 88.5 | 0.930 | 1.85 | O=C(NC1CCC1)c1ccc(-c2ccncc2)c(F)c1 |
| 2 | DPP4 | 88.0 | 0.917 | 1.84 | O=C(NC1CC1)c1ccc(F)c(-c2ccncc2)c1 |
| 3 | GLP1R | 87.7 | 0.924 | 2.12 | COc1ccc(-c2cc(C(=O)N3CCNCC3)on2)cc1 |
| 4 | DPP4 | 87.4 | 0.926 | 2.30 | Nc1cc(N2CCCC2)nn1-c1ccc(C(F)(F)F)cc1 |
| 5 | GLP1R | 87.2 | 0.914 | 2.18 | O=C(NC1CCNCC1)c1cc(-c2ccc(Cl)cc2)on1 |
All top candidates pass Lipinski, Veber, PAINS filters. Mean SA score 3.07 (easy to synthesize). 88% low hERG risk. 79 unique Murcko scaffolds across 120 candidates (66% diversity). Caveat: candidates were generated from v2 pseudo-labeled model; re-screening with v3 real-data model is pending.
Molecular Dynamics (BioHPC)
| System | PDB | Status | Duration |
|---|
| Insulin Monomer | 1MSO | ✅ Running | 100ns done + 500ns extending |
| Insulin Hexamer | 4ZXB | 🔄 Equilibrating | EM → NVT → NPT |
| DPP-4 + Inhibitor | 6PXV | 📋 Queued | — |
Methods
- Data: ChEMBL API, 12,301 IC50 records (standard_relation='=', units=nM). Deduplicated by (SMILES, target).
- Split: Murcko scaffold-based (4,568 unique scaffolds → 80/10/10). No scaffold overlap between train/val/test.
- v3 (FP+Desc): Morgan FP (1024-bit) + 11 descriptors → dual-channel NN. Multi-task (BCE + MSE + SupCon). 791K params.
- v4 (GAT): Molecular graph → 3-layer Graph Attention Network (atoms=nodes, bonds=edges). 4-head attention, global mean/max pooling. 307K params. Multi-task (BCE + MSE). 150 epochs, AdamW, CosineAnnealing, AMP FP16.
- v4 (Transformer): SMILES tokenization (character-level, vocab=41) → 6-layer Transformer encoder (d=256, 8 heads). Mean pooling, multi-task. 3.3M params.
- v4 (Ensemble): Weighted average (0.75×GAT + 0.25×Transformer) optimized on validation set.
- Evaluation: ROC-AUC, PRC-AUC, pIC50 Pearson R, Enrichment Factor @ 1%. Drug-likeness: SA score, PAINS, Lipinski/Veber/Egan.
- Hardware: 2× Tesla V100-PCIE-32GB (research1), dual-GPU parallel training (GAT on GPU 0, Transformer on GPU 1).
- References: Rep3Net (2025), CL-MFAP (2025), Graph Attention Networks (Veličković+ 2018).
Limitations & Next Steps
- v4 solved generalization: GAT AUC 0.89 shows graph learning works. But EF@1% = 1.3 (weak ranking) remains unsolved.
- Dataset imbalance (78.6% active): Publication bias inflates metrics. Need better negatives or dose-response curves.
- 2D only: GAT sees bonds/atoms, not 3D geometry. Adding conformers + docking scores would help binding prediction.
- No wet-lab validation: Top candidates need synthesis + in vitro IC50 assay. Computational predictions are hypotheses, not proof.
- What would improve EF@1%? (1) Uncertainty quantification (Bayesian GNN), (2) Active learning (iteratively retrain on assay results), (3) Physics-based scoring (molecular docking, MD free energy).
- Production readiness: GAT is research-grade. For real screening: ensemble with docking, integrate ADMET, add human expert review loop.