1from src.inference.predict import predict_drug_effects
23result = predict_drug_effects(4"OC(=O)C1=CN(C2CC2)c2cc(N3CCNCC3)c(F)cc2C1=O",# Ciprofloxacin5 device="cpu"6)78print(f"MIC: {result['predicted_mic_uM']:.2f} µM")9print(f"MoA: {result['moa_class']} ({result['moa_confidence']:.0%})")10for t in result['top_targets'][:3]:11print(f" Target: {t['gene']} ({t['score']:.3f})")
Model Architecture
SMILES Input
│
├── GINEConv GNN (3 layers, 256-dim) ──────────── molecular graph topology
│ └── 17 atom features + 4 bond features
│ └── triple pooling: mean + max + attention → 768-dim
│
├── Morgan + RDKit Fingerprints (4096-bit) ────── substructure patterns
│ └── 3-layer MLP → 256-dim
│
├── Physicochemical Descriptors (12 features) ─── scaffold-independent properties
│ └── MW, logP, TPSA, HBD, HBA, RotBonds, ArRings, Rings, FrCSP3, MolMR, etc.
│ └── 2-layer MLP → 256-dim
│
└── Data Source Embedding (8 → 16-dim) ────────── corrects for inter-source bias
│
▼
Fusion: 768-dim MLP + LayerNorm + residual connection
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Target Predictor │
│ Bilinear attention: compound emb × ESM-2 protein embs │
│ 17 E. coli genes with pre-computed ESM-2 (650M) vectors │
└──────────────────────────┬──────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Metabolic Integration Layer (differentiable FBA surrogate) │
│ iML1515: 1,516 genes × 2,712 reactions × 40 pathways │
│ Computes: reaction inhibition → flux perturbation → │
│ pathway impact → growth proxy → 256-dim features │
└──────────────────────────┬──────────────────────────────────┘
│
┌─────────────┼─────────────┐
▼ ▼ ▼
MIC Head MoA Head Pathway Head
3-layer MLP 6-class CE 40-dim BCE
→ µg/mL → class → pathway impacts
Total parameters: ~9M (7.6M trainable)
Training
Curriculum Learning
Phase
Epochs
Tasks Active
GNN
LR
A
1–20
MIC only
Training
1.5×10⁻⁴
B
21–40
MIC + Targets
Frozen
1.5×10⁻⁴
C
41+
All tasks
Unfrozen at LR/10
GNN: 1.5×10⁻⁵
Phase A builds strong MIC features. Phase B trains the target head without destabilising the GNN. Phase C brings MoA online with differential learning rates.
Loss Function
Task
Loss
Weight
MIC
Huber (δ=2.0)
1.0
MoA
Cross-entropy (label smoothing 0.1)
25.0
Targets
Focal loss (γ=2.0, α=0.25)
80.0
Pathways
Binary cross-entropy
5.0
Target entropy
Activation + diversity regulariser
0.15
Regularisation
Technique
Details
Dropout
0.35 on all heads and fusion layers
Weight decay
5×10⁻³ (AdamW)
MIC label noise
Gaussian N(0, 0.5) on training MIC values
SMILES augmentation
2× non-canonical enumerations per compound
EMA
Exponential moving average (decay=0.999) for smoother validation
SWA
Stochastic weight averaging in final 20% of training
LR schedule
Linear warmup (5 epochs) → cosine annealing
Early stopping
Patience 30, monitoring val_mic
Gradient clipping
max_norm=5.0
Test Set Results (v21)
Metric
Value
MIC RMSE
2.21 log₂ µM
MIC R²
0.50
MIC Spearman ρ
0.70
Within 2-fold accuracy
54.4%
MoA Macro F1
0.98
MoA Balanced Accuracy
97.0%
Target AUROC
0.997
Evaluated on a held-out scaffold-split test set of 2,743 compounds.
MoA-stratified — MoA-labelled compounds distributed proportionally across all splits
80/10/10 train/validation/test
Biological Integration
ESM-2 Protein Embeddings
Pre-computed embeddings from facebook/esm2_t33_650M_UR50D for all 1,516 iML1515 gene products. Projected from 1280-dim → 512-dim for bilinear attention scoring against compound embeddings.
iML1515 Metabolic Model
The differentiable FBA surrogate uses matrices from the iML1515 genome-scale metabolic model:
Wildtype FBA fluxes: Baseline metabolic state from constraint-based modelling
Biomass coefficients: Growth contribution of each reaction
Learnable sensitivity weights: How much each reaction responds to drug-induced inhibition
File Structure
ecoli-k12-drug-model/
├── app.py # Gradio web interface
├── train.py # Training script
├── install_and_run.py # One-click installer + launcher
├── START_MAC.command # Mac double-click launcher
├── START_WINDOWS.bat # Windows double-click launcher
├── requirements-app.txt # Inference-only dependencies
├── requirements.txt # Full training dependencies
├── configs/
│ └── model_config.yaml # Architecture hyperparameters
├── models/
│ ├── best_model.pt # Best Phase B checkpoint (val_mic=1.09)
│ ├── best_model_phase_c.pt # Best Phase C checkpoint (has MoA)
│ └── final_model.pt # EMA model at end of training
├── src/
│ ├── data/
│ │ ├── merge_datasets.py # Multi-source dataset builder
│ │ ├── dataset.py # PyTorch Dataset + collate
│ │ ├── precompute_features.py # Graph/FP/physchem caching
│ │ ├── compile_moa_labels.py # 180 curated antibiotic MoA labels
│ │ ├── fetch_chembl_moa.py # ChEMBL mechanism API
│ │ ├── fetch_patric_amr.py # PATRIC clinical MIC data
│ │ ├── fetch_bindingdb_targets.py # BindingDB interactions
│ │ └── propagate_moa_labels.py # Tanimoto MoA propagation
│ ├── features/
│ │ └── molecular_graph.py # SMILES → graph + fingerprint + physchem
│ ├── models/
│ │ ├── compound_encoder.py # Multi-branch encoder (GNN+FP+Physchem)
│ │ ├── target_predictor.py # Bilinear compound-protein scoring
│ │ ├── metabolic_layer.py # Differentiable FBA surrogate
│ │ └── multitask_head.py # Top-level model with task heads
│ ├── training/
│ │ ├── losses.py # Multi-task loss function
│ │ └── evaluation.py # Test set metrics
│ └── inference/
│ └── predict.py # Single-compound prediction API
└── data/
├── raw/ # Source data (.parquet files)
└── processed/ # Feature caches, splits, gene lists
Interpreting Results
MIC Values
MIC (µg/mL)
Interpretation
≤ 0.5
Highly potent — strong antibiotic candidate
0.5 – 4
Potent — comparable to clinical antibiotics
4 – 16
Moderate — may need optimisation
16 – 64
Weak — significant barriers to activity
> 64
Likely inactive against E. coli
The model predicts in log₂ µM internally. To convert:
µM → µg/mL: multiply by molecular weight / 1000
log₂ µM → µM: compute 2^(predicted value)
MoA Confidence
The model outputs softmax probabilities across 6 classes. Predictions with >80% confidence on the top class are generally reliable. Lower confidence suggests the compound may have a novel or mixed mechanism.
Target Scores
Target interaction probabilities range 0–1. Most compounds show low activation (<0.1) across all targets. Scores above 0.3 indicate meaningful predicted binding. The model is most confident for well-characterised targets like folA (dihydrofolate reductase) and lpxC (LPS biosynthesis).
Limitations
Prediction accuracy: MIC predictions have ~2.5-fold average error (RMSE 2.21 log₂ µM). This is comparable to inter-laboratory MIC variability but should be treated as order-of-magnitude estimates, not precise values.
E. coli K12 only: Trained on laboratory K12 strain data. Clinical isolates with acquired resistance (ESBL, carbapenemases, plasmid-mediated resistance) will have different MIC profiles.
Scaffold bias: Best performance on compounds structurally similar to the training set (fluoroquinolones, beta-lactams, aminoglycosides, tetracyclines, sulfonamides). Novel scaffolds may have higher prediction error.
Target coverage: Only 17 of ~1,500 E. coli proteins have sufficient training labels. The model cannot predict interactions with uncharacterised targets.
MoA propagation: 91% of MoA labels are inferred by Tanimoto similarity (threshold 0.40), not experimentally confirmed. Novel chemotypes may be misclassified.
Not for clinical use. This is a research tool for prioritising compounds in early-stage drug discovery. All predictions require experimental validation.
