ML Drug Discovery

Leveraging machine learning to identify small molecules that stabilize mutant G6PD enzymes and restore catalytic function.

Drug Discovery Pipeline

Data Collection

ChEMBL, PubChem compound libraries

Featurization

Molecular fingerprints, descriptors

Model Training

Graph neural networks, transformers

Virtual Screening

Score millions of candidates

Validation

MD simulations, experimental tests

Neural Network Architecture

Graph neural network for molecular property prediction

GNN Architecture

Input: Molecular Graph (atoms, bonds)

Message Passing Layers (3x)

Global Pooling

Dense Layers (256, 128)

Output: Binding Affinity Score

Graph Attention Network

Primary model using attention mechanisms to weight atomic contributions to binding affinity predictions.

0.89

ROC-AUC

0.76

Precision

0.82

Recall

ChemBERTa Transformer

SMILES-based transformer model for molecular property prediction, pretrained on 77M compounds.

0.91

ROC-AUC

0.79

Precision

0.85

Recall

Model Training

Real-time training progress for G6PD binding affinity prediction

Top Candidate Compounds

Predicted stabilizers for G6PD Mediterranean (S188F) variant

🧪

AG-348 Analog

Pyruvate kinase activator derivative

Score: 0.94

⚗️

NADP+ Mimetic

Structural cofactor stabilizer

Score: 0.91

🔬

Flavonoid-7

Natural product derivative

Score: 0.88

💊

Sulfonamide-12

Active site binder

Score: 0.85

* Compounds are computational predictions requiring experimental validation

Methodology & Tools

PyTorch Geometric

Graph neural network implementation for molecular representation learning and property prediction.

RDKit

Cheminformatics toolkit for molecular fingerprints, descriptors, and 3D conformer generation.

ChEMBL Database

Curated bioactivity data for model training, including enzyme inhibition and binding assays.