GNNs for Molecules: Drug Discovery and Material Design
Published:
TL;DR: Drug discovery requires predicting how molecules interact with biological targets — a task that historically required either expensive experiments or domain-expert features. GNNs learn directly from molecular graphs, outperforming Morgan fingerprints on most property prediction benchmarks and enabling virtual screening of billions of compounds.
The Drug Discovery Pipeline
Drug discovery takes 10-15 years and $2B+ per approved drug. GNNs accelerate three key stages:
- Virtual screening: filter billions of candidate molecules to thousands using property predictions
- Lead optimisation: predict ADMET (absorption, distribution, metabolism, excretion, toxicity) properties
- De novo design: generate novel molecules with desired properties
What GNNs Predict
ADMET properties:
- Solubility: how much dissolves in water (affects bioavailability)
- Lipophilicity (LogP): determines membrane permeability
- Toxicity (hERG, AMES): cardiac toxicity, mutagenicity
- Metabolic stability: how quickly the liver degrades the drug
- Blood-brain barrier penetration: reaches the brain?
Binding affinity:
- Predicted IC50, Ki, Kd for specific protein targets
- Virtual screening: rank candidates by predicted affinity
Quantum chemistry (QM9 benchmarks):
- HOMO-LUMO gap (electronic excitation energy)
- Dipole moment, polarisability
- Zero-point energy
The GNN Pipeline for Molecules
SMILES string → RDKit graph → Atom/bond features
↓
GNN (2-4 layers)
↓
Node embeddings
↓
Global pooling (sum/attention)
↓
Graph embedding
↓
MLP → property prediction
Atom features: atomic number, formal charge, number of Hs, hybridisation (sp/sp²/sp³), aromaticity, chirality
Bond features: bond type (single/double/triple/aromatic), is-conjugated, is-ring, stereo
Key Models for Molecular Property Prediction
MPNN (Gilmer et al., 2017): introduced the message passing neural network framework for molecules. First systematic study showing GNNs outperform Morgan fingerprints on QM9.
AttentiveFP (Xiong et al., 2019): adds graph attention for molecular property prediction. Handles multi-task learning across different ADMET endpoints.
Grover (Rong et al., 2020): self-supervised pre-training on 10M unlabelled molecules, then fine-tune on small labelled datasets. Solves the labelled data scarcity problem in drug discovery.
MolBERT / ChemBERTa: treat SMILES as a sequence, apply BERT-style pre-training. Competitive with graph-based methods on many benchmarks.
Why pre-training matters: Labelled molecular data is expensive — assaying binding affinity for 1000 compounds costs $100K+. GNNs trained from scratch on small datasets overfit. Pre-training on 10M+ unlabelled molecules (ChEMBL, PubChem) provides a strong starting point. Fine-tuning on 1000 labelled examples then reaches performance previously requiring 100K+ labels. </insight> ## Virtual Screening at Scale **The challenge:** DrugBank has 13,000 approved drugs. PubChem has 100M compounds. Synthesisable chemical space has ~10^{60} molecules. Which to test? **GNN-based screening:** 1. Train GNN on known active/inactive pairs for target protein 2. Run inference on virtual library (billions of molecules) 3. Select top-k predicted actives for experimental validation Companies like Insilico Medicine, Schrödinger, and Recursion use GNN-based virtual screening as a core workflow. ## Protein-Ligand Interaction Beyond single-molecule property prediction: predicting how a small molecule (ligand) binds to a protein target. **Input:** protein structure (graph of residues) + ligand structure (graph of atoms) + 3D binding pose **Model:** heterogeneous GNN with protein nodes, ligand nodes, and protein-ligand interaction edges. Equivariant GNNs (EGNN, SE3-Transformers) respect 3D symmetry. **Output:** binding affinity score (docking score) ## Benchmarks - **MoleculeNet:** 17 datasets covering classification and regression across ADMET endpoints - **OGB-molhiv:** HIV activity (41,127 molecules) - **OGB-molpcba:** 128 PCBA assays (437,929 molecules) - **QM9:** 12 quantum chemistry properties (134k molecules) - **MD17:** molecular dynamics trajectories for force field learning ## Summary GNNs have become the default molecular representation learning method in computational drug discovery, replacing handcrafted Morgan fingerprints. The key advantages: end-to-end learning, generalisation across chemical space, and compatibility with both 2D connectivity and 3D geometric information. With pre-training on large unlabelled databases, GNN-based models now approach expert-level performance on standard ADMET prediction benchmarks. ## References - Gilmer, J., Schütt, K. T., Ramsundar, B., Ramakrishnan, R., Bronskill, M., Gomes, C., & Dahl, G. E. (2017). [Neural Message Passing for Quantum Chemistry](https://arxiv.org/abs/1704.01212). *ICML 2017* (MPNN: unified framework for molecular GNNs, benchmarked on QM9 properties). - Rong, Y., Bian, Y., Xu, T., Xie, W., Wei, Y., Huang, W., & Huang, J. (2020). [Self-Supervised Graph Transformer on Large-Scale Molecular Data](https://arxiv.org/abs/2007.02835). *NeurIPS 2020* (GROVER: large-scale pre-training of molecular GNNs on 10M unlabelled molecules for drug property prediction). - Hu, W., Liu, B., Gomes, J., Zitnik, M., Liang, P., Pande, V., & Leskovec, J. (2020). [Strategies for Pre-training Graph Neural Networks](https://arxiv.org/abs/1905.12265). *ICLR 2020* (systematic study of GNN pre-training strategies for molecular property prediction and other biological tasks).