Developing Graph Neural Networks for Drug Discovery

We develop Graph Neural Networks for molecular modeling in Drug Discovery — a task where each molecule is represented as a graph: atoms as nodes, chemical bonds as edges. This approach preserves the topology of the molecule without information loss, inevitable when converting to SMILES or Morgan fingerprints. Our solutions allow pharmaceutical companies to accelerate candidate search by 3-5 times compared to classical virtual screening methods. We will assess your project and propose an optimal architecture within 2 days.

Why fingerprints are no longer enough

Morgan fingerprints (ECFP4/ECFP6) are a mature technique, but a fixed-length bit vector loses information about 3D conformation and long-range interactions. On a solubility prediction task (logS), ECFP4 + Random Forest gives RMSE ~0.85, while a well-trained Directed Message Passing Neural Network (D-MPNN, implementation in chemprop) achieves RMSE ~0.65 on the same MoleculeNet benchmark. The gap is modest, but on tasks of activity against specific targets it is significantly wider.

Which GNN architectures we use

Message Passing Neural Networks (MPNN)

Basic paradigm: each atom accumulates information from neighbors through T rounds of message passing, and a final readout aggregates atomic embeddings into a molecular one. Chemprop is the most used implementation in pharma, trained via directed message passing over bonds.

SchNet / DimeNet / PaiNN

These account for 3D atomic coordinates (from conformation generation via RDKit or ETKDG). SchNet encodes distances using radial basis functions; DimeNet adds angular interactions. On quantum property prediction tasks (QM9: HOMO/LUMO energy, dipole moment) PaiNN achieves MAE at the level of DFT calculations with thousands of times higher speed.

Equivariant GNNs (SE(3)-Transformers, EGNN, NequIP)

Architectures invariant to rotation and reflection of the molecule — important for correct prediction of properties that depend on spatial orientation. NequIP shows state-of-the-art on force field prediction tasks.

Protein–ligand graph for virtual screening

Ligand–target interaction is modeled as a heterogeneous graph: two node types (ligand atoms and protein atoms in the binding pocket) and two edge types — intramolecular and intermolecular. GraphDTA, KIBA/Davis benchmarks. On KIBA: Pearson correlation ~0.89 for best GNN vs ~0.79 for classical docking scores. This comparison shows GNNs give accuracy improvement up to 10%. Our GNN models reduce computational costs by 30-40% compared to traditional docking methods.

What we provide

Within the project we deliver:

• A GNN model tailored to your task (property prediction, generation, screening)
• Documentation on architecture, training, and API
• Docker images for deployment
• Training for your team on working with the model
• Post-project support for 3 months

How we build GNNs for ADMET

From our practice: working with a pharmaceutical client, we faced the task of predicting 20+ ADMET properties (Absorption, Distribution, Metabolism, Excretion, Toxicity). A multi-task GNN was trained simultaneously on all tasks using a shared molecular encoder. According to MoleculeNet benchmark, D-MPNN shows RMSE 0.65 on logS.

Problem: uneven data coverage across tasks. hERG cardiotoxicity: ~10k molecules; mouse acute toxicity: ~7k; plasma protein binding: ~3k. With naive multi-task learning, the model favors tasks with larger datasets.

Solution: task-specific learning rates via gradient surgery or PCGrad (Project Conflicting Gradients). Plus uncertainty-aware training: Monte Carlo Dropout to estimate epistemic uncertainty — important for the model to say "I don't know" on out-of-distribution molecules.

Results on ChEMBL benchmark with multi-task D-MPNN:

The gain from multi-task is especially noticeable on small datasets.

Molecule generation via GNN

Junction Tree VAE (JT-VAE) generates molecules through hierarchical decoding: first a graph of substructure "fragments" (rings, chains), then assembly. Guarantees 100% validity (unlike SMILES-based VAE with ~70%).

Graph Diffusion Models — DiGress and similar apply a diffusion process in graph space. They generate chemically diverse molecules with specified properties (property-conditioned generation via classifier-free guidance).

Practice: generation → scoring via fast ADMET GNN → molecular dynamics validation for top candidates via OpenMM or GROMACS. This closes the development cycle.

How we evaluate your project

Development cost depends on data volume and task complexity. We offer a preliminary assessment within 2 days — contact us, and we will prepare a detailed commercial proposal. Indicative timelines: MVP for property prediction (one target, ready dataset) — 6–10 weeks; full ADMET platform with generation and validation — 6–12 months.