Drug Discovery using Equivariant Graph Neural Networks

Equivariant Graph Neural Networks for Drug Discovery

Determining how potential drug molecules interact with target proteins is one of the main challenges in drug discovery. Current approaches demand significant computational resources and frequently require information about the target binding site, which is a problem when researching unfamiliar proteins.

Equivariant Graph Neural Network

While developing AIChemy, our platform for Bayesian Deep Virtual Screening, we explored Equivariant Graph Neural Networks (EGNN) that predict binding sites blindly (no prior knowledge needed) together with ligand position and orientation. We benchmarked the results against Qvina-W, a traditional blind-docking tool.

The test docks Imatinib, a Novartis kinase inhibitor for leukemia and gastrointestinal stromal tumours, against Tyrosine Kinase 6HD6. In the comparison below, EGNN predictions are in green, the ground-truth pose is in red, and another available ligand is in blue. Qvina-W was run with default settings, generating 10 poses in a 50×50×50 search space.

Two findings stand out:

• Speed. The EGNN ran more than seven times faster than Qvina-W.
• Binding site detection. The EGNN found the correct binding site; Qvina-W’s closest prediction diverged further from ground truth and identified the wrong pocket.

Next steps

We are now layering uncertainty quantification on top, both epistemic (from the model itself) and aleatoric (from the data), so each pose comes with a confidence the downstream medicinal chemistry workflow can act on. That direction fed directly into AIChemy as it exists today and into our subsequent benchmarks against classical virtual screening.

References

• E(n) Equivariant Graph Neural Networks (Satorras et al., 2021)
• EquiBind (Stärk et al., 2022)

Related

• Borch, our open-source deep universal probabilistic programming language used underneath AIChemy.
• AIChemy poster at DDC Europe 2024: identifying novel CDK9 inhibitors, a later benchmark on a 24,000-compound library.
• Desupervised & RG Discovery receive Eurostars Grant, the partnership that funded much of the drug-discovery work.