Equivariant Graph Neural Networks
for Drug Discovery
Predicting how a potential drug-like molecule interacts with a target protein is one of the main challenges within Drug Discovery. Currently, available methods are computationally expensive and usually require knowledge of the target site, which can be problematic when dealing with novel proteins.
Equivariant Graph Neural Network
On our way to building AIchemy, a platform for Bayesian Deep Virtual Screening, we investigate the potential of Equivariant Graph Neural Network (EGNN) which predicts not only the receptor binding site (without any prior knowledge), but also the bound pose and orientation of the ligand. We compare the results to Qvina-W, a classical tool for blind docking - predicting the receptor-ligand binding without prior knowledge.
Performance of the EGNN (left) with Qvina-W (right) for docking of an inhibitor drug for leukemia and gastrointestinal stromal tumors from Novartis, Imatinib, against the Tyrosine Kinase 6HD6.
As our test case, we perform docking of an inhibitor drug for leukemia and gastrointestinal stromal tumors from Novartis, Imatinib, against the Tyrosine Kinase 6HD6. We compare the performance of the EGNN (left) with Qvina-W (right). Models' predictions are shown in green, while the ground truth is shown in red. In blue we show another ligand available in the 6HD6 crystal structure from PDB. We use Qvina-W with default settings, generating 10 poses with previously centered receptors and a searching space of 50x50x50. The centroid distance is a measure of how well the model predicted the binding pocket, for Qvina-W it is calculated for the predicted pose closest to the ground truth.
As you can see the EGNN completed over 7 times faster than Qvina-W, while detecting the binding site better! Not only was Qvina-W's closest prediction further from the ground truth than EGNN's but Qvina-W also favored the wrong binding site. These results show promise for faster Drug Discovery, proving the usefulness of Graph Neural Networks for molecular docking.
Our next step is developing the uncertainty capability of the EGNN in terms of epistemic and aleatoric uncertainties.
Stay tuned for more updates! :)
For more details on Equivariant Neural Networks:
https://arxiv.org/abs/2202.05146
https://arxiv.org/abs/2102.09844
As you can see the EGNN completed over 7 times faster than Qvina-W, while detecting the binding site better! Not only was Qvina-W's closest prediction further from the ground truth than EGNN's but Qvina-W also favored the wrong binding site. These results show promise for faster Drug Discovery, proving the usefulness of Graph Neural Networks for molecular docking.
Our next step is developing the uncertainty capability of the EGNN in terms of epistemic and aleatoric uncertainties.
Stay tuned for more updates! :)
For more details on Equivariant Neural Networks:
https://arxiv.org/abs/2202.05146
https://arxiv.org/abs/2102.09844
Read More