VI-SEEM team ranks #1 in international computer-aided drug design competition

Computational tools hold great promise to speed up the discovery of new, safer medications and reduce the cost of the drug discovery process. In recent years, significant advances in molecular modeling have contributed to the discovery of several drugs now in the market. Within the VI-SEEM project, the Cournia lab at the Biomedical Research Foundation, Academy of Athens participated in an international drug design competition. This competition was organized as a blind computational prediction of experimental results, and was coordinated by the Drug Design Data Resource (D3R), University of California, San Diego with the goal to examine the efficiency of current computational methodologies and allow for improved methods to emerge. Under this scope, Grand Challenge 2 was held between September 2016 – February 2017, during which the team had to predict the binding modes of Farnesoid X Receptor (FXR) agonists and their free energies of binding to FXR. The Challenge consisted of two Stages: in Stage 1 prediction of the crystal structures of 36 FXR-agonist complexes and the relative binding affinities of 33 FXR ligands binding to FXR took place. In Stage 2, crystal structures of the 36 protein-ligand complexes (provided by Roche pharmaceuticals) were disclosed, and the free energy predictions, using the experimental crystal structures were repeated.

For predicting the crystal structures, the team utilized a variety of computational techniques including molecular docking, shape & interaction similarity calculations, enhanced by physics-based methods, such as binding pose metadynamics and free energy perturbation calculations. For the binding affinity predictions, we applied free energy perturbation calculations and Molecular Dynamics simulations.

These calculations require advanced computing facilities in order to provide realistic results in a reasonable timeframe. By applying for VI-SEEM resources, we had could access the Supercomputer System “ARIS” in GRNET, Greece. For our calculations, we were awarded 24,000 GPU hours. The ARIS DELL PowerEdge R730 system comprises of 88 NVIDIA Tesla K40M GPU cards (2880 CUDA cores each), distributed over 44 nodes, each one housing two cards. The input and output data from our calculations is stored on the repositories site of GRNET and available to the public.

After evaluating these predictions by comparison with experimental data, brought the Cournia lab ranked 1st in the pose predictions out of 49 completed submissions with a median RMSD of 0.99 Å. A case of a successful pose prediction is illustrated in Figure 1. In the binding affinity predictions, the lab ranked 3rd out of 12 submissions, in which alchemical free energy methods were used.


Figure 1. Successful protein-ligand crystal structure prediction.