In a recent collaborative study, researchers from the University of Washington, the Korea Institute of Science and Technology, Seoul, Republic of Korea, and the University of California developed a highly accurate structure-based virtual screen method, RosettaVS, for drug discovery. Their approach outperforms other state-of-the-art methods on a wide range of benchmarks, partially due to their ability to model receptor flexibility. They incorporate this into a new open-source artificial intelligence accelerated virtual screening platform for drug discovery. 

According to the researchers, using this platform, they can screen multi-billion compound libraries against two unrelated targets, a ubiquitin ligase target KLHDC2 and the human voltage-gated sodium channel NaV1.7. This method allows for the accurate modelling of protein-ligand complexes, accommodating full flexibility of receptor side chains and partial flexibility of the backbone. However, it is not directly applicable for large-scale virtual screening due to: 

1. Its inability to accurately model certain functional groups (as the original method was tested on hundreds of thousands of compounds rather than the billions in this study)
2. It lacks an entropy model for accurately ranking different compounds binding to the same target. Moreover, it is prohibitively expensive to dock each compound in a multi-billion chemical compound library using physics-based virtual screening methods.

The researchers utilized active learning techniques to guide the exploration of the vast chemical space. The greedy strategy was used to select new compounds for each iteration to augment the training dataset without using explicit uncertainty information to reduce the computational cost and inference time.

A comparison

The superior performance of RosettaGenFF-VS and RosettaVS on the CASF2016 and DUD benchmarks establishes it as a leading physics-based method for ligand docking and virtual screening. The notable performance of RosettaVS comes from two major advances. Firstly, the combination of high docking accuracy and sampling efficiency allows the virtual screening protocol equipped with RosettaGenFF-VS to find the correct binding minimum of the protein-ligand complex more effectively than other methods. Secondly, unlike most other virtual screening methods that tend to work well only on more hydrophobic, deeper, and larger protein pockets, our method also demonstrates high performance with more polar, shallower, and smaller pockets, likely due to the better balance of protein-ligand versus intra-ligand molecular energies achieved by RosettaGenFF-VS.

According to the researchers, though their methodology outperforms existing approaches in all aspects, they believe there is room for further improvement. A notable trend in recent years has been the surge in the application of artificial intelligence across various scientific domains, including protein structure prediction, drug discovery, and materials design.

The researchers believe that the future enhancements to their protocol will likely involve the integration of GPU acceleration and deep learning models.

Nature