High performance computing (HPC), in particular 3D shape similarity, has played a significant role and made many important contributions to the drug discovery process for several years.

However, the advent of GPU (graphics processing unit) based HPC is poised to disrupt this field and other traditional applications of computational drug discovery.

Traditionally, the benefits of computational drug discovery were limited to those molecular modelers and computational chemists who had the necessary expertise to set up and analyze these very useful but potential lengthy computations. Over the past few years, there has been an increasing shift to develop new or customize existing tools to make the benefits of computational drug discovery more accessible and useful to the broader audience of all chemists involved in the discovery process.

Scientists from Abbott Labs published a pioneering paper on the subject in 2008.1 The article describes efforts to fuse information obtained from multiple computational sources to provide statistically meaningful probabilities of equipotency between any two molecules. This was done by analyzing pair-wise 2D fingerprint similarity as well as 3D shape similarity coupled with Abbott’s extensive collection of measured activity data. Careful analysis of these results showed them how to fuse the 2D and 3D shape similarity results to provide well-validated activity predictions.

In an effort to deploy this technology to a wide audience, this group built a "Google simple" interface to the underlying system that enabled anyone in the organization to upload a molecule of interest and receive back a detailed list of molecules ranked ordered by their probability of having the same potency as the query molecule. This application became known as Lead Hopper and its usage spread quickly throughout the organization. The success of this initial effort led to a follow up paper by the same group that was published in early 2011.2 This paper further improved the predictive values by adding protein structural information—through molecular docking—to the equation.

Despite the obvious potential of Lead Hopper and the underlying technology, the process was limited due to the speed with which results could be delivered. While 2D similarity can be calculated in real time, 3D similarity calculations could require hours of dedicated computation on a large cluster of processors. Seeing the potential benefits to transforming this application and others like it into real-time predictive engines, pharmaceutical companies collaborated to spur the investigation and implementation of a mechanism to make 3D similarity calculations operate in real time.

After working closely with engineers from NVIDIA, it became clear that a GPU-based solution could deliver the desired performance improvements. Furthermore, the performance was found to scale linearly with the number of GPUs, as more hardware resources are made available to calculation.

A combination of advanced technology and close scientific interactions with collaboration members— including Abbott Labs and Pfizer—culminated in the release of a GPU-based implementation of the core 3D shape similarity algorithm used in the original Abbott paper. As a result, a single computer equipped with four NVIDIA Tesla cards is capable of screening over 2 million molecule conformations per second, enabling real-time 3D shape similarity for use in tools like Lead Hopper. See the graph for a comparison of CPU versus GPU computing performance.

The ability to determine which other compounds in a corporate collection are likely to exhibit similar activity to a known query is now a question that can be answered in seconds rather than hours or days. This capability will allow individual chemists to make quick, rationally justified decisions about which additional leads to pursue. In addition, researchers can use a post high-throughput screening (HTS) analysis tool to identify false negatives and potential leads that were not present in the HTS run.

 As history has shown, speed increases of this magnitude frequently disrupt common and expected usage of a given technology and open the doors to many unforeseen applications.

References

1. Muchmore SW, et al. Application of Belief Theory to Similarity Data Fusion for Use in Analog Searching and Lead Hopping. J Chem Inf Model. 2008;May;48(5):941-8. Epub 2008 Apr 17.

2. Swann S, et al. A Unified, Probabilistic Framework for Structure- and Ligand-Based Virtual Screening. J. Med. Chem. 2011;54(5):pp1223–1232.