Drug discovery is challenging and expensive, especially in the phase of lead optimization which predicts binding to a target. Computational approaches can add substantial value. Computing relative binding free energies (RBFE) between congeneric molecules using molecular dynamics (MD) greatly reduces the search space and considerably improves convergence. The Silicon Therapeutics (STX) INSITE platform uses an RBFE protocol based on MD running on GPUs as part of the lead optimization pipeline. STX has proven the capabilities of INSITE by finding small molecule inhibitors for challenging targets that have no known small molecule inhibitors. The two critical bottlenecks are throughput and timescale. For throughput, while GPUs have substantially improved cost-effectiveness, improvement is still highly desirable, especially for energy cost per simulation. For timescale, current GPU RBFE simulations typically run 10ns or less which often yields unconverged results. We propose to address both problems by accelerating INSITE with FPGA-based clusters and FPGA cloud instances. FPGAs are commodity computation devices whose primary use has been in communication routers; they are also ideal for MD. The hardware adapts to the application, rather than the reverse, and so effects high efficiency. Also, since FPGAs are hybrid communication/computation processors, large-scale communication can proceed with both high bandwidth and low latency. Our preliminary work has shown that MD on clusters of FPGAs approaches the performance of proprietary ASIC-based clusters for several core functions, and that a current 256-node FPGA-accelerated cluster can simulate a 50K particle model at a rate of over 10us per day; this is 20x that of a commodity cluster of any size. Our overall goal is to a create a commercial-quality pipeline for running economical and long-timescale RBFE simulations. The programs will be highly useful to the internal drug discovery efforts of STX. STX will also explore avenues to provide the software as a service either in the cloud (e.g., through AWS) or through an in-house platform. The overall Aim of this Phase I proposal is to prove the concept of FPGA-accelerated RBFE for clouds and clusters in delivery of the modeled performance, in validation of simulation quality, and in other software engineering metrics. Therefore our Aim 1 is accelerate key RBFE functions with FPGAs, with subtasks designed to span the space of target systems and deployment scenarios.
Our Aim 2 is to enhance the STX computational toolflow to create internal and external infrastructure to support the work in Aim 1, to develop new computational methods to take advantage of the target architectures, and to evaluate and specify deployment options.
Computational approaches have been found to substantially aid in the discovery of new drugs. The goal of this proposal is to apply novel and emerging computational methods to improve the cost-effectiveness and quality of these approaches.
