Here at Atomwise, we are really proud of the AtomNet® tool we’ve built to perform virtual screening of targets against billions of compounds to find promising candidates. This approach has already provided new leads for academic labs and drug discovery pipelines (around the world). But we are still on the lookout for ways to improve our approach, whether it’s the machine learning part of our AI tool or the various algorithms that run behind the scenes.
Today on Behind the AI, we will highlight ways to enhance the performance of our AI tools with some clever software engineering. About a year ago, we tasked Senior Software Engineer Adrian Morrison, PhD, with an important development project: reconfiguring key code to improve efficiency. Our AtomNet® technology makes predictions based on chemical structure necessitating a two-step process. The first step is to generate a viable structure of a candidate drug molecule interacting with a target protein, known as docking, and the second step is to run this structure through our neural network and predict the binding affinity of the docked compound. In our original software workflow, the docking step ran on CPUs while the scoring step ran on GPUs. Running the workflow on two different specialized cloud instances made for longer run times, high complexity, and too much maintenance. Morrison’s goal was to reconfigure everything to run on GPUs for a streamlined operation.
To rethink the docking step — a process in which the most suitable orientation and shape of a ligand within a target’s binding site, the pose, is determined computationally — Morrison began with the tools commonly used for this procedure. AutoDock Vina is one of the best known, and Smina is a more efficient, parallel implementation of it. Though they’re tailored for CPUs, these tools are very good at what they do.
CUina takes advantage of the massively parallel GPU operations, as it simultaneously optimizes thousands of poses for dozens of ligands. Thanks to this and a lean implementation, it outperforms CPU alternatives and actually delivers equivalent or better pose quality results. Even better, it does so five times faster than our previous method. It’s exactly what we were hoping for when Morrison embarked on the ambitious project.
But he sees the accomplishment as much bigger than a software engineering endeavor. “The heart and soul of our company are the scientists who actually do the drug discovery work,” Morrison says. “CUina is a sharper tool that allows them to do that work better.”
Learn More
Adrian Morrison, PhD, was selected to present a poster on this project and his work in implementing CUina at the American Chemical Society Fall 2020 Virtual Meeting & Expo. Take a deeper dive into his work by listening to the audio presentation and viewing the poster - Efficient GPU Implementation of AutoDock Vina
About Atomwise
Atomwise is revolutionizing how drugs are discovered with AI. We invented the use of deep learning for structure-based drug discovery, today developing a pipeline of small-molecule drug candidates advancing into preclinical studies. Our AtomNet® technology has been used to unlock more undruggable targets than any other AI drug discovery platform. We are tackling over 600 unique disease targets across 775 collaborations spanning more than 250 partners around the world. Our portfolio of joint ventures and partnerships with leading pharmaceutical, agrochemical, and emerging biotechnology companies represents a collective deal value approaching $7 billion. Atomwise has raised over $174 million from leading venture capital firms to advance our mission to make better medicines, faster.
