An explosion in access to high-power hardware and deep computer learning has made companies like Atomwise Inc. confident that artificial intelligence (AI) can finally deliver on its decades-old promise to accelerate drug discovery. The company plans to put hundreds of its computationally discovered compounds in investigators' hands through its new awards program, but the question is whether the results of awardees’ experiments will validate the platform and win over skeptics jaded by years of AI hype.
Computer algorithms have been touted as revolutionary time- and cost-savers for drug development since the 1980s, but have yet to substantially lower either hurdle. However, Atomwise and others believe growth in computer processing power and in deep-learning methods over the last decade could be the difference-makers that fulfill the technology's promise for molecular design.
Specifically, the deep-learning advances use layers of successive processing units to extract key features from complex data, which improves the way computers learn rules and use them to reason about new data.
Atomwise's platform, dubbed AtomNet, uses structural information to predict binding between molecular targets and small molecules. But unlike widely used molecular docking or quantitative structure-activity relationship (QSAR) models, which scientists program to follow the basic rules of chemistry, AtomNet teaches itself those rules by processing millions of data points capturing successful and unsuccessful ligand-binding interactions (see "Machines Learn Biochem").
Atomwise Inc.'s artificial intelligence platform, dubbed AtomNet, learns the rules of biochemistry by using neural networks to sift through large molecular databases containing successful and unsuccessful