Generative Models Should at Least Be Able to Design Molecules That Dock Well: A New Benchmark

Designing compounds with desired properties is a keyelement ofthe drug discovery process. However, measuring progress in the fieldhas been challenging due to the lack of realistic retrospective benchmarks,and the large cost of prospective validation. To close this gap, wepropose a benchmark based on docking, a widely used computationalmethod for assessing molecule binding to a protein. Concretely, thegoal is to generate drug-like molecules that are scored highly bySMINA, a popular docking software. We observe that various graph-basedgenerative models fail to propose molecules with a high docking scorewhen trained using a realistically sized training set. This suggestsa limitation of the current incarnation of models for de novo drug design. Finally, we also include simpler tasks in the benchmarkbased on a simpler scoring function. We release the benchmark as aneasy to use package available at https://github.com/cieplinski-tobiasz/smina-docking-benchmark. We hope that our benchmark will serve as a stepping stone towardthe goal of automatically generating promising drug candidates.

Automated summary

Designing compounds with desired properties is essential in drug discovery, but measuring progress has been challenging due to lack of benchmarks and validation cost. To address this, a docking-based benchmark was proposed, using the widely used SMINA software. Graph-based generative models fail to produce high-scoring molecules, suggesting limitations in current models for de novo drug design. The benchmark also includes simpler tasks and is available at https://github.com/cieplinski-tobiasz/smina-docking-benchmark.