The availability and accessibility of chemistry-on-demand and ultra-large screening libraries (ULSLs) give rise to an opportunity for mining new chemical spaces for novel scaffolds. This combination of cheap à la carte synthesis that covers tens of billions of molecules may mark the inception of a golden age of computational approaches, where a comprehensive physical high-throughput screen is no longer a viable option for early-stage drug discovery. We have previously presented AtomNet -- our convolutional network based model for predicting binding affinities from the three-dimensional protein-ligand complex. In this talk, we introduce our approach for trawling through billions of compounds using a more performant model (“trawler”) to predict their AtomNet scores. The most promising compounds found by the trawler are then evaluated on AtomNet, and these scores are used to improve the trawler model iteratively. The implementation exploits the recent advances in machine-learning algorithms and scalable on-demand computing resources. We detail the algorithmic and engineering challenges of deploying ULSLs in a production setting, and the tradeoffs between speed and accuracy. We present some early results from a few projects, and discuss avenues for further improvement.

Billion Trawler Answer: A scalable and effective machine learning based solution for the virtual high-throughput screening of ultra-large libraries

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.