We report here the results from more than 100 prospective applications of machine-learning to small-molecule hit-discovery. Physical high-throughput screens are no longer a viable option for early-stage drug discovery because synthesized-on-demand ultra-large screening libraries (ULSL), which cover tens of billions of molecules, are too big for brute-force physical screening. Effectively exploring ULSL requires extraordinarily accurate algorithms, to avoid an overwhelmingly large number of false-positives because of the magnitude of the explored space. While machine learning tools promise to revolutionize drug discovery in precisely this manner, it has been extremely challenging to evaluate their actual prospective performance. We showed that the excellent performance reported on standard medicinal chemistry benchmarks are likely due to weaknesses in the design of these benchmarks, rather than the generalizability of the algorithms. Therefore, we have been running the largest assessment of machine learning for hit discovery in history, comprising over 700 projects on 575 targets with 330 universities in 40 countries. Our first 100+ results empirically demonstrate the viability of our machine-learning technology as a replacement for high-throughput screens.

Results from the world's largest distributed prospective application of machine learning to small-molecule hit discovery

ACS Fall 2020 Virtual Meeting

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