Machine learning and drug discovery: Never the twain shall meet?

COMP 284

Nigel Duffy, nigel@numerate.com, Jessen Yu, Guido Lanza, John Griffin, john@numerate.com, Paul Boardman, Rich McClellan, Brad Dolin, Patrick Linehan, Sean Sylvis, and Brandon Allgood. Numerate, Inc, 1150 Bayhill Drive, Suite 203, San Bruno, CA 94066
Numerate has developed a novel drug engineering platform wherein key design decisions are made by predictive computational models, developed by machine learning techniques, rather than by chemists. Previous attempts at using machine learning in drug discovery have met with limited success. The reasons for this rest on the inability or failure to address challenging statistical problems associated with ligand-based design. We will discuss the problems that have limited previous attempts and present results from two programs based on the solutions we have devised. In the first program we designed dual-acting compounds for addressing multiple aspects of cardiovascular risk. The 19 compounds synthesized represent four proprietary series and display differential pharmacology relative to statins in vitro and in vivo. In the second program we designed broad spectrum non-nucleoside HIV1 reverse transcriptase inhibitors. We identified compounds with whole cell activities comparable to Sustiva in 6 months with only 21 compounds synthesized.
 

Drug Discovery
8:30 AM-12:10 PM, Wednesday, August 19, 2009 Walter E. Washington Convention Center -- 147A, Oral

Division of Computers in Chemistry

The 238th ACS National Meeting, Washington, DC, August 16-20, 2009