The algorithms that use web search trends to predict our tastes in music, literature and other entertainment forms — known as “recommender systems” — are everywhere in e-commerce, often nailing preferences with eerie accuracy.
Could the same approaches used by sites like Amazon and Netflix be applied to more imposing challenges — say, developing new disease-fighting drugs? A new UW 2020: WARF Discovery Initiative project will explore the possibilities.
The project, one of 21 approved this month by the UW-Madison Office of the Vice Chancellor for Research and Graduate Education, is led by Julie Mitchell, a UW-Madison professor of biochemistry and mathematics. It also features significant input from experts at the Morgridge Institute for Research and the Wisconsin Institute for Discovery.
Anthony Gitter, a Morgridge investigator and assistant professor of biostatistics and medical informatics, says the goal will be to create machine learning tools that dramatically reduce the time and cost associated with screening compounds for therapeutic relevance.
“This is for our UW-Madison colleagues who have found a new target that might be linked to disease, and want to do whatever they can to find interesting drug candidates,” says Gitter, a co-principal investigator on the project.
The problem today is when researchers discover a protein that may play a role in disease, they face the daunting task of screening millions of single chemicals that might favorably interact. There are few ways of narrowing the search to a more feasible scale.
“We are hoping to get the number down to thousands, if not hundreds, of likely chemical candidates that gives scientists a reasonable and more cost-effective place to start,” he says.
Two optimization experts from WID — electrical and computer engineering Professor Robert Nowak and computer sciences Professor Stephen Wright — have experience with recommender systems and the ability to push them to new heights.
They will use results from UW-Madison labs that have completed chemical screening projects to help their systems learn. The project is starting with about 75,000 chemicals that have already been screened by UW-Madison partners. The team will also merge and compare existing public datasets with these 75,000 local results.
The recommender-system algorithm must look across many targets to search for the patterns of chemical activity and learn how a target may react to new chemical. “We may also be able to adaptively select specific experiments that will help to improve predictions about such chemical reactions,” says Nowak, “in a fashion similar to the adaptive recommendation systems we’ve developed to improve beer recommendations (beermapper.com).”
“The project leverages the team’s combined expertise in chemical screening and in theoretical and applied machine learning and optimization,” adds Wright. “It is a truly interdisciplinary effort in the tradition of the Wisconsin Institute for Discovery, the Morgridge Institute for Research, and UW-Madison.”
The goal of UW2020 is to stimulate and support cutting-edge, innovative and groundbreaking research at UW–Madison and the acquisition of shared instruments or equipment that will open new avenues for researchers.