Leveraging artificial intelligence could help drug developers more efficiently screen chemical libraries to identify the few chemicals most likely to affect their research targets.
Carbone researchers Anthony Gitter, PhD, Michael Newton, PhD, and Spencer Ericksen, PhD, have partnered to develop AI tools to streamline early investigations and lay the groundwork for increased success in new drug development. Their work at UW feeds into a global effort to share knowledge and advance new purposes for AI as a useful tool in medical research.
“Open science has been a big theme for us,” said Gitter, an associate professor of Biostatistics and Medical Informatics at UW and investigator for the Morgridge Institute for Research. “We can’t do the research we do unless we build on software and datasets. We build on previous work, and then we turn around and share everything we do so now other people can build on it.”
Developing new medications for cancer and other illnesses is a lengthy process, taking significant funding and several years of intensive study at the bench level and in clinical trials to become FDA-approved.
“It’s a very long process … I think 10 years is probably even a lower estimate in actual time it takes to get a new medication approved,” said Newton, who also is chair of the UW Department of Biostatistics and Medical Informatics.
The initial step in creating a new medication involves identifying a biological target, such as a mutated protein in a cancer cell, and determining which types of chemicals can affect that target. Ericksen performs screenings to identify potential effective chemicals in his work at the UW Carbone Cancer Center Small Molecule Screening Facility, one of the shared research cores assisting researchers at UW as well as beyond campus.
Traditionally, these chemical screens have been done in a “brute force” high-throughput method that casts a wide net to test all possible chemicals in a library to identify a small group of possible options for further study.
Gitter, whose expertise is in computer science, collaborated with Ericksen to create a text-based AI model that leverages existing published research, chemical data and AI chat models. When prompted about the biological target and desired effect in a drug, it suggests possible chemicals to study. Yifan Deng, a PhD student in Gitter’s lab, led this work.
“It doesn’t prove anything definitively, but it at least says, ‘that could be promising, you might be on the right track,’” Gitter said of their model.
Newton, Gitter and Ericksen have also developed an AI model that helps narrow down potential chemicals of interest to test when there is little information about a new potential drug target. Peng Yu, a PhD student in Newton’s lab, also contributed to this research.
While AI tools can help with efficiency, they don’t replace key aspects of drug research. Ericksen said scientists still need to review those suggested chemicals to ensure that the proposed chemicals and data behind them are sound.
“I think there are people worried about AI just replacing the human aspect, but it’s a tool that can help guide you to some of the more realistic options to try, but you still need that person with expertise, that training, to understand what’s being put out and train it,” he said. “These are starting points, and I would say most of the work that we do is in very early stage just validating to see is there promise here.”