Each year, scientists identify thousands of compounds that show promise to treat disease. But many fail before becoming medicines, in part because it’s hard to predict how they’ll behave inside the human body.
Researchers at UC San Francisco are building an open-source system of laboratory tests that mimic the body at the level of proteins, cells, and tissues. By generating large amounts of high-quality experimental data, the team hopes to train AI models that can help scientists identify more promising drug candidates.
That effort, called the OpenADMET Consortium, will grow and refine this work with $15 million in support from the OpenAI Foundation. The OpenAI Foundation is a nonprofit dedicated to ensuring artificial general intelligence benefits all of humanity.
OpenADMET began two years ago thanks to a federal grant from the Advanced Research Projects Agency for Health (ARPA-H). The new funding from the OpenAI Foundation will build on OpenADMET’s initial emphasis on drug metabolism to study pharmacokinetics — the properties that determine how chemicals move through the body — while generating large-scale experimental datasets for training AI models.
“Most drug development fails because the drugs have toxicity, get metabolized too quickly, or don’t reach the intended tissue,” said James Fraser, PhD, chair of the Department of Bioengineering and Therapeutic Sciences in the UCSF Schools of Pharmacy and Medicine, and a governing board member of OpenADMET. “If we can solve these problems in a predictive way, we’ll have fewer failures, and the successes will be faster, cheaper, and more plentiful.”
OpenADMET also includes three other UCSF faculty members — Zev Gartner, PhD; Willow Coyote-Maestas, PhD, MS; and Aashish Manglik, MD, PhD — as well as the Bay Area biotech company, Octant, and the nonprofit, Open Molecular Software Foundation (OMSF).
From molecules to medicines
Pharmacokinetics is often broken down into absorption, distribution, metabolism, and excretion, or ADME. The T in OpenADMET refers to toxicity.
The team is devoting some of its efforts to determining which compounds can cross the blood-brain barrier — the network of specialized blood vessels that shield the brain — in part by building a “blood brain barrier on a chip.” It could help scientists prioritize prospective treatments for conditions like Alzheimer’s based on the treatment’s ability to traverse the barrier.
The UCSF laboratories will use techniques like X-ray crystallography and cryo-EM to understand how different drug-like molecules interact with proteins that transport drugs across membranes, including molecular gatekeepers at the blood-brain barrier.
Octant is developing more efficient ways to measure the physical properties that shape a drug’s pharmacokinetics, using techniques like chromatography and mass spectrometry to generate data on thousands of different molecules.
OMSF is developing the software and open infrastructure needed to turn those data into predictive AI models and test them through community challenges — an approach that has already drawn hundreds of participants from academia, biotech, pharma, and the machine-learning community.
“Traditionally, after finding a promising drug candidate, groups would design experiments around that one molecule to try to predict its performance in people,” Fraser explained. “Our goal is to develop a system that works for any drug candidate, so no one has to start their validation from scratch.”