A new set of computer models has successfully predicted
negative side effects in hundreds of current drugs, based on the similarity
between their chemical structures and those molecules known to cause side
effects, according to a paper appearing online in Nature.
The team, co-led by researchers in the University of California,
San Francisco School of Pharmacy, Novartis Institutes for BioMedical Research
(NIBR), and SeaChange Pharmaceuticals, Inc.—a UCSF spinoff company launched by
two of the paper’s authors—set out to test how well a computer model could help
researchers eliminate risky drug prospects by identifying which ones were most
likely to have adverse side effects.
Drugs frequently interact with more than one target, with
hundreds of these targets linked to the side effects of clinically used
therapeutics. Focusing on 656 drugs that are currently prescribed, with known
safety records or side effects, the team was able to predict such undesirable
targets—and thus potential side effects—half of the time.
That’s a significant leap forward from previous work, which
has never tackled hundreds of compounds at once, according to Brian Shoichet,
PhD, a UCSF professor of pharmaceutical chemistry who was the joint advisor on
the project alongside Laszlo Urban, MD, PhD, at Novartis.
As a result, it offers a possible new way for researchers to
focus their efforts on developing the compounds that will be safest for
patients, while potentially saving billions of dollars each year that goes into
studying and developing drugs that fail.
“The biggest surprise was just how promiscuous the drugs
were, with each drug hitting more than 10% of the targets, and how often the
side-effect targets were unrelated to the previously known targets of the
drugs,” said Shoichet, whose laboratory is renowned for its work in using
computational simulations to identify new targets for known drugs. “That would
have been hard to predict using standard scientific approaches.”
Adverse drug effects are the second most common reason,
behind effectiveness, that potential drugs fail in clinical trials, according
to the paper. The cost of developing an approvable drug is frequently cited at
about $1 billion across 15 years, although recent estimates have ranged as high
as $4 billion to $12 billion per drug, depending upon how many of these
failures are included in the estimate.
“This basically gives you a computerized safety panel, so
someday, when you’re deciding among hundreds of thousands of compounds to
pursue, you could run a computer program to prioritize for those that may be
safest,” said Michael Keiser, PhD, co-first author of the paper, who started
working on the project as a doctoral student in Shoichet’s laboratory and co-founded
SeaChange with Shoichet and John Irwin, PhD, also of UCSF, upon graduation.
It also offers the possibility for identifying possible new
uses for medications that are already on the market, according to Peter
Preusch, Ph.D., who oversees structure-based drug design grants at the National
Institutes of Health’s National Institute of General Medical Sciences, which
partly supported the study.
“By providing a way to identify the unintended targets of a
drug, this advance will not only help streamline the drug development pipeline,
but also will provide valuable guidance in efforts to repurpose existing drugs
for new diseases and conditions,” Preusch said. “This work represents a notable
contribution that is likely to find broad applications in the pharmaceutical
The project builds on UCSF’s legacy as a leader in
developing computer-based approaches to efficiently screen millions of
chemicals for those with the best potential for drug development. The UCSF
School of Pharmacy was the first to develop computer-based molecular docking
software, which both public and private researchers use to visualize how
potential drugs might attach to target molecules to inhibit their function. It
also builds upon UCSF’s commitment to industry collaborations that advance
pharmaceutical science. Novartis has one of the strongest and most productive
drug pipelines in the industry, with more than 130 projects in clinical
development, according to the company.
The current project is based on technology UCSF developed,
known as the “similarity ensemble approach” (SEA), which compares the shape of
each drug to thousands of other compounds and uses that to predict which
proteins they might both bind to — essentially, guilt by association. The
technique was named among Wired magazine’s “Top Scientific Breakthroughs of
In this project, the UCSF and SeaChange team ran a computer
screen on 656 drugs that are currently in clinical use to predict which ones
were most likely to bind to the 73 target proteins that appear on Novartis’
safety panel for testing drugs for side effects such as heart attacks. Meanwhile,
NIBR developed a statistical method of relating those targets to known side
The computer model identified 1,241 possible side-effect targets
for the 656 drugs, of which 348 were confirmed by Novartis’ proprietary
database of drug interactions. Another 151 hits revealed potential side effects
that had never been identified for these drugs, yet which Novartis confirmed
through lab testing. Among those was a synthetic form of estrogen that has been
known for years to cause stomach pain, with no known cause. The screen showed
that it binds strongly to a target known as COX-1, which is the protein target
of non-steroidal anti-inflammatory drugs, such as aspirin, which also can cause
stomach pain, ulceration, and bleeding.