PAINS alerts, a widely used screening tool deployed in the early phases of drug discovery to weed out undesirable compounds, are wrong so often they can’t be trusted on its own, according to scientists at the University of North Carolina at Chapel Hill.
PAINS — or pan-assay interference compounds —are a prominent source of false positives in the drug-discovery process. PAINS are biologically active compounds that masquerade as potential drug candidates during the initial high volume screening used to search for possible new drugs. PAINS work by disrupting the assay technology used in the screening to report biological activity, but they are not active against the intended biological target.
Seven years ago, a team of scientists identified a set of a molecular fingerprints that they said could be used to identify PAINS compounds. They found 480 molecular fragments they called PAINS alerts. The presence of a PAINS alert in a compound meant the compound was a PAINS. Pharmaceutical scientist were grateful to finally have an easy-to-use tool to filter PAINS out of the drug discovery process.
Not so fast, says Alex Tropsha, Ph.D., the K.H. Lee Distinguished Professor at the UNC Eshelman School of Pharmacy. His team, led by graduate student Stephen Capuzzi, took a hard look how the 480 PAINS alerts were developed. They found that 328 of the alerts — almost 70 percent — were created from just four or fewer compounds each. No viable conclusions should be made from such small samples, Capuzzi said.
“In some cases, just one compound with a certain substructure exhibited bad behavior, and then an alert was created. In the PAINS screening tool, that alert gets applied to everything with a similar structure,” Capuzzi said. “It’s like someone who was bitten by a dog named Fido assuming that every dog named Fido is dangerous.”
The UNC-Chapel Hill researchers analyzed thousands of compounds that have been tested in hundreds of experiments and showed that majority of compounds flagged with PAINS alerts do not exhibit any tendency to interfere with drug-candidate screening technology and may in fact be consistently inactive despite extensive testing.
“PAINS are real, and they’re a real problem for medicinal chemists who might be screening tens of thousands of compounds a day looking for a drug candidate,” Tropsha said. “However, PAINS alerts flag vast numbers of good compounds as PAINS. We also found that the presence of PAINS alerts, contrary to expectations, did not reflect any heightened activity during the assay process.”
When challenging a popular idea like PAINS alerts, the UNC-Chapel Hill researchers had to be sure they were right in saying the system is critically flawed so they tested their idea extensively.
First, they analyzed thousands of compounds in a publicly available database of thoroughly tested chemicals and showed that many compounds that possess PAINS substructures do not actually appear to interfere with drug-screening technology.
Next, the team applied PAINS filters to a group of chemical compounds known as dark chemical matter. The compounds in this group have all been tested more than 100 times each without showing any biological activity. They, by definition, should not contain any PAINS alerts because PAINS are biologically active, but the UNC team found 109 PAINS alerts present in 3,570 dark chemical matter compounds reviewed.
Finally, UNC and other researchers have also observed that 87 FDA-approved drugs, including 19 from the World Health Organization List of Essential Medicines, contain PAINS alerts. One example is zidovudine, which is used to prevent and treat HIV/AIDS. Between 2008 and 2011, zidovudine was prescribed over 1 million times in the U.S. alone. Eltrombopag, an FDA-approved breakthrough treatment for thrombocytopenia, or low platelet count, contain PAINS alerts. So does Tadalafil, better known as Cialis, which has annual sales topping $2 billion and is approved to treat both erectile dysfunction and enlarged prostate.
“People have been blindly using PAINS alerts, and if they continue to do so, many much-needed drugs may never be found,” Tropsha said. “Computational tools can aid in the triage process for finding new drug compounds, but if they’re used inappropriately or without caution, the same tools may prevent future drugs from ever making it to market.”
Tropsha’s group continues to work on computational approaches that will reliably predict possible drug candidates while eliminating unlikely drug candidates from further consideration. Their study was published in the ACS Journal of Chemical Information and Modeling and was selected as an ACS Editors’ Choice Article.
The study’s authors are
- Stephen J. Capuzzi, a graduate student in the Division of Chemical Biology and Medicinal Chemistry at the UNC Eshelman School of Pharmacy;
- Eugene Muratov, Ph.D., a research assistant professor and associate director of the Laboratory for Molecular Modeling at the UNC Eshelman School of Pharmacy; and
- Alexander Tropsha, Ph.D., the K. H. Lee Distinguished Professor in the Division of Chemical Biology and Medicinal Chemistry and associate dean for pharmacoinformatics and data science at the UNC Eshelman School of Pharmacy.