New Machine Learning Tool Can Predict Adverse Drug Effects

June 29, 2020

A new computer algorithm might be the next step toward accurate prediction of adverse drug reactions.

Researchers from Harvard Medical School and the Novartis Institutes for BioMedical Research announced the creation of an open-source machine learning tool capable of predicting drug adverse effects (AEs).

The study, published in The Lancet journal EBioMedicine, examined 2 databases: 1 that reported adverse drug reactions and another with 184 proteins that specific drugs are known to interact with. Investigators constructed a computer algorithm to develop associations between the drug reactions and the 184 individual proteins.

The algorithm discovered 221 associations, some known and some new. These associations indicated which proteins contribute to certain AEs and which may not.

Adverse reactions to drugs contribute to 2 million US hospitalizations each year, according to the Department of Health and Human Services, and occur during 10% to 20% of hospitalizations, according to the Merck manuals. As such, researchers have attempted numerous tactics to minimize the numbers, but because an individual drug can affect several proteins at once, it is difficult to predict what side effects, adverse or otherwise, may occur. The difficulty extends to identifying the protein responsible.

The new algorithm could help predict these AEs before the drug goes to human clinical trials, as well as before and after it enters the market.

The research team supplied the algorithm with 2 large data sets. The first, provided by Novartis, gave information about the proteins that each of the 2000 drugs interacted with. The second, from the FDA, provided 600,000 physician reports of AEs in patients.

Aside from generating information about the relation between individual proteins and adverse reactions, the algorithm also supported previous observations, such as how a drug binding to the protein hERG could cause cardiac arrhythmias.

To the researchers’ surprise, the algorithm also found an association between the protein PDE3 and over 40 AEs. Previous findings already showed that PDE3 could cause arrhythmias, low platelet counts and elevated levels of enzymes, but this was the first time that PDE3 was predicted to cause AEs in the muscles, bones, connective tissue, kidneys, urinary tract, and ear.

To ensure that other predictions were accurate, researchers fed the algorithm reports from 2014 through 2019, this and other predictions matched recent real-world reports. The team also compared the results to drug labels, scoured through scientific literature, and used other validation techniques.

“What seemed like false-positive predictions proved not to be false at all when the new reports became available,” said Robert Ietswaart, one of the 4 co-first authors of the paper.

The machine learning tool, however, is still limited, able to assess less than 1 percent of the 20,000 genes in the human genome. Consequently, the work should not be thought of as a complete understanding of every adverse drug event, Ietswaart said.

The model is open-source and posted for free online, allowing scientists to use and improve upon it.

References

  1. Dutchen, Stephanie. Predicting side effects. News Release. Harvard Medical School; June 18, 2020. Accessed June 22, 2020.https://hms.harvard.edu/news/predicting-side-effects.
  2. R. Ietswaart, Arat S, Chen AZ, et al. Machine learning guided association of adverse drug reactions with in vitro target-based pharmacology. EBioMedicine. 2020. doi: https://doi.org/10.1016/j.ebiom.2020.102837