Breakthrough Machine Learning Algorithm Predicts Drug Adverse Events

Breakthrough Machine Learning Algorithm Predicts Drug Adverse Events

During a ‘Hackathon,’ researchers from Harvard Medical School along with the Novartis Institutes for BioMedical Research developed a compelling tool that predicts how drug candidates may trigger adverse reactions before they actually reach human clinical trials or access the market. By developing a new open-source machine learning-based algorithm, this Cambridge, MA-based team could transform how drugs are developed in the future.

Published recently in the Lancet journal EBioMedicine, heretofore not possible ways of developing safer therapies could transform drug development. 

The Challenge in Drug Development

According to the Department of Health and Human Services (HHS), adverse drug relations are unfortunately associated with 2 million U.S. hospitalizations per year. As reported in Merck Manuals, adverse reactions occur during up to 20 percent of all hospitalizations. These adverse drug reactions, known as well as “side effects,” can occur due to a number of reasons from incorrect dosages to interaction of multiple medicines to off-label use.

Not an Easy Problem to Solve

Despite great advances in science and technology the problems associated with adverse reactions aren’t easily solved. That’s because one drug can result to countless interactions with various proteins in the human body. Just because a drug is designed to target and interact with a certain protein—an intended target—doesn’t mean that the drug will actually do so. Hence, it can be very challenging to actually predict resultant side effects. Once an interaction occurs, researchers may not be able to easily diagnose the responsible protein targets.

The Study

The Harvard-led research team leveraged an existing data repository of known adverse drug reactions in addition to a database involving 184 proteins associated with known drug interactions. Specifically, the team leveraged first a U.S. Food and Drug Administration (FDA) data base of 600,000 physician reports of patient drug reactions as well as a proprietary database of Swiss-based Novartis involving 184 proteins and how they interact with up to 2,000 drugs.

The team designed and developed a computer algorithm to start to decipher meaning out of the various connections and relations. This machine learning algorithm was developed to actually learn from the data. Put to use in experiments, the algorithm was able to identify 221 associations connecting various individual proteins and specific drug reactions. In some cases, the researchers were aware of these connections but, in some other cases, the finding was in fact novel.

Thanks to this powerful algorithm, researchers can start to better pinpoint which proteins are likely to be associated with drug targets contributing to specific side effects. Moving forward, this breakthrough can help researchers actually forecast if a new therapy candidate, for example, has a high probability of triggering side effects either by itself or when combined with specific medicines. Hence, the machine learning software can be run prior to testing the drug in humans.

Not a “Silver Bullet” but Powerful Contributor to Drug Dev Acceleration

Robert Ietswaart, a research fellow in genetics working in the lab of Stirling Churchman at Harvard Medical School’s Blavatnik Institute, and a co-first author of the study, declared, “Machine learning is not a silver bullet for drug discovery, but I do believe it can accelerate many different aspects in the difficult and long process of developing new medicines.”

The Process: Hackathon

As reported by Stephanie Dutchen with Harvard Medical School News & Research, this project arose out of a “quantitative science hackathon” in 2018 organized by Novartis Institutes for BioMedical Research (NIBR). This was made possible by Mirjam Trame who at the time was with Novartis and Andy Stein, a Director at Novartis.

Apparently, it started when a Novartis (NIBR) scientific lead named Laszlo Urban introduced some fundamental challenges with analyzing safety risks of new drug candidates. Establishing a compelling Hackathon topic, the group of graduate students and postdocs tapped into and harnessed their data science and machine learning knowledge in a quest to help NIBR solve this fundamental challenge.

Urban, actually an executive director with the company, shared that most of the time the hackathons are for learning projects and not industry-specific problem solving. However, in this case, the study led to a novel application with a study published in a highly respected journal.

Open Source Availability

Data associated with the study, as well as the code, is available in GitHub, while the actual software code is available here

Lead Research/Investigators

Robert Ietswaarta, Postdoctoral Fellow, Harvard Medical School

Laszlo Urban, Executive Director, Novartis

For other researchers, see the source