How Ohio State University Researchers Use Machine Learning to Repurpose FDA-Approved Drugs
Machine learning is a powerful tool for its predictive capabilities in finding tiny needles of meaningful patterns in large haystacks of real-world data.
From this vantage point, AI has enormous potential to improve and advance biotechnology, pharmaceutical, and health care industries. In an AI milestone study published on Monday in Nature Machine Intelligence, researchers at Ohio State University show how deep learning can emulate clinical trials to identify drug candidates for repurposing, a solution that can help improve drug safety and accelerate novel treatments by clinicians.
The practice of repurposing drugs by clinicians is legal and not unusual. When medical doctors prescribe drugs that have been approved by the U.S. Food and Drug Administration (FDA) for purposes different from what is printed on the labels, the drugs are being used “off-label.” Just because a medication is FDA-approved for a specific type of condition does not exclude it from having possible medical benefits for other purposes.
For example, Metformin, a drug that is FDA-approved for treating type 2 diabetes, is also used off-label to treat polycystic ovary syndrome (PCOS), and other conditions.
Trazodone, an antidepressant with FDA-approval to treat depression, is also prescribed by doctors off-label to help treat patients with sleep issues.
In cancer treatment, many drugs with FDA-approval for a certain type of cancer are also used frequently off-label for combination chemotherapy according to the National Cancer Institute at the National Institutes of Health (NIH). An example of this is the drug combination of Taxotere (Docetaxel), Adriamycin (Doxorubicin Hydrochloride), and Cyclophosphamide to treat breast cancer.
Out of 1.74 billion ambulatory pediatric visits, 18.5 percent had one or more off-label drug orders according to research conducted by Dr. Daniel Horton and his team that was published in Pediatrics in 2019. Pediatric off-label drug use was commonly prescribed for adolescents (321.5 orders per 1000 visits), and roughly 83 percent of all neonatal visits during 2006-2015 per the same study. Examples of off-label use include using antidepressants for attention-deficit hyperactivity disorder (ADHD), and antihistamines for upper respiratory tract infections.
The Ohio State University research trio of Ping Zhang (senior author), Lai Wei, and Ruoqui Liu set out to use AI to find drugs that could be repurposed to decrease the risk of stroke and heart failure in coronary artery disease patients.
“We present an innovative study design for the estimation of the drug’s effect from longitudinal observational data,” wrote the researchers.
The team created a deep learning model for predicting treatment probability consisting of an embedding module, a recurrent neural network, and a prediction module. The input data from patient data includes the treatment, outcomes, and values for potential confounders.
Confounders are variables that are related to or associated with the exposure and outcome. As a hypothetical example, say a correlation is identified between music festivals and increases in skin rashes during any given year. Music festivals do not directly cause skin rashes. In this case, one possible confounding variable between the two may be outdoor heat, as music festivals tend to run outdoors when the temperature is high, and heat is a known cause for rashes. But there may be many others such as dust, age of festival attendee, and many other possible confounders. When working with real-world data, the number of confounders could number in the thousands. AI deep learning is well-suited to find patterns in the complexity of potentially thousands of confounders.
For this study, the research team used Python and PyTorch running on four NVIDIA GPUs. For their model, the researchers created a LSTM (long short-term memory), a type of recurrent neural network (RNN) that learns from both input data as well as from “memory” of previous experience. LSTM is often used for temporal data with time sequencing, as learning “memories” help preserve the learning quality.
From the embedding module, embedded confounding vectors are input to the recurrent neural network (the LSTM). Two LSTMs were used, one to model the diagnosis, and the other for prescription codes of patients.
The prediction module is a fully connected neural network that takes the aggregated features from the recurrent neural network, and outputs predictions on the probability of receiving a treatment.
Using data from over 1.178 million patients with coronary artery disease, the deep learning model evaluated the effect of 55 repurposing drug candidates on different disease outcomes. This resulted in six drug candidates (metoprolol, fenofibrate, hydrochlorothiazide, pravastatin, simvastatin, valsartan) that the deep learning model predicted will improve patient outcomes, but have not yet been indicated for treating coronary artery disease.
“This study can be extended in multiple directions in the future,” the researchers wrote. “For the study, we use the hypothesized confounders including demographics, co-morbidities, and co-prescribed drugs.”
With this demonstrated proof-of-concept, now researchers and clinicians have a powerful AI drug discovery tool to rapidly discover new treatments for diseases by repurposing existing medications in the future.
Copyright © 2021 Cami Rosso All rights reserved.
Credit: Google News