Summary: A new machine learning framework helps researchers identify existing drugs that may be repurposed to treat other conditions by emulating clinical trials on large-scale patient data.
Source: Ohio State University
Researchers at Ohio State University have built a deep learning system that analyzes vast real-world medical data to propose drugs that might be repurposed to improve outcomes in diseases for which those drugs are not currently prescribed.
Drug repurposing — finding new uses for approved medications — can accelerate treatment options, lower safety risks, and reduce the time and cost needed to bring therapies into clinical practice. Although repurposing sometimes results from chance observations (for example, botulinum toxin’s evolution from treating crossed eyes to migraine relief and cosmetic wrinkle reduction), rigorous testing is still required to confirm benefits and safety when a drug is applied to a different disease.
To speed this process, the Ohio State team developed a flexible computational framework that integrates large-scale patient care datasets with advanced machine learning. The system evaluates thousands of patient-level variables to emulate randomized clinical trials using real-world data collected from electronic health records, insurance claims, and prescription histories.
In the published study, the framework was applied to patients with coronary artery disease to search for drugs that might reduce the risk of downstream events such as heart failure and stroke. However, the approach is adaptable and can be configured to investigate drug candidates for most other diseases, provided clearly defined clinical outcomes.
“This work shows how artificial intelligence can be used to ‘test’ a drug on patient data, accelerating hypothesis generation and helping prioritize candidates for clinical trials,” said Ping Zhang, assistant professor of computer science and engineering and biomedical informatics at Ohio State and senior author on the paper. “These models are tools to assist clinicians and researchers; they do not replace physician judgment.”
The research appears in Nature Machine Intelligence. The team emphasizes that machine learning is not a substitute for randomized controlled trials, but it can handle the many confounders present in real-world datasets — variables such as age, sex, race, disease severity and coexisting conditions that influence treatment responses. A deep learning architecture lets the model account for hundreds or thousands of these confounding factors simultaneously.
Using claims data from nearly 1.2 million patients with heart disease, the model encoded treatment histories, timing of prescriptions and diagnostic tests, and a wide range of patient characteristics. Drugs were represented by their active pharmaceutical ingredients. By applying causal inference methods, the framework emulated clinical trial conditions: it defined treated and control cohorts, followed patients for a two-year window, and estimated the effect of each drug on specified outcomes at the study endpoint.
Causal inference enabled comparison across multiple candidate treatments and helped the researchers rank which interventions appeared most promising for reducing the risk of heart failure and stroke among patients with coronary artery disease. The model identified nine drugs with a high likelihood of delivering benefit; three are already prescribed for related conditions, while six emerged as novel repurposing candidates within this analysis.
Among notable candidates the model flagged were metformin, a widely used diabetes medication, and escitalopram, an antidepressant often prescribed for anxiety and depression. Both drugs are already under investigation in other studies for potential cardiovascular benefits, which supports the framework’s ability to surface clinically plausible repurposing opportunities.
Zhang emphasized that the significance of this work lies primarily in the method: a reproducible, customizable pipeline that leverages longitudinal real-world data, deep learning, and causal inference to generate ranked lists of repurposing candidates and estimated treatment effects. The approach can be tailored to different diseases by redefining the clinical outcome of interest and importing the appropriate patient records.
Funding: The study was supported by the National Center for Advancing Translational Sciences through the Center for Clinical and Translational Science at Ohio State University.
Graduate student Ruoqi Liu and research assistant professor Lai Wei, both at Ohio State, contributed to the study alongside Ping Zhang.
About this artificial intelligence research news
Source: Ohio State University
Contact: Ping Zhang – Ohio State University
Image: The image is in the public domain
Original Research: Closed access. “A deep learning framework for drug repurposing via emulating clinical trials on real-world patient data” by Ruoqi Liu, Lai Wei & Ping Zhang. Nature Machine Intelligence
Abstract
A deep learning framework for drug repurposing via emulating clinical trials on real-world patient data
Drug repurposing is an efficient strategy to discover new uses for approved medications, enabling faster translation from research to clinical care. Large-scale real-world datasets, such as electronic health records and insurance claims, capture longitudinal information on many drug users. We present a practical, customizable framework that combines deep learning with causal inference to emulate randomized clinical trials using retrospective observational data. Applied to a coronary artery disease cohort drawn from millions of patients, the framework identifies drugs and combinations that appear to improve outcomes and that are not currently indicated for treating coronary artery disease, thereby supporting prioritization of candidates for further clinical evaluation.