In a recent proof-of-concept study published in the journal npj Digital Medicine, researchers from the United States of America used generative artificial intelligence (GAI) in the form of ChatGPT-4 (short for Chat generative pre-trained transformer 4) to identify drug repurposing candidates against Alzheimer’s disease (AD).
They found that GAI technologies can be successfully leveraged to integrate online scientific knowledge, aiding in prioritizing drug repurposing candidates for treating diseases.
Study: Leveraging generative AI to prioritize drug repurposing candidates for Alzheimer’s disease with real-world clinical validation. Image Credit: SuPatMaN/Shutterstock.com
Background
AD, a highly prevalent and irreversible neurodegenerative disorder, poses significant healthcare challenges with limited treatment options.
Drug repurposing, exploring existing drugs for novel therapeutic applications, offers a faster, cost-effective alternative to new drug development for such health conditions.
This approach leverages established safety profiles, expediting clinical translation and improving patient accessibility. Success relies on efficiently identifying promising candidates from a diverse pool of drugs.
The conventional drug repurposing process involves an extensive literature review navigating diverse sources for evidence. This labor-intensive process necessitates interdisciplinary expertise and faces challenges in synthesizing meaningful hypotheses from the vast search space. Streamlining the process is crucial to enhance its efficiency.
GAI showcases notable comprehension and response capabilities, particularly in medical contexts like answering medical exam questions, clinical decision-making, and drug discovery.
Despite promising applications, its deployment in healthcare requires rigorous verification of its functional utility and reliability with real-world clinical data due to concerns about potential information fabrication.
Researchers in the present study examined the feasibility of using GAI to identify drug repurposing candidates and clinically validated the recommended drug options using real-world data.
About the study
The present study conducted ten independent queries using ChatGPT-4 to generate drug repurposing candidates for AD. The queries were structured with prompts instructing the tool to list the top 20 drugs based on potential efficacy, excluding those originally designed for AD.
The JavaScript Object Notation (JSON) format was specified for clarity, and a subsequent prompt was used to verify and correct the list for distinctness and proper ranking.
The queries aimed to encourage differentiation between drugs developed for AD and those for other diseases, focusing on potential effectiveness.
Clinical validation studies employed electronic health record (EHR) data from two datasets: Vanderbilt University Medical Center (VUMC, n >3,000,000), and the National Institutes of Health All of Us program (n = 235,000).
The data from VUMC was de-identified. For each candidate drug, a retrospective cohort study commenced at age 65, excluding those with prior AD diagnosis, non-Alzheimer's dementia, missing demographic data, or lacking EHR follow-up after age 65.
AD diagnosis was based on specific International Classification of Diseases (ICD) codes. Propensity score (PS) matching (2:1) was employed, considering sex, race, EHR length after 65 years, and drug-specific comorbidities to create comparable drug-exposed and unexposed cohorts.
The effects of multiple drug exposures on participants were not considered. Statistical analysis involved Cox proportional hazards regression models for survival analyses and a meta-analysis of hazard ratios (HR).
Results and discussion
As per the study, the top three drug recommendations from ChatGPT-4 were metformin (antidiabetic), losartan (antihypertensive), and minocycline (antibiotic). Analysis of EHR data suggested that these drugs showed a significantly lower risk of AD after ten years.
Although the findings were limited by small sample sizes, metformin demonstrated treatment effects in the positive direction (HR <1). Furthermore, simvastatin and pioglitazone exhibited potential beneficial effects but without statistical significance based on analysis of the VUMC and All of Us data.
In the meta-analysis, metformin showed a protective effect against Alzheimer's disease (HR = 0.67), while simvastatin and losartan also demonstrated significant protective effects.
However, conflicting directionality in effect estimates for losartan was observed between VUMC and All of Us datasets.
Insufficient case counts hindered the evaluation of bexarotene, nilotinib, minocycline, candesartan, rapamycin, and lithium. Further research is needed to confirm these findings.
ChatGPT-4 did not suggest any Food and Drug Administration (FDA)-approved drugs for AD. The findings show the tool’s effectiveness in drug repurposing based on its capacity to follow instructions and rapidly synthesize relevant information from the literature.
However, the findings are limited by the reliance on frequency for drug prioritization, potential EHR data issues, statistical power constraints, challenges in establishing primary indications for some drugs, covariate imbalances, inability to establish causation, and the need for ongoing monitoring of tool’s performance in drug repurposing.
Conclusion
In conclusion, the present study demonstrates the potential and efficiency of ChatGPT-4 as an AI-based tool for drug repurposing, efficiently generating a promising drug list for testing in EHRs, with AD as a case study.
The findings indicate that the tool can effectively retrieve and integrate information from diverse literature sources, providing a streamlined framework potentially applicable to various diseases to uncover novel therapeutic uses of existing drugs.
Rethinking the germ theory: A balanced take on microbes and disease outcomes