Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that raises major concerns in healthcare due to its irreversibility and high prevalence among older adults1. Despite decades of research, treatment options for AD remain limited, leaving patients and families with little hope. Drug repurposing to identify novel therapeutic applications for existing drugs is an attractive additional approach to discovering treatment options compared to the costly and time-consuming process of new drug development alone, particularly for serious, widespread conditions that continue to have few effective treatments, such as AD2. In addition to accelerated timelines and lower costs throughout the discovery-to-market process, the approach offers well-established drug safety profiles and expedited clinical translation with enhanced patient accessibility. Nevertheless, the success of drug repurposing hinges on the prompt and accurate identification of promising candidates among a large collection of drugs.

The search for drug repurposing candidates typically relies on a comprehensive review of the scientific literature, focusing on studies that offer evidence of efficacy for certain drugs or their constituent ingredients. Mechanistic insights, preclinical experiments, clinical reports, large-scale observational studies, and drug repurposing databases collectively form the space within which searches are conducted. However, this review process is labor- and time-intensive, requiring researchers to incorporate interdisciplinary expertise in disease mechanisms, molecular biology, pharmacology, clinical research, and bioinformatics. For example, a search with the keywords “(Alzheimer’s disease OR neurodegenerative disorder) AND (treatment OR drug OR repurposing)” yields around 233,000 published papers on PubMed and over a million on Google Scholar. It is highly unlikely for domain experts to comprehensively read and synthesize meaningful hypotheses from such a massive space. As such, approaches that streamline this process offer an advantage in repurposing efforts.

Recent advancements in generative artificial intelligence (GAI), exemplified by OpenAI’s ChatGPT3, have showcased the remarkable capability of AI to understand and respond to diverse inquiries. The comprehension and response capabilities of GAI derive from extensive exposure to a vast corpus from the Internet, nuanced encoding of knowledge, and subsequent optimization of responses that display reasoning processes4,5. Beyond answering general questions, GAI has demonstrated effectiveness in specialized medical contexts6, including answering U.S. Medical Licensing Examination questions7, clinical decision-making consultations8, and drug discovery9. For example, ChatGPT has been recognized for producing informed consent documents for medical procedures that are more patient-readable, complete, and accurate than those written by surgeons10. Moreover, ChatGPT has shown promise as a clinical decision-support tool for offering patient management suggestions11. Additionally, ChatGPT is already being leveraged by biotechnology companies to suggest novel pathways for drug targets12.

However, given its nascent stage and concerns regarding fabrication of information7,13, responsible deployment of this tool in the medical setting necessitates comprehensive verification of its functional utility and reliability with clinical data in the real world.

We hypothesized that ChatGPT can function as an AI-driven screening tool to generate drug repurposing candidates for AD. To assess this hypothesis, we provided ChatGPT (model GPT-4) with two sequential prompts. First, we prompted ChatGPT to provide the twenty most promising drug repurposing candidates for AD. Next, we prompted ChatGPT to confirm its previous output and return a final list of drugs (Fig. 1a). To account for the probabilistic nature of ChatGPT’s responses, we repeated this process ten times, resulting in a total of 59 unique drug candidates (Supplementary Table 1). We confirmed that each candidate appeared in at least one publication discussing their potential use in AD. We then identified the ten most frequently appearing drugs for subsequent testing with clinical data (minimum frequency N = 7, maximum frequency N = 10).

Fig. 1: An illustration of the study design.
figure 1

a Employing iterative queries of ChatGPT to recommend twenty drugs for AD repurposing. b Evaluating the potential efficacy of the ten most frequently suggested drugs using electronic health records (EHR) data from two large clinical databases.

Full size image

For each generated candidate, we composed two cohorts using de-identified electronic health record (EHR) data from large clinical datasets: (1) Vanderbilt University Medical Center (VUMC), and (2) the National Institutes of Health (NIH) All of Us Research Program14 (Fig. 1b). We employed Cox proportional hazards regression to compare the risk of developing AD between individuals with prior drug exposure and individuals never exposed to the drug. We used age 65 as time zero; prior drug exposure was defined by medication use ≤65 years of age. Each drug-exposed cohort was matched to an unexposed group based on propensity score (PS), using sex, race, EHR length after age 65, and drug-specific comorbidities at age 65 (i.e., at the time of cohort entry) as covariates. Drug-specific comorbidities were selected based on primary clinical indication. Given that the cohort size for a particular drug might not be sufficiently large in the independent datasets, we also performed a meta-analysis to derive a statistically robust estimate of each drug’s hazard ratio.

We observed that three of the top ten ChatGPT recommendations were associated with a significantly reduced risk of AD after ten years of follow-up using VUMC data: the antidiabetic medication metformin (hazard ratio (HR) = 0.67, 95% confidence interval (CI): 0.54–0.82, p < 1.5 × 10−4), the antihypertensive agent losartan (HR = 0.73, 95% CI: 0.57–0.92, p = 0.009), and the antibiotic minocycline (HR = 0.34, 95% CI: 0.13–0.89, p = 0.028) (Fig. 2). Though our studies with All of Us were limited by smaller sample sizes, metformin showed treatment effects in the expected direction (i.e., HR < 1). While not statistically significant at p < 0.05, the lipid-lowering medication simvastatin and the antidiabetic medication pioglitazone also exhibited beneficial treatment effects in both the VUMC and All of Us data.

Fig. 2: Associations between exposure to ChatGPT-suggested drug repurposing candidates and AD risk.
figure 2

Hazard ratios (HR) and 95% confidence intervals (CI) are shown for VUMC (blue squares), the NIH All of Us Research Program (red squares), and the combined meta-analysis (gray squares). ** indicates drugs associated with significantly reduced AD risk using VUMC data (p < 0.05); * indicates drugs associated with significantly reduced AD risk in the meta-analysis (p < 0.05). To ensure adequate statistical power, we did not report drugs with fewer than five AD cases in the study cohort (i.e., bexarotene and nilotinib in both VUMC and All of Us; minocycline, candesartan, rapamycin, and lithium in All of Us).

Full size image

In the meta-analysis, we confirmed the protective effect of metformin (HR = 0.67, 95% CI: 0.55–0.81, p = 6.4 × 10−5). The meta-analysis also revealed a statistically significant protective treatment effect for simvastatin (HR = 0.84, 95% CI: 0.73–0.98, p = 0.024) that had not been identified in either the VUMC or All of Us data in isolation. Losartan was found to have a significant protective treatment effect in meta-analysis as well (HR = 0.76, 95% CI: 0.60–0.95, p = 0.017); however, the effect estimates from VUMC and All of Us were opposing in their directionality.

Inadequate AD case counts (N < 5) prevented the evaluation of bexarotene and nilotinib in both VUMC and All of Us. The effects of minocycline, candesartan, rapamycin, and lithium could not be tested in All of Us for the same reason.

We found that ChatGPT’s utility as a drug repurposing tool resides in its ability to follow instructions pertaining to drug repurposing and rapidly synthesize information from relevant literature. ChatGPT did not propose any FDA-approved drugs for AD, suggesting that it accurately interprets the premise of drug repurposing. In this study, the drugs suggested with the highest frequency by ChatGPT were not novel repurposing candidates for AD, but rather drugs frequently mentioned together with AD in the literature. Antidiabetic drugs such as metformin and pioglitazone have received considerable attention as potential therapeutic candidates for AD, driven by increasing evidence implicating insulin resistance in the pathogenesis of AD15,16,17. Similarly, reported associations between AD and cardiovascular disease have sparked numerous investigations into the repurposing of cardiovascular drugs for AD, including statins and antihypertensive agents such as losartan and candesartan18,19,20. Rapamycin, nilotinib, lithium, and bexarotene have also been heavily explored in AD drug repurposing studies21,22,23.

We observed protective effects against AD for three of the ten drugs most frequently suggested by ChatGPT–metformin, simvastatin, and losartan–in meta-analysis combining data from two large-scale EHRs. Use of metformin, which produced the strongest signal in our meta-analysis, was associated with a 33% decreased risk of incident AD after age 65. Simvastatin and losartan produced more modest effects. In meta-analysis, simvastatin was associated with a 16% decreased risk of AD, while losartan was associated with a 24% decreased risk of AD. Whereas metformin and simvastatin were found to have consistent treatment effects (HR < 1) in both VUMC and All of Us, losartan had conflicting treatment effects (statistically significant HR < 1 using VUMC data, non-significant HR > 1 using All of Us data). This suggests that losartan’s protective treatment effect in meta-analysis may have been driven by the larger sample size from VUMC. Despite supporting findings for these three drugs in previous studies, much remains unknown about the mechanisms by which these drugs affect AD pathophysiology and pathology, and population-based studies have not provided conclusive results24,25,26. Further investigation in preclinical and clinical studies will be needed to ascertain the viability of these drugs in decreasing risk of AD.

Our findings suggest that ChatGPT can generate quality hypotheses for drug repurposing. ChatGPT expedites the process of extensive literature review, which has become infeasible for humans to perform alone. Since 2012, various strategies have led to repurposing suggestions for over 573 drugs for AD27. Variability in the types of data used to suggest these repurposing candidates, particularly with regards to data quality and relevance to AD, has made it difficult to quickly and accurately pinpoint the most promising candidates among this long list of drugs, slowing progress in drug repurposing. Using ChatGPT, we were able to quickly generate a short list of drugs for testing using real-world clinical data, with three out of ten drugs demonstrating protective effects against AD according to our meta-analysis. These findings suggest that ChatGPT has the potential to effectively prioritize among a large drug space, demonstrating advantages over traditional methods for identifying promising drug repurposing hypotheses. With minimal costs, ChatGPT has the capacity and scalability to substantially accelerate the review process, allowing researchers to focus on testing and validating the hypotheses. Moreover, the anticipated regular updates of ChatGPT (which provide access to new Internet content) and its search engine plugins allow for consistently up-to-date and uninterrupted drug repurposing research. Furthermore, combining ChatGPT-powered hypotheses with robust verification using real-world clinical datasets provides a cost-effective pipeline to investigate preliminary signals before allocating additional resources to extensive research and clinical trials. This validation process serves as a critical balancing force to disprove invalid hypotheses, thereby enhancing the trustworthiness of the results and assuaging concerns about adverse consequences of AI hallucinations–a major criticism of ChatGPT use. Despite these advantages, any pipelines incorporating ChatGPT must account for the possibility of overlooked, but promising, repurposing candidates, which can transpire when candidates exhibit low occurrence in the literature or necessitate complex reasoning ability based on indirect evidence that surpasses ChatGPT’s capabilities.

Our study has several limitations of note. First, we relied upon frequency to prioritize drug candidates; however, the number of times a repurposing candidate appears in ChatGPT queries may not be directly related to its promise in treating disease. Second, EHRs can contain missing or incomplete data28, and discontinuities in medication adherence may not be reported with perfect fidelity, creating possibilities for misclassification of outcome or exposure. Third, despite the use of two large EHRs, we still did not have adequate statistical power for hypothesis testing of less common drugs (e.g., nilotinib). Fourth, while our study evaluated drug exposure broadly as any-time, any-dose exposure ≤65 years of age, there exist many opportunities for deeper phenotyping in characterizing drug exposure. Fifth, we sought to control for a single primary indication for each drug using MEDI; however, we were unable to establish a clear primary indication for several drugs (i.e., nilotinib, bexarotene, minocycline, and rapamycin). Furthermore, a fully balanced covariate distribution was not achieved for metformin and simvastatin (standardized mean difference >0.1 for EHR length after 65 and drug-specific comorbidities), suggesting there may be some residual confounding (although likely to bias towards the null). Like all EHR-based observational studies, our study is susceptible to unmeasured confounding, which may bias the detected associations between drug exposure and AD. While we sought to control for confounding by indication by accounting for each drug’s primary clinical indication when PS matching, this was not an exhaustive list of all possible drug indications and drug-associated comorbidities. We were also not able to control for covariates that are not consistently reported in the EHR that may impact AD risk, such as social determinants of health. Sixth, this study cannot establish causal effects or mechanisms as might be the case in a clinical trial. Additionally, although ChatGPT exhibits exceptional response quality for general queries, further research is required to benchmark a range of GAI models and their fine-tuned variants for greatest effectiveness and reliability in supporting biomedical tasks, particularly drug repurposing. Lastly, the quality of LLM-generated responses for drug repurposing may vary over time29, influenced by the continuous updates in data, learning techniques, and evaluating metrics in model training. Continuous actions in monitoring the performance of LLMs, paired with validations using more large-scale datasets, are essential to ensure the sustained effectiveness of this approach for drug repurposing.

ChatGPT’s effectiveness for drug repurposing suggestions can be attributed to its ability to encode the vast literature on the topic into its parameters and subsequently decode relevant information to produce responses. As research in this area evolves, it will be important for future investigations to focus on how best to leverage LLMs to discover novel insights in conjunction with explicit knowledge representation and reasoning technologies, such as knowledge graphs30. Specifically, biomedical knowledge graphs that are representative of validated relationships among entities like genes, proteins, and existing drugs can be provided to LLMs to guide biomedical hypothesis generation. On the other hand, the traditional reasoning process over biomedical knowledge graphs may also benefit from using LLMs to derive more informative latent representations for the aforementioned entities. The full potential and scalability of these two technical paths warrant further investigation.

Still, this proof-of-concept study showcases the feasibility of employing ChatGPT as an AI-driven hypothesis generator for drug repurposing, enabling the prompt generation of a promising list of drugs for subsequent testing in EHRs, using AD as a case study. Our findings suggest that ChatGPT is able to encode valuable insights concerning novel potential therapeutic utilities for existing drugs by comprehensively synthesizing literature, and can subsequently decode this knowledge when responding to queries. Pipelines that leverage the capabilities of ChatGPT offer a streamlined new framework for drug repurposing that can be applied to numerous diseases.

Methods

Interactions with ChatGPT

In this study, we interacted with OpenAI’s ChatGPT (GPT-4) to generate promising drug repurposing candidates for AD. These interactions were conducted in May 2023, at which time the technology had access to information accumulated until September 2021. Ten independent queries were performed, ensuring that each query did not serve as the context for another. Each query consisted of two prompts. The first prompt described the instructions for generating drug repurposing candidates, whereas the second prompt asked ChatGPT to self-correct its output from the previous prompt.

  • Prompt 1: Please provide a list of the 20 most promising drugs for repurposing in the treatment of Alzheimer’s disease based on their potential efficacy, and indicate the diseases they were originally developed to treat. Please rank them in descending order of potential effectiveness and use the JSON format to include the “Drug” and “Disease” keys.

  • Prompt 2: Please check if the generated list meets the requirement: (1) exclude the drugs that were originally designed for Alzheimer’s disease, (2) 20 distinct drugs, (3) in JSON format, (4) rank in descending order according to the potential effectiveness. If not, please regenerate the list that meets the requirement.

In our queries, we intentionally emphasized drugs’ original purposes to encourage ChatGPT to distinguish between the drugs originally intended to treat AD and those used in treating other diseases. This helped to limit the possibility that candidates with original use in AD were returned in the final list of each query. We also imposed a specific format for the drugs returned in the queries to facilitate subsequent processing. We also asked ChatGPT to rank drugs according to their potential effectiveness. While ChatGPT claimed that it “cannot rank the generated drugs with respect to their potential effectiveness since the data is not definitive and is constantly evolving” in multiple responses, we sought to emphasize the notion of effectiveness during the drug generation process. It is important to note that we did not use the order of drugs in the generated lists for subsequent drug selection. It is notable that in the situation where ChatGPT occasionally returned a list of drug items with placeholder names like “Drug A, Drug B, …”, indicative of a failure to extract relevant knowledge, we terminated the particular conversation instance and initiated a new one. We provided screenshots of an example of ChatGPT’s response in Supplementary Figs. 1 and 2.

Data source

We performed our clinical validation studies using de-identified EHR data from (1) Vanderbilt University Medical Center (VUMC), a major academic medical center in Nashville, Tennessee, and (2) the All of Us Research Program run by the National Institutes of Health (NIH), a U.S. nation-wide clinical database. VUMC’s de-identified EHR database contains longitudinal clinical data including diagnosis codes, lab values, and medications for over three million patient records31. The NIH All of Us Research Program database contained de-identified EHR data for over 235,000 participants at the time of this study14. The EHR data in both resources is standardized according to the Observational Medical Outcomes Partnership (OMOP) Common Data Model32, allowing for reproducible cohort formation and characterization of drug exposures in the two databases.

Usage of All of Us data was approved by the NIH All of Us Research Program. All EHR data from VUMC was de-identified, such that this study was deemed to be exempt by the Institutional Review Board.

Study cohort

For each candidate drug, we conducted a retrospective cohort study using age 65 as time 0. Each study was limited to individuals aged 65 or older with no prior diagnosis of AD. We excluded individuals with a diagnosis of non-Alzheimer’s dementia (vascular dementia, diffuse Lewy body disease, frontotemporal dementia, mixed dementia, and dementia associated with Parkinson’s disease), individuals without EHR follow-up after age 65, and individuals with missing demographic characteristics.

We defined a confirmed diagnosis of AD as patients with at least one AD diagnosis code in their EHR using ICD-9-CM code 331.0 and ICD-10-CM codes G30.1, G30.8, and G30.9. We have previously shown that using ICD codes to phenotype AD patients has a high PPV (94%) in VUMC’s de-identified EHR database33.

To capture all relevant drug exposures when creating the drug-exposed group, medications were mapped to their ingredients using RxNorm34. Individuals with at least one recorded exposure to the drug of interest occurring at ≤65 years of age were considered to be exposed. Individuals whose first record of drug exposure occurred after age 65 were excluded from the analysis.

We gathered demographic characteristics (sex and race), remaining chart length, and comorbidities at age 65 to generate a propensity score (PS) for matching. The comorbidities were selected to mitigate potential confounding by indication. We used MEDI35, an ensemble medication indication resource, to identify the primary clinical indication for each drug repurposing candidate. MEDI contains over 63,000 medication-indication pairs with indication prevalence evaluated using EHR data. We queried MEDI for the highest prevalence indications for each drug and used these to define a single primary indication for the drug. If there was no consensus among the top indications, a primary indication for the drug was not defined. MEDI reports medication indications using only ICD-9-CM; as such, we mapped the ICD-9-CM code(s) comprising the primary indications to ICD-10-CM codes using the General Equivalence Mappings developed by the Centers for Medicare & Medicaid Services. Supplementary Table 2 reports the set of ICD-9-CM and ICD-10-CM codes relied upon to define the comorbidities and the drugs they pertain to. A confirmed comorbidity status was defined as disease diagnosed at the start of follow-up (i.e., at ≤65 years of age).

We applied 2:1 PS matching (nearest-neighbor algorithm, caliper = 0.1) with sex, race, length of EHR after age 65, and relevant drug-specific comorbidities as covariates to form comparable drug-exposed and unexposed cohorts for each suggested drug repurposing candidate. PS matching was performed using the MatchIt R package36. The participant counts for each drug after matching (AD/exposed, no AD/exposed, AD/unexposed, and no AD/unexposed) are provided in Supplementary Table 3. The covariate balance between the drug-exposed and unexposed groups after matching is provided in Supplementary Table 4.

Based on our study design, an individual with a history of exposure to multiple drug repurposing candidates could be included in more than one drug-exposed cohort. We did not consider potential compound effects resulting from multiple drug exposures.

Statistical analysis

All survival analyses were performed using Cox proportional hazards regression models. Each model compared the risk of AD in individuals exposed to a drug repurposing candidate and PS-matched individuals never exposed to the drug. Follow-up ended at the first of (1) AD diagnosis, (2) last recorded EHR observation, or (3) ten years. We censored observations after ten years of EHR follow-up (i.e., at age 75) to minimize differential loss to follow-up. To ensure adequate statistical power, we did not report drugs with fewer than five AD cases included in the final study cohort. We used p < 0.05 as our significance threshold given the small number of tests (N = 10).

Meta-analysis of hazard ratios was performed using NCSS statistical software37. Cochran’s Q test was used to assess heterogeneity. Meta-analysis was performed under a fixed-effects model.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.