We manually screened the 353 articles and excluded review articles (N = 72) and out-of-scope articles (N = 157). The literature review workflow is summarized in Figure 1. Articles outside the scope of the research question included studies investigating drug repurposing opportunities for diseases other than AD, studies focusing on designing new compounds or modifying existing compounds (for example, synthesizing and evaluating nitazoxanide-based derivatives for potential use in AD treatment (Li et al., 2020) or preparing and assessing a novel ibuprofen microemulsion (Wen et al., 2021)), studies suggesting possible gene targets for AD therapies without mention of existing drugs acting on those targets, and study protocols and other descriptions of planned work. We applied a broad definition of AD in identifying relevant studies, including not only those studies with the primary aim of identifying drug repurposing candidates for AD, but also studies with indirect relationships to AD. For instance, we included studies developing general drug repurposing tools and making predictions for AD (Wu et al., 2013; Croset et al., 2014; Issa et al., 2016; Liu et al., 2016; Jiang et al., 2019; Yu and Gao, 2019; Zeng et al., 2019; Cai et al., 2021; Chen et al., 2021; Tsuji et al., 2021; Muslu et al., 2022), as well as studies suggesting drug repurposing candidates with uses not limited to AD (for example, drugs with potential for treating both AD and depression (Fukuchi, 2020)).

After removing non-qualified articles, we included 124 relevant articles for this review. For each of the 124 studies, we collected information on the repurposing strategy, suggested drug repurposing candidates, and validation plan. The type of repurposing strategy (experimental, computational, or combination) was identified based on the preliminary evidence used to suggest drug candidates (e.g., in vitro screening and in vivo assays were considered experimental drug repurposing approaches; virtual screening, network models, machine learning models, gene signatures, and data mining were considered computational drug repurposing approaches). The seven drugs currently approved for the treatment of AD (donepezil, rivastigmine, memantine, galantamine, memantine/donepezil, aducanumab, and lecanemab) were excluded from the analysis.

For each study, we documented positive and negative findings. We considered positive findings to be drugs reported as potentially effective in AD (with or without validation of these effects). We interpreted negative findings as drugs found to increase AD risk or drugs with no impact on AD risk, depending on the study authors’ interpretation provided in the discussion of the study findings. We did not independently attribute negative findings based on the study validation (or lack thereof).

We also recorded the journal that each study was published in and the publication year. We collected the most recent impact factor of each journal from Journal Citation Reports (JCR); for journals without an impact factor in JCR, we used the most recent Scopus 2-year impact score, as JCR and Scopus impact factors have been found to be similar (Gray and Hodkinson, 2008). One journal did not have an impact factor available (Nature Aging).

The data extracted during literature review is available in Supplementary Table S2.

We defined validation as confirmation of a previously identified signal using an intentionally designed study with well-defined outcome measures. We categorized the validation as in vitro, in vivo, or other. In vivo validation included animal models, randomized controlled trials, and real-world data (defined as patient data collected outside of clinical trials—e.g., data from electronic health records [EHRs] and health insurance claims data (Park, 2021)). The “other” category of validation included computational experiments such as molecular docking simulations or queries of drug perturbation databases. We did not consider use of existing literature support to fall within our definition of validation, as this shows replication of results rather than true validation. We also did not consider gene set enrichment analysis and pathway analysis to serve as validation, requiring a more direct link between the suggested drugs and AD to be considered validation. For studies with validation, we assessed whether the validation was specific to AD, defined as validation studies using cellular or animal models of AD or evaluating clinical outcomes of AD. Validation not specific to AD included animal or cellular models unrelated to AD and molecular investigations of proposed drugs.

We used RxNorm (Nelson et al., 2011), a standardized nomenclature for clinical drugs maintained by the National Library of Medicine, to normalize the names of the drug repurposing candidates so that all drugs were referred to by their generic names. As drug repurposing involves identifying novel therapeutic applications for existing drugs, we used the Prescribable RxNorm API (Prescribable RxNorm API, 2022) to identify the currently prescribable drugs among the suggested drug candidates. We first mapped drug names to RxNorm Concept Unique Identifiers (RxCUIs). To account for differences in reporting of drug names across studies (for example, two studies suggesting different salt forms of the same drug), we mapped the drug RxCUIs contained in the RxNorm Prescribable subset to their ingredient RxCUIs. For example, we mapped the drug valproic acid (RxCUI = 11118) to its ingredient valproate (RxCUI = 40254). Similarly, we mapped the drug glatiramer acetate (RxCUI = 84375) to its ingredient glatiramer (RxCUI = 214582). While some drugs map to multiple ingredients in RxNorm, we only encountered drugs that had a one-to-one mapping between the original drug RxCUI and the ingredient RxCUI.

We then used the Anatomical Therapeutic Chemical (ATC) Classification System to group the drugs at five different levels (Anatomical Therapeutic Chemical ATC, 2022). The Level 1 ATC codes group drugs into fourteen groups based on anatomical system of action (e.g., C for cardiovascular system) and pharmacological properties (e.g., L for antineoplastic and immunomodulating agents); ATC Levels 2–5 further categorize drugs into therapeutic, pharmacological, and chemical subgroups. We extracted ATC codes for the ingredient RxCUIs using the RxNorm API. The drug-ATC code mapping is a one-to-many relationship, as a single drug with multiple therapeutic uses can map to multiple ATC codes.

The drug repurposing strategies utilized by the reviewed studies varied widely. Out of the 124 studies reviewed, 71 employed computational (in silico) repurposing strategies, 48 involved experimental repurposing approaches (in vitro and/or in vivo), and five used a combined approach (Figure 2). There was a marked increase in the number of studies using computational drug repurposing strategies between 2012 and 2022, with 26 computational studies published in 2021, compared to sixteen experimental studies and two combination studies.

Out of the 48 studies using experimental drug repurposing approaches, 22 used in vitro methods and 20 used in vivo methods, while the remaining six studies used a combination of the two. Out of the computational drug repurposing studies, network modeling was the most common approach (N = 21 studies). Other common computational approaches were machine learning (N = 15), genetic signatures (N = 14), structure-based analyses (N = 14), and non-NLP data mining (N = 13). Mendelian randomization was the most infrequently used computational approach (N = 4). Eight studies used a combination of different computational approaches.

We found that most of the reviewed studies (98/124) focused exclusively on drug repurposing for AD, with the exception of eleven studies which developed general tools for drug repurposing and fifteen studies which suggested drug repurposing candidates for other diseases in addition to AD. The studies that used computational repurposing approaches identified an average of 13 ± 18 drug candidates, compared to an average of 2 ± 2 drug candidates for experimental studies and 9 ± 9 drug candidates for combination studies.

61% (76/124) of the studies used additional data to validate the potential efficacy of their proposed AD repurposing candidates. Most of the experimental repurposing studies (45/48, or 94%) performed a validation, whereas only 44% (31/71) of the computational repurposing studies reported a validation.

The 76 studies with validation employed a wide range of validation methods, categorized as in vitro, in vivo, and other (Figure 7A). In vivo validation in the form of animal models was the most popular validation method, used by 31 (25%) of the reviewed studies. Figure 7B shows the validation distribution for the six studies that suggested clozapine as a potential AD repurposing candidate. Only two of these studies performed a validation, one using in vitro data and one with in vivo data.

Finally, we explored the relationship between journal impact and signal quality to determine whether drug repurposing studies published in journals with higher impact factors were more likely to perform validation. Considering all journals with an impact factor available, we computed the mean impact factor for studies with validation compared to studies without validation. We found that studies without validation (N = 48) were published in journals with an average impact factor of 6.33 ± 3.61, while studies with validation (N = 75) were published in journals with an average impact factor of 6.17 ± 3.61. Using the categorized validation, we found that the studies with in vivo validation had an impact factor of 7.24 ± 4.27, whereas studies with in vitro validation had an impact factor of 4.94 ± 2.04. Studies with AD-specific validation were published in journals with only a slightly higher impact factor than studies without AD-specific validation (average impact factor 6.99 ± 3.94 compared to 5.89 ± 3.40).

Determining high-priority candidates remains a challenging task in drug repurposing for all diseases and was clearly demonstrated in this review. In terms of ATC classification, drugs acting on the nervous system (ATC code N, 17% of the suggested candidates) may be a particularly favorable class of drugs for AD repurposing—AD is a disease of the brain, and effective treatments for AD will need to cross the blood-brain-barrier. On the other hand, antineoplastic and immunomodulatory drugs may be less favorable for treatment of AD (ATC code L, 16% of the suggested candidates), given that many anticancer therapies and immunosuppressive agents carry significant adverse effects that may cause more harm than benefit in a long prescription period. However, these are still very broad categories, as demonstrated in Figure 5–drugs acting on the nervous system include antipsychotics, antidepressants, antiepileptics, and opioids; antineoplastic and immunomodulatory drugs encompass protein kinase inhibitors, monoclonal antibodies, hormone antagonists, and immunosuppressants, among others.

While the wide array of repurposing candidates for AD offers multiple potential therapeutic options to explore, it also presents challenges for identifying the most promising drugs among the many possibilities. Over half of the drugs (65%) had only a single study suggesting their utility in AD. Among all 124 reviewed studies, clozapine was most frequently proposed as a repurposing candidate for AD (N = 6); other frequently suggested drugs were aripiprazole, risperidone, sunitinib, vandetanib, vorinostat, pioglitazone, tamoxifen, verapamil, and adenosine (N = 5). Even among this high frequency subset, there is substantial variety—three are antipsychotics acting on the nervous system (clozapine, aripiprazole, risperidone); four are antineoplastic and immunomodulating agents (two protein kinase inhibitors: sunitinib and vandetanib, one hormone antagonist: tamoxifen, and one other antineoplastic agent: vorinostat); two act on the cardiovascular system (one calcium channel blocker: verapamil and one other cardiac preparation: adenosine); and one acts on the alimentary tract and metabolism (the blood glucose lowering drug pioglitazone). However, when using validation as an indicator of high-quality drug repurposing signals, the list changes—the most frequently suggested drugs with AD-specific validation were doxycycline (an antibiotic), etodolac (a non-steroidal anti-inflammatory drug), fingolimod (an immunomodulating medication used in the treatment of multiple sclerosis), and granisetron (used as an antiemetic to treat nausea and vomiting) [N = 2]. Further, when focusing solely on the drugs with real-world data validation, yet another list of drugs emerges: fluticasone, mometasone, dexamethasone, pioglitazone, febuxostat, atenolol, and sildenafil. The only point of overlap between these three lists is the antidiabetic drug pioglitazone.

We found that validation was inconsistently performed among the reviewed studies, with only 61% (76/124) of the studies conducting additional investigations to support their preliminary drug repurposing candidates. Animal models and in vitro studies were most commonly used for validation, which may explain the particularly high validation rate among the experimental studies (94%). Interestingly, no combination studies performed a further validation, although this may have been an artifact resulting from how we defined this repurposing strategy (computational and experimental studies performed in tandem to suggest two different types of drugs, rather than performed sequentially to filter down a drug list).

The large number of AD repurposing studies presents possibilities for negative and contradictory results, which must be carefully considered. Only one of the reviewed studies reported evidence of increased AD risk associated with drug use, specifically for the PCSK9 inhibitors evolocumab and alirocumab (Williams et al., 2020). These results were not directly contradicted by any of the other reviewed studies, which did not suggest either evolocumab or alirocumab as potential AD repurposing candidates. Several studies reported evidence of drug candidates likely to have little to no effect on AD risk. These drugs included lipid-lowering agents (ezetimibe, mipomersen, and statins) (Williams et al., 2020), antihypertensives (Walker et al., 2020), dimethyl fumarate (approved for the treatment of multiple sclerosis) (Möhle et al., 2021), and drugs commonly used in treatment of rheumatoid arthritis (tofacitinib, tocilizumab, and TNF inhibitors) (Desai et al., 2022). Still, statins (atorvastatin, simvastatin, rosuvastatin, fluvastatin) and ezetimibe were suggested as viable AD repurposing candidates by other studies, suggesting that reported negative outcomes may not entirely discredit the candidacy of certain drugs (Cheung et al., 2014; Croset et al., 2014; Kondo et al., 2017; Bauzon et al., 2020; Wu et al., 2021).

Our study has several limitations. Importantly, we focused our literature search on PubMed, which is largely regarded as the primary database for biomedical literature. As a result, we may have missed some AD drug repurposing candidates discussed in journals from other fields, such as computer science (e.g., IEEE Journals, which are indexed by Google). In addition, we found that very few studies reported explicitly negative outcomes (i.e., a drug candidate found to have paradoxical effects or even no effect), suggesting potential selective reporting bias, which may have impacted our results.

Given the large number of AD drug repurposing studies and suggested repurposing candidates, consolidating the evidence to identify high-priority drugs for subsequent investigation in clinical trials has proven to be a major challenge. Validation represents one approach to prioritizing among the suggested drugs, as the drugs with high-quality validation data should intuitively be more promising repurposing candidates. However, as demonstrated in this study, validation is inconsistently performed—39% of the reviewed studies did not perform a validation, and the studies that did perform a validation used a variety of different approaches, roughly half of which had limited relevance to AD. Furthermore, the complexity of the mechanisms underlying the pathogenesis of AD and its clinical heterogeneity have made it incredibly difficult to develop robust models of AD for use in validation, potentially limiting the generalizability of the drug repurposing studies that sought to perform AD-specific validation. Finally, the clinical viability of the suggested drugs requires deeper exploration, as many studies neglected to consider possible safety concerns associated with drug use when providing repurposing candidates (e.g., immunosuppressants and antineoplastic agents).

Despite its challenges, drug repurposing for AD still holds much promise, particularly in expanding the scope of possible AD therapies beyond drugs targeting amyloid β. However, future drug repurposing studies will need to make a concerted effort to narrow down the list of candidates, which will require thorough validation. Real-world data, particularly from EHRs, represents an especially valuable tool for investigating long-term drug effects in real-time patient health outcomes but was rarely used for validation in the reviewed studies. Going forward, leveraging this big data from diverse datasets (e.g., the National Institutes of Health All of Us Research Program, UK Biobank, and local EHRs) will be critical to identify clinically meaningful drug repurposing candidates for AD. The incorporation of EHR data into drug repurposing pipelines may also transform the process of identifying drug repurposing candidates, transitioning from primarily hypothesis-driven studies to non-hypothesis-driven approaches relying on pattern identification in large EHRs.