CauReL: Dynamic Counterfactual Learning for Precision Drug Repurposing in Alzheimer's Disease

CauReL: Dynamic Counterfactual Learning for Precision Drug Repurposing in Alzheimer's Disease | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article CauReL: Dynamic Counterfactual Learning for Precision Drug Repurposing in Alzheimer's Disease Yanfei Wang, Minghao Zhou, Zijia Tang, Chenxi Xiong, Breton Asken, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8206648/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Alzheimer’s disease has few effective therapies, and decades of amyloid- and tau-focused trials have delivered only modest benefit with substantial toxicity. Drug repurposing using real-world data offers a faster and lower-risk route to new treatments, yet current approaches typically average effects across populations, model disease onset and progression separately, and provide little insight into which patients are most likely to benefit. We present CauReL, a dynamic counterfactual representation learning framework that enables transparent, patient specific estimation of treatment effects from large-scale electronic health records for precision drug repurposing in AD. CauReL first learns balanced latent representations of treated and untreated patients using Integral Probability Metric regularization, then jointly predicts two clinically linked outcomes, incident AD and time from mild cognitive impairment (MCI) to AD, to generate paired counterfactual outcomes for every individual. A counterfactual explanation module quantifies how clinical features shape benefit at the patient level, and uplift trees transform complex heterogeneity into simple, rule-based subgroups suitable for trial enrichment and clinical decision support. Using independent cohorts from OneFlorida + and All of Us, we screened outpatient prescriptions with at least 20 percent exposure among 28,605 individuals with mild cognitive impairment, of whom 4,990 progressed to Alzheimer’s disease. CauReL substantially improved covariate balance and distributional overlap across drug cohorts and achieved strong predictive accuracy for both incidence (AUC greater than 0.90) and progression timing (C index 0.81 to 0.84; Spearman 0.80 to 0.86). Twenty drugs showed consistent protective associations, with four emerging as highly reproducible across both networks, the metabolic agents liraglutide and empagliflozin and the neuroactive agents entacapone and amantadine. These drugs were associated with meaningful absolute risk reductions and clinically significant delays in progression from mild cognitive impairment to Alzheimer’s disease. Metabolic drugs produced the strongest benefits in individuals with diabetes, obesity, or cardiovascular disease, whereas neuroactive drugs provided broadly consistent protection across most subgroups. CauReL is available as an open source Python package with a companion web server for direct application to new cohorts or disease settings ( https://caurel.site/ ). This work delivers a scalable and interpretable framework for prioritizing repurposable drugs and designing targeted clinical trials for the patients most likely to benefit. Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Medical research Health sciences/Neurology Biological sciences/Neuroscience Causal AI Counterfactual representation learning Individualized Treatment Effects (ITE) Drug Repurposing Electronic Health Records (EHRs) Alzheimer’s Disease (AD) Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Full Text Additional Declarations No competing interests reported. Supplementary Files SFig.1.jpg Supplementary Figure 1. Feature importance in All of Us (a)visualizes subgroup-level heterogeneity stratified by race and sex. The x-axis represents the probability that individuals within a subgroup experience a beneficial treatment effect, while the y-axis represents the probability of high variability or inconsistent effects within that subgroup. Subgroups in the lower-right quadrant exhibit both high benefit and low variability, indicating the most reliable and clinically promising populations. (b)shows ATEs stratified by comorbidity status, including diabetes, hypertension, heart failure, obesity, depression, and stroke. The x-axis lists comorbidities, while the y-axis shows the ATE, with darker bars indicating patients with the comorbidity and lighter bars indicating patients without it. Negative values represent protective associations. SFig.2.jpg Supplementary Figure 2. Subgroup analyses in All of Us (a)depicts the mean ATEs by age group, where the x-axis represents age categories (50-59, 60-69, 70-79, ≥80), and the y-axis represents the mean ATE. Error bars indicate variability within each age group. (b)shows directional feature importance, where the x-axis lists baseline features and the y-axis shows their directional impact on predicted AD risk. Features with bars pointing to the left are associated with enhanced drug protection, while those pointing to the right are linked to attenuated drug effects. SFig.3.jpg Supplementary Figure 3. Time of AD conversion in All of Us (a)presents Kaplan-Meier survival curves comparing treated patients and matched controls for time to conversion from MCI to AD. The x-axis represents months since MCI diagnosis, and the y-axis represents the probability of remaining free from AD. Red curves represent treated patients, and blue curves represent matched controls. Clear separation between curves indicates slower progression among treated patients, with all log-rank tests yielding p < 0.05. (b)shows the counterfactual model predictions for each drug. Violin plots depict the predicted distribution of conversion times under two scenarios: no treatment (blue, left) and treatment (red, right). Wider sections indicate higher densities of patients at a given predicted conversion time, and vertical black bars mark the median prediction. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8206648","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":559065025,"identity":"935449ef-7385-42c5-af24-f14622e114a5","order_by":0,"name":"Yanfei Wang","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Yanfei","middleName":"","lastName":"Wang","suffix":""},{"id":559065026,"identity":"97dcadb2-dfd0-44fa-9960-b4d76bdde7d3","order_by":1,"name":"Minghao Zhou","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Minghao","middleName":"","lastName":"Zhou","suffix":""},{"id":559065028,"identity":"356d54f4-09eb-43b6-bc8c-18dd2f620b92","order_by":2,"name":"Zijia Tang","email":"","orcid":"","institution":"Duke University","correspondingAuthor":false,"prefix":"","firstName":"Zijia","middleName":"","lastName":"Tang","suffix":""},{"id":559065030,"identity":"4e6408df-9050-4523-a8ee-252714ffffee","order_by":3,"name":"Chenxi Xiong","email":"","orcid":"","institution":"Purdue University West Lafayette","correspondingAuthor":false,"prefix":"","firstName":"Chenxi","middleName":"","lastName":"Xiong","suffix":""},{"id":559065032,"identity":"6f54ff7e-9814-4104-bd3e-503fc8c1f140","order_by":4,"name":"Breton Asken","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Breton","middleName":"","lastName":"Asken","suffix":""},{"id":559065033,"identity":"0d3ae80b-f3e4-4663-af81-6ea7e4ae9fec","order_by":5,"name":"Baijian Yang","email":"","orcid":"","institution":"Purdue University West Lafayette","correspondingAuthor":false,"prefix":"","firstName":"Baijian","middleName":"","lastName":"Yang","suffix":""},{"id":559065035,"identity":"5e90029c-fabd-4280-8c62-48e2dfd46269","order_by":6,"name":"Jing Su","email":"","orcid":"","institution":"Indiana University – Purdue University Indianapolis","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Su","suffix":""},{"id":559065039,"identity":"7e7798d3-105f-459d-8117-dab64cb68363","order_by":7,"name":"Xiaobo Zhou","email":"","orcid":"","institution":"The University of Texas Health Science Center at Houston","correspondingAuthor":false,"prefix":"","firstName":"Xiaobo","middleName":"","lastName":"Zhou","suffix":""},{"id":559065040,"identity":"6cee694a-4f44-42ae-8221-62e716da9e33","order_by":8,"name":"Qianqian Song","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAsUlEQVRIiWNgGAWjYNACAxsG9nYgzUO8loI0Bp7DpGn5cJgELXznDx97+MXgfGIPMwPjg7dtRGiRvJGWbixjcBukhdlwLjFaDG7wmElLALXsZ2Zgk+YlSsv5MyAt50C2sP8mTsuBHDPJDwYHQFrYmInSAvRLmjSDQbJxDzNjs+Scc0RoAYWY5I8/drI97M0HP7wpI0ILwwEGBmZIdDA2EKMeooXxB5FqR8EoGAWjYIQCADmtM4ReoMVBAAAAAElFTkSuQmCC","orcid":"","institution":"University of Florida","correspondingAuthor":true,"prefix":"","firstName":"Qianqian","middleName":"","lastName":"Song","suffix":""}],"badges":[],"createdAt":"2025-11-25 20:53:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8206648/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8206648/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98431998,"identity":"ecedbe53-9c4a-44e6-aae4-bd8280c40b2d","added_by":"auto","created_at":"2025-12-17 16:48:46","extension":"jpg","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3994762,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/39280ccb6d8fb7053217943a.jpg"},{"id":98193826,"identity":"8de363eb-19d1-42f2-897a-33a76cad8a6e","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":142125,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/d3e9f0d6e89178dd2dad2cca.docx"},{"id":98431700,"identity":"bb455c70-2d5d-4311-9fad-5cf670fceb20","added_by":"auto","created_at":"2025-12-17 16:48:11","extension":"jpg","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2505539,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/c48081ad37a124d440e2697a.jpg"},{"id":98193831,"identity":"cc8962ea-e1ff-43ea-bfda-c0a8008433ce","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1694954,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/e78a9dbf251cfa6cd1ff5435.jpg"},{"id":98193832,"identity":"4dec51f8-ba30-48b7-afae-b80e73d64417","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2947562,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/a7f9759a47453cf0aca2b81a.jpg"},{"id":98430821,"identity":"168a5d59-b029-4481-a4cd-240f4b7ab30c","added_by":"auto","created_at":"2025-12-17 16:46:18","extension":"jpg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3168700,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/d7866b31cc4e3f0972279d6a.jpg"},{"id":98430750,"identity":"c3467046-8217-405b-aabe-c9a3d13f24e3","added_by":"auto","created_at":"2025-12-17 16:46:09","extension":"jpg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3159930,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/260f97baf72d389bab4f1645.jpg"},{"id":98193848,"identity":"998b1ec5-460f-42d3-b2a4-c31855bfe1d6","added_by":"auto","created_at":"2025-12-15 06:22:21","extension":"json","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11163,"visible":true,"origin":"","legend":"","description":"","filename":"41bf1fed71cd45d28fab0b94ae36a770.json","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/6b56a6cf1a4bee9665df2ff6.json"},{"id":98432460,"identity":"704ef025-7597-4bff-aa4b-5ee539c6f2b8","added_by":"auto","created_at":"2025-12-17 16:49:35","extension":"jpg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1413742,"visible":true,"origin":"","legend":"","description":"","filename":"SFig.1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/c1be5592a5e61b9543bc2e41.jpg"},{"id":98431483,"identity":"bddd43ea-8e69-46ab-aeb0-fba87505a282","added_by":"auto","created_at":"2025-12-17 16:47:46","extension":"jpg","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1558981,"visible":true,"origin":"","legend":"","description":"","filename":"SFig.2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/e53695ab06067d71de2f9212.jpg"},{"id":98193849,"identity":"9ed0995d-7758-4178-9b76-7aba48b9479b","added_by":"auto","created_at":"2025-12-15 06:22:21","extension":"jpg","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2363023,"visible":true,"origin":"","legend":"","description":"","filename":"SFig.3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/5fa6a3611fdc721945785484.jpg"},{"id":98193839,"identity":"70be4e0d-db68-4dd6-9779-ddc4da930d9e","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"xml","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":174125,"visible":true,"origin":"","legend":"","description":"","filename":"41bf1fed71cd45d28fab0b94ae36a7701enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/70f83e5be8175d0333005040.xml"},{"id":98431725,"identity":"afb94296-4610-4a6d-8605-555463075c00","added_by":"auto","created_at":"2025-12-17 16:48:13","extension":"jpg","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3994762,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/1d36873ceb4b10a80bc96d48.jpg"},{"id":98193836,"identity":"3a3452f9-01b3-4c56-91ab-56e2794bf472","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2505539,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/eed280b4e6220c0d54ccce57.jpg"},{"id":98193846,"identity":"41b6ca9a-205b-495e-90a6-827cc30c5d3b","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1694954,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/ca0ff63f5fab8920457539b0.jpg"},{"id":98193855,"identity":"1541775e-9734-4060-951a-427128419c3f","added_by":"auto","created_at":"2025-12-15 06:22:21","extension":"jpg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2947562,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/f761efec54923600bccb62bd.jpg"},{"id":98193847,"identity":"c0884100-5910-41e2-9aaa-62a5bc0eb2f7","added_by":"auto","created_at":"2025-12-15 06:22:21","extension":"jpg","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3168700,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/ccff6284b04e73ada7a61509.jpg"},{"id":98432796,"identity":"82d48fd7-b084-47f0-80fc-bc813de62258","added_by":"auto","created_at":"2025-12-17 16:49:58","extension":"jpg","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3159930,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/4d44e95473e6a6725caf667e.jpg"},{"id":98432153,"identity":"63c98e8e-b149-4e9b-b5e4-ffbc22c440f6","added_by":"auto","created_at":"2025-12-17 16:49:07","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1424866,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/d4fb2e71e83d3cd6fa6dee08.png"},{"id":98193852,"identity":"823d66ca-5a68-465e-9b35-74601c9f138f","added_by":"auto","created_at":"2025-12-15 06:22:21","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":846502,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/b06a19dec0e82a6be2b40731.png"},{"id":98432225,"identity":"9cfb3e57-c78d-44b7-bf78-8bc004a2672d","added_by":"auto","created_at":"2025-12-17 16:49:15","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":547099,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/7167bb058f5826abf4c0bf4d.png"},{"id":98193840,"identity":"bc4fec20-4b2d-4626-aa8b-0bb1844cf6eb","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1017055,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/8e08eb84e958fa0066d352b1.png"},{"id":98193858,"identity":"fc6ccc85-51c7-46f0-b990-96cdebb6a6d6","added_by":"auto","created_at":"2025-12-15 06:22:21","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1192477,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/9d00bc9b643f5471fffebf3a.png"},{"id":98193853,"identity":"f126fdf5-7e31-4268-9e87-01ecf0dc8044","added_by":"auto","created_at":"2025-12-15 06:22:21","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1073472,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure6.png","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/da9029d8dcd6eeb43588e14d.png"},{"id":98193844,"identity":"a30b810d-1d0f-492d-9c1a-b476cae1c85b","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"xml","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":172635,"visible":true,"origin":"","legend":"","description":"","filename":"41bf1fed71cd45d28fab0b94ae36a7701structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/d8cab623ea70d7e692bc9bbc.xml"},{"id":98193850,"identity":"afaee26a-cb14-4e7d-b351-4326ddcf9f67","added_by":"auto","created_at":"2025-12-15 06:22:21","extension":"html","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":195117,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/beec6758b6b0a56d24e61d96.html"},{"id":98193828,"identity":"1bc08fa2-0965-4798-a558-b6a329f426bf","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":3994762,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy Cohort Assembly Flowchart.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e Study design and population. Electronic health record (EHR) data from the OneFlorida+ network (24,593 MCI patients, 4,102 AD conversions) and the All of Us program (4,012 MCI patients, 888 AD conversions) were analyzed. Drugs prescribed to at least 20 % of MCI patients were retained, yielding 186 candidates. Treated patients (T = 1) initiated the candidate drug after MCI diagnosis and maintained ≥ 3 refills during ≥ 1 year of follow-up; controls (T = 0) had no exposure within 5 years of MCI diagnosis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(b)\u003c/strong\u003e Model architecture. CauReL combines a representation network that balances treated and untreated groups via Integral Probability Metric (IPM) regularization with a dual-prediction network that models AD onset and progression time under both treatment and control.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(c)\u003c/strong\u003e Quantifying treatment effects. The framework estimates individualized treatment effects (ITEs) aligned with observed outcomes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(d)\u003c/strong\u003e Interpretable subgroup discovery. CFR-based uplift trees reveal rule-based subgroups associated with benefit or harm.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(e)\u003c/strong\u003e Clinical translation. Benefit-risk mapping identifies four reproducible for AD repurposing.\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/0c53dfa6718e6c2e2a9d6b44.jpg"},{"id":98431927,"identity":"b7c87b73-4721-4e11-ae5d-ed7ea3b47490","added_by":"auto","created_at":"2025-12-17 16:48:39","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":2505539,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEvaluation of covariate balancing and predictive performance across IPM regularization strategies.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e shows the standardized mean differences (SMDs) for baseline covariates between treated and control groups before and after representation learning using three Integral Probability Metric (IPM) configurations: MMD-Linear (yellow), MMD-RBF (blue), and Wasserstein distance (red). The x-axis represents individual drug candidates, and the y-axis represents the average SMD across all baseline covariates. Each drug is displayed as a line segment, where the circle denotes the pre-regularization SMD and the triangle denotes the post-regularization SMD. Longer line lengths indicate greater improvement in covariate balance, while higher starting SMDs reflect more substantial initial imbalance between treated and control groups.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(b)\u003c/strong\u003e presents the Kolmogorov-Smirnov (KS) distance between treated and control group distributions in the learned representation space. The x-axis lists individual drugs, while the y-axis shows the mean KS distance. Lower KS values indicate better alignment of the two distributions. Bars represent the mean KS value across 10-fold cross-validation, and error bars denote the standard deviation, reflecting the stability of each method.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(c)\u003c/strong\u003e displays the predictive performance of the counterfactual outcome model for identifying which MCI patients progress to AD, measured by the area under the receiver operating characteristic curve (AUC). The x-axis represents individual drug, and the y-axis represents the AUC value. All three IPM configurations achieved AUC values above 0.90, indicating strong predictive performance.\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/4c9ec7a3a35662ed80d5166e.jpg"},{"id":98431155,"identity":"d924ca90-ac16-4909-935f-c64e84526873","added_by":"auto","created_at":"2025-12-17 16:47:10","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1694954,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIdentification of drug candidates and cross-cohort validation of treatment effects.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e and \u003cstrong\u003e(b)\u003c/strong\u003e show the ranked average treatment effects (ATEs) for progression from MCI to AD across 20 drug candidates. The x-axis lists the drugs in order of effect size, while the y-axis shows the estimated ATE, where more negative values indicate stronger protective associations. \u003cstrong\u003e(a)\u003c/strong\u003e shows results from the OneFlorida+ discovery cohort, and \u003cstrong\u003e(b)\u003c/strong\u003e shows results from the All of Us validation cohort. \u003cstrong\u003e(c-f)\u003c/strong\u003eprovide ATEs and 95% confidence intervals (CIs) of the four top candidate drugs: liraglutide \u003cstrong\u003e(c)\u003c/strong\u003e, empagliflozin \u003cstrong\u003e(d)\u003c/strong\u003e, entacapone \u003cstrong\u003e(e)\u003c/strong\u003e, and amantadine \u003cstrong\u003e(f)\u003c/strong\u003e. FL- and AOU- prefixes indicate estimates from the OneFlorida+ and All of Us datasets, respectively. Bar colors represent the regularization method: yellow for MMD-Linear, blue for MMD-RBF, and red for Wasserstein distance. Negative ATEs indicate that patients receiving the drug were less likely to progress from MCI to AD compared with untreated individuals.\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/657788d3adeea945d7b73b1d.jpg"},{"id":98193837,"identity":"7a234a76-5c28-43f5-be4a-3a21d176a7b1","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":2947562,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSubgroup analyses reveal heterogeneity of treatment effects\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e presents waterfall plots of individualized treatment effects (ITEs) for each drug, sorted from most to least beneficial. The x-axis represents individual patients, while the y-axis shows the ITE value. Negative ITEs indicate a protective effect against progression to AD, with the proportion of negative values reflecting the consistency of benefit across patients.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(b)\u003c/strong\u003e visualizes subgroup-level heterogeneity stratified by race and sex. The x-axis represents the probability that individuals within a subgroup experience a beneficial treatment effect, while the y-axis represents the probability of high variability or inconsistent effects within that subgroup. Subgroups in the lower-right quadrant exhibit both high benefit and low variability, indicating the most reliable and clinically promising populations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(c)\u003c/strong\u003e shows ATEs stratified by comorbidity status, including diabetes, hypertension, heart failure, obesity, depression, and stroke. The x-axis lists comorbidities, while the y-axis shows the ATE, with darker bars indicating patients with the comorbidity and lighter bars indicating patients without it. Negative values represent protective associations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(d)\u003c/strong\u003e depicts the mean ATEs by age group, where the x-axis represents age categories (50-59, 60-69, 70-79, ≥80), and the y-axis represents the mean ATE. Error bars indicate variability within each age group.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/0bda82ff9cf59318c62279aa.jpg"},{"id":98432012,"identity":"0f2948e2-ac13-4a34-90f9-54380e447419","added_by":"auto","created_at":"2025-12-17 16:48:50","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":3168700,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFeature importance and interaction toward AD\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e shows directional feature importance, where the x-axis lists baseline features and the y-axis shows their directional impact on predicted AD risk. Features with bars pointing to the left are associated with enhanced drug protection, while those pointing to the right are linked to attenuated drug effects.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(b)\u003c/strong\u003e provides a heatmap of pairwise interaction effects between clinical and demographic features. The x- and y-axes represent individual features. Red cells indicate synergistic effects, where the combined risk is greater than additive, while blue cells indicate antagonistic or overlapping effects. Cells outlined in black represent clinically relevant interactions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(c)\u003c/strong\u003e displays uplift decision trees that segment patients into groups with varying levels of benefit. Deep blue nodes represent subgroups with the largest predicted benefit, light blue nodes represent moderate benefit, yellow nodes indicate minimal benefit, and red nodes indicate potential harm.\u003c/p\u003e","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/87a012f570c24354e9221e21.jpg"},{"id":98193842,"identity":"613c5f57-d65f-493a-9189-6516ae19ddbd","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":3159930,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEffects of top candidate drugs on delaying AD progression\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e presents Kaplan-Meier survival curves comparing treated patients and matched controls for time to conversion from MCI to AD. The x-axis represents months since MCI diagnosis, and the y-axis represents the probability of remaining free from AD. Red curves represent treated patients, and blue curves represent matched controls. Clear separation between curves indicates slower progression among treated patients, with all log-rank tests yielding p \u0026lt; 0.05.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(b)\u003c/strong\u003e summarizes the original observed outcomes, showing the distribution of times to AD conversion for treated (red) versus untreated (blue) patients. The x-axis lists the four candidate drugs, and the y-axis represents time in months.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(c)\u003c/strong\u003e shows the counterfactual model predictions for each drug. Violin plots depict the predicted distribution of conversion times under two scenarios: no treatment (blue, left) and treatment (red, right). Wider sections indicate higher densities of patients at a given predicted conversion time, and vertical black bars mark the median prediction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(d)\u003c/strong\u003e compares predicted versus observed treatment effects for delaying AD conversion. The x-axis represents the model-predicted delay, and the y-axis represents the observed delay. Each point represents a patient, with concordance index (C-index) values between 0.81 and 0.84 and Spearman correlations between 0.80 and 0.86, demonstrating strong agreement between predicted and actual effects.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(e)\u003c/strong\u003e illustrates the directional feature importance for each drug’s counterfactual model. The x-axis represents the magnitude of the feature’s influence on predicted time to AD conversion, while the y-axis lists key clinical and demographic features (e.g., age, comorbidities). Red bars indicate features associated with increased AD risk (shorter time to conversion), whereas green bars indicate protective features associated with delayed progression. Each subpanel corresponds to one of the four candidate drugs: liraglutide, empagliflozin, entacapone, and amantadine.\u003c/p\u003e","description":"","filename":"Figure6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/92d476af3da18778ad2e4dda.jpg"},{"id":100369562,"identity":"cb4af6c5-87bd-4509-a87c-4d3a94637a43","added_by":"auto","created_at":"2026-01-16 07:59:08","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":18170484,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1_covered_bdd8afb8-fb95-4267-a16f-b950116fdc65.pdf"},{"id":98193824,"identity":"382779a5-372a-4d83-bf49-a2bc3dcb5a63","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1413742,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Figure 1. Feature importance in All of Us\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003evisualizes subgroup-level heterogeneity stratified by race and sex. The x-axis represents the probability that individuals within a subgroup experience a beneficial treatment effect, while the y-axis represents the probability of high variability or inconsistent effects within that subgroup. Subgroups in the lower-right quadrant exhibit both high benefit and low variability, indicating the most reliable and clinically promising populations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(b)\u003c/strong\u003eshows ATEs stratified by comorbidity status, including diabetes, hypertension, heart failure, obesity, depression, and stroke. The x-axis lists comorbidities, while the y-axis shows the ATE, with darker bars indicating patients with the comorbidity and lighter bars indicating patients without it. Negative values represent protective associations.\u003c/p\u003e","description":"","filename":"SFig.1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/b84d215f4fb00a836dcea36b.jpg"},{"id":98430663,"identity":"9c2e0778-40ac-45b0-9d2c-bdbe4ecd8623","added_by":"auto","created_at":"2025-12-17 16:46:01","extension":"jpg","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":1558981,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Figure 2. Subgroup analyses in All of Us\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003edepicts the mean ATEs by age group, where the x-axis represents age categories (50-59, 60-69, 70-79, ≥80), and the y-axis represents the mean ATE. Error bars indicate variability within each age group.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(b)\u003c/strong\u003eshows directional feature importance, where the x-axis lists baseline features and the y-axis shows their directional impact on predicted AD risk. Features with bars pointing to the left are associated with enhanced drug protection, while those pointing to the right are linked to attenuated drug effects.\u003c/p\u003e","description":"","filename":"SFig.2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/0511c5f177eedf4832358cb6.jpg"},{"id":98193838,"identity":"4a4fc329-a32c-4574-8f99-10a8a8f58a91","added_by":"auto","created_at":"2025-12-15 06:22:20","extension":"jpg","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":2363023,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Figure 3. Time of AD conversion in All of Us\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003epresents Kaplan-Meier survival curves comparing treated patients and matched controls for time to conversion from MCI to AD. The x-axis represents months since MCI diagnosis, and the y-axis represents the probability of remaining free from AD. Red curves represent treated patients, and blue curves represent matched controls. Clear separation between curves indicates slower progression among treated patients, with all log-rank tests yielding p \u0026lt; 0.05.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(b)\u003c/strong\u003eshows the counterfactual model predictions for each drug. Violin plots depict the predicted distribution of conversion times under two scenarios: no treatment (blue, left) and treatment (red, right). Wider sections indicate higher densities of patients at a given predicted conversion time, and vertical black bars mark the median prediction.\u003c/p\u003e","description":"","filename":"SFig.3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8206648/v1/08b92c2b212c12066193865a.jpg"}],"financialInterests":"No competing interests reported.","formattedTitle":"CauReL: Dynamic Counterfactual Learning for Precision Drug Repurposing in Alzheimer's Disease","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Causal AI, Counterfactual representation learning, Individualized Treatment Effects (ITE), Drug Repurposing, Electronic Health Records (EHRs), Alzheimer’s Disease (AD)","lastPublishedDoi":"10.21203/rs.3.rs-8206648/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8206648/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAlzheimer\u0026rsquo;s disease has few effective therapies, and decades of amyloid- and tau-focused trials have delivered only modest benefit with substantial toxicity. Drug repurposing using real-world data offers a faster and lower-risk route to new treatments, yet current approaches typically average effects across populations, model disease onset and progression separately, and provide little insight into which patients are most likely to benefit. We present CauReL, a dynamic counterfactual representation learning framework that enables transparent, patient specific estimation of treatment effects from large-scale electronic health records for precision drug repurposing in AD. CauReL first learns balanced latent representations of treated and untreated patients using Integral Probability Metric regularization, then jointly predicts two clinically linked outcomes, incident AD and time from mild cognitive impairment (MCI) to AD, to generate paired counterfactual outcomes for every individual. A counterfactual explanation module quantifies how clinical features shape benefit at the patient level, and uplift trees transform complex heterogeneity into simple, rule-based subgroups suitable for trial enrichment and clinical decision support.\u003c/p\u003e\u003cp\u003eUsing independent cohorts from OneFlorida\u0026thinsp;+\u0026thinsp;and All of Us, we screened outpatient prescriptions with at least 20 percent exposure among 28,605 individuals with mild cognitive impairment, of whom 4,990 progressed to Alzheimer\u0026rsquo;s disease. CauReL substantially improved covariate balance and distributional overlap across drug cohorts and achieved strong predictive accuracy for both incidence (AUC greater than 0.90) and progression timing (C index 0.81 to 0.84; Spearman 0.80 to 0.86). Twenty drugs showed consistent protective associations, with four emerging as highly reproducible across both networks, the metabolic agents liraglutide and empagliflozin and the neuroactive agents entacapone and amantadine. These drugs were associated with meaningful absolute risk reductions and clinically significant delays in progression from mild cognitive impairment to Alzheimer\u0026rsquo;s disease. Metabolic drugs produced the strongest benefits in individuals with diabetes, obesity, or cardiovascular disease, whereas neuroactive drugs provided broadly consistent protection across most subgroups.\u003c/p\u003e\u003cp\u003eCauReL is available as an open source Python package with a companion web server for direct application to new cohorts or disease settings (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://caurel.site/\u003c/span\u003e\u003cspan address=\"https://caurel.site/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). This work delivers a scalable and interpretable framework for prioritizing repurposable drugs and designing targeted clinical trials for the patients most likely to benefit.\u003c/p\u003e","manuscriptTitle":"CauReL: Dynamic Counterfactual Learning for Precision Drug Repurposing in Alzheimer's Disease","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-15 06:22:15","doi":"10.21203/rs.3.rs-8206648/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8295cce9-07a2-4dc9-9976-bc3c7de05383","owner":[],"postedDate":"December 15th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":59517539,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":59517540,"name":"Health sciences/Diseases"},{"id":59517541,"name":"Health sciences/Medical research"},{"id":59517542,"name":"Health sciences/Neurology"},{"id":59517543,"name":"Biological sciences/Neuroscience"}],"tags":[],"updatedAt":"2026-01-13T18:53:43+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-15 06:22:15","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8206648","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8206648","identity":"rs-8206648","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}