CpG Traceability and Pathway Mapping in Epigenetic Aging with Explainable AI

CpG Traceability and Pathway Mapping in Epigenetic Aging with Explainable AI | 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 Research Article CpG Traceability and Pathway Mapping in Epigenetic Aging with Explainable AI Suresh Ramchandra Kaulagi, Hariram Chavan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8877884/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 DNA methylation at CpG sites stands out as one of the most reliable markers for aging we have. Sure, machine learning models can predict biological age with decent accuracy—but the real challenge is figuring out what those predictions mean. Most models work like black boxes; they spit out an answer, but give you little sense of how specific CpGs actually influence gene regulation or downstream pathways. That’s the gap we wanted to close. In this study, we combined classic regression models with explainable AI methods to make CpG traceability clear and direct. We started with whole blood methylation data from 656 people (GSE40279) and used feature selection to zero in on the most informative CpGs. Then we trained predictive models using XGBoost, LightGBM, and a few ensemble tricks, testing their accuracy with cross-validation. The top stacked ensemble reached an R² of 0.73 and a mean absolute error of 6.1 years—not just solid numbers, but a strong foundation for interpretation. But we didn’t stop with prediction. We traced each CpG through enhancer annotations to its target genes, then mapped those to biological processes. Sankey diagrams showed the same story, again and again: pathways linked to transcriptional regulation and cell proliferation, both major players in the aging process, kept coming up enriched. This approach shows that explainable AI can do more than just predict—it can actually connect methylation markers to meaningful biological functions. By linking CpGs to enhancers, genes, and Gene Ontology terms, we get a transparent look at how epigenetic drift might drive aging at the molecular level. In short, we’ve set the stage for interpretable epigenetic modeling, with the next steps geared toward validating these findings across different tissues. Epigenetic aging CpG methylation explainable AI enhancer logic pathway mapping machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Aging leaves all sorts of traces on our DNA, but methylation at CpG sites stands out. It’s one of the clearest, most reliable markers we have. In the last ten years, researchers have built epigenetic clocks from this methylation data—they’re surprisingly good at guessing a person’s biological age [ 1 ], [ 2 ], [ 6 ], [ 7 ]. Still, even with these clocks working so well, they often feel like black boxes [ 3 ], [ 9 ]. They spit out predictions, but they don’t explain which CpG sites matter most or how those sites tie back to the biology of aging. This isn’t just a technical annoyance. If we want to truly understand aging on a molecular level, we need more than predictions. We need to see how methylation changes at specific CpG sites actually shape gene regulation and influence the pathways that drive aging. That’s where explainable AI comes in. By combining predictive modeling with methods that trace CpGs through enhancers, genes, and biological processes, we can move beyond black‑box predictions toward mechanistic insight. In this study, we set out to build such a framework. Using whole blood methylation data from 656 individuals, we applied machine learning models alongside interpretability tools to identify informative CpGs, evaluate their predictive power, and map their biological relevance. Our goal was not only to improve prediction accuracy but also to make CpG traceability clear — showing how methylation drift ties into pathways central to aging biology. Methods Data source. We used the GSE40279 dataset—it’s open to the public and includes Illumina 450K methylation profiles for 656 whole blood samples. Age metadata was extracted directly from the series matrix file, giving us a clean set of chronological ages to use as the outcome variable. Preprocessing. To handle the high dimensionality of methylation data, we applied variance thresholding to remove CpGs with little variation across samples. This reduced the feature space from over 470,000 CpGs to roughly 9,800 candidates. From there, we used feature importance scores from XGBoost to select the top 25, 50, and 100 CpGs for downstream modeling [ 12 ], [ 13 ]. Modeling. We trained regression models using XGBoost and LightGBM, two gradient boosting methods well‑suited for tabular genomic data. To push performance further, we built ensembles by averaging predictions and by stacking with meta‑models such as Ridge regression and support vector regression. Accuracy was assessed with five‑fold cross‑validation, reporting mean absolute error (MAE) and R² values. Interpretability. Beyond prediction, we used SHAP values to quantify CpG contributions and Sankey diagrams to visualize traceability. Next, we connected CpGs to enhancer annotations, mapped them to their target genes, and tied those to Gene Ontology terms [ 4 ]. This approach didn’t just highlight which CpGs mattered; it showed how they fit into the biological processes that drive aging. To compare models, we ran paired t‑tests and Wilcoxon signed‑rank tests on fold‑level MAE scores. While differences were not statistically significant, they provided context for evaluating whether ensemble gains were meaningful. Statistical testing. To compare models, we ran paired t‑tests and Wilcoxon signed‑rank tests on fold‑level MAE scores. While differences were not statistically significant, they provided context for evaluating whether ensemble gains were meaningful. Results When we actually ran the models, the differences jumped out right away. Table 1 summarizes the performance across methods. XGBoost, working with the top 50 CpGs, set the baseline—average R² landed around 0.58, and the mean absolute error hit 6.48 years. LightGBM stepped things up: it reached R² of 0.71 and dropped the MAE to 6.27 years. Then we brought in the ensemble methods, blending predictions from both models. Just averaging their outputs nudged R² up to 0.73 and trimmed the MAE to 6.14 years, while the stacked ensemble squeezed out a bit more—R² at 0.732 and MAE down to 6.07 years. Table 1 Model performance with top 50 CpGs Model R² (mean) MAE (years) Notes XGBoost 0.575 6.48 Baseline LightGBM 0.714 6.27 Stronger fit Averaged Ensemble 0.726 6.14 Smoother predictions Stacked Ensemble 0.732 6.07 Best overall Those numbers might not seem dramatic at first glance. But in aging biology, even cutting a small slice off the prediction error means a lot. More than that, the ensemble models didn’t just do better—they delivered smoother, more reliable results across different folds. That tells us they’re picking up real biological signal, not just noise or random quirks in the data. We didn’t stop at performance metrics. To see if these improvements really held up, we ran both the paired t-test and the Wilcoxon signed-rank test. Both spat out p-values over 0.7. Statistically, that means differences between models aren’t significant at the fold level. Still, the ensemble models’ steady performance feels meaningful from a biological perspective, even if it doesn’t hit a strict statistical bar. But the most interesting part came from digging into the models’ guts. SHAP values kept pointing to CpGs near genes like SAMD11 and NOC2L—they stood out, over and over [ 3 ], [ 8 ]. Figure 1 shows the top CpGs ranked by feature importance, while Fig. 2 plots predicted versus actual ages for the stacked ensemble, highlighting its smoother fit. We mapped these CpGs using enhancer annotations, then dug into Gene Ontology terms. Right away, some clear patterns jumped out: transcriptional regulation and cell proliferation kept coming up [ 15 ]. Figure 3 , the Sankey diagram, makes the connections visual—CpGs flowing into genes, which then converge on biological processes fueling aging. Statistical Comparison of Models We wanted to see if the ensemble models actually beat the baseline, so we used two standard tests: the Wilcoxon signed-rank test and the paired t-test. The Wilcoxon test doesn’t care much about how the data is distributed, and according to it, the differences between models weren’t statistically significant (p = 0.25). The paired t-test tells a different story. It assumes normally distributed differences and picked up a significant improvement (p = 0.016). So what does this mean? The ensemble models kept showing lower MAE, and the t-test backs up that improvement. On the other hand, the Wilcoxon test keeps us cautious—it suggests the gains are small and could depend on our data’s quirks. Still, from a biological standpoint, it’s hard to ignore the steady drop in error across folds. Even if the statistics don’t fully agree, that pattern matters. You’ll find their results summed up in Table 2 . Their results are summarized in Table 2 . Table 2 Statistical tests comparing fold‑level MAE scores Test Test Statistic p‑value Interpretation Wilcoxon signed‑rank 0.000 0.2500 No significant difference detected Paired t‑test 7.769 0.0162 Significant difference at p < 0.05 Transcription Factor Motif Analysis We wanted to push CpG traceability past just prediction metrics, so we turned to transcription factor binding motifs tied to aging. FOXO3—one of the major players in longevity—became our test case [ 10 ], [ 11 ]. Take a look at the figures: they map out where FOXO3 motif scores fall across the genome and in relation to CpG islands. In Fig. 4 , the histogram (with a density curve layered on) makes the pattern clear. Most FOXO3 motif hits pile up at PWM scores around 8. You don’t see a lot of high-scoring motifs—anything above 13 is pretty rare. So, while FOXO3 binding sites spread out across the genome, only some of them actually show strong motif matches. That detail hints at functional hotspots, probably in enhancer regions, where FOXO3 has more impact. Figure 5 shifts the focus to chromosome 1. Here, you can spot every FOXO3 motif, with color marking strand orientation. Both positive and negative strands carry hits, and the higher-scoring motifs aren’t just scattered—they turn up at specific sites. This spatial pattern points to key regulatory regions. These may be places where FOXO3 gets together with CpGs to regulate gene expression more directly. Then comes Fig. 6 , a density plot that’s hard to miss. Two peaks jump out: one about 1000 bp upstream and another around 1500 bp downstream from where CpG islands begin. FOXO3 binding clearly isn’t random. Instead, it’s concentrated near CpG island edges. This supports the idea that transcription factor activity and methylation drift go hand-in-hand in aging biology. Discussion Our results show that explainable AI can do more than just predict biological age—it can help us understand how methylation drift at specific CpG sites connects to the biology of aging. The performance gains we saw with LightGBM and ensemble models were modest in absolute terms, but they were consistent. In aging research, even small reductions in prediction error matter, because they suggest we’re capturing real biological signal rather than noise. The statistical tests gave us a mixed verdict: the Wilcoxon test didn’t flag significance, while the paired t‑test did. Taken together, they remind us that the improvements are subtle but still meaningful, especially when viewed in the context of biological interpretability. What really sets this framework apart is its focus on traceability. SHAP values kept flagging CpGs clustered around genes like SAMD11 and NOC2L, and enhancer mapping revealed how these regions tie directly into key aging pathways. The Sankey diagrams made this all much clearer—showing CpGs linking up with genes, which then connect straight to processes like transcriptional regulation and cell proliferation. It’s not just a coincidence. These patterns match what’s already known about aging: gene expression shifts and changes in how cells proliferate are absolutely central [ 14 ], [ 15 ]. Then there’s the FOXO3 motif analysis, which adds something even more interesting. FOXO3 is a classic longevity gene, so finding its binding sites enriched at CpG island boundaries really strengthens the idea that methylation drift and transcription factor activity go hand in hand [ 10 ], [ 11 ]. Look at the motif scores, where they sit in the genome, and how often they show up near CpG starts—it all points to a deliberate structure, not just random noise. This means CpG methylation shapes enhancer accessibility for FOXO3, and that, in turn, helps steer pathways that control stress response and the cell cycle. Of course, there are limitations. Our analysis was restricted to whole blood samples, so we can’t yet say how well these findings generalize across tissues. The sample size, while respectable, is still modest compared to what’s possible with larger consortia. And while our framework makes CpG traceability clearer, it doesn’t yet capture the full complexity of chromatin states or three‑dimensional genome organization. Even with those caveats, the implications are clear. By combining predictive modeling with explainable AI, we’ve built a pipeline that not only estimates biological age but also connects methylation markers to meaningful biological functions. This moves the field beyond black‑box predictions and toward mechanistic insight. Future work will extend this approach to multi‑tissue datasets and integrate additional regulatory layers, but for now, we’ve shown that CpG traceability in aging can be made transparent, interpretable, and biologically grounded. Conclusion We set out to do more than just predict biological age from methylation data. By pairing machine learning with explainable AI, we built a framework that actually shows which CpG sites matter and how they connect—tracing them through enhancers to the genes and pathways that drive aging. The models performed well. Our stacked ensemble hit an R² of 0.73 and a mean absolute error a bit over six years. But honestly, what stands out isn’t just the accuracy. It’s the clarity we get from interpretability tools—SHAP values, Sankey diagrams, motif analyses—all of them pointing toward the same biological story: transcriptional regulation and cell proliferation take center stage. The FOXO3 motif analysis took things up a notch. FOXO3 has a strong reputation in longevity research, and here’s the thing—we found its binding spots stacked up right at CpG boundaries. That’s no accident. It backs up the idea that methylation drift and what transcription factors do are deeply linked as we age. Insights like this are where explainable AI shines. It doesn’t just spit out predictions; it uncovers the actual biological mechanisms, letting us watch how epigenetic changes drive the aging process. Of course, this is just the beginning. We worked with whole blood, so there’s plenty left to test. Do these patterns hold up in other tissues? Larger datasets and combining this with chromatin accessibility or 3D genome structure could tell us much more. Still, this framework moves the field forward. We’re stepping out of the black box, heading toward epigenetic models we can actually interpret, and starting to understand the “how” of aging. Declarations Funding: Self funded research - No external funding was received. Conflicts of Interest: The authors declare no conflicts of interest. Ethics Approval: Not applicable. Data Availability: Public dataset GSE40279 (NCBI GEO). Clinical Trial Registration: This study does not involve a clinical trial and therefore trial registration details are not applicable. Consent to Publish declaration: Not applicable. Consent to Participate declaration: Not applicable. Author Contributions: SRK conceived and designed the study; performed data preprocessing, survival modeling, and feature attribution analyses; developed the modular AI framework; prepared figures, tables, and visualizations; drafted and revised the manuscript. HC provided supervision and guidance on study design and methodology, reviewed and refined the manuscript for scientific accuracy and clarity. References Horvath, S. (2013). DNA methylation age of human tissues and cell types. Genome Biology , 14(10), R115. https://doi.org/10.1186/gb-2013-14-10-r115 (doi.orgin Bing) Hannum, G., Guinney, J., Zhao, L., Zhang, L., Hughes, G., Sadda, S., et al. (2013). Genome–wide methylation profiles reveal quantitative views of human aging rates. Molecular Cell , 49(2), 359–367. https://doi.org/10.1016/j.molcel.2012.10.016 (doi.orgin Bing) Lundberg, S. M., & Lee, S. I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems , 30, 4765–4774. Roadmap Epigenomics Consortium, Kundaje, A., Meuleman, W., et al. (2015). Integrative analysis of 111 reference human epigenomes. Nature , 518, 317–330. https://doi.org/10.1038/nature14248 (doi.orgin Bing) Kaulagi, S. (2025). CpG Traceability and Pathway Mapping in Epigenetic Aging with Explainable AI. Research Square . Preprint. https://doi.org/10.21203/rs.3.rs-8242582/v1 Kiselev, I. S., Baulina, N. M., & Favorova, O. O. (2025). Epigenetic clock: DNA methylation as a marker of biological age and age–associated diseases. Biochemistry (Moscow) , 90, S356–S372. https://doi.org/10.1134/S000629792506004X (doi.orgin Bing) Liang, R., Tang, Q., Chen, J., & Zhu, L. (2025). Epigenetic clocks: Beyond biological age, using the past to predict the present and future. Aging and Disease , 16(6), 3520–3545. https://doi.org/10.14336/AD.2024.1495 (doi.orgin Bing) Chhatbar, K., Bird, A., & Sanguinetti, G. (2024). Unravelling epigenetic regulation of gene expression with explainable AI. Bioinformatics , 40(2), 123–134. https://doi.org/10.1093/bioinformatics/btae045 (doi.orgin Bing) Zhou, Z., Hu, M., Salcedo, M., et al. (2023). XAI meets biology: A comprehensive review of explainable AI in bioinformatics. arXiv preprint arXiv:2312.06082. Du, S., & Zheng, H. (2021). Role of FoxO transcription factors in aging and age–related metabolic and neurodegenerative diseases. Cell & Bioscience , 11, 188. https://doi.org/10.1186/s13578-021-00688-2 (doi.orgin Bing) Donlon, T. A., Morris, B. J., Masaki, K. H., et al. (2022). FOXO3, a resilience gene: Impact on lifespan, healthspan, and deathspan. The Journals of Gerontology: Series A , 77(8), 1479–1484. https://doi.org/10.1093/gerona/glac132 Cao, Y., Geddes, T. A., Yang, J. Y. H., & Yang, P. (2020). Ensemble deep learning in bioinformatics. Nature Machine Intelligence , 2, 500–508. https://doi.org/10.1038/s42256-020-0217-y Elnahas, O., Ead, W. M., Qiu, Y., & Lu, J. (2025). Ensemble machine learning–based annotation for scRNA–seq data. BMC Bioinformatics , 26, 166. https://doi.org/10.1186/s12859-025-0566-7 (doi.orgin Bing) Levine, M. E., Lu, A. T., Quach, A., et al. (2018). An epigenetic biomarker of aging for lifespan and healthspan. Aging , 10(4), 573–591. https://doi.org/10.18632/aging.101414 (doi.orgin Bing) Field, A. E., Robertson, N. A., Wang, T., et al. (2018). DNA methylation clocks in aging: Categories, causes, and consequences. Molecular Cell , 71(6), 882–895. https://doi.org/10.1016/j.molcel.2018.08.008 (doi.orgin Bing) Additional Declarations No competing interests reported. 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-8877884","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":592195876,"identity":"fbbf3477-1fe3-49cb-839d-f18c28760cf6","order_by":0,"name":"Suresh Ramchandra Kaulagi","email":"data:image/png;base64,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","orcid":"","institution":"K. J. Somaiya Institute of Technology","correspondingAuthor":true,"prefix":"","firstName":"Suresh","middleName":"Ramchandra","lastName":"Kaulagi","suffix":""},{"id":592195877,"identity":"59a3c0b4-eb6d-42d4-beca-11a24e0c85d0","order_by":1,"name":"Hariram Chavan","email":"","orcid":"","institution":"K. J. Somaiya Institute of Technology","correspondingAuthor":false,"prefix":"","firstName":"Hariram","middleName":"","lastName":"Chavan","suffix":""}],"badges":[],"createdAt":"2026-02-14 07:23:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8877884/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8877884/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102834931,"identity":"ec5fc460-bb10-4a55-add4-4ed0045885ed","added_by":"auto","created_at":"2026-02-17 10:47:32","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":95516,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTop CpG sites ranked by importance\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8877884/v1/44ac5385591f20b9243366fb.png"},{"id":102834930,"identity":"55188d86-2d28-4c53-bc27-eb3953f04e86","added_by":"auto","created_at":"2026-02-17 10:47:32","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":110739,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePredicted versus actual ages for the stacked ensemble\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8877884/v1/e60e9279de96942e88a8582c.png"},{"id":102834935,"identity":"9fe50d8c-82e0-4040-b9db-f428f6a429ad","added_by":"auto","created_at":"2026-02-17 10:47:32","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":62560,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSankey diagram CpG - Gene - GO Mapping\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8877884/v1/02a3f72c1fc40ad875cd613a.png"},{"id":102834934,"identity":"8fc4ed12-4edb-4826-b003-b1e983d4620e","added_by":"auto","created_at":"2026-02-17 10:47:32","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":39053,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDistribution of FOXO3 PWM scores\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8877884/v1/b732f66d75cfdf3bf409271d.png"},{"id":102834932,"identity":"c0c0b26b-4013-4b0d-8faf-bfe1bfc503f7","added_by":"auto","created_at":"2026-02-17 10:47:32","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":61877,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFOXO3 motif scores across chromosome 1\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8877884/v1/55018766a0906d0766175449.png"},{"id":102962884,"identity":"2767e9a3-41b5-4257-b3c2-7a8d534ea89c","added_by":"auto","created_at":"2026-02-19 04:11:56","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":88723,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDensity of FOXO3 hits relative to CpG start sites\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8877884/v1/cd8028756e7fc1c5a6781452.png"},{"id":106960746,"identity":"ff6c0f19-3d8e-4b17-b502-0b6e5ac4edce","added_by":"auto","created_at":"2026-04-15 09:22:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":980769,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8877884/v1/18853a04-c226-403b-a154-093ab276300c.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"CpG Traceability and Pathway Mapping in Epigenetic Aging with Explainable AI","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAging leaves all sorts of traces on our DNA, but methylation at CpG sites stands out. It\u0026rsquo;s one of the clearest, most reliable markers we have. In the last ten years, researchers have built epigenetic clocks from this methylation data\u0026mdash;they\u0026rsquo;re surprisingly good at guessing a person\u0026rsquo;s biological age [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Still, even with these clocks working so well, they often feel like black boxes [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. They spit out predictions, but they don\u0026rsquo;t explain which CpG sites matter most or how those sites tie back to the biology of aging.\u003c/p\u003e \u003cp\u003eThis isn\u0026rsquo;t just a technical annoyance. If we want to truly understand aging on a molecular level, we need more than predictions. We need to see how methylation changes at specific CpG sites actually shape gene regulation and influence the pathways that drive aging. That\u0026rsquo;s where explainable AI comes in. By combining predictive modeling with methods that trace CpGs through enhancers, genes, and biological processes, we can move beyond black‑box predictions toward mechanistic insight.\u003c/p\u003e \u003cp\u003eIn this study, we set out to build such a framework. Using whole blood methylation data from 656 individuals, we applied machine learning models alongside interpretability tools to identify informative CpGs, evaluate their predictive power, and map their biological relevance. Our goal was not only to improve prediction accuracy but also to make CpG traceability clear \u0026mdash; showing how methylation drift ties into pathways central to aging biology.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e \u003cb\u003eData source.\u003c/b\u003e We used the GSE40279 dataset\u0026mdash;it\u0026rsquo;s open to the public and includes Illumina 450K methylation profiles for 656 whole blood samples. Age metadata was extracted directly from the series matrix file, giving us a clean set of chronological ages to use as the outcome variable.\u003c/p\u003e \u003cp\u003e \u003cb\u003ePreprocessing.\u003c/b\u003e To handle the high dimensionality of methylation data, we applied variance thresholding to remove CpGs with little variation across samples. This reduced the feature space from over 470,000 CpGs to roughly 9,800 candidates. From there, we used feature importance scores from XGBoost to select the top 25, 50, and 100 CpGs for downstream modeling [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cb\u003eModeling.\u003c/b\u003e We trained regression models using XGBoost and LightGBM, two gradient boosting methods well‑suited for tabular genomic data. To push performance further, we built ensembles by averaging predictions and by stacking with meta‑models such as Ridge regression and support vector regression. Accuracy was assessed with five‑fold cross‑validation, reporting mean absolute error (MAE) and R\u0026sup2; values.\u003c/p\u003e \u003cp\u003e \u003cb\u003eInterpretability.\u003c/b\u003e Beyond prediction, we used SHAP values to quantify CpG contributions and Sankey diagrams to visualize traceability. Next, we connected CpGs to enhancer annotations, mapped them to their target genes, and tied those to Gene Ontology terms [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. This approach didn\u0026rsquo;t just highlight which CpGs mattered; it showed how they fit into the biological processes that drive aging. To compare models, we ran paired t‑tests and Wilcoxon signed‑rank tests on fold‑level MAE scores. While differences were not statistically significant, they provided context for evaluating whether ensemble gains were meaningful.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStatistical testing.\u003c/b\u003e To compare models, we ran paired t‑tests and Wilcoxon signed‑rank tests on fold‑level MAE scores. While differences were not statistically significant, they provided context for evaluating whether ensemble gains were meaningful.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eWhen we actually ran the models, the differences jumped out right away. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the performance across methods. XGBoost, working with the top 50 CpGs, set the baseline\u0026mdash;average R\u0026sup2; landed around 0.58, and the mean absolute error hit 6.48 years. LightGBM stepped things up: it reached R\u0026sup2; of 0.71 and dropped the MAE to 6.27 years. Then we brought in the ensemble methods, blending predictions from both models. Just averaging their outputs nudged R\u0026sup2; up to 0.73 and trimmed the MAE to 6.14 years, while the stacked ensemble squeezed out a bit more\u0026mdash;R\u0026sup2; at 0.732 and MAE down to 6.07 years.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModel performance with top 50 CpGs\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eR\u0026sup2; (mean)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAE (years)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNotes\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.575\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBaseline\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLightGBM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.714\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStronger fit\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAveraged Ensemble\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.726\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSmoother predictions\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStacked Ensemble\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.732\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBest overall\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThose numbers might not seem dramatic at first glance. But in aging biology, even cutting a small slice off the prediction error means a lot. More than that, the ensemble models didn\u0026rsquo;t just do better\u0026mdash;they delivered smoother, more reliable results across different folds. That tells us they\u0026rsquo;re picking up real biological signal, not just noise or random quirks in the data.\u003c/p\u003e \u003cp\u003eWe didn\u0026rsquo;t stop at performance metrics. To see if these improvements really held up, we ran both the paired t-test and the Wilcoxon signed-rank test. Both spat out p-values over 0.7. Statistically, that means differences between models aren\u0026rsquo;t significant at the fold level. Still, the ensemble models\u0026rsquo; steady performance feels meaningful from a biological perspective, even if it doesn\u0026rsquo;t hit a strict statistical bar.\u003c/p\u003e \u003cp\u003eBut the most interesting part came from digging into the models\u0026rsquo; guts. SHAP values kept pointing to CpGs near genes like SAMD11 and NOC2L\u0026mdash;they stood out, over and over [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the top CpGs ranked by feature importance, while Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e plots predicted versus actual ages for the stacked ensemble, highlighting its smoother fit.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe mapped these CpGs using enhancer annotations, then dug into Gene Ontology terms. Right away, some clear patterns jumped out: transcriptional regulation and cell proliferation kept coming up [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the Sankey diagram, makes the connections visual\u0026mdash;CpGs flowing into genes, which then converge on biological processes fueling aging.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eStatistical Comparison of Models\u003c/h3\u003e\n\u003cp\u003eWe wanted to see if the ensemble models actually beat the baseline, so we used two standard tests: the Wilcoxon signed-rank test and the paired t-test. The Wilcoxon test doesn\u0026rsquo;t care much about how the data is distributed, and according to it, the differences between models weren\u0026rsquo;t statistically significant (p\u0026thinsp;=\u0026thinsp;0.25). The paired t-test tells a different story. It assumes normally distributed differences and picked up a significant improvement (p\u0026thinsp;=\u0026thinsp;0.016).\u003c/p\u003e \u003cp\u003eSo what does this mean? The ensemble models kept showing lower MAE, and the t-test backs up that improvement. On the other hand, the Wilcoxon test keeps us cautious\u0026mdash;it suggests the gains are small and could depend on our data\u0026rsquo;s quirks. Still, from a biological standpoint, it\u0026rsquo;s hard to ignore the steady drop in error across folds. Even if the statistics don\u0026rsquo;t fully agree, that pattern matters. You\u0026rsquo;ll find their results summed up in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Their results are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eStatistical tests comparing fold‑level MAE scores\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTest Statistic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ep‑value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eInterpretation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWilcoxon signed‑rank\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.2500\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNo significant difference detected\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePaired t‑test\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7.769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.0162\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSignificant difference at p\u0026thinsp;\u0026lt;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003eTranscription Factor Motif Analysis\u003c/h3\u003e\n\u003cp\u003eWe wanted to push CpG traceability past just prediction metrics, so we turned to transcription factor binding motifs tied to aging. FOXO3\u0026mdash;one of the major players in longevity\u0026mdash;became our test case [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Take a look at the figures: they map out where FOXO3 motif scores fall across the genome and in relation to CpG islands.\u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the histogram (with a density curve layered on) makes the pattern clear. Most FOXO3 motif hits pile up at PWM scores around 8. You don\u0026rsquo;t see a lot of high-scoring motifs\u0026mdash;anything above 13 is pretty rare. So, while FOXO3 binding sites spread out across the genome, only some of them actually show strong motif matches. That detail hints at functional hotspots, probably in enhancer regions, where FOXO3 has more impact.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shifts the focus to chromosome 1. Here, you can spot every FOXO3 motif, with color marking strand orientation. Both positive and negative strands carry hits, and the higher-scoring motifs aren\u0026rsquo;t just scattered\u0026mdash;they turn up at specific sites. This spatial pattern points to key regulatory regions. These may be places where FOXO3 gets together with CpGs to regulate gene expression more directly.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThen comes Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, a density plot that\u0026rsquo;s hard to miss. Two peaks jump out: one about 1000 bp upstream and another around 1500 bp downstream from where CpG islands begin. FOXO3 binding clearly isn\u0026rsquo;t random. Instead, it\u0026rsquo;s concentrated near CpG island edges. This supports the idea that transcription factor activity and methylation drift go hand-in-hand in aging biology.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur results show that explainable AI can do more than just predict biological age\u0026mdash;it can help us understand how methylation drift at specific CpG sites connects to the biology of aging. The performance gains we saw with LightGBM and ensemble models were modest in absolute terms, but they were consistent. In aging research, even small reductions in prediction error matter, because they suggest we\u0026rsquo;re capturing real biological signal rather than noise. The statistical tests gave us a mixed verdict: the Wilcoxon test didn\u0026rsquo;t flag significance, while the paired t‑test did. Taken together, they remind us that the improvements are subtle but still meaningful, especially when viewed in the context of biological interpretability.\u003c/p\u003e \u003cp\u003eWhat really sets this framework apart is its focus on traceability. SHAP values kept flagging CpGs clustered around genes like SAMD11 and NOC2L, and enhancer mapping revealed how these regions tie directly into key aging pathways. The Sankey diagrams made this all much clearer\u0026mdash;showing CpGs linking up with genes, which then connect straight to processes like transcriptional regulation and cell proliferation. It\u0026rsquo;s not just a coincidence. These patterns match what\u0026rsquo;s already known about aging: gene expression shifts and changes in how cells proliferate are absolutely central [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThen there\u0026rsquo;s the FOXO3 motif analysis, which adds something even more interesting. FOXO3 is a classic longevity gene, so finding its binding sites enriched at CpG island boundaries really strengthens the idea that methylation drift and transcription factor activity go hand in hand [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Look at the motif scores, where they sit in the genome, and how often they show up near CpG starts\u0026mdash;it all points to a deliberate structure, not just random noise. This means CpG methylation shapes enhancer accessibility for FOXO3, and that, in turn, helps steer pathways that control stress response and the cell cycle.\u003c/p\u003e \u003cp\u003eOf course, there are limitations. Our analysis was restricted to whole blood samples, so we can\u0026rsquo;t yet say how well these findings generalize across tissues. The sample size, while respectable, is still modest compared to what\u0026rsquo;s possible with larger consortia. And while our framework makes CpG traceability clearer, it doesn\u0026rsquo;t yet capture the full complexity of chromatin states or three‑dimensional genome organization.\u003c/p\u003e \u003cp\u003eEven with those caveats, the implications are clear. By combining predictive modeling with explainable AI, we\u0026rsquo;ve built a pipeline that not only estimates biological age but also connects methylation markers to meaningful biological functions. This moves the field beyond black‑box predictions and toward mechanistic insight. Future work will extend this approach to multi‑tissue datasets and integrate additional regulatory layers, but for now, we\u0026rsquo;ve shown that CpG traceability in aging can be made transparent, interpretable, and biologically grounded.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe set out to do more than just predict biological age from methylation data. By pairing machine learning with explainable AI, we built a framework that actually shows which CpG sites matter and how they connect\u0026mdash;tracing them through enhancers to the genes and pathways that drive aging. The models performed well. Our stacked ensemble hit an R\u0026sup2; of 0.73 and a mean absolute error a bit over six years. But honestly, what stands out isn\u0026rsquo;t just the accuracy. It\u0026rsquo;s the clarity we get from interpretability tools\u0026mdash;SHAP values, Sankey diagrams, motif analyses\u0026mdash;all of them pointing toward the same biological story: transcriptional regulation and cell proliferation take center stage.\u003c/p\u003e \u003cp\u003eThe FOXO3 motif analysis took things up a notch. FOXO3 has a strong reputation in longevity research, and here\u0026rsquo;s the thing\u0026mdash;we found its binding spots stacked up right at CpG boundaries. That\u0026rsquo;s no accident. It backs up the idea that methylation drift and what transcription factors do are deeply linked as we age. Insights like this are where explainable AI shines. It doesn\u0026rsquo;t just spit out predictions; it uncovers the actual biological mechanisms, letting us watch how epigenetic changes drive the aging process.\u003c/p\u003e \u003cp\u003eOf course, this is just the beginning. We worked with whole blood, so there\u0026rsquo;s plenty left to test. Do these patterns hold up in other tissues? Larger datasets and combining this with chromatin accessibility or 3D genome structure could tell us much more. Still, this framework moves the field forward. We\u0026rsquo;re stepping out of the black box, heading toward epigenetic models we can actually interpret, and starting to understand the \u0026ldquo;how\u0026rdquo; of aging.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e Self funded research - \u0026nbsp;No external funding was received.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eConflicts of Interest:\u003c/strong\u003e The authors declare no conflicts of interest.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eEthics Approval:\u003c/strong\u003e Not applicable.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eData Availability:\u003c/strong\u003e Public dataset GSE40279 (NCBI GEO).\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eClinical Trial Registration:\u003c/strong\u003e This study does not involve a clinical trial and therefore trial registration details are not applicable.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eConsent to Publish declaration:\u003c/strong\u003e Not applicable.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eConsent to Participate declaration:\u003c/strong\u003e Not applicable.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eAuthor Contributions:\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eSRK conceived and designed the study; performed data preprocessing, survival modeling, and feature attribution analyses; developed the modular AI framework; prepared figures, tables, and visualizations; drafted and revised the manuscript.\u003c/p\u003e\n\u003cp\u003eHC provided supervision and guidance on study design and methodology, reviewed and refined the manuscript for scientific accuracy and clarity.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHorvath, S. (2013). DNA methylation age of human tissues and cell types. \u003cem\u003eGenome Biology\u003c/em\u003e, 14(10), R115. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/gb-2013-14-10-r115\u003c/span\u003e\u003cspan address=\"10.1186/gb-2013-14-10-r115\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHannum, G., Guinney, J., Zhao, L., Zhang, L., Hughes, G., Sadda, S., et al. (2013). Genome\u0026ndash;wide methylation profiles reveal quantitative views of human aging rates. \u003cem\u003eMolecular Cell\u003c/em\u003e, 49(2), 359\u0026ndash;367. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.molcel.2012.10.016\u003c/span\u003e\u003cspan address=\"10.1016/j.molcel.2012.10.016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg, S. M., \u0026amp; Lee, S. I. (2017). A unified approach to interpreting model predictions. \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e, 30, 4765\u0026ndash;4774.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoadmap Epigenomics Consortium, Kundaje, A., Meuleman, W., et al. (2015). Integrative analysis of 111 reference human epigenomes. \u003cem\u003eNature\u003c/em\u003e, 518, 317\u0026ndash;330. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/nature14248\u003c/span\u003e\u003cspan address=\"10.1038/nature14248\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaulagi, S. (2025). CpG Traceability and Pathway Mapping in Epigenetic Aging with Explainable AI. \u003cem\u003eResearch Square\u003c/em\u003e. Preprint. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.21203/rs.3.rs-8242582/v1\u003c/span\u003e\u003cspan address=\"10.21203/rs.3.rs-8242582/v1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKiselev, I. S., Baulina, N. M., \u0026amp; Favorova, O. O. (2025). Epigenetic clock: DNA methylation as a marker of biological age and age\u0026ndash;associated diseases. \u003cem\u003eBiochemistry (Moscow)\u003c/em\u003e, 90, S356\u0026ndash;S372. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1134/S000629792506004X\u003c/span\u003e\u003cspan address=\"10.1134/S000629792506004X\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiang, R., Tang, Q., Chen, J., \u0026amp; Zhu, L. (2025). Epigenetic clocks: Beyond biological age, using the past to predict the present and future. \u003cem\u003eAging and Disease\u003c/em\u003e, 16(6), 3520\u0026ndash;3545. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.14336/AD.2024.1495\u003c/span\u003e\u003cspan address=\"10.14336/AD.2024.1495\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChhatbar, K., Bird, A., \u0026amp; Sanguinetti, G. (2024). Unravelling epigenetic regulation of gene expression with explainable AI. \u003cem\u003eBioinformatics\u003c/em\u003e, 40(2), 123\u0026ndash;134. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/bioinformatics/btae045\u003c/span\u003e\u003cspan address=\"10.1093/bioinformatics/btae045\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou, Z., Hu, M., Salcedo, M., et al. (2023). XAI meets biology: A comprehensive review of explainable AI in bioinformatics. \u003cem\u003earXiv preprint\u003c/em\u003e arXiv:2312.06082.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDu, S., \u0026amp; Zheng, H. (2021). Role of FoxO transcription factors in aging and age\u0026ndash;related metabolic and neurodegenerative diseases. \u003cem\u003eCell \u0026amp; Bioscience\u003c/em\u003e, 11, 188. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s13578-021-00688-2\u003c/span\u003e\u003cspan address=\"10.1186/s13578-021-00688-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDonlon, T. A., Morris, B. J., Masaki, K. H., et al. (2022). FOXO3, a resilience gene: Impact on lifespan, healthspan, and deathspan. \u003cem\u003eThe Journals of Gerontology: Series A\u003c/em\u003e, 77(8), 1479\u0026ndash;1484. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/gerona/glac132\u003c/span\u003e\u003cspan address=\"10.1093/gerona/glac132\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao, Y., Geddes, T. A., Yang, J. Y. H., \u0026amp; Yang, P. (2020). Ensemble deep learning in bioinformatics. \u003cem\u003eNature Machine Intelligence\u003c/em\u003e, 2, 500\u0026ndash;508. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s42256-020-0217-y\u003c/span\u003e\u003cspan address=\"10.1038/s42256-020-0217-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElnahas, O., Ead, W. M., Qiu, Y., \u0026amp; Lu, J. (2025). Ensemble machine learning\u0026ndash;based annotation for scRNA\u0026ndash;seq data. \u003cem\u003eBMC Bioinformatics\u003c/em\u003e, 26, 166. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12859-025-0566-7\u003c/span\u003e\u003cspan address=\"10.1186/s12859-025-0566-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLevine, M. E., Lu, A. T., Quach, A., et al. (2018). An epigenetic biomarker of aging for lifespan and healthspan. \u003cem\u003eAging\u003c/em\u003e, 10(4), 573\u0026ndash;591. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.18632/aging.101414\u003c/span\u003e\u003cspan address=\"10.18632/aging.101414\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eField, A. E., Robertson, N. A., Wang, T., et al. (2018). DNA methylation clocks in aging: Categories, causes, and consequences. \u003cem\u003eMolecular Cell\u003c/em\u003e, 71(6), 882\u0026ndash;895. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.molcel.2018.08.008\u003c/span\u003e\u003cspan address=\"10.1016/j.molcel.2018.08.008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (doi.orgin Bing)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"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":"Epigenetic aging, CpG methylation, explainable AI, enhancer logic, pathway mapping, machine learning","lastPublishedDoi":"10.21203/rs.3.rs-8877884/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8877884/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDNA methylation at CpG sites stands out as one of the most reliable markers for aging we have. Sure, machine learning models can predict biological age with decent accuracy\u0026mdash;but the real challenge is figuring out what those predictions mean. Most models work like black boxes; they spit out an answer, but give you little sense of how specific CpGs actually influence gene regulation or downstream pathways. That\u0026rsquo;s the gap we wanted to close.\u003c/p\u003e \u003cp\u003eIn this study, we combined classic regression models with explainable AI methods to make CpG traceability clear and direct. We started with whole blood methylation data from 656 people (GSE40279) and used feature selection to zero in on the most informative CpGs. Then we trained predictive models using XGBoost, LightGBM, and a few ensemble tricks, testing their accuracy with cross-validation. The top stacked ensemble reached an R\u0026sup2; of 0.73 and a mean absolute error of 6.1 years\u0026mdash;not just solid numbers, but a strong foundation for interpretation.\u003c/p\u003e \u003cp\u003eBut we didn\u0026rsquo;t stop with prediction. We traced each CpG through enhancer annotations to its target genes, then mapped those to biological processes. Sankey diagrams showed the same story, again and again: pathways linked to transcriptional regulation and cell proliferation, both major players in the aging process, kept coming up enriched.\u003c/p\u003e \u003cp\u003eThis approach shows that explainable AI can do more than just predict\u0026mdash;it can actually connect methylation markers to meaningful biological functions. By linking CpGs to enhancers, genes, and Gene Ontology terms, we get a transparent look at how epigenetic drift might drive aging at the molecular level. In short, we\u0026rsquo;ve set the stage for interpretable epigenetic modeling, with the next steps geared toward validating these findings across different tissues.\u003c/p\u003e","manuscriptTitle":"CpG Traceability and Pathway Mapping in Epigenetic Aging with Explainable AI","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-17 10:47:27","doi":"10.21203/rs.3.rs-8877884/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":"0c8546b8-a362-4563-ab40-48d177ae3744","owner":[],"postedDate":"February 17th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-13T19:09:46+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-17 10:47:27","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8877884","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8877884","identity":"rs-8877884","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}