Guidelines on optimizing DNA methylation reference panels for cell-type deconvolution

Xiaolong Guo; Andrew E Teschendorff

doi:10.1038/s42003-026-09745-1

. 2026 Feb 19;9:454. doi: 10.1038/s42003-026-09745-1

Guidelines on optimizing DNA methylation reference panels for cell-type deconvolution

Xiaolong Guo ^1,^✉, Andrew E Teschendorff ^1,^✉

PMCID: PMC13031392 PMID: 41708983

Abstract

Accurate cell-type deconvolution is critical for correct interpretation of Epigenome-Wide Association Studies. For all cell-type deconvolution tasks, it is necessary to estimate underlying cell-type fractions in a sample, which is usually accomplished using a DNA methylation reference panel built from sorted or single-cell DNAm data. Two competing approaches have emerged to build such reference panels, one which uses machine-learning, and another based on optimizing effect size and cell-type specificity. Here we demonstrate that the latter approach is preferable, because, owing to the relatively small number of sorted samples used in building panels, standard machine learning does not optimize effect size and cell-type specificity, causing the model to overfit and underperform when tested in independent data. Furthermore, adult blood panels built from cell-type specific hypomethylated markers improves estimation of cell-type fractions when compared to panels built from hypermethylated ones. These insights provide important guidelines for optimizing the construction of future DNAm reference panels. To aid this task, we have added a function for building an optimized DNAm reference panel to our EpiDISH R-package.

Subject terms: Microarrays, Epigenomics

A comparative analysis of methods for building DNA methylation reference panels for cell-type deconvolution highlight important guidelines for optimizing their construction.

Introduction

The goal of epigenome wide association studies (EWAS) is to identify DNA methylation (DNAm) changes linked to traits, disease risk factors or disease itself^1,2. Because DNAm is highly cell-type-specific^2,3, measuring it in bulk tissue, as is most often done in DNA methylome studies, can cause statistical confounding⁴. This is because DNAm changes in bulk-tissue can arise from shifts in cell-type composition and not because of underlying DNAm changes in individual cell-types^3–6. This challenge can be addressed by estimating the actual cell-type fractions in the samples, which could be achieved either experimentally (e.g., blood cell-counts) or statistically^4,7–9, the latter being more feasible and cheaper. Statistically, the procedure involves using a cell-type deconvolution algorithm that first estimates the underlying cell-type proportions in a bulk-tissue sample, subsequently using these estimated fractions as covariates in linear models to identify differentially methylated cytosines (DMCs) that are not driven by changes in cell-type composition. The estimation of these cell-type fractions requires the construction of a DNAm reference panel, which contains representative DNAm profiles for all main underlying cell-types in the tissue of interest, and which is normally built by processing genome-wide DNAm profiles of sorted cell-types^9–12 or single-cells^13,14. Although studies have shown that the inference of cell-type fractions from such a DNAm reference panel is a relatively robust procedure¹⁵, it is nevertheless important to ensure that DNAm reference panels are built in the most optimal way. This is particularly important in view of the growing need to build such panels for a wider range of different tissue-types at increased cellular resolution^14,16–18. In this regard, a key challenge is posed by the relatively small number of sorted samples available for each cell-type (typically less than 10)^10,11, as generating larger numbers of sorted samples (say 25) for each of 10–20 cell types would require the generation of several hundred samples. Similar limitations apply to single-cell DNAm (scDNAm) data, which is plagued by high sparsity and not yet widely available, with existing scDNAm datasets only profiling samples from a small number of independent donors^13,19,20. Thus, it is important to set up guidelines for the optimized construction of DNAm reference panels from what are typically small DNAm reference datasets.

Ideally, a DNAm reference panel should comprise DNAm markers that are (i) highly discriminatory between different cell types, (ii) stable within each cell type, and (iii) robust to variable factors such as age, race or genotype of the donors. As the deconvolution process essentially projects the methylation information of a bulk-tissue sample onto a feature space defined by the reference markers, if this space is poorly defined (i.e., markers lack true cell-type specificity or discriminatory power), the accuracy of the projection and estimation of cell-type fractions will be compromised. Traditionally, the selection of cell-type-specific markers for a reference panel is based on identifying methylation sites with significant statistical differences between the various cell types. The first approach, due to Houseman et al.⁹, independently fitted one-way analysis of variance (ANOVA) models to each CpG to test for differences in mean methylation levels across K purified cell types, selecting CpGs with the largest F-statistics to constitute the reference library/panel. However, the F-statistic prioritizes sites with large overall DNAm differences across all cell types, which can lead to the over-selection of CpGs that strongly differentiate one or a few cell types while neglecting others, resulting in imbalanced discriminatory power within the panel. To address the global limitation of the ANOVA method, subsequent studies shifted to using t-statistics (now called “Legacy method”, implemented in the Bioconductor package minfi using the function pickCompProbes) to select CpGs that are significantly hyper- or hypo-methylated in a specific cell type compared to all others²¹. Subsequent optimized feature selection frameworks have largely built upon the library generated by this “Legacy method” as their initial candidate pool of marker CpGs. Examples include iterative refinement strategies like IDOL and RESET^22,23, as well as machine learning approaches such as Elastic Net and Random Forest²⁴. However, marker CpGs selected purely on statistical significance ranking and machine learning may potentially overlook the magnitude of the effect size, as a small effect can still be very informative if standard deviations are small. We reasoned that such small effect size CpGs are more likely to represent false positives (i.e., features that do not validate in independent data), because based on the available evidence from large WGBS atlases²⁵, true cell-type specific marker CpGs have large or reasonably large effect sizes. Moreover, selecting candidate marker CpGs using t-statistics does not guarantee maximum specificity for a given cell-type, especially if the number of sorted samples per cell-type varies significantly between cell-types.

To overcome these limitations, we recently introduced a novel marker selection strategy based on a joint t-statistic and “gap specificity score” (GSS)^26,27. This GSS metric is designed to directly prioritize the magnitude of the methylation difference—the ‘gap’—between a target cell type and all other cell types, thus maximizing effect size and specificity, while also ensuring that markers are selected based on statistical significance and hence a high within-cell-type stability. Here, we directly compare this GSS-based approach to the competing method described above based on using IDOL, as well as other machine-learning methods, including Elastic Net and Random Forests. We demonstrate that the GSS-based approach builds a more accurate and robust deconvolution reference panel. We further validate this approach by demonstrating its enhanced ability to detect known age-related changes in immune cells and to yield more biologically meaningful or significant findings in epigenome-wide association studies (EWAS) of schizophrenia and type 1 diabetes. As such, the insights gained from our comparative analysis may serve as useful guidelines for the construction of future DNAm reference panels.

Results

Contamination of low-effect-size DMCs

The construction of a DNAm reference panel according to the IDOL algorithm begins with a feature selection strategy followed by an optimization designed to select an optimal subset. Here, we demonstrate that the use of t-statistics in the feature selection step can lead to contamination of undesirable low-effect size DMCs that are retained after optimization. To show this, we utilized the sorted immune-cell library generated by Salas et al., which encompasses EPICv1 DNAm profiles of 12 sorted immune cell types¹⁰ (Methods). Initially, for each CpG site, a two-sample t-test was performed to compare the mean DNAm of a given cell type against the remaining 11 cell types. For each cell type, CpGs exhibiting a statistically significant difference, defined by a false discovery rate (FDR) less than 0.05, were identified as cell type-specific DMCs. These DMCs were then ranked, followed by the selection of N marker CpGs per cell type, comprising the N/2 DMCs with the largest positive t-statistics and the N/2 DMCs with the largest negative t-statistics. Using N = 100 and following application of IDOL, we observed that there were many low effect size DMCs among the selected markers (Fig. 1a), thus highlighting a potential problem with using t-statistics to build DNAm reference panels.

A novel gap-specificity-score (GSS) to improve DNAm reference panels

To address the above problem, we devised a novel approach to rank DMCs by a gap-specificity score (GSS) (Fig. 1b). This score quantifies the methylation ‘gap’: for hypermethylated markers, it is the target cell type’s minimum methylation value minus the maximum methylation value taken over all other samples representing the other cell types; for hypomethylated markers, it is defined by the minimum methylation value over all other cell types minus the target cell type’s maximum methylation value (Fig. 1b). Thus, in both cases, the GSS should ideally be positive and as large as possible, as this ensures specificity of the marker to the one given cell-type. GSS is designed to identify robust cell type-specific markers by prioritizing this ‘gap’ (reflecting effect size) alongside the marker’s cross-sample stability within the target cell type.

In order to objectively compare GSS to the previous t-statistic-based (‘legacy’) approach, we used the same cell-type-specific DMCs identified earlier using t-statistics, but now ranking hypermethylated and hypomethylated DMCs separately by the GSS and subsequently selecting the top N/2 from each category (Fig. 1b). The effect size of each DMC was defined as the absolute difference between the mean beta value of its target cell type and the mean beta value of the remaining 11 cell types. As required, top DMCs ranked by the GSS displayed significantly larger effect sizes than those ranked purely by t-statistics, and this was stable for variable choices of N (Fig. 1c, Supplementary Data 1). Fixing N = 100, we next compared the effect sizes from the two methods for each cell type separately. We observed that ranking based only on t-statistics often included DMCs with notably low effect sizes, an issue especially prominent for lymphocyte subtypes (Fig. 1d, Supplementary Data 1).

To further demonstrate the value of the GSS, we compared it to standard effect size measures, including Cohen’s F and Cohen’s D²⁸. Using Cohen’s F, which is a global effect size measure across all cell-types, top-ranked markers could only distinguish between the myeloid and lymphoid lineages, as the variation between these lineages is much larger than the variation within the myeloid and lymphoid lineages, hence failing to discriminate more similar cell-types from each other (e.g., naïve vs memory CD4+ T-cells) (SI Fig. S1a). Cohen’s D, which estimates effect sizes by comparing one cell-type to all others, also proved suboptimal as it fails to guarantee specificity, i.e., it frequently selected probes with high effect size but negative GSS (SI Fig. S1b, Supplementary Data 1). In contrast, ranking features by positive GSS values ensures specificity (SI Fig. S1c, Supplementary Data 1). A concrete example of a marker for memory CD4+ T-cells selected using Cohen’s D illustrates how its GSS is negative, since T-regulatory cells display even lower DNAm levels than memory CD4+ T-cells (SI Fig. S1d).

Low-effect-size DMCs can negatively impact cell-type deconvolution accuracy

To confirm that low effect size DMCs can negatively impact downstream cell-type deconvolution, we built DNAm reference matrices using the two different approaches (Fig. 1b), subsequently applying them to deconvolve twelve experimental mixtures, each being a mixture of 12 immune-cell-types with known cell type proportions¹⁰ (Supplementary Data 2). We used the robust partial correlations (RPC) method within EpiDISH^29,30 to estimate the underlying cell-type fractions, which were then compared to the true fractions by calculating the Pearson correlation coefficient (PCC) and root mean square error (RMSE) (Methods, Fig. 2a, Supplementary Data 1). While both ranking strategies yielded similar PCCs, the GSS-strategy consistently achieved a lower global RMSEs (Fig. 2a, b, Supplementary Data 1). Furthermore, this improvement was not dependent on the optimization algorithm, as it consistently outperformed the t-statistic-ranked panel when fractions were evaluated using two other optimization methods (Constrained Projection⁹ and Support Vector Regression³¹) (SI Fig. S2, Methods, Supplementary Data 1).

Notably, the RMSE for memory CD4 T cells was markedly higher than that of other cell types (Fig. 2a, Supplementary Data 1), consistent with their DMCs exhibiting the smallest effect sizes relative to other cell types (Fig. 1d, Supplementary Data 1). To further investigate the variable performance across cell types, we calculated the mean t-statistic (representing average significance) and the mean gap specificity score (representing average effect size) for the top 100 marker DMCs selected (by t-statistics) for each cell type. This revealed that deconvolution performance (i.e., a low RMSE, high PCC) of each cell-type was strongly and significantly correlated with the cell-type’s mean gap specificity score, while the correlation with the mean t-statistic used to select them was weaker and not statistically significant (Fig. 2c, Supplementary Data 1). This suggests that the accuracy of a deconvolution reference panel is determined by the effect size of inter-cell-type differences (i.e., gap specificity score) rather than their statistical significance (i.e., t-statistic). To further confirm this, we devised a sliding window approach, where we compared deconvolution accuracies for a total of 46 windows of size 50 DMCs, with DMCs ranked either by t-statistic or GSS (Methods). References ranked by GSS demonstrated a strong performance dependence on marker rank for both PCC and RMSE, while references ranked by t-statistics exhibited only a weak correlation for PCC and showed no significant association with RMSE (Fig. 2d, Supplementary Data 1).

Hypomethylated DMCs improve cell-type deconvolution accuracy

While the legacy method selected both hyper- and hypomethylated DMCs impartially, hypomethylated DMCs are more likely to be true cell-type-specific markers, since markers that are uniquely hypomethylated for a given cell-type are much more frequent than uniquely hypermethylated ones²⁵. To further refine the GSS-based strategy, we thus asked if a reference constructed solely from top-ranked hypomethylated DMCs could outperform a reference combining both hyper- and hypomethylated DMCs, and a reference constructed using only top-ranked hypermethylated DMCs. Using a range of N values, the three distinct reference matrices were then compared in terms of their accuracy to deconvolve the twelve experimental 12-cell-type mixtures (Fig. 3a, Supplementary Data 1). The reference constructed solely from hypermethylated DMCs yielded the lowest PCC values across all N values and the highest RMSE values when N > 100, compared to the other two reference types, indicating that only a small proportion of hypermethylated DMCs are effective for cell-type discrimination (Fig. 3b, Supplementary Data 1). Conversely, using only top hypomethylated DMCs at N = 600 emerged as the best-performing reference, achieving the highest PCC and lowest RMSE (Fig. 3b, Supplementary Data 1). Compared to the combined reference type, this ‘hypomethylated-only’ strategy resulted in a modest but clear reduction in RMSE, and notably improved deconvolution accuracy for memory CD4 T cells, which had previously exhibited poorer deconvolution accuracy (Fig. 3a, Supplementary Data 1). While the ‘hypomethylated-only’ strategy performed best, the underlying combined reference type also achieved excellent performance (PCC ~ 1, RMSE ~ 0.01), demonstrating how the inclusion of many accurate markers can offset the presence of low effect size markers (Fig. 3b, Supplementary Data 1).

Fig. 3 — a Line plots display the Pearson correlation coefficient (PCC) and root mean square error (RMSE) of the true and estimated proportions for each cell type of 12 experimental mixtures. The “Combined” strategy refers to ranking solely based on the gap specificity score, without considering the direction of DMCs, to select the top DMCs for constructing the reference. The “All hypermethylated DMCs” and “All hypomethylated DMCs” strategies refer to ranking by gap specificity score and then using only hypermethylated DMCs or only hypomethylated DMCs, respectively, to build the reference. Lines are colored by cell type. b Boxplots summarize the distribution of PCC and RMSE for the 12 cell types at each library size, as displayed in (a). Individual data points overlaid on the plots represent the n = 12 cell types.

ML-optimization for marker selection can overfit to small training data

Next, we sought to determine if the IDOL algorithm, which is often used to construct and optimize a DNAm reference panel^10–12,23, could further improve the performance of a reference matrix built exclusively from top hypomethylated DMCs ranked by the GSS. Since both the legacy method and the GSS-strategy reached their best performance at N = 600 (Fig. 2b, Fig. 3b, Supplementary Data 1), we applied IDOL to the two reference matrices defined at this N-value, labeling them as ‘T-Both’ and ‘GSS-Hypo’, respectively.

We note that IDOL also requires as input a training set comprising experimental mixtures generated from sorted immune cell types. Using the leave-one-out method, the probability of each CpG being selected is then updated based on whether the average R² and average RMSE of the estimated proportions for all cell types improve, continuing this process until convergence²³. Thus, we applied the IDOL algorithm to optimize the T-Both and GSS-Hypo references, adopting the same training set and libSize = 1200 parameter as used by Salas et al.¹⁰ (Methods), resulting in refined T-Both-IDOL and GSS-IDOL DNAm reference panels. Surprisingly, comparing the effect sizes of markers before and after IDOL optimization showed that IDOL has a tendency to select DMCs with lower effect sizes (SI Fig. S3a, b, Supplementary Data 1), irrespective of whether we started out using the T-Both or GSS-Hypo reference panel. To explore this further, we assessed the performance of the four reference matrices in the experimental mixtures from the training set as well as those of an independent validation set¹⁰ (Methods, see Supplementary Data 2 for known cell-type proportions). Interestingly, the IDOL-optimized references demonstrated comparable PCCs in the training set but somewhat higher RMSEs (SI Fig. S3c, Supplementary Data 1). Moreover, IDOL-optimized references performed worse in terms of both PCC and RMSE in the independent validation set (SI Fig. S3d, Supplementary Data 1). Overall, ‘GSS-Hypo’ retained its superior deconvolution performance, even when compared to its IDOL-optimized version, indicating that IDOL may be overfitting slightly due to the relatively small number of experimental mixtures used in training.

To demonstrate that this result is robust to different library sizes and generalizes to other machine-learning algorithms, we performed a similar analysis, but now using Elastic Net (EN)³² and Random Forests (RF)³³ instead of IDOL (Methods). As before, effect sizes of selected features using EN or RF were substantially smaller compared to GSS-Hypo (Fig. 4a, b, Supplementary Data 1), translating into a lower accuracy for the cell-type deconvolution on experimental mixtures (Fig. 4c, d, Supplementary Data 1), as well as on an independent DNAm dataset with matched FACS counts for 4 cell-types (Methods, Fig. 4e, f, Supplementary Data 1 and 2). Because the improvement observed with GSS-Hypo could be driven by the selection of hypomethylated markers, we also benchmarked GSS-Hypo against an analogous approach that uses hypomethylated markers selected using t-statistics (T-hypo) (Methods). GSS-hypo generally outperformed T-hypo across the experimental mixture datasets and the dataset with matched FACS counts (Fig. 4c–f, Supplementary Data 1). Collectively, these results support the view that the improvements seen are not only due to the selection of hypomethylated markers but also due to improved specificity, and that, given the relatively small number of sorted samples, any machine learning method is likely to overfit to random variation.

Fig. 4 — a Violin plots comparing the distribution of effect sizes for DMCs derived from different reference matrices. The left plot compares the ‘T-Both’ reference (N = 100, containing 50 hyper- and 50 hypomethylated DMCs) against the ‘IDOL’ reference (libsize = 1200, trained from a reference library (N = 1000) sorted by t-statistics). The right plot compares the ‘GSS-Hypo’ reference (N = 100, containing only hypomethylated DMCs) against the same ‘IDOL’ reference. Significance levels determined by a one-tailed Wilcoxon rank-sum test are indicated as: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 1e−4. Exact P-values are provided in Supplementary Data 1. b as a, but comparing DMCs from GSS-Hypo with DMCs selected by ELN (Elastic Net, left plot) and DMCs selected by RF (Random Forest, right plot). Exact P-values are provided in Supplementary Data 1. c Line plots illustrate the PCC and RMSE calculated between true and estimated proportions for each cell type across the 12 experimental mixtures in the training set of IDOL. The corresponding box plots summarize the distribution of PCC and RMSE across all 12 cell types. d as c, but for 12 independent experimental mixtures. e Boxplot shows the FACS-measured proportions of T cell naive and memory subtypes in a peripheral blood dataset (n = 76). f Line plots illustrate the PCC and RMSE calculated between FACS-measured and estimated proportions for the 76 peripheral blood samples. The corresponding dot plots summarize the performance of each method across the n = 4 T cell subtypes. Diamonds indicate the mean value, error bars represent the standard deviation (SD), and individual small points represent the n = 4 cell types.

GSS-reference detects the known age-related increase in NK cells

The immune cell composition of peripheral blood undergoes significant changes with age, characterized by shifts in the proportions of various subgroups, including the ratios of myeloid to lymphoid cells and memory to naive T cells^34–36. To further compare the GSS-Hypo and t-statistic-based reference panels, we examined their ability to capture known age-related changes in immune-cell type composition. Performing a meta-analysis across 15 adult blood cohorts, while adjusting for sex, we observed that for all 3 reference panels (i.e., GSS-hypo, T-hypo and T-both), naive B-cell, naïve CD8+ T-cell and naïve CD4+ T-cell fractions decreased with age, whilst corresponding memory fractions increased (Fig. 5a, Methods, Supplementary Data 3). These results align with findings from previous flow cytometric studies, particularly the pronounced decline of naive CD8⁺ T cells with age^34,35. Less trivial is the positive association of NK fractions with age³⁰. Only the GSS-Hypo and T-Hypo panels identified a significant age-associated increase in NK-cells (Fig. 5a, b), consistent with previous reports^35,37. Thus, in this context, the GSS-Hypo panel is more accurate in the estimation of cell-type fractions, but only because of its exclusivity for hypomethylated markers.

Fig. 5 — a Barplots with 95% confidence intervals display the association of the 12 immune cell-type fractions estimated by three reference panels with chronological age, under a random effect model over 15 whole blood cohorts. b Forest plots provide a more detailed illustration of the association of NK fractions with chronological age estimated by the three references in each cohort. The P-value of a random effect meta-analysis model is given. Error bars represent 95% confidence intervals. Heterogeneity I² indices and P-values are also given on top of each panel.

GSS-reference panel improves interpretability of EWAS findings

Small differences in estimated cell-type fractions could also impact EWAS findings²⁹. To test this, we first analyzed two independent schizophrenia (SZ) case-control EWAS whole blood datasets from Hannon et al.^38,39, aiming to determine which reference matrix yielded more consistent disease-associated signals across cohorts (Methods). In both the Hannon1 (discovery) and Hannon2 (validation) cohorts, we applied a multivariate linear regression model, adjusting for covariates (e.g., age and sex) and cell-type fractions estimated by one of the 3 reference panels. In the discovery cohort, we obtained 328, 352, and 351 SZ-DMCs (FDR < 0.05) using T-Both, T-Hypo, and GSS-hypo references, respectively. This result was robust to additional adjustment for chip and position, as well as top principal methylation components (SI Fig. S4, Supplementary Data 1). In Hannon2, we observed a strong batch effect, which we successfully adjusted for using Combat⁴⁰ (SI Fig. S5, Supplementary Data 1). A total of 260 (T-Both), 274 (T-Hypo) and 271 (GSS-hypo), replicated in the validation cohort (Fig. 6a). Because there was substantial overlap between the SZ-DMCs from each of the 3 panels (Fig. 6b, Supplementary Data 1), any downstream differences in GSEA would be relatively minor. Indeed, in all 3 cases, we observed enrichment for the steroid biosynthesis pathway, which has been implicated in SZ (SI Fig. S6, Methods, Supplementary Data 1)^41,42. Thus, we focused on those SZ-DMCs exclusively replicated by each of the 3 panels, i.e., a total of 16 (T-Both), 18 (T-Hypo) and 20 (GSS-hypo) SZ-DMCs, respectively (Fig. 6b, Supplementary Data 1). Interestingly, in the case of GSS-hypo, the 20 SZ-DMCs were enriched for inflammation-related pathways, whilst this enrichment was absent or less evident for T-Both or T-Hypo (Fig. 6c, Supplementary Data 4, Supplementary Data 1). Given that inflammation is well-known to be associated with SZ⁴³, this suggests that the 20 SZ-DMCs identified using GSS-hypo are likely to be true positives. Supporting their biological significance, among the 20 SZ-DMCs, were CpGs mapping to the chemokine CCL18, consistent with reports of elevated CCL18 chemokine levels in the serum of SZ patients⁴⁴, and to the chemokine receptor gene CXCR5, a gene implicated in neuroinflammation and schizophrenia^45,46 (Fig. 6d, Supplementary Data 4, Supplementary Data 1). Another CpG mapped to the KDM2B gene (Fig. 6d, Supplementary Data 4, Supplementary Data 1), which is known to promote IL-6 production and inflammatory responses through chromatin remodeling⁴⁷. Notably, mutations in KDM2B have also been directly linked to schizophrenia and are associated with neurodevelopmental disorders^48,49. Thus, overall, these results support the view that the GSS-hypo panel improves inference.

Fig. 6 — a Scatter plots show the t-statistics of each CpG in EWAS conducted in two independent schizophrenia (SZ) case–control studies using fractions estimated with T-Both reference (left), T-Hypo reference (middle) and GSS-Hypo reference (right). The horizontal gray dashed line represents the threshold where P < 0.05, while the vertical gray dashed line represents the threshold where FDR < 0.05. Red dots indicate CpGs significant in both studies. Pearson correlation coefficients and corresponding P values are also provided. b Venn diagram shows the number of intersecting and unique SZ-DMCs identified in (a). c Barplots list inflammation-related KEGG pathways, ranked by statistical significance, with the bars depicting the fraction of inflammation-related KEGG pathway genes that were found enriched among SZ-DMCs. Exact P-values are provided in Supplementary Data 1. d Violin plots show three representative schizophrenia-associated DMCs detected only by GSS-Hypo from (b). P-values from a two-tailed Wilcoxon rank sum test are provided. e Barplots list the top-ranked enriched KEGG pathway genes, with bars indicating the fraction of KEGG pathway genes that were found enriched among T1D-DMCs. The number of T1D-DMCs in the top two cases has been matched to the number obtained using the GSS-Hypo reference. *FDR < 0.05.

As a second validation, we analyzed a longitudinal DNAm dataset (Illumina 450k + EPICv1) representing a prospective study of type 1 diabetes (T1D)⁵⁰ (Methods). This dataset, which has also been previously analyzed by us²⁷, comprises 395 blood samples collected from infants and children at multiple timepoints during the progression of islet autoimmunity (IA) prior to T1D diagnosis. Following a similar analytical approach as the SZ study, we used linear models to identify T1D-associated DMCs, adjusting for confounders including age, sex, timepoint, technical factors (chip and position), and the 12 immune cell fractions derived from either the T-Both, T-Hypo or GSS-Hypo reference. Analysis using T-Both, T-Hypo and GSS-hypo identified 11,264, 10,012 DMCs, and 8020 T1D-DMCs, respectively (FDR < 0.05). To avoid bias and ensure a fair downstream comparison, we restricted the T-Both and T-Hypo lists to their respective top 8020 DMCs (ranked by P-value), subsequently comparing their enrichment for KEGG pathways using missMethyl⁵¹. This revealed that while all three lists showed highly significant enrichment for the ‘Type I diabetes mellitus’ pathway, the DMC list derived from the GSS-Hypo reference was more enriched for T1D-related genes (Fig. 6e). Additionally, as T1D is an autoimmune disease, the ‘Antigen processing and presentation’ pathway should also be enriched: notably, GSS-Hypo also displayed the most significant enrichment for this pathway and contained the most genes (Fig. 6e). Overall, these data indicate that using the GSS-Hypo reference, we can obtain improvements in the enrichment of biologically relevant pathways.

Discussion

Our study highlights the importance of a long-standing question in omic data science, namely, when selecting features, do we rely on statistics only, or should we also take the magnitude of the effect into account? Generally speaking, using statistics and P-values to select features is of paramount importance. However, once features have been selected based on P-values, a less trivial question is whether significant features should be ranked by the statistic (or P-value) or whether they should be re-ranked according to effect size. The answer to this question largely depends on the biological problem and context. As demonstrated here, in the context of building DNAm reference panels for cell-type deconvolution, where one has to select features that are highly discriminatory of different cell-types, relying only on statistics and P-values to select and rank features is a suboptimal procedure, as it can inadvertently place low-effect size markers at the top of the ranked list, which, if subsequently not removed, can negatively impact downstream cell-type deconvolution. Indeed, our results clearly suggest that these low-effect-size markers are either false positives or lower-quality markers that arise because of the typically low number of sorted cell samples. The number of independent sorted samples is typically small because of the high cost and difficulty of generating such samples for a relatively large number of donors and cell types. Our GSS-strategy, which re-ranks features previously selected by statistics and P-values by their cell-type specific effect size, can efficiently remove these false positive or low-quality markers, improving the quality of the downstream cell-type deconvolution and statistical inference, as demonstrated here, not only in terms of the accuracy of the cell-type deconvolution itself, but also by improving the interpretability of EWAS findings.

Interestingly, we found that the IDOL algorithm may not optimize DNAm reference panels. The reduced performance of the IDOL-derived reference panels was true irrespective of whether we re-ranked significant DMCs by the GSS or not, and that performance dropped further when tested in independent datasets. Thus, a likely explanation as to why IDOL may reduce performance is that it overfits to the experimental mixtures used in training due to their small sample size. In support of this, a reduced performance was also seen using other machine learning algorithms such as Elastic Net or Random Forests. Of note, this overfitting happens despite leave-one-out strategies, because the number of experimental mixtures (n = 12) in the training set is too small. Thus, we envisage that IDOL would be more valuable if it could be trained on a much larger set of experimental mixtures.

Another important insight from this work is that building DNAm reference panels using exclusively cell-type-specific hypomethylated DMCs may further improve inference. This is consistent with the observation that the most cell-type-specific CpGs are those that are unmethylated in the given cell type, as indeed cell-type-specific hypermethylated CpGs are relatively infrequent²⁵. Thus, including hypermethylated markers may inadvertently increase the number of low-effect-size markers, compromising downstream inference. However, we caution that this likely applies only to adult tissue-types and terminally differentiated cell-types^14,25, because for other tissue-types like cord-blood, the presence of immature and progenitor cell-types may require hypermethylated markers to improve discrimination, as demonstrated recently by us²⁷. Indeed, DNAm-atlases like that of Loyfer et al. are focused on adult tissue and cell-types²⁵, which may lead to the misleading impression that most cell-type specific markers are always unmethylated in the given cell-type. However, in early development, where DNAm undergoes substantial remodeling, comparing immature and progenitor/stem cell-types to each other, the amount of cell-type specific hypo and hypermethylation is much more comparable.

One potential caveat to our GSS strategy is that we do not provide a concrete recommendation as to the number of markers that should enter a DNAm reference panel. This, however, is a complex question, as we now explain. There are 4 important criteria that will determine what a reasonable or optimal number of markers is. First, it is desirable to have a balanced number of markers for each cell type, as this ensures that the inference is not biased to the cell types with more markers. When building a panel, it is a priori not known if this is even possible, and may require using different GSS thresholds per cell-type. Second, the number of markers per cell-type should be large enough so that one can still reliably estimate fractions when, say, a proportion of the markers fail QC in independent datasets. Thus, redundancy in a panel is desirable as it confers robustness. On the other hand, too many markers will eventually also compromise their quality as measured by the GSS. Third, it is important to maximize the fraction of high GSS markers entering the panel (say, >70%), because this circumvents the need for further optimization using machine learning. Indeed, the purpose of any subsequent optimization, say via cross-validation, would be to remove false positives or low-quality markers, yet if the original selection of markers is characterized by high GSS-values (say >0.5), then these are extremely unlikely to be false positives, in which case any optimization may only be counterproductive, specially if the training or optimization is performed on small datasets. Although potential confounders like age, smoking, genotype or ethnicity generally only account for much smaller effect sizes compared to the DNAm differences between cell-types^27,52, their effect may be non-negligible in the context of cross-validations performed on small datasets. To make this point clear, a method like IDOL would be advantageous only if the original marker selection contains a significant number of false positives driven by such confounders, yet as shown here, IDOL would also require a larger training set for this advantage to become apparent. Fourth, the number of markers per cell-type is also a function of the number of cell-types in the panel. The more cell-types we have, the more markers we may need, since distinguishing more similar cell-types from each other (say classical from non-classical monocytes) may require a larger number of lower GSS features. In this regard, it is worth pointing out that even the current highest resolution panels for blood, encompassing 12 immune cell-types, fall short, since the number of biologically or clinically relevant immune cell-types is thought to be around 50 or more⁵³. Despite this clear limitation, performing an adjustment at the current resolution of 12 immune cell-types is better than no adjustment at all⁵⁴. In summary, the number of markers per cell-type is a complex function of many factors, and not necessarily amenable to an optimization procedure, especially if the number of training set samples is small. However, based on our experience building and successfully validating DNAm reference panels^{7,15,27,29,30}, we advise on the order of around 50 GSS > 0.3 markers per cell-type, so ~600 markers for a 12-cell-type panel.

Although our GSS-strategy led to improvements in statistical inference in the context of 3 different EWAS analyses (age, SZ, and T1D), we acknowledge that improvements in the case of SZ and T1D were marginal. That improvements are marginal is entirely expected, given that the cell-type deconvolution estimation procedure is relatively robust, as it can tolerate up to 30% of errors in the deconvolution panel¹⁵. Nevertheless, as future efforts will need to generate DNAm reference matrices for larger numbers of cell-types¹⁸ using, say, sorted or single-cell DNAm data^14,19,20, this will require accurate deconvolution for an increasing number of cell-types, which will place more stringent demands on the reliability and quality of a DNAm reference panel. Thus, the insights gained here can provide guidelines to help build future and more accurate DNAm reference panels. To aid the community, we now provide within our EpiDISH R-package²⁹, an R-function to build DNAm reference panels from sorted or single-cell DNAm data using our GSS-strategy.

In summary, sole reliance on statistics and P-values to select features for building DNAm reference panels can lead to suboptimal cell-type deconvolution, even if followed up with a machine-learning optimization strategy. This is because of the inclusion of low effect size markers, as well as overfitting to the inevitably small datasets comprising independent sorted or single-cell DNAm data. As shown here, the GSS-strategy, which selects a reasonable number of markers with high specificity and large effect size, optimizes cell-type deconvolution tasks and will prove valuable for future efforts to build higher-resolution deconvolution panels.

Methods

Statistics and reproducibility

Many of the analyses in this study involve computation of root mean square errors (RMSE) and Pearson correlation coefficients (PCC) or R-values over mixtures, either experimentally generated ones or in-silico, in both cases, the mixing fractions are known. Robustness was addressed by running the deconvolution for different values of parameters. In the applications to EWAS data, we performed a random effects meta-analysis over 15 independent whole blood cohorts, a statistical procedure that directly addresses reproducibility. In the context of identifying SZ-DMCs, reproducibility was assessed by using two independent cohorts.

The DNAm reference library for panel construction

Salas et al. (GSE167998)¹⁰: This dataset comprises EPICv1 profiles of 12 sorted immune-cell subsets, totaling 56 sorted samples: 6 basophils (Baso), 6 memory B-cells (Bmem), 4 naive B-cells (Bnv), 4 memory CD4+ T-cells (CD4Tmem), 5 naive CD4+ T-cells (CD4Tnv), 4 memory CD8+ T-cells (CD8Tmem), 5 naive CD8+ T-cells (CD8Tnv), 4 eosinophils (Eos), 5 monocytes (Mono), 6 neutrophils (Neu), 4 natural killer (NK) cells, and 3 regulatory T-cells (Treg). This dataset had previously been processed by us in Luo et al.³⁰.

Experimental mixture DNAm and matched FACS datasets

Salas et al. (GSE167998)¹⁰: This EPICv1 study also generated 12 experimental mixture samples composed of the aforementioned 12 cell types, with known mixing proportions for each cell type. This dataset had previously been processed by us in Luo et al.³⁰.

Salas et al. (GSE182379)¹⁰: This dataset comprises EPICv1 profiles of 12 experimental mixture samples of the same 12 immune cell types, with known mixing proportions for each cell type. idat files were downloaded and processed using the minfi R-package. We retained 775,940 probes with significantly detected values across all 12 samples. The resulting beta-valued data matrix was then adjusted for type-2 probe bias using BMIQ.

Salas et al. (GSE180683)¹⁰: This dataset comprises EPIC profiles of 76 peripheral blood samples from 76 glioma patients, including FACS cell count information for naïve and memory CD4+ and CD8+ T cells. Data from this cohort were processed using the identical pipeline described above.

Cell mixture deconvolution

Cell-mixture deconvolution refers to the process of inferring the cellular composition of a mixed sample, which is typically accomplished using a multivariate regression, either in constrained or unconstrained form^9,10,22,29. Ridder et al. have demonstrated that the robust partial correlation (RPC) method of EpiDISH is currently among the best-performing algorithms among 16 tested deconvolution algorithms⁵⁵. Therefore, we estimated corresponding cell-type fractions using the EpiDISH^29,56. Specifically, we ran the epidish function with “RPC” as the method and maxit = 500. We also evaluated performance using the Constrained Projection (method = “CP”) and CIBERSORT (method = “CBS”) methods, both available in the same package. For Constrained Projection, we tested both the ‘equality’ and ‘inequality’ constraint parameters (constraint = “equality” and constraint = “inequality”).

Sliding window analysis

To assess the robustness of the t-statistic and gap specificity score ranking methods, a sliding window analysis was performed. This analysis was based on two distinct N = 1000 reference libraries. The first was the t-statistics library (500 hyper- and 500 hypo-methylated DMCs per cell type), and the second was the GSS library (500 hyper- and 500 hypo-methylated DMCs per cell type). For each of these two 12*1000 libraries, the hyper- and hypo-methylated DMC lists for each cell type were independently sorted in descending order according to their respective metric (t-statistic or GSS). A sliding window was then applied to each sorted library to generate sequential reference matrices. The window size was set to 50 DMCs, composed of the top 25 hyper- and top 25 hypo-methylated DMCs from the current position. The window was moved with a step size of 10 DMCs (5 hyper- and 5 hypo-methylated DMCs). This process generated 46 sequential reference matrices for each of the two ranking methods, where each subsequent matrix was composed of progressively lower-ranked DMCs. Each of the resulting 92 (46 ×2) reference matrices was used to deconvolve the independent 12-mixture validation dataset using the RPC method. The mean PCC and mean RMSE across all 12 cell types were calculated for each window. Finally, Spearman’s rank correlation was used to quantify the association between the window index (1–46) and the mean deconvolution performance (PCC and RMSE).

IDOL algorithm

Koestler et al. have provided the source code for the core IDOL function, IDOLoptimize, and the IDOL R package in their GitHub repository https://github.com/immunomethylomics/IDOL²³. As performed by Salas et al.¹⁰, we first used the legacy method to construct a feature pool containing 1,000 DMCs for each cell type using the sorted immune cell types from GSE167998, then optimized them using the experimental mixtures from GSE167998, and subsequently conducted independent validation using the experimental mixtures from GSE182379. The library size parameter, libSize = 1200, was determined as the optimal size by Salas et al. based on testing ranging from 250 to 3000 CpGs¹⁰.

Elastic net marker selection

As a primary linear benchmark, we trained an Elastic Net (ELN) model using the 12 purified cell types from the GSE167998 dataset as the training set. To ensure a direct and fair comparison, the analysis was restricted to the identical pre-screened DMC feature pool (CpGs) also used by the IDOL method. We implemented a multinomial logistic regression model using the glmnet R package (family = “multinomial”)³², treating the 12 sorted immune cell types as the categorical response variable. Hyperparameter optimization was conducted to select both the elastic net mixing parameter (α) and the regularization strength (λ). We performed a grid search for α (testing values from 0.1 to 1.0 in 0.1 increments). For each α value in the grid, the optimal λ was determined using threefold cross-validation via the cv.glmnet function. This two-dimensional tuning process identified a global optimal α of 0.9. For this final model, we selected the lambda.min value, which represents the model with the minimum cross-validated error. Finally, we extracted all DMCs (N = 348) that had non-zero coefficients from this optimized model to construct the final ELN-derived reference matrix.

Random Forest marker selection

To provide a robust non-linear benchmark, we implemented a Random Forest (RF) classifier³³. To ensure a direct and fair comparison, this analysis utilized the same pre-screened DMC feature pool that was used for both the Elastic Net benchmark and as the input for IDOL. The analysis leveraged the ranger R package, which is optimized for high-dimensional data⁵⁷. Hyperparameter tuning for the model was conducted using the caret R package⁵⁸. We focused on optimizing the key hyperparameter, mtry, which defines the number of variables randomly sampled as candidates at each split. A pre-defined grid of mtry values was evaluated using threefold cross-validation. After identifying the optimal mtry value, a final native multiclass classification model was trained on the entire training dataset using the ranger package. This model was built using 500 trees and fixed the mtry parameter to the optimal value determined during tuning. The model treated the 12 sorted immune cell types as the categorical response variable. For this final model, feature importance was quantified using the Mean Decrease in Accuracy (MDA), calculated via permutation on the out-of-bag (OOB) samples. This process yielded a single, global ranking of all DMCs based on their contribution to the multiclass model’s accuracy. From this list, we selected the top 1200 features to construct the final RF-derived reference matrix.

Human adult whole blood DNA methylation datasets

All the cohorts used here have previously been normalized and analyzed by us³⁰, including: (1) LiuMS (GSE106648): 279 HM450k peripheral blood (PB) samples, with age range from 16 to 66 (140 multiple sclerosis patients + 139 controls)⁵⁹; (2) Song (GSE169156), 2052 normal EPIC PB samples, with age range from 18 to 66⁶⁰; (3) HPT‑EPIC & HPT‑450k (GSE210255 & GSE210254): 1394 normal PB samples (EPIC set, age range from 21 to 87) and 418 normal PB samples (450k set, age range from 34 to 91)⁶¹; (4) Barturen (GSE179325): 574 normal EPIC PB samples, with age range from 19 to 103⁶²; (5) Airwave (GSE147740): 1129 normal EPIC PB samples, with age range from 25 to 60⁶³; (6) VACS (GSE117860): 529 HM450k samples from whole blood in HIV-positive men, with age range from 25 to 75⁶⁴; (7) Ventham (GSE87648): HM450k PB samples from 204 newly-diagnosed IBD cases and 178 controls, with age range from 17 to 79, and 2 samples have no age information⁶⁵; (8) Hannon‑1 and 2 (GSE80417 & GSE84727): 675 HM450k PB samples with age range from 18 to 90 (353 schizophrenia cases and 322 controls), and 847 HM450k PB samples with age range from 18 to 81 (414 schizophrenia cases and 433 controls)^38,39; (9) Zannas (GSE72680): 422 normal HM450k PB samples, with age range from 18 to 77⁶⁶; (10) Flanagan/FBS (GSE61151): 184 normal HM450k PB samples, with age range from 35 to 83⁶⁷; (11) Johansson (GSE87571): 732 normal HM450k PB samples, with age range from 14 to 94, and 3 samples have no age information⁶⁸; (12) TZH (OEP000260): 710 normal EPIC PB samples, with age range from 19 to 71⁶⁹; (13) HNM (GSE40279): 656 normal HM450k PB samples, with age range from 19 to 101⁷⁰.

Meta-analyses

In each cohort, we standardized the fractions of each cell type to unit variance. We then assessed associations between the standardized cell type fractions and chronological age using a multivariate linear regression. For specific cohorts where additional covariates were available, such as gender, disease and smoking status, the multivariate regression models included these factors as covariates. The specific covariates adjusted for in each of the 15 cohorts are detailed in Supplementary Data 3. For each regression, we extracted the corresponding effect size, standard error, Student’s t test, and P-value. We then performed a fixed and random effect inverse variance meta-analysis using the metagen function implemented in the meta R-package (version 7.0-0)⁷¹.

Analysis of the schizophrenia (SZ) EWAS datasets

For the Epigenome-Wide Association Study (EWAS) of Schizophrenia (SZ), we utilized the Hannon1 (GSE80417) and Hannon2 (GSE84727)^38,39 datasets. These Illumina 450k DNAm datasets have previously been normalized by us³⁰. Briefly, this processing and normalization involved probe filtering, which included the exclusion of 17,541 SNP-associated probes, 11,458 sex chromosome probes, and 26,569 cross-reactive probes. Only probes with detection P-values < 0.05 across all samples were kept for further analysis. The filtered beta-valued data were subsequently normalized with BMIQ⁷². To systematically investigate major sources of variation within this dataset and identify potential confounding variables, we performed a Singular Value Decomposition (SVD) analysis on the normalized beta-value matrix of the Hannon1 and Hannon2 datasets. The matrices were first mean-centered by subtracting the row-wise mean from each CpG. We then estimated the number of significant singular vectors using Random Matrix Theory, as implemented in the EstDimRMT function from the isva R package⁷³. SVD was subsequently performed using the base svd() function in R. To identify which factors contributed to the principal axes of variation, we tested for associations between the top singular vectors (or PCs) and the recorded sample covariates. Specifically, linear regression (via the lm function) was used to assess the association between each SV (PC) and continuous variables (e.g., age, cell-type fractions), while analysis of variance (ANOVA) was used for categorical variables (e.g., disease status, sex, SentrixID, and SentrixPos). To identify DMCs associated with SZ, we employed the limma R package separately in the discovery (Hannon1) and validation (Hannon2) cohorts⁷⁴. For each cohort, a design matrix was constructed modeling disease status while adjusting for age, sex, and the 12 estimated immune cell-type fractions. We fitted this model to the methylation data for every CpG site using the lmFit function. Subsequently, empirical Bayes moderation was applied using the eBayes function to compute moderated t-statistics, log-fold changes, and P-values. DMCs were identified in the discovery cohort (Hannon1) at an FDR < 0.05 (derived from the eBayes P-values) and subsequently carried forward for replication in the validation cohort (Hannon2), with significance defined as a nominal P-value < 0.05.

Following the DMC analysis, we performed pathway enrichment analysis for KEGG pathways using the gometh function from the missMethyl R package⁵¹.

Analysis of type-1 diabetes DNAm dataset

DAISY study⁵⁰: The Illumina DNAm dataset was downloaded from the GEO data repository (GEO accession GSE142512). This DNAm dataset was processed and normalized as described by us previously²⁷, and consists of 395 samples from 174 individuals. The ages of the individuals range from 6 months to 22 years, with the majority being under 10 years old. A total of 211 samples were profiled using the EPICv1 array, and 184 samples were sequenced using the 450k array. Half of the individuals were eventually diagnosed with type 1 diabetes, but all samples were taken before diagnosis. For the T1D EWAS, we also applied the limma pipeline (lmFit and eBayes), fitting a multiple linear model that adjusted for multiple confounders, including age, sex, sample timepoint, technical factors such as chip and position, and the 12 estimated immune cell fractions, to identify T1D-associated DMCs. DMCs were considered significant at an FDR < 0.05. Following the DMC analysis, we performed pathway enrichment analysis for KEGG pathways using the gometh function from the missMethyl R package⁵¹.

List of all software used

All data analysis was conducted within the R software environment v4.5.0 (https://cran.r-project.org/). Raw idat files were first processed using minfi v1.54.1 (https://bioconductor.org/packages/release/bioc/html/minfi.html), followed by an adjustment for type-2 probe bias using the BMIQ method from CHAMP v2.21.1 (https://www.bioconductor.org/packages/release/bioc/html/ChAMP.html). Cell-type fractions were subsequently estimated using the epidish function (“RPC” method) from EpiDISH v3.21 (https://bioconductor.org/packages/release/bioc/html/EpiDISH.html), while IDOL v0.0.0.9 (https://github.com/immunomethylomics/IDOL) was used for comparative library optimization. For machine learning benchmarks, Elastic Net models were implemented using glmnet v4.1 (https://cran.r-project.org/web/packages/glmnet/index.html), and Random Forest models were constructed using ranger v0.16.0 (https://cran.r-project.org/web/packages/ranger/index.html) with hyperparameter tuning performed via caret v6.0 (https://cran.r-project.org/web/packages/caret/index.html). Downstream EWAS analyses utilized limma v3.59.1 (https://bioconductor.org/packages/release/bioc/html/limma.html) for differential methylation and isva v1.9 (https://cran.r-project.org/web/packages/isva/index.html) for singular value decomposition analysis. Meta-analyses were performed with the metagen function from meta v7.0.0 (https://cran.r-project.org/web/packages/meta/index.html) and KEGG pathway enrichment analyses conducted with missMethyl v1.42.0 (https://bioconductor.org/packages/release/bioc/html/missMethyl.html). The core GSS-strategy developed in this study was implemented as a custom R function, ConstructDNAmPanel.R, which has been integrated into our EpiDISH R package and made publicly available on GitHub (https://github.com/sjczheng/EpiDISH).

Reporting summary

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

Supplementary information

Transparent Peer Review file^{(1.9MB, pdf)}

42003_2026_9745_MOESM2_ESM.pdf^{(31.1KB, pdf)}

Description of Additional Supplementary files

Supplementary Data 1^{(3.3MB, xlsx)}

Supplementary Data 2-4^{(29.4KB, xlsx)}

Reporting Summary^{(2.7MB, pdf)}

Acknowledgements

This project was funded by the National Natural Science Foundation of China grants RFIS W2431024, 32370699 and 32570775.

Author contributions

The study was conceived by AET. Statistical and bioinformatic analyses were performed by X.G. The paper was written by XG and AET.

Peer review

Peer review information

Communications Biology thanks Keegan Korthauer, who co-reviewed with Erick I. Navarro-Delgado and other anonymous reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Dr. Nilanjan Banerjee and Dr. Rosie Bunton-Stasyshyn. A peer review file is available.

Data availability

The following DNAm datasets analyzed here are publicly available from the NCBI GEO website https://www.ncbi.nlm.nih.gov/geo/ under the accession numbers GSE40279 (HNM), GSE106648 (LiuMS), GSE169156 (Song), GSE210255 (HPT-EPIC), GSE210254 (HPT-450k), GSE179325 (Barturen), GSE147740 (Airwave), GSE117860 (VACS), GSE87648 (Ventham), GSE84727 (Hannon2), GSE80417 (Hannon1), GSE72680 (Zannas), GSE61151 (Flanagan/FBS), GSE87571 (Johansson), GSE142512 (DAISY study), GSE167998 (Salas), GSE182379 (Salas), GSE180683 (Salas). The Illumina EPIC DNAm data for the TZH cohort can be viewed at NODE under accession number OEP000260, or directly at (https://www.biosino.org/node/project/detail/OEP000260), and accessed by submitting a request for data access. The remaining data are available within the Article, Supplementary Tables (Supplementary Data 1), or Source Data file (Supplementary Data 2).

Code availability

A function ConstructDNAmPanel.R has been added to our EpiDISH R-package, freely available from https://github.com/sjczheng/EpiDISH. This new version of EpiDISH has also been published on ZENODO: 10.5281/zenodo.18478028.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Contributor Information

Xiaolong Guo, Email: guoxiaolong2022@sinh.ac.cn.

Andrew E. Teschendorff, Email: andrew@sinh.ac.cn

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-026-09745-1.

References

1.Rakyan, V. K., Down, T. A., Balding, D. J. & Beck, S. Epigenome-wide association studies for common human diseases. Nat. Rev. Genet.12, 529–541 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
2.Lappalainen, T. & Greally, J. M. Associating cellular epigenetic models with human phenotypes. Nat. Rev. Genet.18, 441–451 (2017). [DOI] [PubMed] [Google Scholar]
3.Teschendorff, A. E. & Relton, C. L. Statistical and integrative system-level analysis of DNA methylation data. Nat. Rev. Genet.19, 129–147 (2018). [DOI] [PubMed] [Google Scholar]
4.Liu, Y. et al. Epigenome-wide association data implicate DNA methylation as an intermediary of genetic risk in rheumatoid arthritis. Nat. Biotechnol.31, 142–147 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Jaffe, A. E. & Irizarry, R. A. Accounting for cellular heterogeneity is critical in epigenome-wide association studies. Genome Biol.15, R31 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Zhang, Z., Christensen, B. C. & Salas, L. A. Recalibrate concepts of epigenetic aging clocks in human health. Aging (Albany NY)16, 11125–11127 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Zheng, S. C. et al. Correcting for cell-type heterogeneity in epigenome-wide association studies: revisiting previous analyses. Nat. Methods14, 216–217 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Teschendorff, A. E. & Zheng, S. C. Cell-type deconvolution in epigenome-wide association studies: a review and recommendations. Epigenomics9, 757–768 (2017). [DOI] [PubMed] [Google Scholar]
9.Houseman, E. A. et al. DNA methylation arrays as surrogate measures of cell mixture distribution. BMC Bioinforma.13, 86 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Salas, L. A. et al. Enhanced cell deconvolution of peripheral blood using DNA methylation for high-resolution immune profiling. Nat. Commun.13, 761 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Gervin, K. et al. Systematic evaluation and validation of reference and library selection methods for deconvolution of cord blood DNA methylation data. Clin. Epigenetics11, 125 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Salas, L. A. et al. An optimized library for reference-based deconvolution of whole-blood biospecimens assayed using the Illumina HumanMethylationEPIC BeadArray. Genome Biol.19, 64 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Cai, M., Zhou, J., McKennan, C. & Wang, J. scMD facilitates cell type deconvolution using single-cell DNA methylation references. Commun. Biol.7, 1 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Zhu, T. et al. A pan-tissue DNA methylation atlas enables in silico decomposition of human tissue methylomes at cell-type resolution. Nat. Methods19, 296–306 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Zheng, S. C. et al. A novel cell-type deconvolution algorithm reveals substantial contamination by immune cells in saliva, buccal and cervix. Epigenomics10, 925–940 (2018). [DOI] [PubMed] [Google Scholar]
16.Zhang, Z. et al. HiTAIC: hierarchical tumor artificial intelligence classifier traces tissue of origin and tumor type in primary and metastasized tumors using DNA methylation. NAR Cancer5, zcad017 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Zhang, Z. et al. HiTIMED: hierarchical tumor immune microenvironment epigenetic deconvolution for accurate cell type resolution in the tumor microenvironment using tumor-type-specific DNA methylation data. J. Transl. Med.20, 516 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Pike, S. C. et al. Glioma immune microenvironment composition calculator (GIMiCC): a method of estimating the proportions of eighteen cell types from DNA methylation microarray data. Acta Neuropathol. Commun.12, 170 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Chien, J. F. et al. Cell-type-specific effects of age and sex on human cortical neurons. Neuron112, 2524–2539 e2525 (2024). [DOI] [PubMed] [Google Scholar]
20.Lee, D. S. et al. Simultaneous profiling of 3D genome structure and DNA methylation in single human cells. Nat. Methods16, 999–1006 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Aryee, M. J. et al. Minfi: a flexible and comprehensive Bioconductor package for the analysis of Infinium DNA methylation microarrays. Bioinformatics30, 1363–1369 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Bell-Glenn, S., Thompson, J. A., Salas, L. A. & Koestler, D. C. A Novel Framework for the identification of reference DNA methylation libraries for reference-based deconvolution of cellular mixtures. Front. Bioinforma.ume 2, 2022 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Koestler, D. C. et al. Improving cell mixture deconvolution by identifying optimal DNA methylation libraries (IDOL). BMC Bioinforma.17, 120 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Arneson, D., Yang, X. & Wang, K. MethylResolver—a method for deconvoluting bulk DNA methylation profiles into known and unknown cell contents. Commun. Biol.3, 422 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Loyfer, N. et al. A DNA methylation atlas of normal human cell types. Nature613, 355–364 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Teschendorff, A. E., Zhu, T., Breeze, C. E. & Beck, S. EPISCORE: cell type deconvolution of bulk tissue DNA methylomes from single-cell RNA-seq data. Genome Biol.21, 221 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Guo, X. et al. Unified high-resolution immune cell fraction estimation in blood tissue from birth to old age. Genome Med.17, 63 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Cohen, J. Statitical Power Analysis for the Behavorial Sciences. 2nd edn. (Lawrence Erlbaum Associates, Publishers, 1988).
29.Teschendorff, A. E., Breeze, C. E., Zheng, S. C. & Beck, S. A comparison of reference-based algorithms for correcting cell-type heterogeneity in Epigenome-Wide Association Studies. BMC Bioinforma.18, 105 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Luo, Q. et al. A meta-analysis of immune-cell fractions at high resolution reveals novel associations with common phenotypes and health outcomes. Genome Med.15, 59 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Newman, A. M. et al. Robust enumeration of cell subsets from tissue expression profiles. Nat. Methods12, 453–457 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Friedman, J., Hastie, T. & Tibshirani, R. Regularization paths for generalized linear models via coordinate descent. J. Stat. Softw.33, 1–22 (2010). [PMC free article] [PubMed] [Google Scholar]
33.Breiman, L. Random Forests. Springer, (2001).
34.Li, M. et al. Age related human T cell subset evolution and senescence. Immun. Ageing16, 24 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Lin, Y. et al. Changes in blood lymphocyte numbers with age in vivo and their association with the levels of cytokines/cytokine receptors. Immun. Ageing13, 24 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Zhang, L., Mack, R., Breslin, P. & Zhang, J. Molecular and cellular mechanisms of aging in hematopoietic stem cells and their niches. J. Hematol. Oncol.13, 157 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Brauning, A. et al. Aging of the immune system: focus on natural killer cells phenotype and functions. Cells11, 1017 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Hannon, E. et al. An integrated genetic-epigenetic analysis of schizophrenia: evidence for co-localization of genetic associations and differential DNA methylation. Genome Biol.17, 176 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Hannon, E. et al. DNA methylation meta-analysis reveals cellular alterations in psychosis and markers of treatment-resistant schizophrenia. Elife10, e58430 (2021). [DOI] [PMC free article] [PubMed]
40.Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics8, 118–127 (2007). [DOI] [PubMed] [Google Scholar]
41.Carletti, B., Banaj, N., Piras, F. & Bossù, P. Schizophrenia and glutathione: a Challenging Story. J. Personal. Med.13 (2023). [DOI] [PMC free article] [PubMed]
42.Hill, M. et al. Steroidomics in men with schizophrenia. Int. J. Mol. Sci.25, 8729 (2024). [DOI] [PMC free article] [PubMed]
43.Miller, B. J., Buckley, P., Seabolt, W., Mellor, A. & Kirkpatrick, B. Meta-analysis of cytokine alterations in schizophrenia: clinical status and antipsychotic effects. Biol. Psychiatry70, 663–671 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Hong, S. et al. Abnormalities in chemokine levels in schizophrenia and their clinical correlates. Schizophr. Res.181, 63–69 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Ermakov, E. A., Mednova, I. A., Boiko, A. S., Buneva, V. N. & Ivanova, S. A. Chemokine dysregulation and neuroinflammation in schizophrenia: a systematic review. Int. J. Mol. Sci.24, 2215 (2023). [DOI] [PMC free article] [PubMed]
46.Harrer, C. et al. The CXCL13/CXCR5-chemokine axis in neuroinflammation: evidence of CXCR5+CD4 T cell recruitment to CSF. Fluids Barriers CNS18, 40 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Zhou, Q. et al. KDM2B promotes IL-6 production and inflammatory responses through Brg1-mediated chromatin remodeling. Cell Mol. Immunol.17, 834–842 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.van Jaarsveld, R. H. et al. Delineation of a KDM2B-related neurodevelopmental disorder and its associated DNA methylation signature. Genet. Med.25, 49–62 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Yokotsuka-Ishida, S. et al. Positional cloning and comprehensive mutation analysis identified a novel KDM2B mutation in a Japanese family with minor malformations, intellectual disability, and schizophrenia. J. Hum. Genet.66, 597–606 (2021). [DOI] [PubMed] [Google Scholar]
50.Johnson, R. K. et al. Longitudinal DNA methylation differences precede type 1 diabetes. Sci. Rep.10, 3721 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
51.Phipson, B., Maksimovic, J. & Oshlack, A. missMethyl: an R package for analyzing data from Illumina’s HumanMethylation450 platform. Bioinformatics32, 286–288 (2016). [DOI] [PubMed] [Google Scholar]
52.Peng, Q. et al. Analysis of blood methylation quantitative trait loci in East Asians reveals ancestry-specific impacts on complex traits. Nat. Genet.56, 846–860 (2024). [DOI] [PubMed] [Google Scholar]
53.Osumi-Sutherland, D. et al. Cell type ontologies of the Human Cell Atlas. Nat. Cell Biol.23, 1129–1135 (2021). [DOI] [PubMed] [Google Scholar]
54.Qi, L. & Teschendorff, A. E. Cell-type heterogeneity: why we should adjust for it in epigenome and biomarker studies. Clin. Epigenetics14, 31 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
55.De Ridder, K., Che, H., Leroy, K. & Thienpont, B. Benchmarking of methods for DNA methylome deconvolution. Nat. Commun.15, 4134 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
56.Zheng, S. C. et al. EpiDISH web server: epigenetic dissection of intra-sample-heterogeneity with online GUI. Bioinformatics36, 1950–1951 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Wright, M. N. & Ziegler, A. ranger: A fast implementation of random forests for high dimensional data in C++ and R. J. Stat. Softw.77, 17 (2017). [Google Scholar]
58.Kuhn, M. Building predictive models in R using the caret package. J. Stat. Softw.28, 1–26 (2008).27774042 [Google Scholar]
59.Kular, L. et al. DNA methylation as a mediator of HLA-DRB1*15:01 and a protective variant in multiple sclerosis. Nat. Commun.9, 2397 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
60.Song, N. et al. Persistent variations of blood DNA methylation associated with treatment exposures and risk for cardiometabolic outcomes in long-term survivors of childhood cancer in the St. Jude Lifetime Cohort. Genome Med. 13, 53 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
61.Shang, L. et al. meQTL mapping in the GENOA study reveals genetic determinants of DNA methylation in African Americans. Nat. Commun.14, 2711 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
62.Barturen, G. et al. Whole blood DNA methylation analysis reveals respiratory environmental traits involved in COVID-19 severity following SARS-CoV-2 infection. Nat. Commun.13, 4597 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
63.Robinson, O. et al. Determinants of accelerated metabolomic and epigenetic aging in a UK cohort. Aging Cell19, e13149 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
64.Zhang, X. et al. Machine learning selected smoking-associated DNA methylation signatures that predict HIV prognosis and mortality. Clin. Epigenetics10, 155 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Ventham, N. T. et al. Integrative epigenome-wide analysis demonstrates that DNA methylation may mediate genetic risk in inflammatory bowel disease. Nat. Commun.7, 13507 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
66.Zannas, A. S. et al. Epigenetic upregulation of FKBP5 by aging and stress contributes to NF-kappaB-driven inflammation and cardiovascular risk. Proc. Natl. Acad. Sci. USA116, 11370–11379 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
67.Flanagan, J. M. et al. Temporal stability and determinants of white blood cell DNA methylation in the breakthrough generations study. Cancer Epidemiol. Biomark. Prev.24, 221–229 (2015). [DOI] [PubMed] [Google Scholar]
68.Johansson, A., Enroth, S. & Gyllensten, U. Continuous Aging of the Human DNA Methylome Throughout the Human Lifespan. PLoS One8, e67378 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
69.You, C. et al. A cell-type deconvolution meta-analysis of whole blood EWAS reveals lineage-specific smoking-associated DNA methylation changes. Nat. Commun.11, 4779 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
70.Hannum, G. et al. Genome-wide methylation profiles reveal quantitative views of human aging rates. Mol. Cell49, 359–367 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
71.Balduzzi, S., Rücker, G. & Schwarzer, G. How to perform a meta-analysis with R: a practical tutorial. Evid. Based Ment. Health22, 153–160 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
72.Teschendorff, A. E. et al. A beta-mixture quantile normalization method for correcting probe design bias in Illumina Infinium 450 k DNA methylation data. Bioinformatics29, 189–196 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
73.Teschendorff, A. E., Zhuang, J. & Widschwendter, M. Independent surrogate variable analysis to deconvolve confounding factors in large-scale microarray profiling studies. Bioinformatics27, 1496–1505 (2011). [DOI] [PubMed] [Google Scholar]
74.Ritchie, M. E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res.43, e47–e47 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Transparent Peer Review file^{(1.9MB, pdf)}

42003_2026_9745_MOESM2_ESM.pdf^{(31.1KB, pdf)}

Description of Additional Supplementary files

Supplementary Data 1^{(3.3MB, xlsx)}

Supplementary Data 2-4^{(29.4KB, xlsx)}

Reporting Summary^{(2.7MB, pdf)}

Data Availability Statement

[CR1] 1.Rakyan, V. K., Down, T. A., Balding, D. J. & Beck, S. Epigenome-wide association studies for common human diseases. Nat. Rev. Genet.12, 529–541 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR2] 2.Lappalainen, T. & Greally, J. M. Associating cellular epigenetic models with human phenotypes. Nat. Rev. Genet.18, 441–451 (2017). [DOI] [PubMed] [Google Scholar]

[CR3] 3.Teschendorff, A. E. & Relton, C. L. Statistical and integrative system-level analysis of DNA methylation data. Nat. Rev. Genet.19, 129–147 (2018). [DOI] [PubMed] [Google Scholar]

[CR4] 4.Liu, Y. et al. Epigenome-wide association data implicate DNA methylation as an intermediary of genetic risk in rheumatoid arthritis. Nat. Biotechnol.31, 142–147 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR5] 5.Jaffe, A. E. & Irizarry, R. A. Accounting for cellular heterogeneity is critical in epigenome-wide association studies. Genome Biol.15, R31 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR6] 6.Zhang, Z., Christensen, B. C. & Salas, L. A. Recalibrate concepts of epigenetic aging clocks in human health. Aging (Albany NY)16, 11125–11127 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR7] 7.Zheng, S. C. et al. Correcting for cell-type heterogeneity in epigenome-wide association studies: revisiting previous analyses. Nat. Methods14, 216–217 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR8] 8.Teschendorff, A. E. & Zheng, S. C. Cell-type deconvolution in epigenome-wide association studies: a review and recommendations. Epigenomics9, 757–768 (2017). [DOI] [PubMed] [Google Scholar]

[CR9] 9.Houseman, E. A. et al. DNA methylation arrays as surrogate measures of cell mixture distribution. BMC Bioinforma.13, 86 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR10] 10.Salas, L. A. et al. Enhanced cell deconvolution of peripheral blood using DNA methylation for high-resolution immune profiling. Nat. Commun.13, 761 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR11] 11.Gervin, K. et al. Systematic evaluation and validation of reference and library selection methods for deconvolution of cord blood DNA methylation data. Clin. Epigenetics11, 125 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR12] 12.Salas, L. A. et al. An optimized library for reference-based deconvolution of whole-blood biospecimens assayed using the Illumina HumanMethylationEPIC BeadArray. Genome Biol.19, 64 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR13] 13.Cai, M., Zhou, J., McKennan, C. & Wang, J. scMD facilitates cell type deconvolution using single-cell DNA methylation references. Commun. Biol.7, 1 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Zhu, T. et al. A pan-tissue DNA methylation atlas enables in silico decomposition of human tissue methylomes at cell-type resolution. Nat. Methods19, 296–306 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR15] 15.Zheng, S. C. et al. A novel cell-type deconvolution algorithm reveals substantial contamination by immune cells in saliva, buccal and cervix. Epigenomics10, 925–940 (2018). [DOI] [PubMed] [Google Scholar]

[CR16] 16.Zhang, Z. et al. HiTAIC: hierarchical tumor artificial intelligence classifier traces tissue of origin and tumor type in primary and metastasized tumors using DNA methylation. NAR Cancer5, zcad017 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR17] 17.Zhang, Z. et al. HiTIMED: hierarchical tumor immune microenvironment epigenetic deconvolution for accurate cell type resolution in the tumor microenvironment using tumor-type-specific DNA methylation data. J. Transl. Med.20, 516 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR18] 18.Pike, S. C. et al. Glioma immune microenvironment composition calculator (GIMiCC): a method of estimating the proportions of eighteen cell types from DNA methylation microarray data. Acta Neuropathol. Commun.12, 170 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR19] 19.Chien, J. F. et al. Cell-type-specific effects of age and sex on human cortical neurons. Neuron112, 2524–2539 e2525 (2024). [DOI] [PubMed] [Google Scholar]

[CR20] 20.Lee, D. S. et al. Simultaneous profiling of 3D genome structure and DNA methylation in single human cells. Nat. Methods16, 999–1006 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR21] 21.Aryee, M. J. et al. Minfi: a flexible and comprehensive Bioconductor package for the analysis of Infinium DNA methylation microarrays. Bioinformatics30, 1363–1369 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR22] 22.Bell-Glenn, S., Thompson, J. A., Salas, L. A. & Koestler, D. C. A Novel Framework for the identification of reference DNA methylation libraries for reference-based deconvolution of cellular mixtures. Front. Bioinforma.ume 2, 2022 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR23] 23.Koestler, D. C. et al. Improving cell mixture deconvolution by identifying optimal DNA methylation libraries (IDOL). BMC Bioinforma.17, 120 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR24] 24.Arneson, D., Yang, X. & Wang, K. MethylResolver—a method for deconvoluting bulk DNA methylation profiles into known and unknown cell contents. Commun. Biol.3, 422 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR25] 25.Loyfer, N. et al. A DNA methylation atlas of normal human cell types. Nature613, 355–364 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.Teschendorff, A. E., Zhu, T., Breeze, C. E. & Beck, S. EPISCORE: cell type deconvolution of bulk tissue DNA methylomes from single-cell RNA-seq data. Genome Biol.21, 221 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR27] 27.Guo, X. et al. Unified high-resolution immune cell fraction estimation in blood tissue from birth to old age. Genome Med.17, 63 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR28] 28.Cohen, J. Statitical Power Analysis for the Behavorial Sciences. 2nd edn. (Lawrence Erlbaum Associates, Publishers, 1988).

[CR29] 29.Teschendorff, A. E., Breeze, C. E., Zheng, S. C. & Beck, S. A comparison of reference-based algorithms for correcting cell-type heterogeneity in Epigenome-Wide Association Studies. BMC Bioinforma.18, 105 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR30] 30.Luo, Q. et al. A meta-analysis of immune-cell fractions at high resolution reveals novel associations with common phenotypes and health outcomes. Genome Med.15, 59 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR31] 31.Newman, A. M. et al. Robust enumeration of cell subsets from tissue expression profiles. Nat. Methods12, 453–457 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR32] 32.Friedman, J., Hastie, T. & Tibshirani, R. Regularization paths for generalized linear models via coordinate descent. J. Stat. Softw.33, 1–22 (2010). [PMC free article] [PubMed] [Google Scholar]

[CR33] 33.Breiman, L. Random Forests. Springer, (2001).

[CR34] 34.Li, M. et al. Age related human T cell subset evolution and senescence. Immun. Ageing16, 24 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR35] 35.Lin, Y. et al. Changes in blood lymphocyte numbers with age in vivo and their association with the levels of cytokines/cytokine receptors. Immun. Ageing13, 24 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR36] 36.Zhang, L., Mack, R., Breslin, P. & Zhang, J. Molecular and cellular mechanisms of aging in hematopoietic stem cells and their niches. J. Hematol. Oncol.13, 157 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR37] 37.Brauning, A. et al. Aging of the immune system: focus on natural killer cells phenotype and functions. Cells11, 1017 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR38] 38.Hannon, E. et al. An integrated genetic-epigenetic analysis of schizophrenia: evidence for co-localization of genetic associations and differential DNA methylation. Genome Biol.17, 176 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR39] 39.Hannon, E. et al. DNA methylation meta-analysis reveals cellular alterations in psychosis and markers of treatment-resistant schizophrenia. Elife10, e58430 (2021). [DOI] [PMC free article] [PubMed]

[CR40] 40.Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics8, 118–127 (2007). [DOI] [PubMed] [Google Scholar]

[CR41] 41.Carletti, B., Banaj, N., Piras, F. & Bossù, P. Schizophrenia and glutathione: a Challenging Story. J. Personal. Med.13 (2023). [DOI] [PMC free article] [PubMed]

[CR42] 42.Hill, M. et al. Steroidomics in men with schizophrenia. Int. J. Mol. Sci.25, 8729 (2024). [DOI] [PMC free article] [PubMed]

[CR43] 43.Miller, B. J., Buckley, P., Seabolt, W., Mellor, A. & Kirkpatrick, B. Meta-analysis of cytokine alterations in schizophrenia: clinical status and antipsychotic effects. Biol. Psychiatry70, 663–671 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR44] 44.Hong, S. et al. Abnormalities in chemokine levels in schizophrenia and their clinical correlates. Schizophr. Res.181, 63–69 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR45] 45.Ermakov, E. A., Mednova, I. A., Boiko, A. S., Buneva, V. N. & Ivanova, S. A. Chemokine dysregulation and neuroinflammation in schizophrenia: a systematic review. Int. J. Mol. Sci.24, 2215 (2023). [DOI] [PMC free article] [PubMed]

[CR46] 46.Harrer, C. et al. The CXCL13/CXCR5-chemokine axis in neuroinflammation: evidence of CXCR5+CD4 T cell recruitment to CSF. Fluids Barriers CNS18, 40 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR47] 47.Zhou, Q. et al. KDM2B promotes IL-6 production and inflammatory responses through Brg1-mediated chromatin remodeling. Cell Mol. Immunol.17, 834–842 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR48] 48.van Jaarsveld, R. H. et al. Delineation of a KDM2B-related neurodevelopmental disorder and its associated DNA methylation signature. Genet. Med.25, 49–62 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR49] 49.Yokotsuka-Ishida, S. et al. Positional cloning and comprehensive mutation analysis identified a novel KDM2B mutation in a Japanese family with minor malformations, intellectual disability, and schizophrenia. J. Hum. Genet.66, 597–606 (2021). [DOI] [PubMed] [Google Scholar]

[CR50] 50.Johnson, R. K. et al. Longitudinal DNA methylation differences precede type 1 diabetes. Sci. Rep.10, 3721 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR51] 51.Phipson, B., Maksimovic, J. & Oshlack, A. missMethyl: an R package for analyzing data from Illumina’s HumanMethylation450 platform. Bioinformatics32, 286–288 (2016). [DOI] [PubMed] [Google Scholar]

[CR52] 52.Peng, Q. et al. Analysis of blood methylation quantitative trait loci in East Asians reveals ancestry-specific impacts on complex traits. Nat. Genet.56, 846–860 (2024). [DOI] [PubMed] [Google Scholar]

[CR53] 53.Osumi-Sutherland, D. et al. Cell type ontologies of the Human Cell Atlas. Nat. Cell Biol.23, 1129–1135 (2021). [DOI] [PubMed] [Google Scholar]

[CR54] 54.Qi, L. & Teschendorff, A. E. Cell-type heterogeneity: why we should adjust for it in epigenome and biomarker studies. Clin. Epigenetics14, 31 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR55] 55.De Ridder, K., Che, H., Leroy, K. & Thienpont, B. Benchmarking of methods for DNA methylome deconvolution. Nat. Commun.15, 4134 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR56] 56.Zheng, S. C. et al. EpiDISH web server: epigenetic dissection of intra-sample-heterogeneity with online GUI. Bioinformatics36, 1950–1951 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR57] 57.Wright, M. N. & Ziegler, A. ranger: A fast implementation of random forests for high dimensional data in C++ and R. J. Stat. Softw.77, 17 (2017). [Google Scholar]

[CR58] 58.Kuhn, M. Building predictive models in R using the caret package. J. Stat. Softw.28, 1–26 (2008).27774042 [Google Scholar]

[CR59] 59.Kular, L. et al. DNA methylation as a mediator of HLA-DRB1*15:01 and a protective variant in multiple sclerosis. Nat. Commun.9, 2397 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR60] 60.Song, N. et al. Persistent variations of blood DNA methylation associated with treatment exposures and risk for cardiometabolic outcomes in long-term survivors of childhood cancer in the St. Jude Lifetime Cohort. Genome Med. 13, 53 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR61] 61.Shang, L. et al. meQTL mapping in the GENOA study reveals genetic determinants of DNA methylation in African Americans. Nat. Commun.14, 2711 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR62] 62.Barturen, G. et al. Whole blood DNA methylation analysis reveals respiratory environmental traits involved in COVID-19 severity following SARS-CoV-2 infection. Nat. Commun.13, 4597 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR63] 63.Robinson, O. et al. Determinants of accelerated metabolomic and epigenetic aging in a UK cohort. Aging Cell19, e13149 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR64] 64.Zhang, X. et al. Machine learning selected smoking-associated DNA methylation signatures that predict HIV prognosis and mortality. Clin. Epigenetics10, 155 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR65] 65.Ventham, N. T. et al. Integrative epigenome-wide analysis demonstrates that DNA methylation may mediate genetic risk in inflammatory bowel disease. Nat. Commun.7, 13507 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR66] 66.Zannas, A. S. et al. Epigenetic upregulation of FKBP5 by aging and stress contributes to NF-kappaB-driven inflammation and cardiovascular risk. Proc. Natl. Acad. Sci. USA116, 11370–11379 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR67] 67.Flanagan, J. M. et al. Temporal stability and determinants of white blood cell DNA methylation in the breakthrough generations study. Cancer Epidemiol. Biomark. Prev.24, 221–229 (2015). [DOI] [PubMed] [Google Scholar]

[CR68] 68.Johansson, A., Enroth, S. & Gyllensten, U. Continuous Aging of the Human DNA Methylome Throughout the Human Lifespan. PLoS One8, e67378 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR69] 69.You, C. et al. A cell-type deconvolution meta-analysis of whole blood EWAS reveals lineage-specific smoking-associated DNA methylation changes. Nat. Commun.11, 4779 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR70] 70.Hannum, G. et al. Genome-wide methylation profiles reveal quantitative views of human aging rates. Mol. Cell49, 359–367 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR71] 71.Balduzzi, S., Rücker, G. & Schwarzer, G. How to perform a meta-analysis with R: a practical tutorial. Evid. Based Ment. Health22, 153–160 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR72] 72.Teschendorff, A. E. et al. A beta-mixture quantile normalization method for correcting probe design bias in Illumina Infinium 450 k DNA methylation data. Bioinformatics29, 189–196 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR73] 73.Teschendorff, A. E., Zhuang, J. & Widschwendter, M. Independent surrogate variable analysis to deconvolve confounding factors in large-scale microarray profiling studies. Bioinformatics27, 1496–1505 (2011). [DOI] [PubMed] [Google Scholar]

[CR74] 74.Ritchie, M. E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res.43, e47–e47 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Guidelines on optimizing DNA methylation reference panels for cell-type deconvolution

Xiaolong Guo

Andrew E Teschendorff

Abstract

Introduction

Results

Contamination of low-effect-size DMCs

Fig. 1. Gap specificity score avoids low-effect-size DMCs obtained by ranking of t-statistics.

A novel gap-specificity-score (GSS) to improve DNAm reference panels

Low-effect-size DMCs can negatively impact cell-type deconvolution accuracy

Fig. 2. Low-effect-size DMCs negatively impact cell-type deconvolution accuracy.

Hypomethylated DMCs improve cell-type deconvolution accuracy

Fig. 3. Exclusive use of hypomethylated markers improves deconvolution accuracy.

ML-optimization for marker selection can overfit to small training data

Fig. 4. GSS-Hypo reference outperforms machine learning based ones.

GSS-reference detects the known age-related increase in NK cells

Fig. 5. GSS-hypo detects increased NK fraction with age.

GSS-reference panel improves interpretability of EWAS findings

Fig. 6. GSS-Hypo reference panel improves interpretability of EWAS hits.

Discussion

Methods

Statistics and reproducibility

The DNAm reference library for panel construction

Experimental mixture DNAm and matched FACS datasets

Cell mixture deconvolution

Sliding window analysis

IDOL algorithm

Elastic net marker selection

Random Forest marker selection

Human adult whole blood DNA methylation datasets

Meta-analyses

Analysis of the schizophrenia (SZ) EWAS datasets

Analysis of type-1 diabetes DNAm dataset

List of all software used

Reporting summary

Supplementary information

Acknowledgements

Author contributions

Peer review

Peer review information

Data availability

Code availability

Competing interests

Footnotes

Contributor Information

Supplementary information

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases