Skip to main content
NCBI home page
Briefings in Bioinformatics logoLink to Briefings in Bioinformatics
. 2025 Oct 20;26(5):bbaf551. doi: 10.1093/bib/bbaf551

Read-level DNA methylation deconvolution enhances circulating tumor DNA detection

Ting Qi 1,2, Lakshmi Narayanan Lakshmanan 3, Yuwei Yang 4, Ying Zhou 5, Min Pan 6, Anders Jacobsen Skanderup 7,✉, Qinyu Ge 8,✉
PMCID: PMC12536878  PMID: 41115210

Abstract

DNA methylation is a key epigenetic modification underlying cellular identity. Conventional methods based on CpG site-level data often lack sensitivity in detecting low-frequency methylation signals. Here, we present Alpha, a novel method combining unbiased segmentation with robust read-level identification of low frequency cell-type-specific methylation signals. Methylation markers identified by Alpha exhibited significant enrichment in regulatory genomic elements such as enhancers, active promoters, and transcription factor binding sites. In simulated cell-type admixtures, Alpha-derived markers demonstrated improved deconvolution performance, exhibiting lower error metrics compared to beta-value based methods (DSS), even with limited marker numbers (N < 50). We combined Alpha with a non-negative least squares approach (Alpha-NNLS) to enable sensitive detection of circulating tumor DNA (ctDNA) in simulated cell-free DNA from breast and colon cancers, outperforming existing read-level methylation-based tumor fraction estimation methods (CelFEER and UXM). We applied Alpha-NNLS to targeted bisulfite sequencing data from early-stage colon cancer plasma samples and demonstrated strong concordance with existing approaches (R2 = 0.98), supporting its potential for sensitive detection of ctDNA.

Keywords: DNA methylation, deconvolution, cell-free DNA, circulating tumor DNA

Introduction

DNA methylation is a key epigenetic modification that can influence gene expression without altering the DNA sequence [1]. This modification mainly occurs in promoter regions and CpG islands, and plays an important regulatory role in gene expression [2], embryonic development [2], cell proliferation [3], cell differentiation [4], and other processes. Studies have shown that a majority of cell-type-specific methylation regions are hypomethylated and coincide with transcriptional regulatory elements [5, 6]. Changes in methylation patterns of specific regions are associated with the occurrence of many diseases, including cancer [7]. Therefore, cancer cell-type-specific methylation regions could serve as biomarkers for cancer detection and diagnosis.

In addition to serving as markers for cancer diagnosis, selecting the most appropriate cell-type-specific markers in reference-based tissue deconvolution helps distinguish similar cell types and reduce the computational complexity [8]. Furthermore, compared to single specific methylation sites, markers based on methylation regions can improve the accuracy of deconvolution [9]. Therefore, identifying appropriate cell-type-specific methylation markers is crucial for tissue and cell-free DNA (cfDNA) components deconvolution [10].

Methylation levels are frequently represented using the β-value, which represents the ratio of methylated reads to total reads overlapping a CpG site. β-values are calculated individually for each CpG site. The most representative β-value based method is Dispersion Shrinkage for Sequencing data (DSS), which is a method to identify differentially methylated CpG sites and regions [11]. It has been reported that methylation states at neighboring CpG sites are highly correlated [12]. In these cases, analyzing β value at each individual CpG site can be insensitive due to failing to consider neighboring CpG sites.

To overcome these limitations and amplify the methylation signal, an alternative read-level α-value metric has been proposed [13]. Instead of focusing on individual CpG sites over all reads, the α-value is calculated by aggregating methylation levels of adjacent CpG sites for each individual read. More importantly, the α-value is expected to perform better than β-value when the signal is weak. Previously, CelFEER [14] utilized these read-level methylation signals and showed better performance compared to β-value-based methods (MethAtlas [15]; CelFIE [16, 17]). In CelFEER, the genome is segmented into fixed-size windows (500 bp), and mean methylation levels are calculated for each region. However, it is not clear how these fixed region lengths impact the performance in the setting of low circulating tumor DNA (ctDNA) levels.

Here, we developed a computational method, Alpha, to identify cell/tissue-specific methylation regions from whole-genome methylation sequencing data. Using a combination of alpha values and an efficient dynamic programming-based genome segmentation (proposed by Loyfer et al. [18]), we investigated the distribution of tissue-specific methylated regions in the genome, their enrichment at active regulatory elements, and explored the ability to deconvolute cell- and tissue-types. Using simulated breast and colon cancer cfDNA data, we demonstrated that markers identified by Alpha outperformed existing cell-type deconvolution methods, especially at low ctDNA levels. Finally, we compared the concordance of our method with results from UXM and CelFEER using cfDNA methylation data from early-stage cancer patients.

Materials and Methods

Data source

The whole genome bisulfite sequencing (WGBS) data of five cell types used in the article are from the NCBI GEO data repository (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE186458). We chose liver hepatocytes, vein endothelial, erythrocyte progenitors, granulocytes, and monocytes, which are the most abundant cell types in plasma except for Megakaryocytes [18]. We have 15 samples in total (three replicates per cell type and five cell types). The enhancer markers (H3K4me1: https://www.encodeproject.org/experiments/ENCSR162RMW/), the active transcription markers (H3K27ac: https://www.encodeproject.org/experiments/ENCSR197UDN/), and the promoter markers (H3K4me3: https://www.encodeproject.org/experiments/ENCSR406DUQ/) of monocytes were downloaded from the ENCODE database (encodeproject.org).

Bulk sample simulation

We mixed the WGBS data from the five cell types above in different, randomly chosen proportions to simulate bulk tissue samples with an average depth of 30x (Fig. S9). We used “mix_pat” from “wgbstools” [19] to simulate 100 mixture samples. The steps that we took to simulate samples can be found in Fig. S3. The proportions of each cell type in the simulated samples are shown in Table S1. Most of the simulated proportion values are centered in the range 0.1–0.3 (Fig. S2).

Breast and colon cancer cfDNA sample simulation

We generated in silico tumor samples by mixing WGBS data from healthy cfDNA, breast luminal epithelium, and colon epithelium in different tumor proportions. Healthy cfDNA, breast and colon samples came from the GSE186458 dataset [GEO Accession viewer (nih.gov)] of the GEO database. We mixed healthy plasma samples with breast and colon samples to generate in silico samples with seven tumor proportions: 0.001, 0.005, 0.01, 0.03, 0.05, 0.10, and 0.15. The total cfDNA content (sequencing coverage) was kept constant across all in silico samples. Using five replicates for each proportion, we simulated a total of 35 breast cancer samples and 35 colon cancer samples separately.

Early cancer plasma dataset processing and analysis

We downloaded the targeted BS-seq dataset for colon cancer from GEO data repository (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE149438). After adapter removal, reads were aligned to hg38 using Bismark [20]. Tumor fraction was predicted using Alpha-NNLS, CelFEER (https://github.com/pi-zz-a/CelFEER) [14], and UXM (GitHub—nloyfer/UXM_deconv) [18] methods separately.

Identification of cell-/tissue-specific methylation regions

Our approach for identifying differentially methylated regions involves three steps.

In the first step, the genome is segmented into distinct blocks that show a similar methylation profile. The “segment” command from “wgbstools” was used for segmentation, which is based on a dynamic programming segmentation algorithm [19]. The algorithm uses a Maximum Likelihood approach to identify the segmentation that minimizes the within-segment variation in methylation levels (see Methods, Fig. S1). Identified segments can be of variable length but should contain at least four CpG sites. All the CpG sites within a single segment will have similar methylation levels (Fig. 1A).

Figure 1.

Figure 1

Open in a new tab

Overview of cell-type-specific methylation markers identification at read level. (A) the genome was segmented into regional blocks (≥4 CpG sites) with similar methylation levels using a dynamic programming-based segmentation algorithm from the “wgbstools” package [19], Figure 1A is reproduced from Loyfer et al., Nature 2023, under CC BY 4.0 license. (B) For each segment, the reads located within it were identified, and the methylation alpha value for each read was calculated. The average alpha value for all reads within the segment was then calculated. To identify target cell-type-specific hypo- and hyper-methylated segments, the average alpha values were compared between a set of target cell type and reference samples using the Wilcoxon rank sum test.

In step 2, the set of reads located within each segment (from step 1) was identified (Fig. 1B). Alpha value was calculated for each read using (equation 1). Finally, alpha values of all the “n” reads within a segment were averaged to get a mean alpha value for the segment (equation 2). This process was repeated for each sample.

graphic file with name DmEquation1.gif (1)
graphic file with name DmEquation2.gif (2)

In the last step, segment mean alpha values were compared between target group and reference group to identify target group-specific differentially methylated segments. Non-parametric Wilcoxon rank-sum test was used to assess the statistical significance of the differences in methylation levels between the two groups. “ranksums” function from “Scipy.stats” package was used for performing Wilcoxon rank-sum test. A P-value threshold of 0.05 was used to assess statistical significance. In addition to P-values, difference in mean alpha values between the two groups (Δ mean alpha) was also calculated (equation 3). Similar to existing studies [13, 21], blocks with Δ mean alpha values ≥0.5 were selected.

graphic file with name DmEquation3.gif (3)

We define blocks with the P-value < .05 and |Δ mean alpha| > 0.5 as specific methylation regions for this cell type. Regions with Δ mean alpha >0.5 are considered as hypermethylated, while regions with Δ mean alpha < −0.5 are considered as hypomethylated.

Distribution of cell- and tissue-specific methylation regions in the genome

“ChIPseeker” R package was used to perform genome annotations and visualization of cell- and tissue-specific methylation regions. “ggplot2” R package was used for visualizations.

Enrichment of methylation markers on functional elements

The “computeMatrix” and “plotProfile” routines from “deeptools” [22] package were used for assessing and visualizing the enrichment of cell-type-specific methylation regions at regulatory elements in the genome such as promoters, enhancers, and active transcription regions. “HOMER” [23] package was used for assessing the enrichment of cell-type-specific methylation regions at transcription factor binding sites (TFBS). The “motifStack” R package was used for visualizing TFBS motifs.

Classification of samples based on cell-type-specific methylation regions

For each cell type, the cell-type-specific methylation markers were sorted in descending order using the absolute Δ mean alpha values, and the top 50 markers were selected. The “pheatmap” package was used for clustering and “prcomp” function was used for classification analysis using the top 50 markers from each cell type as features.

Mixture deconvolution

Non-negative least squares (NNLS) [15] and expectation maximization (EM) deconvolution [16] algorithms were used for deconvoluting cell-free DNA mixtures with cell-type-specific methylation markers as features. Marker selection was performed using two methods, Alpha and DSS [11]. Performance of these markers for deconvolution, in simulated mixtures, was assessed using three metrics, mean squared error (equation 5), average absolute error (equation 6), and root mean squared error (equation 7) calculated between predicted and true mixture proportions. We performed mixture deconvolutions using feature spaces of five different sizes (10, 20, 50, 100, and 150 markers per cell type).

graphic file with name DmEquation4.gif (5)
graphic file with name DmEquation5.gif (6)
graphic file with name DmEquation6.gif (7)

Alpha-NNLS for cell-free DNA cell-type deconvolution

We generated a comprehensive reference atlas of cell-type-specific methylation markers for the major cell types found in plasma (liver hepatocytes, vein endothelial cells, WBC, pancreases, breast luminal epithelial, and colon epithelial cells), the top 300 reference markers of each cell type is shown in Table S2. Using these cell-type-specific methylation regions identified by our Alpha method as input features for NNLS-based deconvolution, we estimated the proportion of breast or colon cells (i.e. tumor fraction) in the simulated breast and colon cancer cfDNA samples. The reference atlas and script can be found in https://github.com/QiT-methods/alpha-marker.

Results

Distribution of cell-type-specific methylation markers on the genome

We identified cell-type-specific methylation markers for five cell types (liver hepatocytes, vein endothelial cells, erythrocyte progenitors, granulocytes, and monocytes) using our current approach (Alpha) based on segment mean alpha values (see Methods). These are the most abundant cell types in plasma except for megakaryocytes [18].

We compared the proportions of cell-type-specific markers located within different regions of the genome such as promoters, UTRs, exons, introns, and other distal intergenic regions. Proportions of cell-type-specific markers within different genomic regions are comparable among the three cell types of hematopoietic origin, erythrocyte progenitors, monocytes, and granulocytes (Fig. S4A). Intronic regions exhibit the largest fraction of markers followed by distal intergenic regions and promoter regions.

We have investigated the length of the Alpha-derived marker region. Alpha-derived marker regions are 400 bp on average (Fig. S5). Besides, most of the identified markers are hypomethylated (Fig. S11). In erythrocyte progenitors, 98% of markers are hypomethylated (Fig. S4B). Hepatocytes and vein endothelial cells show higher proportions of hypermethylated markers in comparison to the three cell types of hematopoietic origin (Fig. S4B).

Enrichment of cell-type-specific methylation markers on functional elements

Many alpha markers are distributed within promoter regions, possibly participating in gene expression regulation [24]. To verify this hypothesis, we investigated the enrichment of markers in functional regulatory elements. We selected enhancer markers marked by H3K4me1, promoter markers marked by H3K4me3, and active transcriptional element marked by H3K27ac for monocyte cells. The enrichment of the top 1000 alpha markers surrounding these three histone markers was analyzed.

The results revealed that the monocyte-specific alpha markers are enriched within enhancers and active transcriptional elements. More importantly, we found no enrichment of monocyte-specific markers in the other four cell types indicating that the markers identified by Alpha are biologically meaningful and unique to the cell type (Fig. 2A). Enrichment analysis found that the markers were enriched in cancer-related genes (e.g. BRCA1) and pathways such as endocytosis and regulation of actin cytoskeleton (Fig. S6). Previous studies have linked these pathways with cancer cell metastasis and immune cell phagocytosis [25–27].

Figure 2.

Figure 2

Open in a new tab

The enrichment of cell-type-specific alpha markers in functional elements. (A) Enrichment of the top 1000 cell-type-specific markers from five cell types within key chromatin regions of monocytes. The regions include active enhancers (H3K4me1), active promoters (H3K4me3), and active regulatory elements (H3K27ac). Enrichment analysis was performed using DeepTools. (B) Motif analysis of the top 1000 markers for each of the five cell types. The logos display overrepresented TFBS motifs. Motif enrichment analysis was performed using Homer.

To further analyse the important biological functions of cell-type-specific markers, we explored the enrichment of alpha markers at TFBSs. The results demonstrated that cell-type-specific alpha markers are enriched at multiple TFBSs and vary across cell types (Fig. 2B). The transcription factors GATA3, TRPS1, and Gata4, which are enriched in erythrocyte progenitors, play important roles in biological processes such as cell development, differentiation, and organ formation [28, 29]. The transcription factor CEBP enriched in granulocytes is involved in immune response and inflammatory response in cells [30], which is consistent with the function of granulocytes. The IRF3 TF of monocytes is involved in cell immune response, tumor immunity, and autoimmune diseases [31]. Moreover, TF ERG enriched in vascular endothelium participates in angiogenesis [32], and its abnormal expression is closely associated with the development of various cancers [33].

In general, cell-type-specific methylation biomarkers identified by Alpha at read-level are enriched in important cell-type-specific functional elements and transcription factors involved in the cell-type-specific regulatory activities.

Performance of alpha markers on classification and deconvolution

To investigate whether the cell-type-specific alpha markers can classify the WGBS data from five cell types (n = 3 replicates per cell type) samples, we selected top 50 markers from each cell types (a total of 250 markers) as features and performed clustering analysis. Three cell types from the blood clustered together, whereas the samples from vein epithelium and hepatocytes clustered into two separate categories (Fig. 3A). We also performed principal component analysis (PCA) to check whether cell-type-specific markers can clearly separate the WGBS data from individual cell types. As shown in Fig. 3B, we found that the 250 markers allowed an accurate classification of 15 samples, with monocytes and granulocytes closer in distance (Fig. 3B).

Figure 3.

Figure 3

Open in a new tab

Performance of alpha markers in sample classification and deconvolution. (A) Heatmap illustrates the hierarchical clustering of 15 biological samples from five cell types while using the top 50 alpha-derived markers. (B) PCA of the same 15 samples based on the top 50 alpha markers. (C) Comparison of deconvolution accuracy between alpha- and DSS-derived marker sets using 100 simulated samples. Three commonly used performance metrics—MAE, MSE, and RMSE—were calculated based on the deviation between estimated and true cell type proportions. (D) Deconvolution performance of NNLS and EM algorithms using varying numbers of alpha markers. Error lines show how deconvolution error changes with marker set size.

Cell-type-specific markers identified using the DSS method can correctly distinguish samples, but the erythrocyte progenitors and vein epitheliums are inaccurately clustered together with DSS markers (Fig. S7). We have investigated the genome position difference between DSS and Alpha markers. Only 68 overlapped regions were found between both marker lists (Table S3, Fig. S10), which suggests that Alpha markers offer a unique and complementary set for deconvolution.

We also trained a support vector machine (SVM) model using markers identified by Alpha, DSS, and UXM methods separately. The results show that models built with Alpha markers achieve the highest classification accuracy (median AUC = 0.8), outperforming DSS markers (median AUC = 0.78) and significantly surpassing UXM markers (median AUC = 0.62) (Fig. S8).

We tested the performance of cell-type-specific markers identified using Alpha and DSS methods to deconvolute 100 in silico mixtures generated from the five cell types. The mixture proportions can be found in Table S1. We selected the top 10, 20, 50, 100, 150 methylation markers, and NNLS and EM methods were used to deconvolute the simulated samples. By calculating the error between predicted and true proportions, the smallest errors (MSE = 0.0013, MAE = 0.022, RMSE = 0.036) were obtained when using Alpha-derived markers (Fig. 3C). Besides, as more markers were used, the error gradually decreased and then stabilized, indicating that choosing too many markers will not only increase the computational complexity but also did not improve the accuracy of deconvolution.

Finally, we compared the performance of Alpha markers in different deconvolution algorithms. The result suggests that the deconvolution based on NNLS had smaller errors compared to EM regardless of the number of markers used (Fig. 3D). Meanwhile, as the number of markers increases, the error tends to stabilize in the EM method. In view of the above results, 50–100 markers per cell type are most suitable for tissue deconvolution.

In summary, Alpha markers can distinguish individual cell types and have the lowest errors during mixture deconvolution and can be used for accurate cell type deconvolution.

Alpha-NNLS for circulating tumor DNA detection in silico samples

The ability to detect cancer early and during relapse will have a significant impact on the clinical outcome of the treatment. But it is challenging to detect tumors early when the tumor fraction is low, resulting in a weak signal.

We explored the performance of markers identified by our Alpha method in detecting low levels of tumor fraction, which can be helpful for non-invasive early cancer detection. We mixed the breast luminal epithelium cell with normal plasma cfDNA WGBS data to simulate breast cancer samples with varying levels of tumor fraction, more specifically 0.001, 0.005, 0.01, 0.03, 0.05, 0.1, and 0.15. A total of 35 breast and 35 colon cancer plasma samples were simulated, with five replicate samples for each tumor fraction.

Based on the NNLS deconvolution algorithm, our method (Alpha-NNLS) can confidently detect breast cell proportions as low as 0.5% in cfDNA simulation samples (Fig. 4A). When the tumor fraction was less than 0.5%, an obvious deviation can be observed between the predicted and the actual tumor fractions (TF), but as the tumor fraction increased, the bias gradually narrowed, and the predicted ratio is almost consistent with the true ratio even if only 50 alpha markers were used (Fig. 4A). When compared with healthy individuals (TF = 0), Alpha-NNLS can observe a significant difference in the predicted tumor fraction values between the TFs 0% and 0.1% (Fig. 4C).

Figure 4.

Figure 4

Open in a new tab

Alpha-NNLS based quantification of plasma cell-free DNA tumor fraction. (A) Log-scaled line plot comparing predicted tumor fractions obtained using alpha-NNLS, CelFEER, and UXM deconvolution methods with the known tumor fractions across varying numbers of markers. Both true and predicted tumor fractions are log10 transformed. The dashed diagonal line represents perfect concordance between predicted and actual values. (B) Line plot of RMSE, MAE, MSE values between the predicted tumor fraction and true tumor fraction with 30, 50, 100, 150, 200, 250, 300 markers. (C) Boxplot of predicted tumor fraction (TF) across eight defined ctDNA levels: normal (0), 0.001, 0.005, 0.01, 0.03, 0.05, 0.1, and 0.15. (D) Log-scaled line plot comparing predicted and true tumor fractions in the simulated colon cancer dataset using alpha-NNLS, across different numbers of input markers.

We compared Alpha-NNLS with another two cfDNA tumor fraction identification methods (CelFEER [14] and UXM [18]) which were reported to exhibit the best performance in cell type deconvolution for cell-free DNA [17]. For CelFEER, the predicted tumor fraction is lower than the true tumor fraction. However, the situation is the opposite with the UXM method (Fig. 4A). Besides, Alpha-NNLS achieves the lowest error between the predicted TFs and the true TFs compared to CelFEER and UXM (Fig. 4B). In addition, the accuracy of predicting tumor fraction is affected by the number of markers used per cell type, with the lowest error in Alpha-NNLS when using 300 markers (MAE = 0.004, MSE = 0.001, RMSE = 0.006), with the lowest error in CelFEER when using 50 markers (MAE = 0.014, MSE = 0.0015, RMSE = 0.02), with the lowest error in UXM when using 100 markers (MAE = 0.13, MSE = 0.02, RMSE = 0.16) (Fig. 4B).

Finally, we also studied the performance of Alpha-NNLS in a simulated colon cancer cfDNA dataset. The simulated tumor fraction levels are identical between breast cancer and colon cancer. A high level of consistency between the predicted tumor fraction and the true tumor fraction can be observed even when the TF = 0.5% (Fig. 4D).

Application of alpha-NNLS in real early-stage cancer plasma samples

To test the performance of our method (Alpha-NNLS) in real early cancer samples, we predict the tumor fraction in real targeted BS-seq early colon cancer samples. Since the true tumor fraction of the samples are unknown, we compared the results from Alpha-NNLS with those obtained from the other two methods (CelFEER and UXM).

There is a strong correlation between the TF predicted by Alpha-NNLS and the TFs predicted by CelFEER (Fig. 5A) and by UXM (Fig. 5B). Besides, our method can detect cancer samples with tumor fraction lower than 0.1, but the tumor fraction of those samples predicted by CelFEER and UXM is near zero. Our results demonstrate that, in real samples, the predictions of our method not only show strong consistency with existing methods but also can detect early-stage cancer when the TF is low, compared to the other two methods.

Figure 5.

Figure 5

Open in a new tab

Application of alpha-NNLS in real targeted BS-seq early plasma cancer samples. (A) Scatter plot comparing TF estimates generated by alpha-NNLS and CelFEER for 32 plasma samples from colon cancer patients. A blue line indicates the line of best linear fit. (B) Scatter plot comparing TF estimates from alpha-NNLS and UXM for the same 32 colon cancer plasma samples. The line represents the best fit line. (C) Boxplot of TF values predicted by alpha-NNLS for target BS-seq colon cancer plasma samples stratified by clinical stage (0, I, II, and IV). Statistical comparisons between groups were performed using the Wilcoxon rank-sum test.

Additionally, we predicted the tumor fraction for plasma samples associated with different colon cancer stages. The Alpha-NNLS model estimated significantly higher tumor fractions for late stage (IV) as compared to early stages (0, I, II) tumors (Wilcoxon rank sum test, P = 1.0e-4, Fig. 5C). Besides, statistical analysis revealed significant inter-stage differences (Wilcoxon rank test) (Fig. S12), supporting the method’s ability to stratify disease stages.

To further evaluate the performance of Alpha-NNLS in early-stage cancer detection, we assessed the detection accuracy for early-stage tumors (Stage I/II) using receiver operating characteristic (ROC) analysis in the targeted bisulfite sequencing (BS-seq) colon cohort. Alpha-NNLS achieved an area under the curve (AUC) of 0.78, suggesting promising utility for early detection (Fig. S13).

Discussion

The ability to detect tumors in low tumor fraction levels can improve early cancer screening and the early detection of relapses after surgery. Unlike genomic mutations, methylation data provides consistent cell-type and tissue-type-specific signals. Several methods have been proposed to quantify tumor fraction from methylation data. These methods generally differ in the approach they use to select cell-type-specific methylation markers. Better marker selection can improve the ability to detect weak signals in low tumor fraction settings.

In this study, we report a new protocol, Alpha. Our work identifies cell-type-specific methylation regions at the read level across the whole genome and explores the enrichment of marker regions at regulatory elements in the genome. Compared with the markers identified by DSS, using Alpha markers can better distinguish samples and deconvolute tissue. Using alpha markers can accurately predict the proportion of cancer cells in plasma at a low tumor fraction, enhancing ctDNA detection.

Most of the Alpha markers are hypomethylated in the five cell types (Fig. S4B). Our results confirm that unmethylated specific regions are enriched near active promoters and enhancers (Fig. 2A). The unmethylated regions relate to the gene expression and the level of methylation in the promoter region is inversely proportional to gene expression [34].

The accuracy of deconvolution is influenced by many factors. For reference-based deconvolution methods, the method of identifying features and the number of features will have unpredictable effects on the performances of deconvolution [9], which is confirmed in our results. The markers calculated based on Alpha have better performance in sample classification and deconvolution compared to the methods based on beta values (Fig. 3). But Alpha is designed for genome-wide BS/EM-seq data and may not be compatible with chip/array-based methylation data, which only provides targeted and narrow methylation signals.

Our research indicates that alpha-based specific biomarkers have a higher sensitivity for low-abundance features (Fig. 4). For early cancer, the tumor fraction in plasma is very low, making it difficult to detect the presence of cancer non-invasively [34]. Alpha-NNLS can identify the proportion of low-abundance tumor cells through effective markers and achieve the consistent TF prediction result compared to CelFEER and UXM despite using targeted sequencing data which has lower cost and are more targeted (Fig. 5). Additionally, we noted that it would be an exciting prospect to perform comprehensive orthogonal comparisons which is also not presented in other methylation-based deconvolution methods [12, 15, 16, 21, 35]. We intend to pursue this in our future work by collecting new cancer samples to acquire the necessary multi-modal cfDNA data.

Conclusion

In conclusion, we developed an improved and read-level method for identification of cell-type specific methylated regions, increasing the available options for the discovery of methylation biomarkers. We also proposed the Alpha-NNLS method, which can reduce errors in tissue deconvolution and enhance ctDNA detection. These approaches could enable improvements to non-invasive early cancer detection and disease monitoring.

Key Points

  • The cell-type-specific methylation markers identified at read-level are most hypo-methylated and enriched in active regulatory elements.

  • The Alpha markers can well distinguish samples and have the better performance in deconvolution compared with beta markers.

  • Alpha-NNLS method can confidently identify cancer information, which enhances circulating tumor DNA detection.

Supplementary Material

Supplementary_Material_bbaf551
supplementary_material_bbaf551.docx (885.8KB, docx)
Table_S1_bbaf551
table_s1_bbaf551.xlsx (18.2KB, xlsx)
Table_S2_bbaf551
table_s2_bbaf551.xlsx (287.2KB, xlsx)
Table_S3_bbaf551
table_s3_bbaf551.xlsx (13.3KB, xlsx)

Acknowledgements

We thank the authors of the tools and dataset used in this work for making the software, scripts, and data freely available for community usage.

Contributor Information

Ting Qi, State Key Laboratory of Digital Medical Engineering, School of Biological Science and Medical Engineering, Southeast University, Sipailou No. 2, Nanjing 210096, China; Genome Institute of Singapore (GIS), Agency for Science, Technology and Research, 60 Biopolis Street, Singapore 138672, Singapore.

Lakshmi Narayanan Lakshmanan, Genome Institute of Singapore (GIS), Agency for Science, Technology and Research, 60 Biopolis Street, Singapore 138672, Singapore.

Yuwei Yang, State Key Laboratory of Digital Medical Engineering, School of Biological Science and Medical Engineering, Southeast University, Sipailou No. 2, Nanjing 210096, China.

Ying Zhou, State Key Laboratory of Digital Medical Engineering, School of Biological Science and Medical Engineering, Southeast University, Sipailou No. 2, Nanjing 210096, China.

Min Pan, School of Medicine, Southeast University, Sipailou No. 2, Nanjing 210096, China.

Anders Jacobsen Skanderup, Genome Institute of Singapore (GIS), Agency for Science, Technology and Research, 60 Biopolis Street, Singapore 138672, Singapore.

Qinyu Ge, State Key Laboratory of Digital Medical Engineering, School of Biological Science and Medical Engineering, Southeast University, Sipailou No. 2, Nanjing 210096, China.

Author contributions

Ting Qi (Investigation, Methodology, Design, Writing—original draft, Editing, Visualization), Lakshmanan Lakshmi Narayanan (Writing—review & editing), Yuwei Yang (Methodology), Ying Zhou (Editing), Min Pan (Writing—review & editing), Anders Jacobsen Skanderup (Supervision, Writing—review & editing), and Qinyu Ge (Conceptualization, Project administration [lead], Supervision, Writing—review & editing)

Conflict of interest: The authors declare that they have no conflict of interests.

Funding

This work was funded by the National Key Research and Development Program of China (2022YFF0710800).

Data availability

The FACS WGBS data of involved cell types and normal cfDNA WGBS data are publicly available in the NCBI GEO data repository (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE186458). The target BS-seq HCC data are publicly accessible in GEO data repository (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE149438).

The script for identifying alpha marker at read-level is available at GitHub (https://github.com/QiT-methods/alpha-marker). Please refer to the README.md for the user manual. The software for sample simulation and genome segmentation is available at GitHub (https://github.com/nloyfer/wgbs_tools) [19].

References

  • 1. Anastasiadi  D, Wellenreuther  M. A conserved genomic code underpins animal DNA methylation patterns. Trends Ecol Evol  2023;38:1016–8. 10.1016/j.tree.2023.08.010 [DOI] [PubMed] [Google Scholar]
  • 2. Yuan  S, Gao  L, Tao  W. et al.  Allelic reprogramming of chromatin states in human early embryos. Natl Sci Rev  2024;11:nwad328. 10.1093/nsr/nwad328 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Guo  H, Vuille  JA, Wittner  BS. et al.  DNA hypomethylation silences anti-tumor immune genes in early prostate cancer and CTCs. Cell  2023;186(13): 2765-2782. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Meissner  A, Mikkelsen  TS, Gu  H. et al.  Genome-scale DNA methylation maps of pluripotent and differentiated cells. Nature  2008;454:766–70. 10.1038/nature07107 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. Sharma  M, Verma  RK, Kumar  S. et al.  Computational challenges in detection of cancer using cell-free DNA methylation. Comput Struct Biotechnol J  2022;20:26–39. 10.1016/j.csbj.2021.12.001 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Yasui  K, Toshima  T, Inada  R. et al.  Circulating cell-free DNA methylation patterns as non-invasive biomarkers to monitor colorectal cancer treatment efficacy without referencing primary site mutation profiles. Mol Cancer  2024;23:1. 10.1186/s12943-023-01910-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Das  PM, Singal  R. DNA methylation and cancer. J Clin Oncol Off J Am Soc Clin Oncol  2004;22:4632–42. 10.1200/JCO.2004.07.151 [DOI] [PubMed] [Google Scholar]
  • 8. Ferro, Dos Santos  MR, Giuili  E, De Koker  A. et al.  Computational deconvolution of DNA methylation data from mixed DNA samples. Brief Bioinform  2024;25(3): bbae234. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. De Ridder  K, Che  H, Leroy  K. et al.  Benchmarking of methods for DNA methylome deconvolution. Nat Commun  2024;15:4134. 10.1038/s41467-024-48466-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Barefoot  ME, Loyfer  N, Kiliti  AJ. et al.  Detection of cell types contributing to cancer from circulating, cell-free methylated DNA. Front Genet  2021;12:671057. 10.3389/fgene.2021.671057 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Feng  H, Conneely  KN, Wu  H. A Bayesian hierarchical model to detect differentially methylated loci from single nucleotide resolution sequencing data. Nucleic Acids Res  2014;42:e69. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Li  W, Li  Q, Kang  S. et al.  CancerDetector: ultrasensitive and non-invasive cancer detection at the resolution of individual reads using cell-free DNA methylation sequencing data. Nucleic Acids Res  2018;46:e89–9. 10.1093/nar/gky423 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Stackpole  ML, Zeng  W, Li  S. et al.  Cost-effective methylome sequencing of cell-free DNA for accurately detecting and locating cancer. Nat Commun  2022;13:5566. 10.1038/s41467-022-32995-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Keukeleire  P, Makrodimitris  S, Reinders  M. Cell type deconvolution of methylated cell-free DNA at the resolution of individual reads. NAR Genom Bioinform  2023;5:lqad048. 10.1093/nargab/lqad048 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Moss  J, Magenheim  J, Neiman  D. et al.  Comprehensive human cell-type methylation atlas reveals origins of circulating cell-free DNA in health and disease. Nat Commun  2018;9:5068. 10.1038/s41467-018-07466-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Caggiano  C, Celona  B, Garton  F. et al.  Comprehensive cell type decomposition of circulating cell-free DNA with CelFiE. Nat Commun  2021;12:2717. 10.1038/s41467-021-22901-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Sun  T, Yuan  J, Zhu  Y. et al.  Systematic evaluation of methylation-based cell type deconvolution methods for plasma cell-free DNA. Genome Biol  2024;25:318. 10.1186/s13059-024-03456-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Loyfer  N, Magenheim  J, Peretz  A. et al.  A DNA methylation atlas of normal human cell types. Nature  2023;613:355–64. 10.1038/s41586-022-05580-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Loyfer  N, Rosenski  J, Kaplan  T. wgbstools: a computational suite for DNA methylation sequencing data representation, visualization, and analysis. bioRxiv  2024. 10.1101/2024.05.08.593132 [DOI] [Google Scholar]
  • 20. Krueger  F, Andrews  SR. Bismark: a flexible aligner and methylation caller for Bisulfite-Seq applications. Bioinformatics  2011;27:1571–2. 10.1093/bioinformatics/btr167 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Li  S, Zeng  W, Ni  X. et al.  Comprehensive tissue deconvolution of cell-free DNA by deep learning for disease diagnosis and monitoring. Proc Natl Acad Sci U S A  2023;120:e2305236120. 10.1073/pnas.2305236120 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Ramírez  F, Ryan  DP, Grüning  B. et al.  deepTools2: a next generation web server for deep-sequencing data analysis. Nucleic Acids Res  2016;44:W160–5. 10.1093/nar/gkw257 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Heinz  S, Benner  C, Spann  N. et al.  Simple combinations of lineage-determining transcription factors prime cis-regulatory elements required for macrophage and B cell identities. Mol Cell  2010;38:576–89. 10.1016/j.molcel.2010.05.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Razin  A, Riggs  AD. DNA methylation and gene-function. Science  1980;210:604–10. 10.1126/science.6254144 [DOI] [PubMed] [Google Scholar]
  • 25. Banushi  B, Joseph  SR, Lum  B. et al.  Endocytosis in cancer and cancer therapy. Nat Rev Cancer  2023;23:450–73. 10.1038/s41568-023-00574-6 [DOI] [PubMed] [Google Scholar]
  • 26. Pena-Martinez  C, Rickman  AD, Heckmann  BL. Beyond autophagy: LC3-associated phagocytosis and endocytosis. Sci Adv  2022;8:eabn1702. 10.1126/sciadv.abn1702 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Jia  X, Yan  B, Tian  X. et al.  CD47/SIRPalpha pathway mediates cancer immune escape and immunotherapy. Int J Biol Sci  2021;17:3281–7. 10.7150/ijbs.60782 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28. Trujillo-Ochoa  JL, Kazemian  M, Afzali  B. The role of transcription factors in shaping regulatory T cell identity. Nat Rev Immunol  2023;23:842–56. 10.1038/s41577-023-00893-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Elster  D, Tollot  M, Schlegelmilch  K. et al.  TRPS1 shapes YAP/TEAD-dependent transcription in breast cancer cells. Nat Commun  2018;9:3115. 10.1038/s41467-018-05370-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. Tara  S, Isshiki  Y, Nakajima-Takagi  Y. et al.  Bcor insufficiency promotes initiation and progression of myelodysplastic syndrome. Blood  2018;132:2470–83. 10.1182/blood-2018-01-827964 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Li  S, Mirlekar  B, Johnson  BM. et al.  STING-induced regulatory B cells compromise NK function in cancer immunity. Nature  2022;610:373–80. 10.1038/s41586-022-05254-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Birdsey  GM, Dryden  NH, Amsellem  V. et al.  Transcription factor erg regulates angiogenesis and endothelial apoptosis through VE-cadherin. Blood  2008;111:3498–506. 10.1182/blood-2007-08-105346 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Guo  C, Figueiredo  I, Gurel  B. et al.  B7-H3 as a therapeutic target in advanced prostate cancer. Eur Urol  2023;83:224–38. 10.1016/j.eururo.2022.09.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Stein  R, Razin  A, Cedar  H. In vitro methylation of the hamster adenine phosphoribosyltransferase gene inhibits its expression in mouse L cells. Proc Natl Acad Sci U S A  1982;79:3418–22. 10.1073/pnas.79.11.3418 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Erger  F, Norling  D, Borchert  D. et al.  cfNOMe—a single assay for comprehensive epigenetic analyses of cell-free DNA. Genome Med  2020;12:54. 10.1186/s13073-020-00750-5 [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

Supplementary_Material_bbaf551
supplementary_material_bbaf551.docx (885.8KB, docx)
Table_S1_bbaf551
table_s1_bbaf551.xlsx (18.2KB, xlsx)
Table_S2_bbaf551
table_s2_bbaf551.xlsx (287.2KB, xlsx)
Table_S3_bbaf551
table_s3_bbaf551.xlsx (13.3KB, xlsx)

Data Availability Statement

The FACS WGBS data of involved cell types and normal cfDNA WGBS data are publicly available in the NCBI GEO data repository (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE186458). The target BS-seq HCC data are publicly accessible in GEO data repository (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE149438).

The script for identifying alpha marker at read-level is available at GitHub (https://github.com/QiT-methods/alpha-marker). Please refer to the README.md for the user manual. The software for sample simulation and genome segmentation is available at GitHub (https://github.com/nloyfer/wgbs_tools) [19].

Articles from Briefings in Bioinformatics are provided here courtesy of Oxford University Press

RESOURCES