Search

DNA methylation (DNAm)-based cell mixture deconvolution (CMD) has become a quintessential part of epigenome-wide association studieswhere DNAm is profiled in heterogeneous tissue types. Despite being introduced over a decade ago, detection limits, which represent the smallest fraction of a cell type in a mixed biospecimen that can be reliably detected, have yet to be determined in the context of DNAm-based CMD. Moreover, there has been little attention given to approaches for quantifying the uncertainty associated with DNAm-based CMD. Here, analytical frameworks for determining both cell-specific limits of detection and quantification of uncertainty associated with DNAm-based CMD are described. This work may contribute to improved rigor, reproducibilityand replicability of epigenome-wide association studies involving CMD.

The human brain comprises heterogeneous cell subtypes whose composition can be altered with physiological and pathological conditions. New approaches to discern the diversity and distribution of brain cells associated with neurological conditions would significantly advance the study of brain-related pathophysiology and neuroscience. We demonstrate that DNA-based cell-type deconvolution achieves an accurate resolution of seven major cell types. Unlike single-nuclei approaches, DNA methylation-based deconvolution does not require special sample handling or processing, is cost-effective, and easily scales to large study designs. Current methods for brain cell deconvolution are limited only to neuronal and non-neuronal cells. Using DNA methylation profiles of the top cell-type-specific differentially methylated CpGs, we employed a hierarchical modeling approach to deconvolve GABAergic neurons, glutamatergic neurons, astrocytes, microglial cells, oligodendrocytes, endothelial cells, and stromal cells. We demonstrate the utility of our method by applying it to data on normal tissues from various brain regions and in aging and diseased tissues, including Alzheimer's disease, autism, Huntington’s disease, epilepsy, and schizophrenia. We expect that the ability to determine the cellular composition in the brain using only DNA from bulk samples will accelerate understanding brain cell type composition and cell-type-specific epigenetic states in normal and diseased brain tissues.

Additional file 1: Supplementary Figures 1–3.

Additional file 1: Figures. Supplemental figures of reduced NDMI 850 in pilot samples, with flow of array and sample comparisons, Bland–Altman analysis, and distributions of NDMI scores by phenotype in the GEO cord blood datasets.

Additional file 2: Figure S1. Correlation between HiTIMED tumor and InfiniumPurify tumor by tumor type across cholangiocarcinoma, kidney papillary cell carcinoma, pancreatic adenocarcinoma, and stomach adenocarcinoma. Figure S2. Methylation state of CpGs in the HiTIMED tumor specific library (L1) and InfiniumPurify default library. Figure S3. HiTIMED tumor purity vs InfiniumPurify tumor purity in thyroid carcinoma. Figure S4. HiTIMED tumor proportion vs other method predicted tumor proportion. Figure S5. HiTIMED immune cell proportions vs true immune cell proportions in artificial mixtures. Figure S6. HiTIMED T cell proportion vs true T cell proportion in artificial mixtures. Figure S7. HiTIMED cell composition in human normal intestinal epithelium and umbilical vein endothelial cells. Figure S8. Performance comparison across HiTIMED, MethylCIBERSORT, and MethylResolver using artificial mixtures. Figure S9. The distribution of the HiTIMED cell composition in TCGA tumors. Figure S10. Cell composition differs substantially and captures sample heterogeneity using HiTIMED-projected proportions. Seventeen cell types were captured for each sample by tumor type. Figure S11. Sensitive analysis comparing outputs from two Cox models with or without cell type proportions adjusted in kidney clear cell carcinoma. Figure S12. Kaplan-Meier survival curves for HiTIMED cells estimates in TCGA tumors. Figure 13. HiTIMED cell comparison and Kaplan-Meier survival curves across immune/angiogenic hot and cold. Figure S14. HiTIMED immune and angiogenic proportions across C1-C6 subtyped TCGA tumor. Figure S15. HiTIMED cell comparisons between drug-sensitive and -resistant metastasized colorectal cancer. Figure S16. HiTIMED cell comparisons in triple−negative breast cancer w/without chemotherapy. Figure S17. Performance comparison across iterations on CpGs selected in HiTIMED for immune and angiogenic cell projection.

Additional file 7: Supplemental tables and figures.

Additional file 1: Table S1. Baseline characteristics of the discovery data sets. Table S2. Baseline characteristics of the validation and application data sets. Table S3. Statistically significant hazard ratios of HiTIMED-projected cell proportions in cancer patients' 5-year survival adjusted for age, gender, tumor stage, HiTIMED-projected tumor proportion, and other cell-type proportions.

Additional file 6. Supplemental Figures and Tables.

Additional file 1. Supplementary Methods.

Additional file 1: Supplementary Table 1. Male to Female Ratio Across Both Datasets. Supplementary Figure 1. Distribution of Age Across Datasets. Supplementary Table 2. Number of Samples for Each Cancer Subtype in Training, Validation, and Test Sets. Supplementary Figure 2. a) Visual flow diagram of method used to find CpG groupings and recapitulation of DNAm profiles. First, the 300 k CpGs are projected into a 6-dimensional embedding using UMAP. Each point in the low dimensional space represents a CpG and proximity between points denotes a shared methylation profile across all of the training samples (n = 503). Then, KMeans clustering was used to find 25 clusters of CpGs with similar profiles. The number of clusters of CpG features were reduced to 8 by filtering out clusters if their variance was above 1 in the 6D space. After that, the CpG features found in each cluster were used to select CpGs to form independent MethylationArrays across the training, validation and test sets. Finally, one autoencoder was trained per each array and the test samples were recapitulated and compared to the original input data; b) Descriptive statistics for final groupings of CpGs and recapitulation scores for each resultant set of CpGs versus the original methylation profiles input into each model. Supplementary Figure 3. Generated/recapitulated beta values versus original beta values for each CpG per individual of the held-out test set (n = 144); b)-f) corresponds to each of eight chosen clusters in order of low to high cluster variance as previously described; a) is an aggregation of the generated/recapitulated versus true beta values of all of the CpG clusters. Supplementary Figure 4. Hierarchically clustered cosine distance matrix between test samples’ VAE-embedded methylation profiles of the held-out test set for the TCGA cohort, colored by: a) Labels assigned to the hierarchical clustering labels for the samples; b) Original TCGA cancer labels; c) RPMM-derived clustering labels on 20 k CpGs. Agreement scores between the RPMM and hierarchal clustering results and the original cancer subtypes were calculated using the v-measure, which takes into account the homogeneity and completeness of the labeling. Note that the clustering colors are not the same because the number of clusters is different from the number of cancer labels. Supplementary Table 3. MethylNet Results on Training (n = 503) and Validation (n = 72) Sets for Sample Age Prediction. Supplementary Table 4. MethylNet Results on Training (n = 503) and Validation (n = 72) Sets for Cell Type Deconvolution. Supplementary Table 5. MethylNet Results on Training (n = 5860) and Validation (n = 840) Sets for Pan-Cancer Classification. Supplementary Table 6. Tukey’s Studentized Range Tests for identifying which of six hierarchical clusters from unsupervised VAE embeddings differ in cellular proportions; VAE embeddings derived from blood test dataset. Supplementary Table 7. Correlation of MethylNet Cellular Proportions to Other Estimation Methods’ Proportions. Supplementary Figure 5. Proportion of IDOL CpGs that are Overlapped by the Top 1 k CpGs for Each Cell-Type. Supplementary Figure 6. Bar Charts of CpGs with the 10 Largest Shapley Scores for Each Cell-Type, linked by red, blue or green lines if shared across subtype for: a-d) Lymphocytes; e-f) Myeloids. Not sharing top 10 CpGs does not indicate that two cell-types do not share similar CpG profiles. Supplementary Figure 7. Bar Charts of CpGs with the 10 Largest Shapley Scores for Age Groups: a) 14–24 and b) 84–94. These CpGs are linked if shared across the age groups, but this does not indicate that they are not shared outside of this top 10 list of CpGs. The top 10 CpGs that are associated with lower age are similar to the older age group; c) Distribution of Shapley Scores for these two age groups. CpG contributions tend to be negative for the younger age groups and positive for the older age groups. Supplementary Table 8. Confusion Matrix Pan-cancer Classification (Colored Superclass). Supplementary Table 9. Breakdown Pancancer Classification Results (Colored by Superclass). Supplementary Table 10. Average Cosine Distance Between Embeddings of Cancer Subtypes Supplementary Figure 8. Micro F1-Scores of the held-out test samples (n = 1676) of the TCGA cohort as they relate to: a) the fraction of training samples included for the training process, b) the number of CpGs. Test performance scales linearly with the number of training samples and logarithmically with the number of CpGs. Confidence intervals were calculated using a 1 k nonparametric bootstrap of the test results for each dataset size point in the line plot, and the resulting bootstrapped f1-scores were used to compute the confidence interval for each point in the line plots; c) performance of MethylNet, pretrained using a VAE, is compared to performance using an MLP with the same architecture; F1-Score confidence intervals were derived using a 1 k nonparametric bootstrap; validation loss for each model is compared at the first training epoch and their ultimate convergence point. Supplementary Figure 9. V-Measure scores of the held-out test samples (n = 1676) of the TCGA cohort as they relate to: a) the fraction of training samples included for the training process, b) the number of CpGs. V-measure scores were derived by applying and comparing hierarchical clustering on the VAE embeddings to known cancer subtype assignments and using a knee point detection algorithm to identify the ideal number of clusters for each tested dataset. In this figure, scores were smoothed using an exponential moving average smoothing technique to illustrate general trends. Supplementary Figure 10. Smoking EWAS study via MethylNet: a) final embeddings derived when finetuning the MethylNet VAE demonstrates cluster separation of the “never” versus “current” smokers; b) confusion matrix for the true and predicted “never” versus “current” smokers; c) plotted average ranks found using SHAP for the CpGs that intersected with CpGs identified by Liu et al. versus the ranks of those corresponding p-values of the EWAS meta-analysis. Supplementary Table 11. Preliminary Results for PAM50 Breast Tumor Classification (n = 1018); 95% confidence intervals of scores estimated via 1000-sample non-parametric bootstrap. Supplementary Figure 11. Internal and External Validation Cohorts: a) Boxenplot demonstrating distribution of ages for internal and external cohorts; notice how age for external validation cohort is greater than that of the internal validation cohort; b) plotted MethylNet predicted age versus actual age. Supplementary Table 12. Select Hyperparameters for Embedding Tasks. Supplementary Table 13. Select Hyperparameters for Prediction Tasks. Supplementary Figure 12. Fine-tuned embeddings (few parts highlighted) for: a) Age Prediction, b) Cell-Type Deconvolution (colored by Neutrophil Cell Type Proportions), and c) Pan-Cancer Classification (Labeled kidney cancers, lung cancers, brain cancers). Supplementary Figure 13. Model Training Curves for a) Age Predictions, b) Cell-Type Predictions, c) Pan-Cancer Predictions. Please note that the learning rates for the prediction curve of a) oscillates quickly every 10 training epochs as compared to a larger timescale.

Additional file 3: Table S3. Differentially Methylated Regions with p-values

Additional file 5: Table S7. Results of Basic Model and Child BMI Model for CpGs with p-values

Additional file 2. Supplemental tables and figures.

Figure S1. Forest plot for the top 10 Bonferroni-significant CpGs from the meta-analysis on the association between continuous GA and offspring DNA methylation at birth adjusted for estimated cell proportions. Figure S2. Sensitivity analysis: Correlation of the point estimates for the no complications model main association of DNA methylation with gestational age (y-axis representing 3648 participants from 17 cohorts) with point estimates for a meta-analysis after excluding three cohorts (MoBa1, MoBa2 and ALSPAC) that were included in a previous publication1,2 (x-axis representing 2190 participants from 14 cohorts). Figure S3. Correlations between methylation and gene expression levels for selected four pairs. First, we created residuals for mRNA expression and residuals for DNA methylation and used linear regression models to evaluate correlations between expression residuals and methylation residuals. These residual models were adjusted for covariates, estimated white blood cell proportions, and technical variation. Figure S4. Sensitivity analysis: Correlation of the point estimates for the no complications model main association of DNA methylation with gestational age (y-axis representing 3648 participants from 17 cohorts) with point estimates for a meta-analysis after excluding Non-European three cohorts (CBC, CHS and CHAMACOS) (x-axis representing 3290 participants from 14 cohorts).

Supplementary information.

Additional file 6: Table S9. Significantly enriched GO terms among CpGs with p-values

Additional file 4: Table S5. CpGs with p-values

Birthweight is associated with health outcomes across the life course, DNA methylation may be an underlying mechanism. In this meta-analysis of epigenome-wide association studies of 8,825 neonates from 24 birth cohorts in the Pregnancy And Childhood Epigenetics Consortium, we find that DNA methylation in neonatal blood is associated with birthweight at 914 sites, with a difference in birthweight ranging from -183 to 178 grams per 10% increase in methylation (PBonferroni

SummaryThe ability to perform high-throughput preprocessing of methylation array data is essential in large scale methylation studies. While R is a convenient language for methylation analyses, performing highly parallelized preprocessing using Python can accelerate data preparation for downstream methylation analyses, including large scale production-ready machine learning pipelines. Here, we present a methylation data preprocessing pipeline called PyMethylProcess that is highly reproducible, scalable, and that can be quickly set-up and deployed through Docker and PIP.Availability and ImplementationProject Name: PyMethylProcessProject Home Page:https://github.com/Christensen-Lab-Dartmouth/PyMethylProcess. Available on PyPI aspymethylprocess.Available on DockerHub viajoshualevy44/pymethylprocess.Help Documentation:https://christensen-lab-dartmouth.github.io/PyMethylProcess/Operating Systems: Linux, MacOS, Windows (Docker)Programming Language: Python, ROther Requirements: Python 3.6, R 3.5.1, Docker (optional) License: MITContactjoshua.j.levy.gr@dartmouth.edu

Figure S1 Correlations between age and fraction of cells with FCO signal in different types of normal tissues on TCGA. Figure S2 Correlations between monocyte infiltration percentage and fraction of cells with FCO signal in different types of tumors on TCGA. Figure S3 Correlations between lymphocyte infiltration percentage and fraction of cells with FCO signal in different types of tumors on TCGA. Figure S4 Correlations between neutrophils infiltration percentage and fraction of cells with FCO signal in different types of tumors on TCGA. Figure S5 The distribution of tumor purity across different types of tumors on TCGA. Figure S6 Correlations between tumor purity and fraction of cells with FCO signal in different types of tumors on TCGA. Figure S7 The FCO signal decreases as tumor stage increases in kidney renal clear cell carcinoma. Figure S8 Methylation status of EZH2 related CpGs from FCO library in normal pancreatic tissue, pancreatic carcinoma and pancreatic carcinoma stem cell. Figure S9 Normal QQ-plots showing the distribution of residuals from linear regression fits in TCGA tumor projects. Figure S10 Spread-Location plots showing the spread of residuals along the ranges of predictors from linear regression fits in TCGA tumor projects. Table S1 P-values based on comparisons of the predicted FCO (%) and tumor purity after adjusting for age, gender, race and vital status using multiple linear regression models across different TCGA studies. Table S2 FCO in pancreatic ductal adenocarcinoma stem cells from GEO data set GSE80241 and glioma stem cells from GEO data set GSE92462. (DOCX 2395 kb)

Figure S1. Box plots of estimates and FACS counts. Cell type proportions generated by FACS (true values) and CP/QP programming (estimates) using IDOL and pickCompProbes L-DMRs for each UCB reference individually and combined. (PDF 110 kb)

Figure S1. Details of SVM BRCA1-like classifier. (A) Overview of copy number mapping algorithm for generating the input for training the SVM BRCA1-like classifier. (B) Receiver-operation characteristic curves (ROC) of the classifier applied to training and test set (AUC = 1.00 and 0.75, respectively). (C-D) Correlation of SVM BRCA1-like probability scores with published HR-deficiency metrics (HRD-LOH and LST scores). ***P T) deamination events in TCGA hormone-receptor-positive breast tumors. A P value indicates statistical significance from linear model adjusting for age, tumor stage, and ER, PR, and HER2 positivity. **P

Table S4. Top 10 pathways for 803 unique genes associated with top 5% CpGs based on P value in the model without cell type adjustment. (DOCX 16Â kb)

Table S4. Top 10 pathways for 803 unique genes associated with top 5% CpGs based on P value in the model without cell type adjustment. (DOCX 16Â kb)

Epigenetic modifications, including DNA methylation, represent a potential mechanism for environmental impacts on human disease. Maternal smoking in pregnancy remains an important public health problem that impacts child health in a myriad of ways and has potential lifelong consequences. The mechanisms are largely unknown, but epigenetics most likely plays a role. We formed the Pregnancy And Childhood Epigenetics (PACE) consortium and meta-analyzed, across 13 cohorts (n = 6,685), the association between maternal smoking in pregnancy and newborn blood DNA methylation at over 450,000 CpG sites (CpGs) by using the Illumina 450K BeadChip. Over 6,000 CpGs were differentially methylated in relation to maternal smoking at genome-wide statistical significance (false discovery rate, 5%), including 2,965 CpGs corresponding to 2,017 genes not previously related to smoking and methylation in either newborns or adults. Several genes are relevant to diseases that can be caused by maternal smoking (e.g., orofacial clefts and asthma) or adult smoking (e.g., certain cancers). A number of differentially methylated CpGs were associated with gene expression. We observed enrichment in pathways and processes critical to development. In older children (5 cohorts, n = 3,187), 100% of CpGs gave at least nominal levels of significance, far more than expected by chance (p value < 2.2× 10(-16)). Results were robust to different normalization methods used across studies and cell type adjustment. In this large scale meta-analysis of methylation data, we identified numerous loci involved in response to maternal smoking in pregnancy with persistence into later childhood and provide insights into mechanisms underlying effects of this important exposure.

Background: DNA methylation regulates gene transcription. Hypomethylation has been linked to bladder carcinogenesis. Trihalomethanes (THM), a class of disinfection by-product in chlorinated water, ...

Additional file 2: Supplementary Methods. Study-specific funding, acknowledgements and methods in alphabetical order, including references.

Additional file 2: Table S1. CpG with p-values

Table S1. Top hypo- and hyper-methylated CpGs in CF that are associated with known genes. (DOCX 17Â kb)

Table S1. Top hypo- and hyper-methylated CpGs in CF that are associated with known genes. (DOCX 17Â kb)

Table S2. Targeted next-generation bisulfite sequencing of EPIC identified genes. (DOCX 15Â kb)

Table S3. Summary of genomic context related with differentially methylated loci in CF. DHS, DNase I hyper-sensitivity sites. (DOCX 17Â kb)

Table S3. Summary of genomic context related with differentially methylated loci in CF. DHS, DNase I hyper-sensitivity sites. (DOCX 17Â kb)

Additional file 5: Supplementary Information. Figures S1 – 6.

Additional file 5: Supplementary Information. Figures S1 – 6.

Figure S2. Box plots of absolute errors. Absolute errors (estimates minus FACS counts) per method (IDOL and pickCompProbes) for each UCB reference individually and combined. (PDF 108 kb)

Additional file 1: Figure S1a. Epigenome-wide Association Study Results of DNA Methylation in Cord Blood with Liver Fat Fraction in Children. Figure S1b. Epigenome-wide Association Study Results of DNA Methylation in Child Peripheral Blood with Liver Fat Fraction in Children.

Additional file 1: Figure S1a. Epigenome-wide Association Study Results of DNA Methylation in Cord Blood with Liver Fat Fraction in Children. Figure S1b. Epigenome-wide Association Study Results of DNA Methylation in Child Peripheral Blood with Liver Fat Fraction in Children.

Table S2. Targeted next-generation bisulfite sequencing of EPIC identified genes. (DOCX 15Â kb)

Additional file 2: Table S1. CpG with p-values

Additional file 2: Supplementary Methods. Study-specific funding, acknowledgements and methods in alphabetical order, including references.