Supplementary Tables and Figures

Supplementary Table 1. Roadmap Epigenomics Project (REP) biological and technical replicate counts.

Technical Replicates
ID Tissue Tissue Short Name Sources WGBS RNA-seq
E058 Penis Foreskin Keratinocyte Keratin. 1 2 3
E065 Aorta Aorta 1 6 2
E066 Adult Liver Liver 3 2 2
E071 Brain Hippocampus Middle Hippo. 2 3 2
E079 Esophagus Esoph. 1 2 2
E094 Gastric Gastric 1 5 3
E095 Left Ventricle L. Ventr. 1 4 2
E096 Lung Lung 1 2 2
E097 Ovary Ovary 1 2 1
E098 Pancreas Pancreas 1 2 2
E100 Psoas Muscle Psoas 2 3 3
E104 Right Atrium R. Atrium 1 3 1
E105 Right Ventricle R. Ventr. 2 5 2
E106 Sigmoid Colon Colon 2 2 3
E109 Small Intestine Intest. 2 4 3
E112 Thymus Thymus 1 2 1
E113 Spleen Spleen 1 3 3
Supplementary Table 2. Differentially expressed and ME-Class interpolated gene counts from Roadmap
Epigenomics Project (REP).
Supplementary Figure 1. Increasing the number of RF estimators from 100 to 1000 for the ROI classifier
does not substantially increase performance as evaluated by: a) accuracy versus 1-reject rate, b)
precision versus recall (PR AUC; 100 estimators: 0.70, 1000 estimators: 0.71), and c) ROC curve (ROC
AUC; 100 estimators: 0.72, 1000 estimators: 0.73).
Supplementary Figure 2. Alternative full-gene methylation representations do not outperform TSS-
centric representations. a) Heat map indicates methylation status at individual CpG sites – red is fully
methylated, blue is fully unmethylated – for an example gene in two samples (Methyl. 1 and Methyl. 2).
Individual points below indicate differential DNA methylation (Methyl. 2 – Methyl. 1) across the example
gene at individual CpG sites. b) Whole Scaled Gene (WSG), c) Whole Gene (WG) and d) Uniform Gene
Features (UGF) representation of the gene in (a). See additional description of each method in the
Materials and Methods. Performance plots of TSS, WSG, WG, and UGF as reported by: e) accuracy
versus 1-reject, f) ROC curve (ROC AUC; TSS: 0.76, WSG: 0.75, WG: 0.70, UGF: 0.65), and g) precision
versus recall (PR AUC; TSS: 0.75, WSG: 0.74, WG: 0.69, UGF: 0.65). CGI = CpG island.
Supplementary Figure 3. Additional evaluation metrics for each method using 17 REP tissue differential
samples: a) positive predictive value (PPV) versus 1- reject rate, b) negative predictive value (NPV)
versus 1- reject rate, c) precision versus recall (PR AUC; ME-Class: 0.75, ROI: 0.70, DMR: 0.63, SW:
0.66) and d) accuracy versus the classifier’s probability of prediction.
Supplementary Figure 4: Metagene plots of genes identified by ME-Class in REP data at different
probabilities of prediction p. Blue curves represent the average Z-score normalized methylation difference
between each sample for downregulated genes while red curves represent the average for upregulated
genes.
Supplementary Figure 5. ME-Class outperforms classifiers using REP data based on only the most
important methylation features, [+0.5kb, +2.5kb] around the TSS, as evaluated by: a) accuracy versus 1-
reject rate, and b) ROC curve (ROC AUC; ME-Class: 0.76, RF-Most Imp. Feat.: 0.74, SW: 0.67, SW-Most
Imp. Feat.: 0.64). RF-Most Imp. Feature is an ME-Class like classifier built using features from only the
region [+0.5kb, +2.5kb] around the TSS. SW-Most Imp. Feat. is similar to the SW approach, but only
using methylation from [+0.5kb, +2.5kb] around the TSS.
Supplementary Figure 6. The addition of methylated CpG density and gene body features (GF) does not
increase ME-Class performance. a) Performance plots of ME-Class altered to use either mCG/CG,
mCG/bp, or CpG density (200bp resolution, CG/bp) as input. (PR AUC; mCG/CG: 0.75, mCG/bp: 0.75,
CG/bp: 0.50; ROC AUC; mCG/CG: 0.75, mCG/bp: 0.75, CG/bp: 0.50) b) Performance plots of ME-Class
with and without adding gene body features (GF) from the ROI classifier including average internal exons,
introns, and downstream features. ROI features are in Fig. 1d. (PR AUC; ME-Class: 0.75, ME-Class+GF:
0.76, ROI: 0.70; ROC AUC; ME-Class: 0.76, ME-Class+GF: 0.76, ROI: 0.72).
Supplementary Figure 7. CpG-poor genes are more predictive of expression classification. ME-Class
performance for genes overlapping a CpG Island (CGI) by >=1bp is reported as: a) accuracy versus 1-
rejection rate and b) ROC curve analysis (ROC AUC: No-CGI: 0.79; CGI: 0.75) c) Histogram of all genes
with complete start and stop annotation according to RefSeq (n=19,175). Low CpG density genes
comprise 26.0% (4,977 genes) while high CpG density genes comprise 74.0% (14,198 genes). d)
Histogram of differentially expressed RefSeq genes (n=12,064 genes), where low CpG density genes
comprise 18.8% (2,265 genes) while high CpG density genes comprise 81.2% (9,799 genes). e)
Histogram of differentially expressed, interpolated RefSeq genes (n=10,524 genes) after applying our
filtering parameters (see Materials and Methods). Low CpG density genes comprise 17.5% (1,842 genes)
while high CpG density genes comprise 82.5% (8,681 genes). ME-Class performance is reported as: f)
accuracy versus 1-rejection rate and g) ROC curve analysis (ROC AUC: CpG-poor: 0.8; CpG-rich: 0.75)
Cutoff between low and high CpG density genes at 0.35 observed/expected normalized CpGs +/-1500bp
of TSS. ME-Class performance with or without added feature of observed/expected normalized CpG
density +/-1500bp of TSS is reported as h) accuracy versus 1-rejection rate and i) ROC curve analysis
(ROC AUC: ME-Class: 0.76; ME-Class, CpG Density: 0.76).
Supplementary Figure 8. Random Forest classifier performs similarly or outperforms alternatives based
on classification performance as measured by a) ROC curve (ROC AUC; RF: 0.76, LR: 0.76, GBCT: 0.76,
DTW-kNN: 0.73, L2-kNN: 0.73, Naïve Bayes: 0.71) and b) accuracy versus 1-reject rate. LR = Logistic
Regression, GBCT= Gradient Boosted Classification Trees, DTW-kNN = Dynamic Time Warping based k-
Nearest Neighbor, L2-kNN = Euclidean distance (L2) based k-Nearest Neighbor.
Supplementary Figure 9. Effect on ME-Class performance of tuning parameters for smoothing, bin
resolution, and interpolation. Performance is reported as: accuracy versus 1-reject, precision versus
recall, and ROC curve. a) Relationship between ME-Class performance and sigma for Gaussian
smoothing with a constant bin resolution of 20bp. b) Relationship between ME-Class performance and
the size of the bin resolution at a constant sigma of 50bp. c) Relationship between ME-Class performance
and alternative interpolation method (PR AUC; PCHIP: 0.76, Linear: 0.76; ROC AUC; PCHIP: 0.76,
Linear: 0.76).
Supplementary Figure 10. Number of training samples and genes determine ME-Class performance.
The testing ROC AUC as a function of a) the number of training samples. The fraction of correctly
identified genes using ME-Class with 9 evaluation samples as a function of b) the number of training
samples and c) the total number of training genes from the training samples. In (a) and (b), each point
indicates the performance across all genes in an individual sample comparison. In (c), each point
indicates the number of training genes and fraction of genes returned for an individual sample
comparison and set of training samples. Training genes in (c) means the number of genes summed
across all training samples. A gene can be counted multiple times if it shows up in multiple samples.
Although it will likely have different methylation profiles and expression values in each comparison.
Permuted sets of all differential training samples (n=8) and a fixed set of differential evaluation samples
(n=9) are randomly chosen from the REP dataset.
Supplementary Figure 11. Performance of Blueprint neutrophil samples in comparison to other
hematopoietic cell types. ME-Class is trained from the full REP dataset and performance is reported as:
a) accuracy versus 1- reject rate and b) ROC curve analysis (ROC AUC; Lymphoid: 0.65, Megakayrocyte:
0.63, Erythoblast: 0.62, Other Myeloid: 0.55).
Supplementary Figure 12. Performance of Blueprint Epigenome samples using a similar leave-one-out
differential sample evaluation cross-validation framework as used for the REP data (see Fig. 1e). The
performance of ME-Class trained and evaluated solely using Blueprint samples is similar to that of a ME-
class model trained from the REP dataset. Shown are ROC AUC of differential comparisons of randomly
chosen single samples of each of the 14 cell types from Blueprint dataset.