To quantify the amount of variation in DNA methylation explained by genomic context, we considered the correlation between genomic context and principal components (PCs) of methylation levels across all 100 samples (Figure 4). We found that many of the features derived from a CpG site’s genomic context appear to be correlated with the first principal component (PC1). The methylation status of upstream and downstream neighboring CpG sites and a co-localized DNAse I hypersensitive (DHS) site are the most highly correlated features, with Pearson’s correlation r=[0.58,0.59] (P<2.2?10 ?16 ). Ten genomic features have correlation r>0.5 (P<2.2?10 ?16 ) with PC1, including co-localized active TFBSs ELF1 (ETS-related transcription factor 1), MAZ (Myc-associated zinc finger protein), MXI1 (MAX-interacting protein 1) and RUNX3 (Runt-related transcription factor 3), and co-localized histone modification trimethylation of histone H3 at lysine 4 (H3K4me3), suggesting that they may be useful in predicting DNA methylation status (Additional file 1: Figure S3). 67,P<2.2?10 ?16 ) [53,54].
Correlation matrix out-of anticipate provides that have first 10 Pcs regarding methylation membership. New x-axis represents among the many 122 features; the newest y-axis represents Pcs 1 through 10. Tone match Pearson’s relationship, as the found on legend. Pc, dominant parts.
Binary methylation reputation forecast
These observations about patterns of DNA methylation suggest that correlation in DNA methylation is local and dependent on genomic context. Using prediction features, including neighboring CpG site methylation levels and features characterizing genomic context, we built a classifier to predict binary DNA methylation status. Status, which we denote using ? we,j ? <0,1>for i ? <1,...,n> samples and j ? <1,...,p> CpG sites, indicates no methylation (0) or complete methylation (1) at CpG site j in sample i. We computed the status of each site from the ? we,j variables: \(\tau _ = \mathbb <1>[\beta _ > 0.5]\) . For each sample, there were 378,677 CpG sites with neighboring CpG sites on the same chromosome, which we used in these analyses.
Hence, forecast out-of DNA methylation status founded simply towards methylation account on nearby CpG websites may not work well, particularly in sparsely assayed aspects of the fresh genome
The 124 have that people used for DNA methylation standing forecast end up in four additional groups (pick A lot more document step one: Table S2 to own a whole record). For every CpG site, we range from the adopting the feature sets:
neighbors: genomic ranges, binary methylation condition ? and you may membership ? of a single upstream and you to definitely downstream nearby CpG website (CpG sites assayed into variety and you can adjoining in the genome)
genomic status: digital opinions proving co-localization of your CpG website with DNA sequence annotations, plus promoters, gene human anatomy, intergenic area, CGIs, CGI shores and you may cupboards, and you may nearby SNPs
DNA series qualities: continued philosophy representing your regional recombination price off HapMap , GC posts from ENCODE , integrated haplotype results (iHSs) , and genomic evolutionary rates profiling (GERP) phone calls
cis-regulatory factors: digital opinions proving CpG site co-localization having cis-regulatory issues (CREs), and additionally DHS sites, 79 particular TFBSs, 10 histone amendment marks and 15 chromatin states, the assayed throughout the GM12878 telephone range, the latest nearest fits so you can whole bloodstream
We used a RF classifier, which is an ensemble classifier that builds a collection of bagged decision trees and combines the predictions across all of the trees to produce a single prediction. The output from the RF classifier is the proportion of trees blackchristianpeoplemeet in the fitted forest that classify the test sample as a 1, \(\hat <\beta>_\in [0,1]\) for i=<1,...,n> samples and j=<1,...,p> CpG sites assayed. We thresholded this output to predict the binary methylation status of each CpG site, \(\hat <\tau>_ \in \<0,1\>\) , using a cutoff of 0.5. We quantified the generalization error for each feature set using a modified version of repeated random subsampling (see Materials and methods). In particular, we randomly selected 10,000 CpG sites genome-wide for the training set, and we tested the fitted classifier on all held-out sites in the same sample. We repeated this ten times. We quantified prediction accuracy, specificity, sensitivity (recall), precision (1? false discovery rate), area under the receiver operating characteristic (ROC) curve (AUC), and area under the precision–recall curve (AUPR) to evaluate our predictions (see Materials and methods).
