Chapter 2 Flashcards

Question

Approximately how many cross-hybridizing probes are present on the EPIC array and how are they defined?

Answer 1

~43K, >47bp homology to an off-target site

Answer 2

Kept polymorphic and sex probes due to clonal cell line

Answer 3

wateRmelon defaults (detP: probes with >1% of samples with detP > 0.05 [1 sample in LUHMES]; bead count: probes with >5% of samples with bead count < 3 [2 samples in LUHMES])

Answer 4

wateRmelon pfilter: samples where >5% of probes have detP > 0.05 (up from default 1% because of oxBS conversion; none flagged)

Answer 5

Was used to avoid BMIQ, which assumes a 3-state beta-mixture model (unmethylated, methylated, hemimethylated) that is not true when considering mixed mC/hmC signal in BS data from neurons. dasen: correct for difference in type I/II intensities, within M/U. d = type I/II intensity difference bg adjustment, s = quantile normalization of M/U intensities applied separately to type I/II probes, n = no dye bias correction Applied separately to BS/oxBS data in LUHMES dataset.

Answer 6

PCA = a form of dimension reduction, transforming data into PCs that explain the overall variance. 1. beta values for each CpG are centered and scaled (transformed so they are on the same range) 2. covariance matrix constructed to ID correlations between CpGs 3. linear combinations of uncorrelated CpGs (PCs) are calculated from covariance matrix (eigenvector = linear combination, eigenvalue = eigenvector's coefficient (% explained variance)) In DNAm QC, can cluster samples along PCs to identify patterns and outliers, and can correlate PCs with known variables to identify biological/technical signal in the data.

Answer 7

empirical Bayes framework used to adjust data for batch effects. Used M-values in LUHMES since they have a more constant stdev across a range of values than betas do (homoskedastic), so L/S estimate may be more accurate. 1. standardize M-value mean/variance across CpGs 2. assume a model for mean and variance of M-values within each batch 3. adjusts each batch to meet model specifications such that mean and variance are standardized across the batches 4. assuming batch effect is similar across CpGs, mean/variance estimates within a batch are shrunk toward the mean across all batches

Answer 8

PCs 1-3 correlated with chip, chip position, collection date, and passage. Collection date and passage were related to each other so just chose one (passage) to correct. Sequentially correction from the most to least correlation with DNAm PCs, separately in BS/oxBS (chip or row first, then passage).

Answer 9

After separately probe filtering, normalization, and batch correction of BS/oxBS data. 95% quantile of negative hmC values was 0.036.

Answer 10

The same model is fit to each CpG, but information about mean/variance across all CpGs is "borrowed" such that mean/variance estimates for individuals CpGs are shrunk toward the average. This results in a more robust estimation of variance when sample size is small. The variance estimations for each CpG are weighted, so more variable and less variable CpGs can be tested separately.

Answer 11

Filtering the CpGs prior to limma could result in a biased/artifically high estimation of variance across CpGs, affecting the eBayes prior/shrinkage.

Answer 12

A moderated t-test is applied at each CpG, which borrows variance estimation across CpGs. BH adjustment applied.

Answer 13

1. P-values ranked in ascending order (smallest to largest) 2. p-value rank/# of tests x FDR = BH critical value 3. the largest p-value smaller than the corresponding critical value is considered significant, and all p-values smaller than this

Answer 14

In a DNAm titration experiment on the 27K array (Du et al. 2010), 0.05 beta value cutoff achieved good tradeoff between TP/FP rate. It is also larger than technical noise on the array (typically ~2-3%).

Answer 15

Hypergeometric test (one-tailed Fisher's exact): probability of drawing k or more successes without replacement from a population

Answer 16

Algorithm removes the most multifunctional genes and recalculates GO enrichments, iteratively until either enrichments are maximally sensitive to gene removal or until half of genes have been removed. FDR correction is applied to results.

Answer 17

1. Cross-link chromatin and DNA with formaldehyde in culture dishes (create covalent bonds) 2. Shear DNA 3. IP with antibody to H3K4me1 4. Use magnetic beads to extract DNA fragments 5. Reverse crosslinking 6. Library preparation (adapter ligation, PCR amplification) and sequencing

Answer 18

1. Quality check with FastQC 2. Aligned to hg38 (not hg19) using Bowtie2 3. Filtered for MAPQ > 30? This is typically done and was done for hippEE project

Answer 19

Broad peaks called using input (10% of sample after IP) as control. MACS models the peak locations based on chromatin-DNA contact sites (will be in between the tag locations of forward/reverse strands), and finds enriched tags relative to input by sampling 1000X and averaging distance between tags. Consensus peaks present in 2 or more replicates

Answer 20

1. RNA extraction (3 bio reps used for LUHMES) 2. Reverse transcription into cDNA 3. Adapter ligation 4. PCR amplification 5. Sequencing

Answer 21

- FastQC used to check sequencing quality (PCR bias, read quality, etc) - Both data were aligned to the reference genome using STAR, with default parameters and allowing gapped alignment to account for differential splicing. - Quality after alignment was checked with samtools and visualization in a genome browser tool in LUHMES and mice - DESeq2 was used to normalize read counts to library size in both experiments - LUHMES data was filtered for genes with mean read count > 10, mouse was filtered for mean read count > 50 - LUHMES DEGs were considered at padj < 0.05 and log2FC > 0.5, while mouse DEGs were considered at padj < 0.15 and log2FC > 0.3 (citing variability and small effect sizes in mouse brain RNA studies...)

Answer 22

1. Input matrix of library size-normalized, filtered read counts 2. Bayesian estimation of dispersion (variation in read counts between replicates) for each gene, shrunk toward the mean dispersion across the experiment 3. GLM + Wald test for differential expression at each gene (does including the covariate significantly contribute to the model?), accounting for experiment-wide dispersion factor 4. BH multiple test correction

Chapter 2 Flashcards

(46 cards)