R and Deseq2

Question 1

Name some benefits of using R

Answer 1

• opensource
• reproducible research w R markdown
• huge community of developers
• custom packages are available

Question 2

What are the uses of .R projects?

Answer 2

• links all files and outputs to project dir
• imported data looked for in proj dir instead of specifying file path
• can save environment in .RData and reload workspace where you left it

Question 3

What are some challenges when trying to identify differentially expressed genes? (3)

Answer 3

• distinguish technical variation from treatment variation (ex: technical factors that can’t be controlled in library prep)
• majority of genes don’t change b/w treatments (hard to perform stats and get significant)
• only few replicates per treatment, hard to estimate variance (not feasible for price, limited material, and experiment execution)

Question 4

What is count normalization?

Answer 4

The determination of size factors to account for/normalize differences in sample sequencing depth, gene length, and RNA composition

Question 5

Why can’t you use total sampling depth to normalize counts?

Answer 5

In highly expressed genes, taking ratios doesn’t reflect the actual expression. Two genes might be expressed at the same levels, but taking ratios might make them appear at a lower level

Question 6

Why should count normalization be done?

Answer 6

the numerical value of non-differentially expressed genes should not vary due to sampling depth or RNA composition. We need to determine a sample-specific size factor for each sample

This is needed to make accurate comparison of gene expression between samples

Question 7

What are the steps in DESeq2 count normalization? (6)

Answer 7

1. determine natural log of genes counts
2. calculate the geometric mean of each row to use a pseudo-reference sample
3. Remove infinite values
4. subtract geometric mean from log of counts (subtract reference from log of counts which is equivalent to lof ratio of counts to reference)
5. calculate median for each sample
6. convert log of median to number

Question 8

What are 2 limits to FDR-controlling procedures? & what are the solutions?

Answer 8

1. multiple testing causes false positives
2. when FDR correct, the more negatives, the more false negatives

solutions:

• low-expressed genes variance can’t be estimated
• remove low-expressed genes

Question 9

What is the probability distribution if a gene is not differentially expressed in 2 different conditions?

Answer 9

The samples come from the same distribution

the probability distribution is uniform from 0 to 1

Question 10

What is the probability distribution if a gene is differentially expressed in 2 different conditions?

Answer 10

Samples come from 2 different distributions
probability distribution is skewed towards 0
most samples below 0.05

Question 11

What are the p values for true positives and false negatives?

Answer 11

true positives : 0 - 0.05

false negatives: > 0.05

Question 12

What is the Benjamini-hochberg method?

Answer 12

A method to control the FDR and account for the fact that sometimes p-values less than 0.05 happen by chance.
It adjusts p-values by making them larger
It ensures that the false positives never make up more than 5% of all positives

Question 13

In independent filtering how is the filter threshold calculated?

Answer 13

filter threshold = max of fit curve - 1SD

Question 14

How does the Benjamini-hochberg method work?

Answer 14

It sets the p-value to which ever is the lower value of
1. the p-value of the next higher rank (after ranking p-values from lowest to highest); p-value(rank+1)

2. p-value(rank) * [total# of p-values]/rank

Question 15

What happens to the numbers of true positives after the Benjamini-hochberg method is applied?

Answer 15

fewer true positives are identified

Question 16

What does FDR do?

Answer 16

Limit the number of false negatives reported, but lose some true positives

Question 17

Why are genes with low read count very noisy?

Answer 17

small changes will have dramatic changes on the calculated fold change; remove these out of dataset

Question 18

What does DESeq2 do to counteract loss of true positives when there are lots of non-differentially expressed genes?

Answer 18

Removes tests that are unlikely to show significant differences and determines which quantiles return the largest number of rejections

Question 19

What is independent filtering?

Answer 19

The removal of genes with very low counts; Maximizes the number of positives

Question 20

What is gene dispersion?

Answer 20

The variance of the gene

Question 21

What are the steps in independent filtering?

Answer 21

Genes with low counts removed (sample mean > filter threshold)

1. determine significant genes for different threshold (expressed as quantiles) and lot of significant genes vs quantities
2. fit curve
3. determine filter threshold

Question 22

Describe properties of read counts (4)

Answer 22

1. sparse events (i.e. small likelihood p of a read mapping to a specific gene) = read count of a given gene likely small
2. discrete and high number of events n( (sampling depth/# of reads)
3. model raw counts with poison distribution (mean = variance)

Question 23

What distribution is suitable for a large sampling depth (n) and a very small number for p?

Answer 23

poissson

important property: mean = variance

Question 24

what does over-dispersed mean? and what is this caused by?

Answer 24

when the variance of the data increases faster than the mean.
caused by biological variation

Question 25

What factors can make variance difficult to deal with?

Answer 25

• only few replicates/conditions make it hard to estimate variance
• at low expression levels, data is very noisy

Question 26

How does DESeq2 deal with challenges with variation? (3)

Answer 26

• Models count matrix using negative bionomical distribution
• borrows info across genes to estimate dispersion and ultimately variance (to determine variance at an expected mean expression); can calculate dispersion for each gene
• determines the log2fold change and its significance (wald statistic)

Question 27

What is negative binomial distribution?

Answer 27

distribution with overdispersion
K = raw count of gene i in sample J with fitted mean u and gene disperation a

a = value to estimate to determine gene variance
K<i> ~NB(u<i>, a<i>)</i></i></i>

u<i> = sq<i></i></i>

s = sample specific size factor
q = number of fragments in sample (expression level)</i></i></i></i></i>

Question 28

What is the formula for variance?

Answer 28

Var(K<i>) = u<i> + a<i>(u<i> )^2</i></i></i></i>

variance of each gene fitted to line which is assumed to be the true dispersion for any given mean

a<i> used to determine log2 fold change and error of that estimate</i></i></i></i></i>

Question 29

On a plot of estimating gene-wise dispersion what happens to genes that are too far away from the best fit line?

Answer 29

the variance of those genes won’t be adjusted

Question 30

describes the steps in estimating gene-wise dispersion?

Answer 30

1. gene dispersion estimates done on their own
2. information from all genes used to created a best ifr line to give general estimate of what to expected for a given mean
3. estimates of individual genes are shrunk towards the best fit line