Skip to content

Latest commit

 

History

History
67 lines (35 loc) · 4.34 KB

Midterm.md

File metadata and controls

67 lines (35 loc) · 4.34 KB
Raw

Biostat578 Midterm Exam

Name:

Student Number:

This is an open book/computer exam. Feel free to use whatever resources you need, but work on your own. There are 12 questions for a total of 100 points. You have until 11:30am to complete this exam. Good luck!

    1. [5 points] X and Y are two datatables in R, how would do an inner join of X and Y.

Solution: Your first need to use setkey(X,"key"), setkey(Y,"key"). Then you can do X[Y,match=0]. This is faster than merge.

    1. [10 points] Let Z be a datatable in R with three columns: student_name, grade, and assignment_type. Grades are numerical values from 0 to 100. assignment_type can be "assignment" or "in-class exam". Write a one line command to compute the average grade for each student broken down by assignment_type, and sort by grade from lowest to highest.

Solution: Z[,list(mgrade=mean(grade)), by="assignment_type"][order(mgrade)]

    1. [10 points] Let E be an ExpressionSet in R. How would you retrieve the expression matrix, the probe information and the sample information? [State the name of the methods (i.e. R function) you would use].

exprs(E), fData(E), pData(E)

    1. [10 points] State three advantages of next generation sequencing over microarrays. Explain your answer.

Solution: No cross-hybridization, more accurate detection of low expression transcripts, can be used to study many biological processes.

    1. [5 points] What is the main idea behind quantile normalization?

Solution: To make the distribution of two (or more) microarrays the same, which is valid when we assume most gene are not changing.

    1. [5 points] What is the main idea behind lowess normalization?

Solution: To remove the non-linear dependence between the M and A values.

    1. [10 points] You've performed a gene expression experiment using microarrays. Unfortunately, some of your samples were processed in June, while others were processed in October. You suspect a batch effect and would like to correct for it in your limma analysis. How would you do that? You may assume that your contrast of interest is not confounded with batch order

Solution: Simply include the batch variable (time information) in your limma analysis.

    1. [5 points] What is the false discovery rate? Why is it preferable to control this versus the family-wise error rate?

Solution: The false discovery rate (FDR) is the proportion of false positive among your discoveries. In large scale genomics, it is preferable to the FWER because it is less conservative.

    1. [15 points] You have a gene expression experiment looking at expression changes in a cohort of 40 subjects before and after drug treatment. You follow the subjects over time and have samples at day 0 (day of drug treatment), day 14 and day 28. How would do set up your design matrix in limma (write the actual R command), what contrasts would you test?

Solution: mm <- model.matrix(~day+subject, eSet)

Then test for the contrasts: "day28-day0", "day14-day0"

    1. [10 points] What is the main idea behind limma? Why is it preferable over a traditional linear model?

Solution: Limma uses an empirical Bayes approach to regularize variances that are then used to form a modified t-statistics. This approach leads to better performance when the number of replicates is small, which is often the case in genomics.

    1. [5 points] limma was derived for gene microarrays. It's been shown recently that it could also be applied to next generation sequencing data after proper data transformation using voom. What is voom actually doing?

Solution: voom is used to prepare RNA-seq count data for limma analysis. The voom method estimates the mean-variance relationship of the log-counts, generates a precision weight for each observation, and then enters these into a limma empirical Bayes analysis pipeline.

    1. [10 points] Do you need to normalize next generation sequencing data? If yes, what procedure would you use?

Solution: Yes, you need to normalize RNA-seq data due to variation in sequencing depth. Without normalization the results would not be comparable across libraries.

    1. [Bonus, 10 points]. Prove that the Bonferroni mutiple testing procedure controls the familywise error rate at the desired level.

Solution: This comes directly from Boole's inequality.