Epidemiology & Biostatistics • Boston University • BA in Neuroscience, 2012 • Duke University Medical Center, 2012-2017 • Research Analyst, Laboratory for AIDS Vaccine Research & Development
regions • Compute mean base-pair level coverage for all brain samples and for each of the 13 brain sub-tissues in GTEx using data from recount2 (Collado-Torres et al. 2017a, Ellis et al. 2017) • Scaled by 40 million reads of 100 base-pairs
the true exons with a skew to short regions • When the mean region length is then calculated would expect a shift to short values • When the median region length is calculated there is some protection from this effect, but it is still present Measuring noise at the “bottom” (Mina Ryten)
• Many short regions are identified together with some true exons with a skew to short regions, BUT there is less noise at the “top” than at the “bottom” • When the mean region length is then calculated would expect a shift to short values and accounts for the appearance of an optimum in the mean region length • This is dampened when you use median values because there is simply less noise at the “top” Measuring noise at the “top” (Mina Ryten)
when a low cut off is used this will generate regions that effectively merge with true exons to produce long region lengths in some cases, which shift the mean upwards. • Noise at the “top” will not generate as many problems but is more likely to collapse over true exons and so the grey line starts following the rest. Collapsing noise at the “bottom” Now take another 20 noise profiles…. and combine this with a low cut off…. (Mina Ryten)
each region separately • Noise at the “bottom” is a much bigger problem than noise at the “top”. • There should be an optimum solution and it might be expected to vary by tissue/expt because the factors contributing to noise at the “bottom” would be expected to vary (e.g. library prep – total RNA would be expected to increase noise at the “bottom”). • We could use exon-exon junction reads as external information to find this optimum solution. • The cut off which results in the highest usage of exon-exon junction reads is the correct cut off. • Defining “usage” might not be straight forward though and will make a difference to this (Mina Ryten)
at the same cutoff • Adapt derfinder to use variable cutoffs that are tissue and context-specific • Examine part of the genome that shows higher expression • Test higher cutoff, determine noise- appropriate level EXPRESSION
M.A., Hansen, K.D., Jaffe, A.E., Langmead, B. and Leek, J.T. 2017a. Reproducible RNA-seq analysis using recount2. Nature Biotechnology 35(4), pp. 319–321. Collado-Torres, L., Nellore, A., Frazee, A. C., Wilks, C., Love, M. I., Langmead, B., . . . Jaffe, A. E. 2017b. Flexible expressed region analysis for RNA-seq with derfinder. Nucleic Acids Research, 45(2), e9. doi:10.1093/nar/gkw852 Ellis, S.E., ColladoTorres, L. and Leek, J. 2017. Improving the value of public RNA-seq expression data by phenotype prediction. BioRxiv. 25