Expression normalization method - Revision history

Burrough at 08:06, 3 October 2011

2011-10-03T08:06:21Z

← Older revision		Revision as of 17:06, 3 October 2011
Line 156:		Line 156:
	sh %! gunzip -c #{infile} \| grep ^chrM \| awk 'BEGIN{sum=0}{sum += $5}END{print sum}' >> #{t.name}!		sh %! gunzip -c #{infile} \| grep ^chrM \| awk 'BEGIN{sum=0}{sum += $5}END{print sum}' >> #{t.name}!
	end		end


			== small RNA normalization ==

			Similar to above, different normalization strategies for FANTOM5 small RNA libraries have been tested including tpm (tags per million sequences), tpmm (tags per million miRNA sequences), TMM, and RLE (see above for description of TMM and RLE).

			=== Failed small RNA libraries ===

			Before normalization, 10 "failed" libraries were removed. These libraries had less than 300 total sequences, to give some perspective the median number of sequences for the set of libraries was ~1.2 million. Here is the list of removed libraries (sample name;sample id):

			#Tracheal Epithelial Cells, donor1;11292-117A5
			#Astrocyte - cerebellum, donor2;11580-120F5
			#Smooth Muscle Cells - Carotid, donor3;11434-118H3
			#Skeletal muscle cells differentiated into Myotubes - multinucleated, donor3;11431-118G9
			#Mesenchymal Stem Cells - bone marrow, donor2;11616-122A5
			#Preadipocyte - breast, donor1;11467-119B9
			#Endothelial Cells - Microvascular, donor2;11342-117G1
			#Renal Mesangial Cells, donor3;11679-122H5
			#Trabecular Meshwork Cells, donor3;11693-123A1
			#Fibroblast - Villous Mesenchymal, donor2;11615-122A4

			=== Methods ===

			The procedure was essentially the same as above, using edgeR to calculate average common dispersion across all FANTOM5 primary cell snapshot libraries with replicates. The sample whose 75%-ile is closest to the mean of 75%-iles was taken as the reference for TMM normalization (default behavior on edgeR). Counts at miRNA loci were aggregated using mirBASE v.17 hairpin definitions, which means we are not distiguishing between miRNAs derived from the 3' and 5' arms.

			=== Results ===

			Common dispersion:
			* tpm: 0.3989
			* tpmm: 0.2133
			* TMM: 0.2003
			* RLE: 0.1992

			Estimate of the coeffecient of variation (square root of the common dispersion):
			* tpm: 0.6316
			* tpmm: 0.4619
			* TMM: 0.4475
			* RLE: 0.4463

			The second list can be interpreted as the amount of variation (up or down) in a library from the "true" miRNA abundance. In other words, in the tpm method the true abundance varies 63% up or down between replicates while using tpmm normalization the true abundance varies 46% up or down.

			Given this, it appears the preferred method would be something like tpm <<<<<<<<< tpmm ~= TMM ~= RLE. TMM and RLE appear to perform substantially better in at least a couple of libraries (see below). As pointed out by Al on the mailing list, the problem with tpm appears to be inherent to the protocol, perhaps stemming from fluctuations in the sequencing of different non-coding RNA populations (tRNA, rRNA, etc.), even across replicates.

			==== Libraries with relatively poor reproducibility ====

			Just as a warning of sorts, the following replicate sets has greater than 75% variation across replicates (poor reproducibility):

			# Smooth Muscle cells – Carotid
			# Smooth Muscle cells – Pulmonary Artery
			# Smooth Muscle cells – Brachiocephalioc
			# Prostrate Epithelial Cells
			# Fibroblast – Mammary
			# Neural stem cells

			All of these libraries have 3 replicates, so it may be a matter of removing one replicate but this has not been investigated further. In theory, it could also indicate that few miRNA are expressed in these leaving very few data points.

			==== RLE/TMM outperformance of tpmm ====

			It should be noted that while TMM and RLE perform about the same as tpmm, there are a few libraries where performance is substantially better for TMM and RLE. Two example libraries with 10% less variation in TMM/RLE relative to tpmm: smooth muscle cells (uterine) and neural stem cells. The likely reason for this is that TMM/RLE correct for overrepresentation of a few loci in a library (see PMID id:).

			=== Files ===

			I am including links to 1) a file ([[Media:MiRNA expression table noquotes reformat.txt.gz]]) containing a list of raw miRNA counts for all currently-sequenced small primary cell RNA libraries and 2) a file ([[Media:All normalization factors.txt.gz]]) containing the necessary normalization information for each library; these can be used to reconstruct the results here or make quick comparisons, plots, etc. of your own. File #2 includes the library ID, the total number of sequences for all libraries (for tpm calculations), the total number of miRNA sequences (for tpmm calculations), and normalization factors for TMM and RLE. To calculate normalized values under the RLE/TMM methods, multiply the library size (total number of miRNA sequences) by the normalization factor and then divide the number of counts by this factor (and multiply by 1million, if so desired). For example:

			effective.libsize <- libsize * TMM_normalization_factor
			norm.count <- raw.count/effective.libsize*1000000

Kawaji at 05:36, 30 September 2011

2011-09-30T05:36:05Z

← Older revision		Revision as of 14:36, 30 September 2011
Line 12:		Line 12:
	=== Results ===		=== Results ===
	Common dispersion:		Common dispersion:
	* no scaling of sample sizes: 0.2388622		* no scaling of sample sizes: 0.2388622 [ sqrt(common.dispersion) = 0.4887353 ]
	* RLE: 0.2495027		* RLE: 0.2495027 [ sqrt(common.dispersion) = 0.4995025 ]
	* TMM: 0.2749048		* TMM: 0.2749048 [ sqrt(common.dispersion) = 0.5243136 ]

	Interpretation:		Interpretation:

Kawaji at 04:31, 29 September 2011

2011-09-29T04:31:59Z

← Older revision		Revision as of 13:31, 29 September 2011
Line 19:		Line 19:
	* no scaling is the best.		* no scaling is the best.
	* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only if RLE is proven to rescue some 'poor quality' libraries, RLE can be a reasonable choice.		* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only if RLE is proven to rescue some 'poor quality' libraries, RLE can be a reasonable choice.
	* TMM is the worst, but it might be improved by taking a better reference samples (at ~~this~~ ~~moment,~~ ~~the~~ ~~first sample~~ is ~~selected~~ ~~for~~ ~~reference~~). However, the improvement would not be so drastic from our experiences below.		* TMM is the worst, but it might be improved by taking a better reference samples (the library whose upper quartile is closest to the mean upper quartile is used). However, the improvement would not be so drastic from our experiences below.

	Note:		Note:

Kawaji at 02:54, 29 September 2011

2011-09-29T02:54:58Z

← Older revision		Revision as of 11:54, 29 September 2011
Line 4:		Line 4:
	== Common dispersion survey (at reliable TSS cluster level) ==		== Common dispersion survey (at reliable TSS cluster level) ==

	Ultimately we are going to use TSS expressions, and normalization methods have to be examined at the cluster level. RLE (relative log expression) method proposed by DESeq package has been implemented recently in edgeR in addition to TMM. Here we tested (i) no scaling (ii) RLE, (iii) TMM on the cluster expression.		Ultimately we are going to use TSS expressions beyond gene level, and normalization methods have to be examined at the cluster level. RLE (relative log expression) method proposed by DESeq package has been implemented recently in edgeR in addition to TMM. Here we tested (i) no scaling (ii) RLE, (iii) TMM on the cluster expression.

	=== Method ===		=== Method ===

Kawaji at 01:53, 29 September 2011

2011-09-29T01:53:36Z

← Older revision		Revision as of 10:53, 29 September 2011
Line 20:		Line 20:
	* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only if RLE is proven to rescue some 'poor quality' libraries, RLE can be a reasonable choice.		* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only if RLE is proven to rescue some 'poor quality' libraries, RLE can be a reasonable choice.
	* TMM is the worst, but it might be improved by taking a better reference samples (at this moment, the first sample is selected for reference). However, the improvement would not be so drastic from our experiences below.		* TMM is the worst, but it might be improved by taking a better reference samples (at this moment, the first sample is selected for reference). However, the improvement would not be so drastic from our experiences below.

			Note:
			* quantile-based scaling implemented in edgeR crashed with the data.

Kawaji at 01:52, 29 September 2011

2011-09-29T01:52:43Z

← Older revision		Revision as of 10:52, 29 September 2011
Line 18:		Line 18:
	Interpretation:		Interpretation:
	* no scaling is the best.		* no scaling is the best.
	* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only if RLE is proven to rescue some '~~low~~ quality' libraries, RLE can be a reasonable choice.		* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only if RLE is proven to rescue some 'poor quality' libraries, RLE can be a reasonable choice.
	* TMM is the worst, but it might be improved by taking a better reference samples (at this moment, the first sample is selected for reference). However, the improvement would not be so drastic from our experiences below.		* TMM is the worst, but it might be improved by taking a better reference samples (at this moment, the first sample is selected for reference). However, the improvement would not be so drastic from our experiences below.

Kawaji at 01:52, 29 September 2011

2011-09-29T01:52:05Z

← Older revision		Revision as of 10:52, 29 September 2011
Line 18:		Line 18:
	Interpretation:		Interpretation:
	* no scaling is the best.		* no scaling is the best.
	* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only ~~when~~ RLE is proven to rescue some 'low quality' libraries, RLE ~~would~~ be a reasonable choice.		* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only if RLE is proven to rescue some 'low quality' libraries, RLE can be a reasonable choice.
	* TMM is the worst, but it might be improved by taking a better reference samples (at this moment, the first sample is selected for reference). However, the improvement would not be so drastic from our experiences below.		* TMM is the worst, but it might be improved by taking a better reference samples (at this moment, the first sample is selected for reference). However, the improvement would not be so drastic from our experiences below.

Kawaji at 01:51, 29 September 2011

2011-09-29T01:51:13Z

← Older revision		Revision as of 10:51, 29 September 2011
Line 20:		Line 20:
	* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only when RLE is proven to rescue some 'low quality' libraries, RLE would be a reasonable choice.		* RLE is reasonably good, however no scaling is the best in terms of common dispersion. Only when RLE is proven to rescue some 'low quality' libraries, RLE would be a reasonable choice.
	* TMM is the worst, but it might be improved by taking a better reference samples (at this moment, the first sample is selected for reference). However, the improvement would not be so drastic from our experiences below.		* TMM is the worst, but it might be improved by taking a better reference samples (at this moment, the first sample is selected for reference). However, the improvement would not be so drastic from our experiences below.

Kawaji at 01:50, 29 September 2011

2011-09-29T01:50:43Z

← Older revision		Revision as of 10:50, 29 September 2011
Line 7:		Line 7:

	=== Method ===		=== Method ===
	* data - Promotome v1.0 - RC1 (release candidate 1, Sep 26th 2011), from https://fantom5-collaboration.gsc.riken.jp/webdav/home/kawaji/110804-dpi-clusters-expression/		* data - Promotome v1.0 - RC1 (release candidate 1, Sep 26th 2011), from https://fantom5-collaboration.gsc.riken.jp/webdav/home/kawaji/110804-dpi-clusters-expression/ only clusters with max_counts >= 50 were used
	only clusters with max_counts >= 50 were used
	* replicates - determined by recognizing "donor","donation","pool","biol_rep","tech_rep", in the sample names.		* replicates - determined by recognizing "donor","donation","pool","biol_rep","tech_rep", in the sample names.

Kawaji at 01:50, 29 September 2011

2011-09-29T01:50:29Z

@@ Line 1: / Line 1: @@
-There are several ways on how to normalize CAGE expression data. 'default' way of normalization has been TPM, but we could think of recent proposal such as TMM (implemented in edgeR, genomebiology.com/2010/11/3/R25). Another point to be considered is if we should include chromosome M in the library sizes or not. To address this point, Marina surveyed several options with checking common dispersion. Kawaji followed up this survey by chrM / rRNA correlation.
+There are several ways on how to normalize CAGE expression data. 'default' way of normalization has been TPM, but we could think of recent proposal such as TMM (implemented in edgeR, genomebiology.com/2010/11/3/R25). Another point to be considered is if we should include chromosome M in the library sizes or not.
-== Common dispersion survey  ==
+== Common dispersion survey  (at reliable TSS cluster level) ==
 Marina performed a series of experiments on TMM normalization (edgeR, genomebiology.com/2010/11/3/R25) with taking the shallowest, the deepest, or universal RNA as a reference and with/without chromosome M. And she simply checked common dispersion on those combinations.