With gene expression, a biological replicate actually would not be expected to always produce the same value, as the innate biological variation amongst individuals in the population is in fact almost always the single largest source of variability (unless studying highly inbred individuals).  In some cases, that innate biological variation will completely swamp all other sources of variability in your samples.

The point of normalizing any gene expression data is not to bring individual estimates to a notionally common scale, but to bring the entire probability data distribution into alignment (with RNA-seq data usually following a Poisson or similar distribution).  That's at the heart of issues with normalization of RNA-seq data, as the distribution of count data covers such a tremendous dynamic range and can be highly variable across a set of samples (even ones where read depth is very similar), that arriving at a normalization scheme that adequately deals with the entire scope of the data probability distribution is difficult.

it is also why myself and other authors have recommended completely ignoring very low count data prior to normalization and differential gene expression analysis.  Very low count RNA-seq data tends to have massive variance across biological replicates relative to high(er) count data, as low count data represents the most error prone estimates of expression in the data set.  Personally, when analyzing data, I first filter out all transcripts with a minimum raw count < 10 in all samples, and then normalize.  That cutoff is arbitrary, but it helps to remove much of the extremely noisy data in the expression estimates.

Since you are talking about a purely relative measure of change, there is no definitive answer to your question.  Fold change in all cases is essentially the ratio of the sample mean expression of your treatment relative to the sample mean expression of your controls.  Since the estimate of expression, be it RNA-seq or microarray, is a relative measure every tool you mention is correct for their specific method of normalization and estimate of dispersion, includng FPKM or RPKM normalizaed estimates of expression.  RPKM, for example, will be overall biased towards up-regulated genes but that is inherent in the way it treats high versus low count data.

If you compare many methods, you will find that the rank order of estimated fold change tends to correlate well amongst the more robust methods, but the actual magnitude of the estimation does not necessarily correlate at all, and is dependent on normalization and always will be (and is so for array data as well).  That is one of the main reasons why selecting differentially expressed genes solely on fold change is highly discouraged - it is inherently unreliable as far as estimating the actual magnitude of change.  The addition of a statistical threshold and a magnitude threshold is more likely to yield genes that are truly differentially expressed.

I would not think it a good idea though to choose one method of normalization for statistical significance testing, and then another, different normalization for estimating fold change.  For one thing, that introduces bias between the two methods of significance thresholding since we know that both significance testing and fold change estimation are dependent on the particular method of normalization chosen.

If you truly need absolute measures of expression and fold change, then you need to use a technology that actually provides that, such as the various methods for qPCR or digital PCR.

Otherwise, you select a normalization technique that you feel is best for your data, and you accept that the computed fold change from any such analysis is a relative estimate and that the actual magnitude of that estimate is not a highly reliable measure of the actual change in expression (although the direction of change, up or down regulation, should be fairly reliable).

Thank you for this very interesting information. It is very helpful.

Hi Steffen,

Very interesting question. Assuming that you have replicate measurements you can actually evalute the performance of the given normalization method.

A replicate sample should always produce the same value unless there are environmental factors that contribute to your measurement. Now this is the aim of normalization you want to remove environmental effects and bring all  your samples on the same level.

The choice of normalization method should be easy in your case. The method that makes replicate measurements look more similar does the right job. The performance of normalization methods can be estimated using a correlation test for your replicate samples.

Once you are happy with the normalization method you can proceed to the data consolidation step which mean you sum up your replicate measurements and turn them into a single observation in order to avoid that you inflate your test statistics by contaminating it with repeats. 

I am not sure what you are measuring, but the results that you obtained by looking at the combination of fold change and p-value are better than going for markers that perform well in one of the metrics but not in the other. Useful tools for this are volcano plots or feature selection algorithms with a decent validation step. Ideally you want to go for nested cross validation.

Kind regards,

Jens

http://www.biomarker-discovery.com/

With gene expression, a biological replicate actually would not be expected to always produce the same value, as the innate biological variation amongst individuals in the population is in fact almost always the single largest source of variability (unless studying highly inbred individuals).  In some cases, that innate biological variation will completely swamp all other sources of variability in your samples.

The point of normalizing any gene expression data is not to bring individual estimates to a notionally common scale, but to bring the entire probability data distribution into alignment (with RNA-seq data usually following a Poisson or similar distribution).  That's at the heart of issues with normalization of RNA-seq data, as the distribution of count data covers such a tremendous dynamic range and can be highly variable across a set of samples (even ones where read depth is very similar), that arriving at a normalization scheme that adequately deals with the entire scope of the data probability distribution is difficult.

it is also why myself and other authors have recommended completely ignoring very low count data prior to normalization and differential gene expression analysis.  Very low count RNA-seq data tends to have massive variance across biological replicates relative to high(er) count data, as low count data represents the most error prone estimates of expression in the data set.  Personally, when analyzing data, I first filter out all transcripts with a minimum raw count < 10 in all samples, and then normalize.  That cutoff is arbitrary, but it helps to remove much of the extremely noisy data in the expression estimates.

Very important information to calculate the differential gene expression from RNA-seq data.

Thanks to all