First, the threshold of 2 is purely arbitrary and is not at all an accepted convention or rule.  In fact, for many years now it has been considered a best practice to base the categorization of differential expression on both a statistical basis as well as a minimal magnitude of change.  But people use alternate limits on the magnitude of change based on the strength of their data.

Basing differential expression solely on magnitude of fold change is in fact the weakest argument one can make, even with a minimum cutoff of 2.  The original MAQC consortium papers demonstrated that fact - that fold change alone is the least reliable basis for determining differential expression.  Between microarrays and qPCR, the genes that are least likely to verify are those selected solely by fold change.  However, those selected by a robust statistical cutoff and a fold change (even one as low as 1.5fold) will almost always verify by qPCR.

If, for example, you had many biological replicates (e.g. greater than 10), and a gene with an observed fold change of 1.6 was shown to have an FDR corrected p-value of 0.001, for example, then many would have no problem considering such a gene to be significantly differentially expressed.  If you then were able to verify that gene with a robust (i.e. many replicate) qPCR study that also demonstrated strong statistical and reliable minimum fold change, then you would have a very strong argument for differential expression, regardless of the choice of minimum fold change.

But your idea that convention requires a minimum fold change of greater than 2 is simply not true, as many published papers use a lower minimum threshold, combined with a statistical criteria.  The main requirement is having a sufficiently robust study design to justify the use of those measures, and that means having robust replication of the populations in question.

First, the threshold of 2 is purely arbitrary and is not at all an accepted convention or rule.  In fact, for many years now it has been considered a best practice to base the categorization of differential expression on both a statistical basis as well as a minimal magnitude of change.  But people use alternate limits on the magnitude of change based on the strength of their data.

Basing differential expression solely on magnitude of fold change is in fact the weakest argument one can make, even with a minimum cutoff of 2.  The original MAQC consortium papers demonstrated that fact - that fold change alone is the least reliable basis for determining differential expression.  Between microarrays and qPCR, the genes that are least likely to verify are those selected solely by fold change.  However, those selected by a robust statistical cutoff and a fold change (even one as low as 1.5fold) will almost always verify by qPCR.

If, for example, you had many biological replicates (e.g. greater than 10), and a gene with an observed fold change of 1.6 was shown to have an FDR corrected p-value of 0.001, for example, then many would have no problem considering such a gene to be significantly differentially expressed.  If you then were able to verify that gene with a robust (i.e. many replicate) qPCR study that also demonstrated strong statistical and reliable minimum fold change, then you would have a very strong argument for differential expression, regardless of the choice of minimum fold change.

But your idea that convention requires a minimum fold change of greater than 2 is simply not true, as many published papers use a lower minimum threshold, combined with a statistical criteria.  The main requirement is having a sufficiently robust study design to justify the use of those measures, and that means having robust replication of the populations in question.

In microarray during the cRNA formation we amplify the available cDNA, So in microarray you cant exactly measure the quantity of starting mRNA, only it can give you clue for Up and Down expression. While in qPCR the exact estimation takes place, so there is no need of any correlation coefficient to make both microarray and qPCR equal. 

Also it depends on you whether the FC 1.5 is important for your experiment or not. It may important in many times. If you did qPCR then you can go with that.

In array data, usually the logFC is taken directly from the logRatio of the signal intensities. Since the signal intensities are not a simple linear function of the target sequence abundance, this interpretation is actually wrong (although it may serve as a rough estimate - but what for?). The obtained logRatios of the signal intensities also depend on the platform and very strongly on the pre-processing of the RNA (preamplification!). We observerd that for dual-color arrays (compared to single-color) as well as for higher preamplification rates the ratios get increasingly more compressed for higher FCs. This means that the logRatios (

Besides the very valuable comments that were made by Jochen and Michael,

there are experimental architectures that are very difficult to analyse, because e.g. they lack a control to which your values could be related to. Such experiments are simple to perform, but hard to analyse. e.g. a diurnal experiment - which is taking samples throughout the day and night. In such experiments, the identification of differentially expressed genes can not be done on fold-change, but requires linear analysis models, e.g. by using LIMMA. To find out how many different gene expression trajectory patterns you might find in your experiment, you could possibly use figure-of-merit or PCA.

In addition, we use different amount of RNA input depending onthe biological question we ask, e..g  whether we like to analyse lowly expressed genes or if we also want to pick differential expression in highly expressed genes.

Instead, in qRT-PCR we use standard references (one lowly expressed gene, two moderately expresssed genes and a higly expressed gene) and run with cDNA made always from the same amount of RNA-input.

Hence, a direct comparison of absolute values is impossible.

Thus, we only conclude on relative expression changes, but not on absolute fold changes/transcript abundances, because this is very hard to quantify.

The accuracy of differential gene expression analysis is critically dependent on data normalization and the performance of gene-specific probes. RT-qPCR has the advantage that each gene probe can be validated first. So, if gene primers are efficient in PCR the fold-changes should reflect accurately what the real changes are. However, in RT-qPCR, typically only one or a few genes are used for normalization and so if these are not truly invariant (which is difficult to prove unless you have lots of replicates), then you have a problem. External spike-in RNAs are theoretically better normalizing genes, but in practice there is the risk that if they get even slightly degraded, they're not much use for normalization. It's anyone's guess how accurately most microarray probes actually work for measuring gene signals, since it's difficult to measure cross-hybridization to similar genes and the efficiency of hybridization to the intended gene. Have you ever seen microarray data and wondered what to do when 2 probes say significant down regulation, but another 2 for the same gene say significant in the other direction?  The dye-bias and other technical reasons which necessitate aggressive normalization of microarray data tend to dampen the fold-changes, such that generally you will see bigger changes by RT-qPCR. I think that RT-qPCR FC values are closer to the truth since they tend to agree better with RNA-seq than microarray. However, there is always the caveat for RT-qPCR that you need good technique. Even small pipetting errors can lead to quite large variation in FC values across replicates. The 2-fold cutoff for considering if it the change in expression is biologically relevant has a practical value for RT-qPCR as that corresponds to one PCR cycle difference. What is your range of Ct values in RT-qPCR?  

Hi Team, thought I'd wade in on this one because some time ago we published a paper on the correlation between Affy microarray data and TaqMan qRT-PCR data (Dallas et al, 2005, BMC Genomics 6:59).

Anyway, fold change by Affy microarray will generally give you an accurate indication of whether a gene is up or down in one sample relative to another, but I wouldn't get too excited (or depressed) about small differences (like 1.5x or 1.7x), especially for genes that are expressed at relatively low levels. You really need to follow up with qRT-PCR to determine fold change accurately.

I would recommend reading our paper, and citing us prominently in your next publication....ha, ha