A quantitative comparison between microarray, deepSeq, qRTPCR and Northern is in my opinion (and our experience) difficult and sometimes not possible. I trust Northern data most though they are not very sensitive. Keep in mind that these methods assume (!) e.g. equal efficiency in RT and priming (for qPCR), equal efficiency in ligation and RT for deepSeq. A factor of 2 is just one more cycle in qPCR! I agree that changes by a factor of 1.8 could be biologically very important but is it not strtching confidence in microarrays a bit too much?

Ceyhun, I have more questions than answers, but perhaps they will help us to get to the correct answer.

Are you using the same RNA samples for your qPCR that you used for the microarray? Which microarray platform are you using? Are the genes demonstrating downregulation by microarray initially highly expressed? I have used both the mouse and human ST arrays from Affymetrix and I have found them to have good dynamic range. However, if the majority of your downregulated genes initially show low signal intensity (suggesting a low transcript level) then any apparent reduction may just be due to technical error, which of course would not be apparent in qPCR.  So you want to set a relatively high background level cutoff if you want more reliable results.

Although qPCR is more sensitive, if your Ct values are greater than about 32, you will find that detection of less transcript will be difficult since data much greater than 32 Ct will be considerably less reliable.

How many genes are showing the phenomenon that you are describing? Are only the genes that initially showed low signal intensity giving you problems with qPCR or is it all genes that have shown downregulation?

Hopefully a few more details will get us to your answer.

Best.

The reverse transcriptases may be different between the platforms.  Self-priming activity of RTs could cause such a result. Try a control reaction where no exogenous primers are added to the cDNA synthesis reaction.  Increasing the temperature in the cDNA synthesis reaction may help reduce self-priming.

Thank you all for the answers,

Dear Wolfgang, you are right about the efficiencies and the factors, but the biggest factor is 4 for down-regulated genes im my data. 

Dear Ricardo, about your questions;

1) Yes, they the same RNA samples

2) I have used Affymetrix

3) The Ct values of down-regulated genes that I have tried so far are between 19 and 26

4) So far I have tried 6 down-regulated genes, except one they all showed the same problem and the Ct values of the gene that I found it down-regulated in qPCR are approximately 19, but the fold change is near -1.8 while it was -2.5 in microarray.

Both in up-regulated and down-regulated genes, Ct vales were lower than 28 so far. Also, I am going to try more genes.

Dear  Thomas, I am using high fidelity reverse transcriptase in cDNA synthesis for qPCR but I do not know which one Affymetrix uses. 

I want to add that the Ct values I mentioned above came from 3 ng cDNA (theoritically, if efficiency of cDNA synthesis was 100%) in 20 ul qPCR reaction.

1.8-3.5 fold down-regulation is quite a small change in expression and could be difficult to detect by any method. You mentioned the Ct of the candidate downregulated genes in RT-qPCR to be 19 and 26, which suggests that these transcripts are quite low abundance. In my experience, I found microarray data often shows "statistically significant" changes in genes of low abundance that are not detected by other methods (RT-qPCR, northern, and RNA-seq). I would guess the reason to be noise fluctations in microarray, since the background signal being inherently variable across the array is rather difficult to correct for. If these low abundance genes are critical for testing your hypothesis, then the only solution is more replicates in RT-qPCR to get more statistical power.   

Normalization of microarray data may be an issue, e.g., if your cultures are in very different growth state. Standard methods like "quantiles" assume that both samples should have the same signal distribution but this is not necessarily the case. So genes that actually show little changes and are not highly expressed may be "pushed down" by the normalization. See e.g. attached references (go to supplement for the yeast paper) and citations therein. Try e.g. LVS (least-variant set) normalization available in R.

Article The Yin and Yang of Yeast Transcription: Elements of a Globa...

Article Revisiting Global Gene Expression Analysis

Article How cyanobacteria pose new problems to old methods: Challeng...

Another issue to consider is that the qPCR assay may not be measuring the same transcript(s).  That is, with an affy array, the results are averaged over many probes selected to balance sensitivity and selectivity for the desired transcript as best they can.  Everything that binds to that set of probes produces signal.  With PCR, the assay measures everything that can be amplified with the chosen set of primers (2 probes).  For mammalian genomes in particular (but even for yeast), there are often multiple homologous genes that will be read out by array for a single set of probes associated with a single gene.  For PCR, it is also possible to amplify more than one product at the same time with a given set of primers.  The latter can (and should) be tested by at least once running the PCR product out on a gel to ascertain that it is pure and of the correct size.  You might also look at the gene itself, the primer sequences and the probe sequences to see if there are homologous sequences elsewhere in the genome.

As to other suggestions provided above - it would be useful to look at the array data for this gene to determine if it is indeed lowly expressed (as the qPCR data seems to suggest).  If you sort the array data by intensity, what percentile do the genes of interest fall in?  Note that on average about 50% of all mammalian genes are not expressed at all in most tissues so for about 1/2 of all your array data, the comparisons are being done between values that are essentially zero and as others have suggested previously, background noise variation can give rise to false positives.   For yeast, similar considerations apply. I often find it useful to restrict my array analyses to genes above a certain intensity threshold.

While others have indicated that a fold change of 1.8-3 fold is "quite small", in my experience, small fold changes on arrays (for highly expressed genes) are often quite reproducible and often underestimate the true fold change (especially for highly expressed genes).  The reason for this is that arrays (contrary to popular belief) are not linear in response to concentration and highly expressed genes can start to approach saturation of the signal (or probe binding capacity on the array).  Hence, I think it's inappropriate to discount small fold changes on the array as they often show up as larger fold changes when measured by other techniques.  

Finally, as sequencing costs have come down and down, it's time to consider running your samples by RNA-seq as opposed to arrays.  With sequence information, many of the concerns about homologous genes go away as sequence data typically allows one to segregate related genes into the appropriate bins. Also, RNA-seq is inherently more linear in response than array data.