That 's typical microarray , so trust p-value

Well, it doesn't have to be typical. Having said that if your protein of interest is very specific for a particular function and is only needed for a very specific subset of genes, the array may have shown you exactly what is happening. As Intawat said in the post before, identify significantly differentially expressed genes using something like a t-test or SAM or whatever test you are fond of and identify the genes that way. Then do some functional clustering in DAVID (http://david.abcc.ncifcrf.gov/) for example to see what functional groups are represented. Hopefully you will find what you have expected to happen.

How did you quantify the knockdown of "your protein"? Just from the microarray or did you do a proper western blot to verify the knockdown? You cannot judge the knockdown from the mRNA level alone, since the protein may be much more stable (long half-life) compared to your mRNA levels and the biological effect depends on the protein.

One more thing: why did you select 0.5 and 1.5? That is an asymetric interval?! I will try to explain it simple using an example. Lets say your gene ABC has an expression level of 10. If you apply your lower fold-change limit, 0.5, you would consider genes that are 10*0.5 = 5 = 1/2 of the previous level . Your upper limit is set to 1.5, so 10*1.5 = 15 = 150% / 1.5/1 , and not 2/1 (double) the amount. That is why people usually use log2-transformed values: 2^-1 = 0.5, 2^1 = 2 // 2^-2 = 0.25, 2^2 =4, i.e reduction to a quarter or increase to the four-fold level. If you want to stick wit 1.5 fold upregulation, you should use 1/1.5 or 0.666667.

I suggest you a fold change and p-value criteria as Intawat Nookaew said, and a functional enrichment to check the biological sense of your gene selection. I suggest you a Babelomics platform ( http://babelomics.bioinfo.cipf.es/ ) in Differential expression analysis you can select adjusted p-value, and in Fatigo or Fatiscan to analyze the functionality. :-)

Hi

Are your expression values logaritmics? If yes i suppose that your FC is computed (expra-exprb). Ok in this case for your downregulated genes you are assuming that one gene in the first condition is expressed at least one less than in the other.

Regarding the size of FC your threshold is small but it depends by the type of treatment that apply to your cells. For example in my experience the characterization of the effects of bioactive molecules contained in foods are able to produce only small global changes but also small changes are able to induce relevant phenotypic differences .

Regarding the number of modulated genes this could be in my opinion a good number or not depending of the type of protein that you knocked down. If your protein is a transcription factor like a nuclear receptor this number could be small otherwise if your protein is involved in secondary pathways this number could be interesting.

I think that a better picture could be obtained by characterizing a theorical model of what kind of pathways you could affect knocking down your protein

If want to apply similar threshold for regulated genes, for example absolute fold change of 1.5, then you have to consider =1.5 as up-regulated. And to increase the probability of similar expression change when you validate the results by qRT-PCR, you may have to apply statistical test threshold also (p

I agree with the previous answers on many points.

Genes with small fold-change and significant p value have to be taken into account.

Functional annotation (by Gene Ontology) and enrichment analysis can reveal the global induction or repression of specific pathways. Even small changes in several genes related to the same pathway can produce biological effects.

However, if you want to study specific transcripts of interest, they need to be validated by quantitative methods in order to get a more precise measurement of the fold change.

How long have you waited before running the microarray expreriment?

It is possiible that the small variations you observed are due to the fact that the transcription effects are not fully evident at the time you run the experiment.

The other possiiblity is that your gene has a limited effect on trascription regulation.

Since the number of de genes is quite limited I would suggest to validate at least some of them by qPCR. Furthermore you can load all of them in software like ingenuity, asking for network generation, starting with your genes and other genes that are connencted to them by annotation. Then you can check how many of the genes belonging to the networks, but not part of the de you started with, are called significant for p-value only in your de experiment. This will allow you to extend the information you have even though the level of actual information is quite limited.

In general I agree with what has been suggested except for the following:

p-value: if your microarray contains 10,000 probesets and you apply a p

Additionally to the aforementioned suggestions, I recommend to calculate the FDR (false discovery rate) of your genes. This is done by a permutation test and gives additional information how valid the observed expression changes are.

I guess you have been given several good and wise suggestions like checking the protein levels on westerns and also the Real time PCR to check the levels of the knockdown. But one more important thing as I understand according to my experience is that when you do microarray with the Knock downs then you get small number of genes differentially regulated. SO if you get the 90 up and 35 down regulated genes then it is not surprising. One protein will affect not many protein expression of pathways. In case of cancer you expect several genes to be differentially regulated as it is multi factorial disease. But you should also try to look for the genes and pathways you expect to be affected by this protein.

There is nothing odd about a small number of genes being affected. You have given too little experimental details to really help you. My only novel aside from the good suggestions above, would be to try a gain of function experiment by over expressing the gene. You may see genes with the opposite expression changes, depending on what the gene is. If KD is 75% I would not expect dramatic changes, you need better than 90 for a strong KD affect. How clean are the cells? Are they clonal? If this is a transitory KD then you likely have a range of affected cells, then who knows what you are measuring. You need to use clonally pure cell lines with KD. Try a selectable or inducible system. all the best.

As I understand it, the problem is NOT the small number of genes being affected, but the small fold change. There is no point giving you advise until you provide more information as noted by some people here. For instance, did you measure the 75% knock down by protein levels or are you judging my mRNA levels? You protein might be more stable then you think. ANd if I may suggest, before doing a micro-array experiment, it is always a good idea to know a bit what depletion and/or over expression of your protein does in the system you are using. Perhaps you did knock it down sufficiently, but it has little biological effect in MCF7 cells under your experimental condition. If that is the case, no statistical algoritm is going to help you.

I believe the most important in cicroarray is how your genes interact. In one of my experiments I also had only 95 genes upregulated with a 2 or less FC. However, all genes were grouped in only 3 pathways, that were being modified a lot. Another important thing is what is the percentage of confirmation by PCR.

What is the distribution of the changes in expression of all genes in your experiment? Are the two datasets so similar that small changes are significant? If the correlation between experiment and control is extremely tight, small changes may be higly significant. also, they may be relevant, but this indeed is coupled to biological function.

Besides first selecting significantly regulated genes, and next assessing their function, another way of evaluating functionally related changes in expression is by assessing the changes in entire functional groups. techniques such as paramatric analysis of gene enrichment (PAGE) or varaiations thereof (t-profiler) can help you assess whether functions as a whole are significanlty changed in microarray datasets.

As mentioned above, more experimental information is needed. Would you expect many genes to be affected by the knockdown of your target? Looking at genes with lower fold change values and significant p-values might also help, particularly with additional functional/pathway data to determine if those changes are biologically relevant. But when considering lower fold change thresholds, you also have to consider the limits of detection of the array, based on array type and data processing (background correction, normalization, summarization) methods employed.

You dont give much detail, but to ascertain if your fold change is true but not noise, firstly you need to run biological and technical replicates. The minimal technical replicate should be a dye-swap experiment. Then, if you use a whole genome microarray, a normalization step MUST be done to address the dye bias issue. At this point, as someone stated above, you need to check the distribution and choose 2 or 3 statistical tests that can be done, provided that they control the FDR. Finally a boolean comparison of the outputs of these 2-3 tests (cut off at p

In my opinion the fold change is not a very reliable indicator of the microarray results, in fact a fold change of two in a gene encoding a transcription factor could have a very different “effect” than a similar fold change on a gene coding a protein involved in metabolism. In my experience referees are very skeptic when genes having fold change lower than 2 are discussed in a paper. Moreover microarrays tend to underestimate fold change in comparison to the real-time PCR and to the RNA-seq. Finally I suggest to correct for multiple tests or to analyze replicates using software like SAM to evaluate the false discovery rate (http://www-stat.stanford.edu/~tibs/SAM/) (http://www.tm4.org/mev/).

My experience with microarrays is that significantly upregulated genes after treatment were at least fivefold more expressed than control. Yet, if you made 6 replicates and p-values are

Well as per my experience the small fold chain you see might be the actual picture in response to your knocked down protein. To be more certain about it you can do a dye swap experiment (technical replicate) + 1 more array with the repeated experimental sample (biological replicate) + validation of the most significantly deferentially regulated genes (sorted out by log2 ratios and pathway analysis) using qPCR + western blot . I guess it would be enough (having 3 data sets) to satisfy the science fraternity.

Dye swap is not recomended by the statisticians...qPCR and western are golden standard.

You have a number of excellent suggestions to follow from the previous posts. Of course you need to validate with RT-PCR and Westerns for the hits that you do have, and you should look for pathway signals. You haven't told us what you knocked down, but that would help guide your analysis in a very important way. Many quite dramatic cell responses (e.g. oxidative stress response) don't involve large gene expression changes. Your gene may act at the level of protein degradation. Once you've done the appropriate controls to make sure your gene is really knocked down, what you really want to see is a change in the functioning of your cells. For MCF-7s, for example, you might look for changes in proliferation or invasiveness, or tumorgenicity in xenografts. What you look for should be guided by biological plausibility.

The fist thing I would do is to check if the data (expression levels of each chip) really reflects the complexity of the transcriptome (which is of 4 log units). Which means that if the smallest signal is 50 the highest should be 50000. We had something similar once and after checking a series of genes none worked, Afterwards we realized that instead of having a dynamic range of the data of 4 logarithmic units it had 2.5. In that case the best was to throw the data. We repeated the hybridizations and worked fine, having the appropriate dynamic range and all the genes we identified were checked and verified the data

Also note that background-correction (especially the "variance-stabilizing offset") and the normalization lowers the "apparent fold-changes". Note that you do not measure fold-changes or concentrations directly, you measure fluorescence signal intensities, probably corrected and transformed. In dual-color arrays, the ratios of the signal intensities are known to underestimate the ratio of the true mRNA concentrations. Therefore I don't like the interpretation of microarray-data as "fold-change" (unless it is not calibrated by spike-ins on similar techniques). Microarray data just give an indication of the regulation (up or down, moderate or rather strong). Anyway, some "exact" value usually would not tell us more (knowing that the apparent signal ratio is 1.4 is not more informative than knowing that the mRNA concentration is 2.5-fold induced).

I would start by having a look at the dynamic range of the data. If it represents the transcriptome if should have 4 log units (at least), which means that whereas some RNAs are present in 1 copy per cell, other are expressed up to 10000. The only experiment on expression profiling that has given me very small fold changes, did not fulfill the above criteria, and what was worse, experiments to corroborate the changes did not work. From then (10 years ago) the first critera we use is that the data should represent the transcriptome (the data scattered over four logs).

Appart from the statistic and qPCR validation advices, you have to take into account that every tissue/cell type is different and responds differently to treatments... I work with gonads and the FC that I found are subtile if I compare them with my colleagues' results on immunology studies...

Well, I am trying to solve this issue and fold-change can not tell you much about initial levels of the mRNA. Proteins are another dimension.

T-test should be replaced by Wilcoxon signed rank test because based on couple of replicates you may not conclude what is your distribution plus who can guarantee that each gene should have the same expression kinetics.The mRNA degradation is really a black box because it is very hard to predict whether it is due to expression inhibition or mRNA degradation itself or maybe some other masking interaction with proteins.

Great suggestions have already come up. I think statistical validation and confirmation of microarray results using RT-PCR, real time PCR or western blots of pathways molecules will help understand your results. The small fold changes may be different and can be confirmed with further experimentation.

If the gene you knocked down is part of a gene family or has splice variants, you may see small changes when its expression is reduced. Also, if your gene is part of a pathway with lots of cross talk--its tough to expect significant changes when you knock it down. The redundancy in gene function is an issue which needs more attention.