If the same groups of samples are measured, yes (the difference between platforms will increase the noise and make it more difficult to identify systematic differences between groups - but this disadvantage can be compenstated by an increase of sample size). If the datasets are about different groups: no, because you will always have group-differences 100% confounded with platform-differences so that you can never say if the difference you see is a group-effect or an artefact comming from differences between platforms. There is no normalization technique that can eliminate this fact.

Hi Harpreet Kaur

As Jochen Wilhelm pointed out, as long as there are shared groups/variables between the datasets, you can combine them. Otherwise, the measured signals would be the differences between the datasets and not the groups. As for your specific questions.

Good signs that the combined dataset is appropriate are:

• Similar distribution of the intensity values
• Separation of the individual samples by the groups of interest in PCA or MDS plots
• Observing well established biological signals.

Other problems specific to microarrays and need to be paid proper attentions are:

• The type of the platform/chip for example one- or two-channel arrays. These are prepared differently and their raw data means different things.
• The mapping of probes to genes/features. Different chips use probes sets that map to different regions of the gene, so you need to make sure the probe sets are comparable. The collapseRow method (https://www.rdocumentation.org/packages/WGCNA/versions/1.66/topics/collapseRows) might help.

Mahmoud Ahmed thank you this explanation. As per my understanding SVA package is used to combine only all those datasets where same platform is used i.e. all datasets derived from GPL3921 but my situation is otherwise. Let's suppose I have 3-4 different affymetrix datasets (diseased and normal samples) where different platforms employed like GPL3921, GPL571 etc. All datasets have RMA values. So whether I should perform quantile normalization and then combine them using some package or vice versa.

Hi Harpreet Kaur

I don't think there is an inherent reason to not use SVA in such cases. Unless for a reason I am not aware of, the method can be applied in this case. In case all the datasets are based on the Affymetrix chips, you should be able to map a large number of the probes across.

Note. As the RMA values are the normalized values, you might want to try getting the raw intensities of the datasets and apply the normalization on the combined dataset.