You should not, data is not quantified in the same way, therefore you will have a lot of bias.

There are tools such as the R package MOFA which have been developed to tackle this problem of low-dimensional reduction of different -omic technologies and kind of find out which of the data better explain the observed variability and such.

In my opinion, it's worth trying (but be sure you have controls to assess whether clustering is due to rna-seq/microarray differences or biological differences).

Some ideas to reduce bias:

• Try quantile-normalizing your datasets.
• Try using MDS with non-parametric distance metrics (e.g. those based on spearman rather than pearson).

FYI, it's much easier to integrate microarray and rna-seq together when you compare downstream analyses (e.g. differential gene expression) rather than gene expression values.