It is not clear from the question what the variables in the data are. In general, there would be two kinds of known variables. To remove the unwanted variables, one would use something like ComBat/RUV. To get the effect of the design variables, one could use DESeq2 to test for differential expression. It could be that the two kind of variables in this dataset are correlated, so their effect was removed from the data during the batch removal.

It is not clear to me if for the PCA you used gene expression or SNPs (though they are not mentioned) identified in the sequencing.

I am not sure I would even be trying to combine the two datasets. You could just analyze each independently for differential expression, and then compare the results of the significant differentially expressed genes. Look to see if you have evidence for the same MoA in each with an ontology enrichment analyses, or GSEA or some such indication of similar (or different) functional response of each cohort. Or compare rank-ordered gene expression for correlated gene response. There are a number of ways to compare the expression profiles without doing a combined analyses. And if the PCA of normalized expression showed they have nothing in common, of course you're going to drastically reduce the variance of both by trying to compare them in any kind of meta-analyses, and thus would be tossing away everything that actually discriminates them.