I have RSEM-normalized-log2 transform data downloaded from Firehose and I found that there are number of missing data and filled as NAs. However, when I checked the raw counts for the same datasets, it was given as 0. So, for downstream analysis can I convert all the NAs as 0. Please guide me.
I would check the accompanying notes to see exactly what post-processing has been performed on these by the Broad Institute. It would seem likely, based on the information that you provide, that they decided to convert values of 0 to NA to avoid producing a 'negative infinity' (log2(0) == -Inf).
However, if you have raw counts already, then why not use those? - these can easily be used with EdgeR or DESeq2.
TCGA raw HTSeq counts are also held at UCSC's Xena Browser.