Many large studies like TCGA use consensus clustering for transcriptional subtype identification. The most popular package for that seems to be ConsensusClusterPlus. If you want to classify an independent cohort, you could run the clustering using the same set of genes and the same number of groups as the original study. Ideally, you would get comparable results. However, it does not seem possible to assign genes to specific groups, at least in ConsensusClusterPlus. Thus, the new groups do not necessarily have to match up with the original groups. Is that correct? Do we just assume that is unlikely?