We made the decisions in HUMAnN 4 to 1) automatically do sum-normalization to CPMs vs. leaving features in RPK units, 2) keep including the UNMAPPED read count in the output (for consistency), but 3) not include this count in the CPM normalization.
We opted to add the automatic normalization since it was preferable to be computing downstream abundances (e.g. pathway abundance) from pre-normalized gene abundances. It was also inconvenient for users that, barring unusual applications, it was almost always the right step to immediately normalize the method’s outputs after they were generated.
The reason for not including the UNMAPPED reads in this normalization is two-fold. 1) Not doing so is more consistent with other areas of shotgun sequencing analysis, where abundances are usually expressed as fractions of MAPPED reads (vs. total reads). 2) Gene abundances are initially computed in RPK units (to adjust for differences in length) before sum normalizing. This is a critical step, but it results in the gene abundances and UNMAPPED abundance being in different units (RPKs vs. reads). Hence, it’s not strictly correct to include them in the same sum during normalization, although doing so is probably not a terrible approximation in practice.
This is something we plan to revisit in future releases. I think a nice alternative might be to compute CPMs over genes, as we do now, but then scale those abundances to fill the % MAPPED reads of the sample (similar to how MetaPhlAn handles % UNKNOWN estimation). This would then place the UNMAPPED read mass and other features in the same composition.