This manuscript provides needed clarification for some unexpected behaviour we can experience with RGI outputs, and outlines some possible improvements notably for better prediction of beta-lactamases.

My major comment is that although the authors point out that only a very few ARG models return high rate of False positive, they do not consider that these genes are actually a high burden for non-experienced users of RGI. It should more clearly emphasized.

The main issue being efflux pumps which can basically be found in every bacterial genome. In the paper, they show for instance that RGI returns ~50% false positive with the adeF reference gene i.e. with % protein similarity ranging between 60% and 70% (and down to 40% when looking at the protein identity score, cf current CARD-R version). At such a low rate, every input genome will return an adeF resistance gene, even when phylogenetically distant (eg. adeF from Acinetobacter baumanni is detected in species from other phyla such as Bacteroides or Campylobacter).

The second critical issue are detections of genes from VAN operons. These operons are of extreme clinical importance since they confer vancomycin resistance in nosocomial pathogens Enterococci and Staphylococci. Their current bitscore cutoff again return a high rate of false positive, with some detections as low as 30% amino acid identity for van Y (result from CARD-R current version). 90% of RGI users will consider that they effectively have a vancomycin resistance gene in their genome (not even the full operon) and find this alarming.

These two issues are in my opinion more urgent to solve than the under-detection of betalactamases, because there is already countless published studies reporting over-inflated numbers of resistance genes (including numerous vancomycin resistance) in genomic and metagenomic assemblies based on RGI's results.