Dear all,
I think with this final question I have covered all the open ends from
the CoP meeting. Please drop anything I missed on the list!
This interesting question was brought up by several people in the
meeting: to what extent are we able to detect novel resistance genes (or
virulence genes, or plasmids)?
Given that our current tools are all based on mapping sequence data on
reference databases, to what extent does this generalise beyond
relatively similar sequences?
An obvious direction to look in are deep learning models, which appear
to have a remarkable capacity to generalise beyond their training data -
but could they actually predict genes coding for a hitherto unknown
resistance mechanism, or even predict resistance for a highly diverged
gene, while never having seen the phenotypic "ground truth"?
This explains my interest in the concordance of genotypic prediction and
phenotypic AMR, and collecting the false negatives: isolates with
positive AST but negative ResFinder prediction are precisely the ones
we'd like to be able to generalise to!
In my very limited experience (we did one study in 2021, looking at
GPlas / MLPlasmid and Deeplasmid for predicting plasmids), the tools
essentially predicted (poorly) what was already in the PlasmidFinder
database.
Clearly though there has been a lot of development since, and I would be
very interested to hear of your experiences!
Marco