Closed ninsch3000 closed 1 year ago
In the case of AbsQuant I think just filtering out zeros reported from an EWF makes a lot of sense for the reasons you mention above @ninsch3000
Ideally this would be a SWF parameter for filtering out sites with expression less than or equal to a threshold, with a default of zero tpm
.
Currently all sites reported by the participant that do not match a ground truth site are considered false positives. However, QAPA apparently reports back all sites it has encountered in its reference DB, but if it didn't find them expressed the score will be set to zero. By considering all non-matched PD sites false positives, for QAPA that number is thus overestimated. This effect is not relevant for the "core" abs quant metrics, but as we also calculate precision and sensitivity it might still influence how we represent QAPAs performance.
Whether or not we should filter for non-zero sites should be open to debate though, as one could argue that this return of all reference DB sites is a design decision of the developers, and we'd give QAPA an advantage if we filtered out those with a particular (in this case zero) score.
Opinions welcome, the implementation shouldn't be complicated, small adjustment of
compute_metrics.py
. We'd however have to re-run the abs quant SWF for QAPA for all challenges.