Response classification changes the apparent predictive value of Par-6 family cell polarity regulator gamma for...

Xu et al. reported that Par-6 family cell polarity regulator gamma discriminated between response groups after programmed cell death protein 1 blockade, with an area under the receiver operating characteristic curve of 0.823. The source study classified the same 28 patients as nine responders and nineteen nonresponders. Xu et al. instead analyzed sixteen and twelve after including seven minimal responders in the responder group. We reanalyzed pretreatment expression using the target, source, exclusion and ordinal definitions. The area under the curve was 0.807 under the target definition and 0.602 under the source definition, a paired difference of 0.205 (95% confidence interval, 0.008-0.415); excluding minimal responders yielded 0.750. An alternative normalization method produced the same contrast. These findings do not diminish the biological relevance of the marker, but show that its discrimination is classification-sensitive in this public cohort. Prespecified response coding and patient-level resampling of the complete feature-selection pipeline are needed before the marker is advanced as treatment-predictive.