We performed a backward model selection with repeated tenfold cross-validation to reduce the feature set. The final model consisted of log sample size, log-fold change, log-fold change, and the interaction effects between log total sample size and log absolute cell count change as well as log absolute cell count and log-fold change, which was expected given that we observed an interaction effect between these two variables in the raw benchmark results (Supplementary Fig. 4d–f).
