#36 · Single-cell

Correcting RNA-seq scale bias and inflated single-cell perturbation prediction metrics

The three items share a concern with the validity of inference from transcriptomic data rather than a single methodological advance. One identifies sample-specific, expression-level-dependent scale bias in TCGA, SU2C and GTEx RNA-seq and proposes nonlinear transforms to correct it. StabCell addresses reproducibility of single-cell analysis by coupling clustering with marker gene detection through complementary-pair subsampling stability selection. The third shows that correlation and cosine metrics in single-cell perturbation prediction are inflated when models share a common control population. What is new is the explicit evaluation of artifacts at the level of samples, cell partitions and benchmark design, paired with procedures that correct or stabilize the result. The thread is heterogeneous, so these items should be read as complementary validity checks rather than a unified framework.