Аннотация
Single-cell RNA sequencing (scRNA-seq) studies increasingly combine expression-based and mutation-derived signals, but small-cohort designs with repeated runs from the same biological unit can make standard cross-validation overly optimistic. We reanalyzed PRJNA736095 (14 SRR runs from 7 GSM/donor proxies) using a GATK-centered RNA-seq variant-calling and gene-burden workflow, then evaluated mutation-derived, expression-only, and combined feature sets with leakage-safe preprocessing inside each validation fold. Run-level repeated stratified cross-validation showed high within-dataset separability for GATK gene-burden features (balanced accuracy 0.973 +/− 0.113), but GSM-grouped leave-one-GSM-out validation reduced balanced accuracy to 0.708 and exact GSM-level permutation testing was not significant (p = 0.257). Expression-only and combined feature sets did not improve GSM-grouped balanced accuracy over the variant-only branch. Expression–mutation marker overlap was not significant after FDR correction, and public external datasets were used only to define feasibility or processed biological context rather than as strong external classifier validation. These findings position the workflow as an auditable, hypothesis-generating framework and highlight pitfalls of standard cross-validation in small-cohort scRNA-seq machine-learning analyses.