Robustness of Categorical Anticancer Susceptibility in the ATLAS Framework: Bootstrap, Leave-One-Group-Out and Exact Statistical Validation

Robustness is not the same as favourable susceptibility. This study examines that distinction directly within the ATLAS framework by testing whether categorical anticancer susceptibility assignments remain stable when the underlying transcriptomic dataset is deliberately perturbed. The analysis covers 97 active anticancer molecular forms across 1,999 cancer cell-line records and 32 tumour groups. Instead of ranking integrated AIC50A_{IC50} values continuously, the study evaluates six predefined ATLAS susceptibility zones and asks three related questions: how concentrated each categorical profile is, how consistently its dominant zone is reproduced under resampling, and how strongly that dominant category is supported against its nearest competitor. To address these questions, the workflow combines 2,000-replicate row-wise bootstrap resampling, leave-one-cancer-type-out analysis, normalized Shannon entropy, dominance-margin assessment, total variation distance, exact one-sided binomial testing, and Benjamini–Hochberg false-discovery-rate correction. The resulting framework separates favourable profiles from profiles that are statistically stable but predominantly outside the adopted ATLAS susceptibility criterion. The final analysis identifies statistically supported dominant-zone predominance for 89 of 97 AACFs and shows that categorical stability can be both favourable and unfavourable. The principal contribution of the study is therefore not a new potency ranking, but a reproducible method for distinguishing stable, unstable, favourable, and outside-criterion susceptibility patterns within a categorical transcriptomic framework. The work is intended as a methodological and statistical validation layer for future computational prioritization, independent replication, and experimental investigation.