Component Ablation and Rank Stability of the ATLAS AACF Susceptibility Index

Working Paper No. 8 (with FAIR) examines how the four molecular-role contributions—AF, CF, AAF and ACF—affect the numerical behaviour, candidate rankings and lineage contrasts of the ATLAS AACF susceptibility index. The study applies leave-one-component-out ablation to frozen molecular predictions, retaining the original coefficients and remaining contributions. Analyses include index reconstruction, registered-input deduplication, normalized rank displacement, Top-10 retention, lineage-specific comparisons, paired stratified bootstrap diagnostics and a matched PaccMann–DRPreter comparison. The main cross-model population comprises 95 distinct registered configurations and 970 eligible cell-line profiles. Primary lineage summaries cover 21 groups containing 953 profiles, while group-versus-rest contrasts retain the full eligible reference population. DRPreter is evaluated through a ten-checkpoint median and individual checkpoint summaries. Principal findings Under the fixed coefficients, AF removal produces the largest average normalized rank displacement: 0.12415 in PaccMann and 0.08758 in the DRPreter median. Corresponding mean Top-10 retention is 53.06% and 66.79%. AF ranks first and ACF last by average displacement in all ten DRPreter checkpoints. The AAF–CF ordering varies across checkpoints. Agreement in the average ablation hierarchy does not extend to candidate prioritization. For the full index, mean within-profile cross-model Spearman correlation is 0.03526, with 8.03% mean Top-10 overlap. Most observed lineage-contrast reversals do not meet the implemented bootstrap sign-pattern criterion. The unique PaccMann case satisfying both reported classifications does not show the same point reversal in the DRPreter median. Bootstrap diagnostics reveal displacement and interval-method sensitivity for several nonlinear summaries. Nominal interval coverage remains unvalidated. The final computational audit passes 268,703 implemented consistency checks, with no recorded failures. These checks support reconstruction, identity accounting and traceability of the retained outputs; they are not independent experimental validations. WP08 characterizes sensitivity of the existing weighted AACF formulation. Individual deletion effects remain conditional on the coefficients, correlated components, prediction pipeline and analysis population. The study does not isolate unique component information, establish biological causality or validate experimental anticancer efficacy. This working paper continues the ATLAS evaluation series by distinguishing computational integrity, ranking sensitivity, exploratory lineage-contrast stability and cross-model candidate agreement.