Entropy of ATLAS AACF Susceptibility Profiles across Cancer Cell Lines: Dependence on Model, Normalization Scale, and Checkpoint Aggregation
Year: 2026
Type: article
Abstract: This record accompanies Working Paper No. 9 of the Independent FAIR Evaluation Series of the ATLAS Framework, entitled Entropy of ATLAS AACF Susceptibility Profiles across Cancer Cell Lines: Dependence on Model, Normalization Scale, and Checkpoint Aggregation. The study examines the concentration of computational AACF susceptibility profiles derived from frozen PaccMann and DRPreter predictions for 95 unique registered configurations across 970 matched cancer cell-line profiles. Scores are transformed into normalized weights using a softmax transformation favouring lower AACF indices, with fixed model-specific reference scales. Profile characterization includes Shannon entropy, normalized entropy, effective configuration counts, and leading configuration weights. The evaluation addresses five normalization-scale multipliers, the effect of retaining 97 registered labels instead of 95 unique configurations, lineage comparisons covering 953 profiles in 21 groups, and variability across ten DRPreter checkpoints. It also distinguishes entropy calculated after median index aggregation from the median of checkpoint-wise entropies. At the reference scales, median normalized entropy was 0.9346 for PaccMann and 0.9528 for the DRPreter median consensus. These aggregate differences coexisted with weak cross-model agreement in cell-line entropy rankings (Spearman ρ = −0.0788) and only six shared profiles among the 97 most concentrated profiles identified by each model. Normalization and checkpoint aggregation affected concentration summaries and rankings, demonstrating that entropy requires an explicitly specified computational representation. The accompanying computational materials document the six-stage workflow through analysis scripts, frozen input resources, configuration and profile mappings, normalization policies, numerical tables and arrays, saved bootstrap draws, and provenance and validation records. A README explains the directory structure, principal outputs, reproduction workflow, and interpretation limits. Resources inherited from ATLAS Working Papers 06–08 retain their upstream licensing and attribution requirements. The final technical audit recorded 2,363 passed checks and no failed checks, reviewing 206 distinct files in the selected analysis chain. This audit supports technical consistency and traceability within the implemented checks; it does not constitute independent experimental or clinical validation. The results establish entropy as a conditional descriptor of computational profile concentration. They do not establish therapeutic efficacy, measured mixture responses, biological mechanisms, or malignant-versus-normal selectivity. The materials support transparent inspection, reanalysis, and further methodological evaluation of the ATLAS framework.
Author: Tsanov, Vasil
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225619997
FWCI:
Cites:
Cited by:
Related to: