Answers are grounded in this paper and cite their sources.
ATLAS Working Paper No. 7 (with FAIR) examines how drug-response predictions depend on the model, evaluation population, and generalization task. It distinguishes three forms of evidence: faithful computational implementation, agreement between model outputs, and predictive performance against observed experimental responses. The cross-model branch compares inherited PaccMann predictions with locally reconstructed DRPreter predictions and exported pdCSM-Cancer predictions. Analyses cover molecular rankings, derived AACF indices, cell-response profiles, leading-candidate overlap, parent–released molecular forms, and shared cell effects. On the principal cohort of 1,014 cell profiles, PaccMann–DRPreter global concordance across 97 AACF indices was low (Spearman 0.0404), with no shared global Top-10 entries. Moderate entity-profile agreement was substantially reduced after the specified shared-cell subtraction. On 46 matched cell lines, PaccMann–pdCSM-Cancer global concordance was higher for pharmaceutical forms than released forms (0.6549 versus 0.3858), without comparable agreement in entity-specific cell profiles. The experimental branch evaluates Macau through SMURFF using 69,437 observed GDSC2 responses, 150 drug IDs, and 554 cell profiles. Completed five-fold evaluations show better performance under new-cell exclusion (pooled Spearman 0.8494, RMSE 1.2614) than new-drug exclusion (Spearman 0.4194, RMSE 2.4732). New-drug evaluation also revealed weak within-cell drug ranking and MCMC diagnostic failures affecting 2.44% of predictions. Both-new evaluation was restricted to one pilot split; Macau predictions for ATLAS query molecules were not generated. The working paper documents molecular and cellular identity resolution, model-specific preprocessing, checkpoint verification, frozen validation splits, baseline procedures, posterior reconstruction, and diagnostic criteria. It defines the records and dependencies required to reproduce the reported calculations. The findings provide a task-specific assessment of computational support for ATLAS candidate prioritization. Cross-model agreement is interpreted as ranking consistency under stated conditions, rather than experimental validation. The study does not establish anticancer activity, clinical efficacy, or therapeutic safety of the ATLAS candidates.
No takes yet. Share an insight, caveat, or question.
Vasil Tsanov (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: