Benchmarking ATLAS Predicted Susceptibility Profiles against Reference Anticancer Drugs: Paired Model Comparisons and Structural Audits

This working paper (with FAIR data) presents WP06 of the ATLAS evaluation series, connecting predicted susceptibility profiles with a reference anticancer-drug panel and complementary molecular-identity and structural audits. It addresses three questions: how ATLAS profiles compare with reference-drug profiles under a shared inference procedure; what limitations affect expansion of the reference registry; and which structural and precursor-to-active-form relationships are reproducible under explicit computational rules. Paired model comparison. The analysis evaluated 97 AACF entries against methotrexate, doxorubicin, paclitaxel, imatinib, and olaparib across 1,019 CCLE expression profiles using a fixed PaccMann checkpoint and documented preprocessing. The molecular inventory comprised 314 distinct ATLAS input strings and five reference-drug inputs. All 485 integrated AACF–drug comparisons showed positive Spearman correlations, with a median of 0.622 and a range of 0.455–0.835. Methotrexate showed the highest median rank concordance within this panel (ρ = 0.739). Individual molecular-form distributions were reported separately from the integrated AACF index, whose weighted values were not equated with single-drug IC₅₀ values or concentrations. Reference-registry coverage and identity. The broader resource retained 100,531 source-agent records and incorporated 9,563 retrieved PubChem compound records. Completion of the retrieval queue was distinguished from verified chemical identity or comprehensive anticancer-drug coverage. Evidence tracing identified 16,552 candidate links supported exclusively by trial aliases, demonstrating why a unique candidate representation does not necessarily establish the identity of its associated source agent. Structural and transformation audits. Searches of 122 distinct assigned active-form graphs against 18,220 reference graphs identified 243 primary identity or chirality-compatible containment pairs, alongside a separate stereochemistry-ignored sensitivity set. Additional chemical-environment constraints retained 81 primary pairs involving 72 reference graphs and 13 AACF entries. Component localization distinguished whole-reference relationships from matches confined to smaller components. A separate audit recovered 176 of 194 registered precursor-to-active-form correspondences under a restricted graph-cut model. The remaining cases comprised ten stereochemical mismatches, six exact recoveries requiring expanded C–O bond eligibility, and two recoveries involving C–C cleavage in the identified diagnostic construction. These categories were retained separately, without automatic correction of the molecular registry. Scope within the ATLAS programme. WP06 examines a defined computational portion of the broader research concept, which distinguishes pharmacological precursors from molecular forms proposed to arise intracellularly. Profile concordance, structural overlap, and graph consistency provide different kinds of evidence. The present results do not establish clinical efficacy, experimentally observed intracellular activation, transport, toxicity, or tumour selectivity. Possible training overlap, unverified historical preprocessing equivalence, and the absence of a control-compound baseline also constrain interpretation. The contribution is a traceable computational benchmark and a documented basis for further controls, reference-identity adjudication, and experimental investigation. Both resolved findings and unresolved assignments are preserved so that subsequent work can test, refine, or challenge the underlying assumptions.