Abstract
<title>Abstract</title> <p>Large pharmacogenomic screens promise to predict which drugs a tumour will respond to from its molecular profile, yet it remains unclear how much of the apparent predictive skill, that is, performance in excess of a baseline that ignores the molecular profile entirely, reflects genuine molecular personalization rather than simpler structure in the data. Here I integrate two independent drug-sensitivity screens, the Genomics of Drug Sensitivity in Cancer (GDSC) and the Cancer Therapeutics Response Portal (CTRP), with multi-omic profiles from the Cancer Cell Line Encyclopedia (CCLE), resolving cell-line identity through Cancer Dependency Map (DepMap) identifiers and drug identity through chemical structure. The harmonized resource spans 678 drugs across 1,171 cell lines (487,799 measurements), with a cross-cohort-comparable sensitivity metric validated against known biology. Using grouped, leave-cell-line-out nested cross-validation scored against a drug-mean baseline, I decompose prediction into a broad-spectrum component (which drugs are generically potent within a tissue) and a mechanism-specific component (personalization surviving after drug means are removed). The broad-spectrum component is strong within every tissue, which identifies it as tissue-by-drug structure rather than molecular personalization; it necessarily vanishes when tissues are pooled, because the metric is standardized per drug across the whole panel, so the pooled comparison confirms the interpretation rather than establishing it. Pooling all tissues does expose a small personalization signal that survives removal of tissue-of-origin structure (held-out Spearman ρ ≈ 0.05 against a within-tissue permutation null), though held-out skill falls by roughly two-fifths once tissue is removed, a tissue-only baseline outscores the molecular model on the same target, and the null constrains lineage rather than every line-level confound. For per-drug discovery I introduce a two-stage significance protocol, a permutation null followed by independent corroboration, and apply it across all 7,836 drug–tissue panels: the calibrated null retains roughly a threefold smaller set than cross-validation confidence intervals admit, leaving candidate biomarkers for a minority (~12%) of drug–tissue pairs at a false-discovery proportion of roughly 0.4, a set that is enriched over fourfold for established drug–target pharmacology and that surfaces novel, recurrent vulnerabilities, most strikingly a melanoma ferroptosis signal shared by three independent GPX4 inhibitors. Taken together, these results support baseline-anchored, permutation-controlled evaluation and drug recommendation that is tiered by confidence.</p>