Transcriptomic Profiles Carry Most of the Multi-Omics Signal for Drug Sensitivity Prediction in Cancer Cell Lines: A Cross-Lineage and Independent-Screen Evaluation of Twelve Anticancer Agents

Background/Objectives: Combining omics layers is widely expected to improve cancer drug response prediction, but this is rarely tested under family-wise error control or on an independent screen. We measured what somatic mutation and copy number data add to gene expression, and how well models transfer to unseen lineages and an external screen. Methods: We linked the Cancer Dependency Map transcriptomic, somatic mutation and copy number profiles to GDSC2 dose–response data for twelve mechanistically diverse anticancer compounds in 509 cell lines. Lineage-only, single-modality, early-fusion and stacked late-fusion gradient-boosting models were compared by repeated cross-validation (DeLong tests, Holm adjustment), leave-one-lineage-out transfer across eleven solid tumour histologies, and external validation on 77-163 held-out PRISM cell lines per compound. Results: Gene expression models outperformed tissue-of-origin baselines for all twelve agents (AUC 0.69–0.84; gains 0.024–0.169; Holm-adjusted p < 0.001 for seven drugs). Adding mutation and copy number data changed AUC by −0.015 to +0.035, and no per-drug gain survived multiplicity correction. Pooled across compounds, early fusion added 0.008 AUC (95% CI 0.000–0.016), a gain that fold-level re-analysis did not support. Mixed-lineage cross-validation exceeded leave-one-lineage-out AUC by a median of 0.030 (p = 0.0005). External PRISM AUCs ranged from 0.46 to 0.76 and followed the concordance between the two screens (Spearman rho = 0.72, p = 0.008). TreeSHAP attributions recovered known determinants, including SLFN11, TP53, BRAF, PIK3CA and downstream MAPK transcriptional effectors. Conclusions: In this twelve-drug panel, gene expression captured most of the predictable response variation. Accuracy fell on held-out lineages and on an independent screen, which sets empirical limits on how far these models can generalise.