When More Omics Is Not Always Better: Drug- and Dataset-Conditional Value of Multimodal Cancer Pharmacogenomics

Background: Multimodal pharmacogenomic models can use more biological information, but additional assays also increase cost, dimensionality, missingness, and integration complexity. It remains unclear whether copy-number and mutation data provide reliable improvement beyond an expression-only model when validation is organized around held-out drugs and response datasets. Methods: This study compared expression-only and early-fusion Ridge models using expression, copy-number, and damaging-mutation data from GDSC1, GDSC2, and CTD2 across a declared 15-compound benchmark. Feature selection, imputation, and scaling were performed within each training fold to limit leakage. Results: In five grouped drug-held-out splits, fusion achieved a mean RMSE of 0.2322 compared with 0.2328 for expression alone, an improvement of 0.0006; fusion performed better in four splits and worse in one. In leave-one-dataset-out validation, fusion was less favorable in all three source-held-out tests. Modality availability also differed substantially, particularly for copy-number data. Conclusions: These findings suggest that additional omics may provide a small benefit within a defined response system, but the benefit is variable and may not transfer across datasets. For early drug discovery and biomarker-model development, the results support retaining additional assays only when their incremental value is reproducible, their availability matches the intended population, and their benefit is confirmed by external validation. The study evaluates a cell-line-only model and therefore does not establish arbitrary new-drug prediction, clinical treatment selection, or causal biological effects.