Factor-Augmented Neural Conditional Density Estimation with Application to Drug Response Prediction

High-dimensional pharmacogenomic data pair thousands of molecular features with hundreds of cancer cell lines, complicating conditional density estimation. We study factor-augmented CINDES, which estimates conditional response densities from principal-component scores through binary classification. For an honest pilot/main-sample construction, assuming fresh-score PCA recovery, we derive a risk bound separating conditional-density learning, PCA score error, and representation bias. This guarantee does not cover the implemented procedure, which reuses observations for PCA and CINDES. In favorable strong-factor simulations, PCA improved a fixed CINDES pipeline, but PCA-MDN achieved lower negative log-likelihood (NLL) and integrated $L^1$ error in the matched comparison. Stress tests showed losses from omitted response-relevant low-variance directions and weak factor recovery, with sample-size-dependent factor-count effects. Across all 692 eligible source-specific genomics of drug sensitivity in cancer (GDSC) units and five splits, mean held-out NLL was 1.5414 for PCA-CINDES, versus 1.5661 and 1.5557 for ElasticNet and PCA/Ridge Gaussian models; PCA-MDN was better at 1.5156. A previously fixed 35-unit audit favored PCA-CINDES over Raw-X CINDES (mean paired NLL difference $-0.178$ ) and found lower measured computational cost. The evidence supports PCA compression when retained factors capture response information and are accurately estimated; it does not establish robustness to general nonlinear or nonfactor structures, or CINDES as the preferred second-stage estimator. The application concerns preclinical prediction rather than clinical utility.