Transfer single-cell foundation models to predict drug resistance in cancer

Single-cell RNA-seq captures cellular heterogeneity that can underlie drug resistance. Yet it remains uncertain whether single-cell–derived representations, including embeddings from foundation models, outperform carefully processed bulk transcriptomes for drug sensitivity prediction. This thesis addresses that gap through a head-to-head comparison of the two modalities. The analysis quantifies when single-cell signal adds value over bulk and tests whether current single-cell foundation-model embeddings surpass strong classical baselines. The study also evaluates cross modality generalization, examining whether models or embeddings trained on one data type improve prediction in the other. A reproducible pipeline links bulk RNA-seq from cancer cell lines to drug-response labels. Matched single-cell cohorts are constructed from the same lines. Per-drug regressors are trained under cell-line–blind cross-validation with fixed partitions. Bulk features apply standard normalization and dimensionality reduction. Single-cell features include pseudobulk summaries and embeddings from pretrained single-cell foundation models (scGPT, scFoundation). Across experiments, carefully processed bulk paired with simple, well-regularized models establishes a strong baseline. On matched cohorts, single-cell pseudobulk does not consistently exceed bulk; when present, gains are small and limited to a subset of drugs. Embeddings from single-cell foundation models likewise fail to surpass bulk across evaluation settings. Cross-modality transfer is ineffective: models or embeddings trained on one modality do not generalize to the other due to distributional mismatch. These results set practical expectations for modality choice in drug-response prediction.