Assessing the use of foundation model embeddings to improve drug response predictions using scRNA-Seq expression data

Cancer drug response (CDR) prediction is challenging due to tumor heterogeneity and limited amount of high-quality response data. Large foundation models trained on single-cell RNA-Seq data have been reported to improve the performance of a variety of downstream tasks, so in this study we evaluated whether CDR predictions could be improved when integrating foundation model embeddings with DeepCDR, a deep learning model that combines drug structure convolutions with gene expression embeddings. Our results show that using foundation model embeddings tested improved CDR interpolation predictions, with the strongest results obtained when using the scGPT - cancer model embeddings. Performance of cell line extrapolation CDR predictions was strong across diverse cell lines, with small and variable improvements when using these embeddings. Drug extrapolation results across chemically diverse drugs were poor, even for structurally similar drugs, showing the need for larger datasets or for earlier integration of the different data sources. Overall, our results highlight the benefit of using foundation model embeddings for the predictions of CDRs and opportunities for architectures with greater cross-modal integrations but highlight the need for further improvements.