Deconvolution of ex-vivo drug screening data and bulk tissue expression predicts the abundance and viability of cancer cell subpopulations

Ex-vivo drug sensitivity screening allows the prediction of cancer treatment effectiveness in a personalized fashion. However, it only provides a readout on mixtures of cells, potentially occulting important information on clinically relevant cell subtypes. To address this shortcoming, we developed a machinelearning framework to decompose drug sensitivity recorded at the bulk level into cell subtype-specific drug sensitivity. We first determined that our method could decipher the cellular composition of bulk samples with top-ranking accuracy across five cancer types compared to state-of-the-art bulk deconvolution methods. We emphasize its effectiveness in the realm of Acute Myeloid Leukemia, where it appears to offer the most precise estimation of leukemic stem cell fractions across three test datasets and three patient cohorts. We then optimized an algorithm capable of estimating cell subtype- and single-cell-specific drug sensitivity, which we evaluated by performing in-vitro drug studies and in-depth simulations. We then applied our deconvolution strategy to the beatAML cohort dataset, currently the most extensive database of ex-vivo drug screening data. We developed a drug sensitivity profile tailored to specific cell subtypes, focusing on four therapeutic compounds predicted to target leukemic stem cells: the previously known midostaurin and A-674563, as well as SNS-032 and foretinib, which have not been previously linked to leukemic stem cells. Our work provides an attractive new computational tool for drug development and precision medicine.