Fig. 2

CaDRReS-Sc accurately predicts drug response in unseen cell types. a Overview of single-cell RNA-seq workflow to preprocess sequencing data and provide inputs to CaDRReS-Sc (indicated by blue dashed lines). The normalized read count values and cell clustering results are utilized by CaDRReS-Sc for predicting drug response, taking into account transcriptomic heterogeneity within each patient. b Overview of CaDRReS-Sc workflow, where a pre-trained pharmacogenomic space based on drug response and gene expression profiles from cell-line experiments is used to provide cell- or cluster-specific drug response predictions. These are then combined to estimate overall drug response and prioritize drug combinations for a patient. c Comparison of prediction accuracy on unseen cell types between CaDRReS-Sc’s objective function and a naïve function that does not take uncertainty in IC50 values into account. Each dot represents a drug (n=226), and dot colors represent the percentage of sensitive cell lines. As can be seen here, CaDRReS-Sc’s objective function is particularly useful when the percentage of sensitive cell lines is low. d Comparison of median absolute error (MAE) obtained based on predictions using CaDRReS-Sc as well as a naïve objective function. CaDRReS-Sc’s robust objective function results in lower MAE across a majority of drugs (points above the y=x line), especially for drugs with a lower percentage of sensitive cell lines (lighter shades). e Histograms showing the average prediction accuracy (error bars show 1 standard deviation) using different drug response prediction approaches. f Histograms showing MAE (error bars show 1 standard deviation) with different drug response prediction approaches. Overall, CaDRReS-Sc was seen to have high accuracy on the sensitive/non-sensitive classification task while reporting the lowest MAE for the IC50 regression task