Fig. 5. Quantitate comparison of monotherapy efficacy predictions in 12 AML patients.

Drug sensitivity score (DSS)³² distributions of the top-15 drugs predicted as (a) the most effective and (b) the least effective monotherapies by each model for individual patients (n = 12). For each patient, 15 technical replicates (top-15 drugs) were compared between models using two-sided pairwise Wilcoxon rank-sum tests, with p-values adjusted for the False Discovery Rate (FDR) with the Benjamini-Hochberg procedure. *a model’s predictions are significantly different (p < 0.05) compared to that of at least one of the two other methods; **a model’s predictions show a significant difference (p < 0.05) when compared to both of the alternative methods. Box plots show the median (central line), 25th and 75th percentiles (box edges), and the range within 1.5 times the interquartile range from the box (whiskers). c Receiver Operating Characteristic (ROC) curves for each model demonstrating their ability to distinguish between effective and ineffective treatments based on the predictions of the most effective and least effective drugs by each model for 12 individual patients (as shown in panels a and b). The shaded area around each curve represents the 95% confidence interval (CI), calculated around the mean, illustrating the variability of the predictions based on data from 12 patients. d A summary table displays the Area Under the Curve (AUC) values for each model, quantifying their overall predictive accuracy. Statistical comparison was performed with the two-sided DeLong’s test, indicating that the prediction performance of scTherapy was significantly better than that of scDrug and BeyondCell (p = 0.00064 and p = 0.0000054). The other two methods show statistically similar prediction results (p = 0.29). Source data are provided as a Source Data file.