Fig. 3

ComboSC provides accurate response predictions for small-molecule drug combination therapy. a t-SNE plot for immune cells and malignant cells in sample su006 (GSM3511756). The colors on the left represent the different cell groups annotated by comboSC. The circled malignant cells in the left t-SNE plot are further divided into five subclusters in the right circle according to their various drug responses. b Radar map showing distinct drug responses of daporinad and sepantronium bromide in five malignant cell subclusters. c Overview of drug response-driven malignant cell graph clustering. ComboSC repeats the iteration for different graph cluster resolution until the drug response is maximized. d Top-ranked drug combinations in comboSC predictions for su006. The height of the bars represents the drug combination score. e UMAP plot for eleven malignant cell clusters in a breast cancer sample (GSE176078). f Radar map showing distinct drug responses of transfer learning-based method for the top two drugs AZD7762 and gemcitabine in eleven malignant cell subclusters. The drug response score at the black highlight circle is 12, equal to the black highlight circle in g. g Radar map showing the distinct drug responses of comboSC for the top drug combination shikonin and trametinib in eleven malignant cell subclusters. The drug response score at the black highlight circle is 12, equal to the black highlight circle in f. h Comparison of calculation time under different drug numbers between comboSC optimization and the traditional iterative method. i This bubble chart shows the count and enrichment significance of the top 10% of comboSC predictions registered in the ClinicalTrials.gov database. The x-axis is the fold enrichment of clinically validated drugs in the top 10% of comboSC predictions compared to all comboSC predictions, and the y-axis is the dataset. The color represents the p value of the enrichment analysis. j Performance of comboSC, comboSC-bulk, and CMAP predictions in three PDC datasets. The y-axis is the DCG score, with a lower DCG score indicating higher ranking accuracy