Fig. 4 | Nature Communications

Fig. 4

From: Community assessment to advance computational prediction of cancer drug combinations in a pharmacogenomic screen

Fig. 4

Feature impact. Drug target annotation is key in top-performing algorithms, as is the meta information about variants including their functional impact and tumor driver gene status. a Cross validation-based distributions of NAD primary metric of SC1B when replacing or adding drug/cell line label with respective features. NAD baseline model (red) used cell line labels and drug labels only as feature inputs. In the other models different drug specific (drug targets, drug target KEGG pathway memberships, drug target-associated Gene Ontology terms or direct interactions between drug targets in a signaling network) or cell line specific (mutations or CNVs of selected, cancer related genes) features (green and blue, respectively) were added either in place of or in addition to the baseline model features. Ensemble model (cyan) is the averaged prediction of the different models. Single asterisks refers to t-test P < 0.05, double asterisks for P < 0.01, and triple asterisks for P < 0.001 compared to baseline model. b Heatmap of decrease in performance (average weighted Pearson correlation) of SC1B for DMIS support vector regression method when a particular feature type is removed (diagonal) or two feature types are removed at once (off diagonal)

Back to article page