Skip to main content
. 2022 Feb 10;25(3):103910. doi: 10.1016/j.isci.2022.103910

Figure 3.

Figure 3

Top genes related to malaria ART resistance as identified by SHAP feature importance analysis and performances of machine learning model after feature selection using the top ranked genes

(A) Top 30 genes of ART resistance prediction model visualized by SHAP analysis based on in vivo P. falciparum transcriptome.

(B) Top 30 genes of ART resistance prediction model visualized by SHAP analysis based on in vitro P. falciparum transcriptome. Mean SHAP values in both A and B were denoted by red dots. Genes were ordered by mean SHAP contributions across all test examples in a 10-fold cross validation. 25% and 75% percentile of SHAP values were denoted by low and upper limits of boxplots.

(C) Model performances for in vivo and in vitro predictions when including only top genes selected by SHAP analysis, as evaluated by AUROC (for binary labels) and C-index (for continuous labels). “All genes” shows prediction performance without feature selection.

(D) Comparison of in vitro prediction performances between using all top ten genes as features and leaving one gene out at a time. Each dot denotes the performce of each of the ten models in cross validation, and the average perfornences of all ten models were denoted by larger dots.