Fig. 5: Comparison of the predictive performance of the best-performing models using their predictive features, integrated with TME descriptors, immune response scores and the EaSIeR score for atezolizumab response prediction in 320 mUC patients from the merged discovery and validation datasets. | npj Precision Oncology

Fig. 5: Comparison of the predictive performance of the best-performing models using their predictive features, integrated with TME descriptors, immune response scores and the EaSIeR score for atezolizumab response prediction in 320 mUC patients from the merged discovery and validation datasets.

From: Predicting atezolizumab response in metastatic urothelial carcinoma patients using machine learning on integrated tumour gene expression and clinical data

Fig. 5

Boxplots comparing MCCs obtained from five 10-fold CV runs (each dot represents a run) of the best-performing models: LGBM combined 49 predictive genes (A), and CART-OMC combined 63 genes with TMB and TNB (B), to those combining with five TME descriptors, nine published immune response scores (Ock_IS is not available in the EaSIeR tool), and EaSIeR score. These features account for the cellular composition of TMEs together with inter- and intracellular communication, offering a comprehensive view of tumour-immune interactions affecting the efficacy of immunotherapy. Additionally, immune response scores and the EaSIeR score are proposed as predictors of response to ICIs. Random-level performance is delimited by the horizontal dashed lines at an MCC of 0.0. Overall, integrating five TME descriptors, nine immune response scores, and the EaSIeR score into the predictive model improved the predictive performance compared to those built solely on predictive genes. Specifically, nine immune response scores contributed significantly to improving the accuracy of both LGBM and CART models.

Back to article page