Strategies for improving the performance of prediction models for response to immune checkpoint blockade therapy in cancer

Table 2 Summary of perdiction performance

	Build models based on the 135 pairwise relations features of immune checkpoint genes			Build models based on the original 18,878 gene expression features
	Random forest	Lasso	XGBoost	Random forest	Lasso	XGBoost
	AUC	AUC	AUC	AUC	AUC	AUC
Combined	0.728	0.757	0.667	0.595*	0.552*	0.582*
Van Allen et al.	0.723	0.547	0.607	0.410	0.444	0.446
Hugo et al.	0.559	0.464	0.445	0.671	0.559	0.322
Riaz et al.	0.711	0.729	0.622	0.568	0.441	0.447

Models were built using Random forest, Lasso, XGBoost based on the combined dataset or an individual dataset. AUCs were calculated based on tenfold cross validation except for the case of using the Hugo et al. dataset alone, where a fivefold cross validation was performed because the sample size of the dataset was so small that the tenfold validation did not yield robust result
*Data had been normalized based on the Combat method [37, 38] when merging the three datasets

ISSN: 1756-0500