Fig 1.
PLSR modeling workflow applied on 183 cancer cell lines on OncoPanel.
(A). Flow chart on the model building and testing steps. (B). A specially designed splitting strategy divides the training dataset into random training, random validation and balance validation subsets. (C). Representative example of random validation and balance validation. Red points were top performing models on 1000 random splits on this balanced split, based on both AUC and correlation measures. (D). AUC and correlation cutoff selection for the core PLSR model.
Fig 2.
Causal network to depict functional relations between sensitivity-specific and resistance-specific signature genes.
The network was reconstructed from canonical signaling pathways regulated by signature genes and a signature specific direct interaction network. Sensitivity-specific signature genes are highlighted with blue thermometers, resistance-specific genes with red thermometers.
Fig 3.
PLSR models performance in predicting Erlotinib-treated patient survival in the BATTLE trial.
A. Erlotinib model predicting Erlotinib treated patients; B. Sorafenib model predicting Sorafenib treated patients; C. Erlotinib model predicting Sorafenib treated patients; and D. Sorafenib model predicting Erlotinib treated patients. TP: true positive; FP: false positive; TN: true negative; FN: false negative; PPV: positive predictive value.
Fig 4.
Survival analysis on biomarker identified treatment sensitive/resistant sub-groups.
A. Using the Erlotinib model to stratify Erlotinib treated patients; B. Using Sorafenib model to stratify Sorafenib treated patients; C. Using Erlotinib model to stratify Sorafenib treated patients; and D. Using Sorafenib model to stratify Erlotinib treated patients.
Table 1.
Predicted percentage of Erlotinib and Sorafenib sensitive samples for some cancer indications from Gene Expression Omnibus datasets.