1. Predicting drug sensitivity from proteomic data using PLS.
Predicting drug sensitivity from proteomic data using PLS. (A) PLS predicts drug AUC from proteomic (RPPA) data by defining new variables as linear combinations of the original (phospho)proteins, such that the new dimensions are maximally correlated with AUC. The agreement between measured and predicted AUC after repeated 10-fold cross-validation was quantified using Spearman rank correlation. (B) Selected drug AUC predictions for MMAE (cytotoxic), dabrafenib (BRAFi), vemurafenib (BRAFi), and gemcitabine (cytotoxic, failed prediction). Predictions are averaged over 100 repeats. (C) Drug AUC predictions compared with randomly shuffled background distributions (N = 8,192). (D) PLS defines a linear model from proteomic data to AUC for each drug. The (normalized) PLS coefficients suggest associations between (phospho)proteins and drug sensitivity or resistance. Unsupervised hierarchical clustering of the linear drug sensitivity models predominately clusters drugs sharing common targets. (E) Most important markers for MAPK inhibitors, as well as the most important differential markers between MEK and BRAF inhibitors. (F) Most important markers for cytotoxic drugs. (G) Most important markers for HDAC inhibitors.
Mattias Rydenfelt et al. LSA 2019;2:e
© 2019 Rydenfelt et al.