Interpretable deep recommender system model for prediction of kinase inhibitor efficacy across cancer cell lines - PubMed (original) (raw)
Associations between cell lines biological processes (horizontal axes) and all of the 74 analyzed drugs (vertical axes), plotted separately for (a) RTK signaling, (b) PI3K/MTOR signaling, (c) ERK MAPK signaling, (d) Cell cycle and (e) Others target pathways. First, for every drug, the Spearman correlation coefficient between every cell line hidden dimension and the response is computed, as shown in Fig. 3. These correlations are then assigned to biological processes associated with a given hidden dimension (see Fig. 2). If more than one hidden dimension is related to a process, an average of correlation is taken and assigned to the process. Drugs were hierarchically clustered using the Euclidean distance and average linkage within a given target pathway category. The horizontal axis is shared for all panels. The vertical axis tick label is formatted as: Drug Name; Putative Targets. Drugs target pathways and putative targets are taken from GDSC annotations. The labels are color-coded by the target pathway. For some drugs, putative targets have not been listed for readability. Those targets are: * – PI3K (class 1), MTORC1, MTORC2, ** – PI3Kalpha, PI3Kdelta, PI3Kbeta, PI3Kgamma, *** – CDK1,CDK2,CDK5,CDK7,CDK9, PKC, **** – RC, ROCK2, NTRK2, FLT3, IRAK1, others, ***** – BRSK2, FLT4, MARK4, PRKCD, RET, SRPK1. The color scale of correlation heatmaps is the same for all categories. See Fig. 2 for term names abbreviations. For graphics, we used Matplotlib, seaborn, and Inkscape version 1.1 (
).