e. it has a value of 1 if the target is inhibited by the drug and a value of zero if the target is not inhibited by the drug. Then, we define the similarity measure as Note that 1 and similarity between drugs with no overlapping targets selleck inhibitor is zero. If two drugs have 50% targets overlapping with same EC50 s, then the sim ilarity measure is 0. 5. The similarities between the drugs Inhibitors,Modulators,Libraries are shown in Additional file 5. Note that except two drugs Rapamycin and Temsirolimus that have a similar ity measure of 0. 989, all other drugs have significantly lower similarities with each other. The maximum simi larity between two different drugs is 0. 169. This shows that any two drugs Inhibitors,Modulators,Libraries in the drug screen are not Inhibitors,Modulators,Libraries significantly overlapping and the prediction algorithm is still able to predict the response.
The low error rate illustrates the accuracy and effec tiveness of this novel method of modeling and sensitivity prediction. Furthermore, these error rates are signifi cantly lower than those of any other sensitivity predic tion methodology we have found. Consistent with the analysis in, the sensitivity prediction rates improve dramatically when incorporating Inhibitors,Modulators,Libraries more information about drug protein interaction. To more effectively compare the results generated via the TIM framework with the results in, we also present the correlation coefficients between the predicted and experimental drug sensitivity values in Table 6. The correlation coefficients for pre dicted and experimentally generated sensitivities for 24 drugs and more than 500 cell lines ranges from 0. 1 to 0.
8 when genomic characterizations are used to predict the drug sensitivities in the CCLE study. In comparison, our approach based on sensitivity data on training set of drugs and drug protein interaction information Inhibitors,Modulators,Libraries produced correlation coefficients 0. 92 for both leave one out and 10 fold cross validation approaches for error estimation. It should be noted that the sensitivity prediction is per formed in a continuous manner, not discretely, and thus effective dosage levels can be inferred from the predic tions made from the TIM. This shows that the TIM frame work is capable of predicting the sensitivity to anti cancer targeted drugs outside the training set, and as such is viable as a basis for a solution to the complicated problem of sensitivity prediction.
In addition, we tested the TIM framework using syn thetic data generated from a subsection of a human cancer pathway taken from the KEGG database. Here, the objective is to show that the proposed TIM method gener ates models that highly represent the underlying biological network which was sampled via synthetic drug pertur bation data. This experiment selleck chemical Enzalutamide replicates in synthesis the actual biological experiments performed at the Keller lab oratory at OHSU. To utilize the TIM algorithm, a panel of 60 targeted drugs pulled from a library of 1000 is used as a training panel to sample the randomly generated network.