Interest ingly, CDK4 is also prominent in all targets this circuit, which is a primary inhibitor of the tumor suppressor pRb, which is also frequently abnormal in spontaneous human osteosar coma. CDK2 is an important modifier of both p53 and pRb and is also represented in this circuit. The importance of PI3K pathway in osteosarcoma has also been recently reported using high throughput genotyping. Our TIM circuit includes AKT2 which is down stream of PI3K. Also, EDNRA selected in the circuit has been known to interact with PKC and activate ERK signaling. If the circuit models shown in Figures 2 and 3 are used to predict sensitivities for comparison with experimen tally generated data, we will get optimistic results as the models are trained using the entirety of the available data.
Thus, we utilize Leave One Out and 10 fold Cross Validation approaches to test the validity of the TIM framework that we Inhibitors,Modulators,Libraries present Inhibitors,Modulators,Libraries in this paper. For the LOO approach, a single drug among the 44 drugs with known inhibition profiles is removed from the dataset and a TIM is built, using the SFFS suboptimal search algo rithm, from the remaining drugs. The resulting TIM is then used to predict the sensitivity of the withheld drug. The predicted sensitivity value is then compared to its experimental value. the LOO error for each drug is the absolute Inhibitors,Modulators,Libraries value of the experimental sensitivity y minus the predicted sensitivity, i. e. y ? . The closer the predicted value is to the experimentally gener ated sensitivity, the lower the error for the withheld drug.
Tables 1, 2, 3 and 4 provides the complete LOO error tables and the average Inhibitors,Modulators,Libraries LOO error for each primary culture. The average LOO error over the 4 cell cultures is 0. 045 or 4. 5%. For the 10 fold cross validation error estimate, we divided the available drugs into 10 random sets of similar size and the testing is done on each fold while being trained on the Inhibitors,Modulators,Libraries remain ing 9 folds. This is repeated 10 times and average error calculated on the testing samples. We again repeated this experiment 5 times and the average of those mean abso lute errors for the primary cell cultures are shown in Table 5. The detailed results of the 10 fold cross valida tion error analysis are included in Additional file 4.
We note that both 10 fold CV and LOO estimates for all the cultures have errors less than 9%, which is extremely low, especially considering the still experimental nature of the contain drug screening process performed in the Keller laboratory and the available response of only 44 drugs with known target inhibition profile. To provide a measure of the overlap between drugs, we based on the EC50 of the drugs D1 and D2. Let the EC50 s of the drugs D1 and D2 be given by the n length vectors E1 and E2 where n denotes the number of drug targets. The entries for the targets that are not inhibited by the drugs are set to 0.