In addition, we employ scalable bounds around the IC50 not s to determine binariza tion values of the numerous kinase targets for each drug. The bounds can be scaled to allow targets that may have EC50 s higher than the IC50 to be considered as a possi ble treatment mechanism. We extend the bounds to low EC50 levels, and often down to 0, to incorporate the possibility of target collaboration at various different EC50 levels. While a high IC50 indicates the likelihood of drug side targets as therapeutic mechanisms, it does not pre clude the possibility of a joint relationship between a high EC50 target and a low EC50 target. Hence, to incorporate the numerous possible effective combinations implied by the IC50 of an effective drug, the binarization range of tar gets for a drug is the range log log B log where 0 B.

For reliability and validity of the target set that we aim to construct, it is important to keep B in a reasonable range, i. e. B should be a smaller constant such as 3 or 4. For the situation where the above bounds do not result in at least one binarized target, the immediate option is to eliminate the drug from the data set before target selection. This prevents incom plete information from affecting the desired target set. As information concerning the drug screen agents gradually becomes complete with respect to other forms of data, such as gene interaction data, additional mechanisms for unexplained targets can be explored and incorporated more readily into the predictive model. With binarization of the data set as explained, we now present the minimiza tion problem that produces a numerically relevant set of targets, T.

to achieve an IC50 within the allotted dosage are given the score of 0, which means ineffective. The Cmax value is used to apply a variable score to the numerous drugs based on the inherent toxicity of the drug. This will also pre vent bias towards drugs with low IC50s, some drugs may achieve efficacy at higher Brefeldin_A levels solely based on the drug EC50 values. Construction of the relevant target set In this subsection, we present approaches for selection of a smaller relevant set of targets T from the set of all possible targets K. The inputs for the algorithms in this subsection are the binarized drug targets and continuous sensitivity score. With the scaled sensitivities, we can develop a fitness function to evaluate the model strength for an arbitrary set of targets. As has been established, for any set of targets T0, drug Si has a unique representation. This representation can be used to separate the drugs into different bins based on the targets it inhibits under T0. Within each of these bins will be several drugs with identical target profiles but different scaled scores.