UCLA

Multiple drugs is much more effective in a cancer treatment than single drugs; however, it is challenging to find a reliable model that describes the relationship between drugs and the cell activities, as well as to construct efficient designs for reducing the number of drug combinations needed to be tested. We analyze data that consists of cellular ATP-levels on lung cancer cells and normal cells with 3 inhibitor drugs, AG490, U0126 and I-3'-M, each at 8 dosage levels. We construct models with different number of predictors and methods as well as different subsets of the full data. We find that large dataset does not guarantee the accuracy of a model's prediction, but an appropriate design and a good model building method are the keys. Square root transformation on the response gives a model prediction surpassing the performance among all other models we built. A second order model constructed with a 27-run factorial design performs remarkably well for both cancer cells and normal cells. It is in general much harder to model or predict cancer cells than normal cells.