Background: Safer and more effective mixtures of anticancer drugs are needed, and modeling can assist in this endeavor. This paper describes classification models that were constructed to predict which fixed-ratio mixtures created from a pool of 10 drugs would show a high degree of in-vitro synergism against H460 human lung cancer cells. One of the tested drugs was doxorubicin and the others were natural compounds including quercetin, curcumin, and EGCG. Explanatory variables were based on virtual docking profiles. Docking profiles for the 10 drugs were obtained for 1087 proteins using commercial docking software. The cytotoxicity of all 10 drugs and of 45 of the 1,013 possible mixtures was tested in the laboratory and synergism indices were generated using the MixLow method. Model accuracy was assessed using cross validation, as well as using predictions on a new set of 10 tested mixtures. Results were compared to models where explanatory variables were constructed using the pseudomolecule approach of Sheridan.
Results: On this data set, the pseudomolecule and docking data approach produce models of similar accuracy. Leave-one-out precision for the negative (highly synergistic) class and the positive (low- or non-synergistic) class was 0.73 and 0.80, respectively. Precision for a nonstandard leave-many-out cross validation procedure was 0.60 and 0.77 for the negative and positive classes, respectively.
Conclusion: Useful classification models can be constructed to predict drug synergism, even in those situations where a limited subset of component drugs can be tested. Compared to the pseudomolecule approach, the virtual docking approach has the advantage of greater potential for biologic interpretation. This distinction may become important as virtual docking software becomes more accurate and docking results more closely resemble actual binding affinities. This is the first published report of a model designed to predict the degree of in-vitro synergism based on the pseudomolecule or docking data approach.