Using game theory to guide the classification of inhibitors of human iodide transporters

ABSTRACTExperimental screening of small molecules against human iodide transporter, NIS, remains challenging, noisy and costly. Machine learning classifiers trained with data from experimental screens may assist in the prioritization of molecules for the follow-up studies. Driven by the need for improving the accuracy of these classifiers and by the availability of two independently acquired experimental datasets, a new data aggregation method is proposed, implemented and cross-validated. By representing a machine learning process as a collaborative game, importance scores of training molecules can be computed from their marginal contributions to classifier's performance. Truncated Monte Carlo simulations were conducted to compute these marginal contributions for the classifiers of NIS inhibitors. In cross-validation experiments, Random forest and eXtreme Gradient Boosting Tree classifiers trained with the highly valuable molecules, achieved an accuracy between 83% and 88%, sensitivity between 90% and 92% and specificity between 73% and 86%. The proposed approach for data valuation may have practical applications for computational decision making in chemical, environmental, and biomedical sciences.