Illustration of the training process. First, signals corresponding to classified state 1 (one amino acid bound) and state 2 (two of the same amino acid bound) for each type of amino acid were imported and normalized. Then, the state 1 blockade, dwell time and s.d. were extracted. Additionally, 1,000 data points, named feature X0001X1000, were extracted from the current density of each signal (from 0 to 1 with an interval of 0.001). Model performance was tested, including RF, NB, NNet, KNN, bagged CART and AdaBoost. RF outperformed the other models, achieving an AUC of 0.990. A tenfold cross-validation was used to prevent overfitting.
