3.3 PLS-DA Analysis
The PLS-DA models were developed for the three microorganisms using the two optimal pre-processing algorithms determined with the PCA models. Therefore, PLS-DA models were made using SG1+SNV and SNV+SG2, respectively. The best factor for each PLS-DA model was found by comparing the classification parameters. Next, the results from both pre-processing methods are compared.
The optimal results for Ml were obtained with SNV+SG2, Sewith SG1+SNV, and Sa with SNV+SG2 from this analysis. ForMl , the PLS-DA model discriminates all samples perfectly with a %T of 100%, %F of 100%, SEN of 1.0, ξ of 1.0, and MCC of 1.0 for both pre-processing methods. Furthermore, the Ml PLS-DA model classifies Sa without difficulty, while the Ml samples are close to the threshold value of zero. Se appears close to the threshold, indicating slight difficulty; however, the samples are farther from the threshold than Ml . For Se, each group has a larger distance to the threshold than the Ml model; however, the model classifies two samples incorrectly from two different classes. Table 2 indicates that SNV+SG2 has better results than SG1+SNV for Se with a %T of 98.4%, %F of 1.6%, SEN of 1.0, ξ of 1.0, and MCC of 1.0. The PLS-DA models for Sa discriminate all samples perfectly with the same results as Ml , as shown in Table 2. Visualizing the predicted values vs. the number of samples shows that Sa has better classification than Ml and Se due to the values being far from the threshold and the clusters having a more compact structure. On the other hand, if separating all three microorganisms is desired, the Ml PLS-DA model with SNV+SG2 shows the best results since all the clusters do not overlap.
(A)