Fig. 3

Targeted Proteomics and Support Vector Machine Classification Identifies Prospective Biomarkers to Distinguish HGSC vs Benign Disease. 471 peptides corresponding to 240 proteins were analysed in EV-enriched blood plasma from HGSC (n = 10) versus control (n = 9) donors using PRM. (A) Volcano plot highlights peptides that were significantly different between malignant and control donor samples. 21 peptides (p-value < 0.05) and HGSC antigen MUC16 (red) were selected for further analyses. B Unsupervised PCA and k-means clustering of pooled samples. Predicted labels (red and black) partially overlapped with true labels (blue = Benign and orange = HGSC). V-measure = 0.603. C Hyperparameter tuning of the linear SVM was performed by LOOCV, leading to hyperparameters C = 0.025–1 and two principal components selected as the ‘optimized’ SVM based on mean accuracy score (> 0.90). Each point of triangulation indicates an SVM combination/fit that was scored using the training set. Feature selection was performed using 231 combinations of peptides and test data. From this analysis, nine combinations of peptides provided an accuracy score of 1.0 on the test data set (see Figure S4A). D, E For example, the combination of CFHR4 and MUC1 provided a Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) score of 1.0. F Training (red) and test samples (white) were represented by women with Stage I, II, and III EOC