121 Antibody epitope repertoires accurately discriminate between CD and healthy controls Subsequently, we aimed to examine the discriminative capacity of antibody epitope repertoires with regard to the presence of CD vs. healthy controls, the presence of UC vs. healthy controls and the presence of CD vs. UC. A selection of 2,687 antibody-bound peptides (after excluding coagulation-associated peptides, n = 128) was used for classification between CD (n = 256), UC (n = 207) and equally sized age- and sex-matched subsets of healthy controls.22 A logistic regression model with elastic net penalty was fitted based on the antibody epitope repertoires using 80% of the data as training set, with the goal of classifying patients with CD or UC from healthy controls and patients with CD from patients with UC (Figure 4, Table S9). When evaluating model performance on 20% of the data (test set), antibody epitope repertoires demonstrated a highly accurate discrimination between patients with CD and healthy controls (area under the curve (AUC) = 0.89, F1-score = 0.80) (Figures 4A–B). In this classification, antibody epitope repertoires had a sensitivity of 76% and specificity of 84% in predicting the presence of CD at a default probability threshold of 0.5 (Figure 4B), with a positive predictive value (PPV) of 83% and negative predictive value (NPV) of 78%. In contrast to CD, serum antibody epitope repertoires showed less, but still accurate, discrimination between patients with UC and healthy controls (AUC = 0.80, F1-score = 0.70). Here, sensitivity and specificity for the detection of UC were 71% and 68%, respectively, with a PPV and NPV of 69% and 70%, respectively. Finally, we assessed the predictive performance of antibody epitope repertoires between patients with CD and UC, which showed only moderate discriminative capacity (AUC = 0.68, F1-score = 0.66). Classification accuracies of all three discriminations were comparable across different machine learning methods (gradient boosting machine (GBM), support vector machine (SVM) and avNNet models), indicating that the results we observed were not specific to the methodology chosen (logistic regression with elastic net penalty) (Table S9). The relative importance of the antibody-bound peptides contributing to each of the presented classifications following elastic net regression can be found in Table S10. The antibody epitope repertoire in IBD
RkJQdWJsaXNoZXIy MjY0ODMw