Fig. 4

Out-of-distribution detection using calibration methods. A Identifying motifs that are enriched in both pathogenic and self-epitopes. Venn diagram comparing the enriched n-grams (left) or position-specific k-mer motifs from 9aa (middle) or 10aa peptides (right) between pathogenic and self-peptides. B Normalised enrichment ratio of shared n-grams enriched in both pathogenic and self-epitopes. C Clusters of peptides with high sequence similarity in contact positions, demonstrated by pairwise global alignment scores. Network graph illustrating the pairwise global alignment score between 9aa pathogenic peptides. Shown are peptides having ≥ 22 alignment scores with ≥ 3 other peptides. D–G Distribution of different calibration-based metrics on pathogenic (i) and self-(ii) peptides, discriminating correctly vs. incorrectly predicted peptides. The calibration-based metrics include maximum softmax probability (MaxProb) (D), temperature scaling using different levels of temperatures that scale the logit values (E), the maximum on average softmax probability over 10 ensembled models (MaxProb on Avg) (F) and maximum on average softmax probability from Monte Carlo dropout iterations (G). The Monte Carlo models were reiterated 100 times with stochastic dropouts of 0.6. T: temperature. Statistical significance by p-values from Student’s t test. ns: non-significant. H Venn diagram comparing the pathogenic peptide predictions using different calibration metrics. MCDropout: Monte Carlo Dropout. I. ROC curve illustrating the performance of the out-of-distribution (OOD) linear regression classifiers using MCDropout on pathogenic (i) and self-peptides (ii)