Fig. 5

Multimodal epigenetic analysis from MESA improves the performance of cancer detection model. A–C ROC curves showing model performances based on different modalities. Methylation, methylation ratio of all target CpGs; Occupancy, nucleosome occupancy of all TSS and polyadenylation target regions; Fuzziness, nucleosome fuzziness of all TSS and polyadenylation target regions; WPS, average WPS for all target regions; Multimodal, the combination of all four types of features. D–F Heatmaps showing the predicted probabilities of single modality models for each sample. The probability represents the predicted probability of classifying the sample to the cancer group. G–I Heatmaps showing pairwise Spearman correlations of the predicted probability of all samples between different types of features. The Spearman correlation values are labeled on the heatmaps. LOOCV on cohort 1, the leave-one-out cross-validation analysis results from 130 iterations on cohort 1. LOOCV on cohort 2, the leave-one-out cross-validation analysis results from 332 iterations on cohort 2. Cross-cohort validation on cohort 3, results of models trained on cohort 3.1 and validated on cohort 3.2