Fig. 1

Schematic diagram displaying the design of MESA. cfDNA is isolated from blood samples of three cohorts (cohort 3 was split into cohort 3.1 and cohort 3.2 for cross-cohort validation) and then processed to generate targeted EM-seq libraries using three targeted panels. Analysis of the EM-seq data enables the extraction of four modalities: cfDNA methylation (purple), nucleosome occupancy (blue), nucleosome fuzziness (green), and windowed protection score (orange). Then, the feature processing and selection are performed for each modality separately. Firstly, features that contain NA or have low variance are removed. Next, the Boruta algorithm is used for feature ranking, and the top-ranking features (shown in red) are selected for the following analysis. Selected features are used for tenfold inner cross-validation for each modality to get the base predictions. Finally, by stacking and training the base predictions, we get a multimodal machine learning model which outperforms the single-modality models in cancer detection