Background: Cell-free DNA (cfDNA) methylation, fragmentation patterns, chromosome instability, and chromatin accessiblity have been previously shown to be valid plasma biomarkers for non-invasive cancer detection. However, conventional whole-genome bisulfite sequencing (WGBS) is unable to simultaneously profile all these biomarkers due to bisulfite-induced DNA damages. Here we developed a machine learning approach to comprehensively integrate multiple types of cancer genomic markers from enzyme-conversion-based low-pass whole-methylome sequencing (WMS) of plasma cfDNA to non-invasively detect colorectal adenocarcinoma. Methods: Plasma cfDNA sampels from 215 colorectal adenocarcinoma patients and 568 healthy individuals were collected and were split into the discovery and independent testing cohort. The discovery cohort includes 150 cancer patients and 398 healthy individuals and the independent testing cohort includes 65 cancer patients and 170 healthy individuals. Whole methylome sequencing (WMS) libraries were generated from enzymatically converted cfDNA and were subsequently paired-end sequenced at ̃2X coverage. The genome-wide methylation density, fragmentation fingerprints, chromosome instability, and chromatin accessibility were extracted from the WMS data and individually modelled via machine learning methods such as SVM, LR, GBDT, random forest. The final predictive model is an ensemble model integrating all uni-modal models. All models were trained and fitted on the discovery cohort. Results: Data of different modalities provide complementary information in separating the cancer patients from the healthy individuals. Unsupervised clustering of the individuals showed clear separation between cancer patients and healthy individuals. The final predictive model achieved AUC =0.982 in the discovery cohort and AUC =0.9664 in the independent testing cohort. Under a specificity of 94.71% (CI: 90% - 98.82%), sensitivity was 78.46% (CI: 60% - 92.31%) in the independent testing cohort. Separating the cancer patients into different stages, we found that the detection power is usuaul lower for early-stage cancer patients. Conclusions: These results demonstrate the first proof of principle on the feasibility of integrating multiple genomic cancer markers to non-invasively detect colorectal adenocarcinoma from WMS plasma cell-free DNA. A large prospective cohort study is planned to further validate its clinical performance.