Low-Dose Computed Tomography (LDCT) has been demonstrated to reduce lung cancer mortality by 20% for high-risk current and former smokers. However, 25% of the subjects in the NLST demonstrated abnormalities and a large fraction of those lesions were determined to be false-positives. There is an unmet need to accurately and non-invasively identify early-stage aggressive lung cancers and distinguish lesions that are life threatening from those that are not. Recently cell-free DNA (cfDNA) in human blood has emerged as an ideal source for cancer detection. In this proposal, we develop an integrated system, CancerRadar, consisting of (1) an experimental assay, cfMethyl-Seq, for cost-effective genome-wide methylation profiling of cfDNA, offering >10 fold enrichment over Whole Genome Bisulfite Sequencing (WGBS) in profiling CpG islands; and (2) a computational framework to extract various information from cfMethyl-Seq data, including cfDNA methylation, cfDNA fragment size, copy number variation (CNV), and microbial composition, and perform multi-feature ensemble learning for detecting malignant lung nodule and locating its primary tumor sites. We will validate CancerRadar with several clinical cohorts. Compared to the commonly used small panels focusing on one type of markers, CancerRadar profiles and integrates genome-wide profiles of multiple genetic/epigenetic features, therefore can robustly capture the very small proportion of tumor-derived cfDNA fragments, comprehensively diagnose patients with heterogeneous cancer pathogenesis, and learn and exploit newly significant features as training sample size grow.