Brain metastases (BM) are the most common and among the most deadly brain tumors. Currently, there are no reliable predictors of BM development from primary cancer, limiting early intervention. Here, we obtained 364 tumor/plasma samples from a large cohort of lung adenocarcinoma (LUAD) and BM patients (N=309), the most common BM source. LUAD DNA methylation signatures were evaluated to build and validate an accurate model predicting BM development from LUAD, which was integrated with clinical factors to provide comprehensive patient-specific BM risk probabilities in a nomogram. Additionally, immune and cell interaction gene sets were differentially methylated at promoters in BM versus paired primary LUAD and had aligning dysregulation in the proteome. Immune cells were differentially abundant in BM versus LUAD. Finally, liquid biomarkers identified from methylated cell-free DNA sequenced in plasma were used to generate and validate accurate classifiers for early BM detection. Overall, LUAD methylomes can be leveraged to predict and non-invasively identify BM with high precision, moving towards improved patient outcomes with personalized treatment.