Liquid biopsy offers a minimally invasive approach for early cancer detection, enabling early treatment. A liquid biopsy detects cancer-related biomarkers in the blood, including circulating cell-free DNA (cfDNA), which contains tumor-derived fragments. Conventional cfDNA analysis faces challenges due to the low fraction of tumor-derived DNA, the high heterogeneity of cfDNA data, and limited clinical trust stemming from its limited interpretability. This project proposes an approach that uses large language models (LLMs) to analyze cfDNA sequences and associated metadata to accurately identify tumor-derived samples and classify tumor types. By treating cfDNA fragments as text, LLMs can learn complex patterns in the underlying data to differentiate between tumor-derived and normal cfDNA.