Agentic Design of Population-Aware cfDNA Cancer Liquid Biopsy Panels Without Cancer-Positive Samples

Targeted cell-free DNA (cfDNA) cancer assays normally select genomic regions by comparing cancer-positive and control samples. Designing a panel for a specific intended-use population therefore requires cancer-positive samples from that population, which encourages panels to be reused across populations with different cfDNA backgrounds. We present a population-aware agent that selects methylation targets for a specified intended-use population without cancer-positive specimens during panel design. The agent generates literature-based mechanistic claims, explores them locally using a reference atlas of sorted cell types, converts them into testable methylation hypotheses, and validates those hypotheses in cancer-free cfDNA from the target population. We used the agent to design a panel for an intended-use population with chronic liver disease and compared the frozen panel with a commercial pan-cancer panel, a published gastrointestinal cancer panel, two single-source baselines, and three footprint-matched random panels. All panels were evaluated with the same model architecture and training. Using in-silico mixtures of tumor and adjacent-normal tissue methylomes in cancer-free plasma, we evaluated performance across tumor fractions. The agent panel had the highest mean standardized partial AUC at specificities>90\% across detectable tumor fractions. At the highest tumor fraction, partial AUC was 0.805, compared with 0.729 for the best external panel. These results show that literature-based claims, explored separately and validated on intended-population background data, can jointly improve target-region selection without any cancer-positive samples.