Toward generalizable prediction of cancer signal using a cell-free DNA language model.

Abstract

Cell-free DNA can be used for early cancer detection, minimal residual disease monitoring, and post-treatment risk stratification. However, current assays are often designed for a single purpose and rely on deep or broad sequencing panels that capture only a small fraction of tumor-derived signals, limiting transferability, increasing cost, and reducing scalability. Fragmentia-AI is an artificial intelligence language model that learns fragment-level sequence patterns in tumor-derived cell-free DNA. Instead of focusing on mutations, it uses the structure of cell-free DNA to detect cancer signals in a partially panel-agnostic manner from ultra-low sequencing input, approximately 0.1%-1% of conventional depth. The model performs well across cancer types and clinical settings, including monitoring after surgery or immunotherapy, and in samples with low variant allele frequencies or no detected mutations. Fragment-level analyses identify shorter fragments and tumor-derived sequence patterns across panels of different sizes and ultra-low-pass whole-genome sequencing in multiple cohorts.