FragMIL: A Fragment-Token Multiple-Instance Learning Framework for Cancer Detection from Cell-Free DNA Fragmentomics

Background/Objectives: Cell-free DNA (cfDNA) fragmentomics from low-coverage whole-genome sequencing (lcWGS) is a promising approach for non-invasive cancer detection. However, most existing methods summarize millions of cfDNA fragments into handcrafted sample-level features before classification, potentially discarding informative patterns carried by individual molecules. We aimed to develop and benchmark a framework that learns directly from fragment-level signals. Methods: We developed FragMIL, a fragment-token multiple-instance learning framework that represents each cfDNA molecule as a learnable token integrating end-motif, genomic-position, and fragment-length information, and aggregates thousands of fragment tokens into a patient-level cancer score. FragMIL was evaluated on 934 samples from three public lcWGS cohorts (pan-cancer, lung cancer, and hepatocellular carcinoma) using five repeated subject-level splits, a 288-configuration design-space benchmark, and comparison with four established fragmentomic approaches (coverage, DELFI, end motif, and fragment size). Results: FragMIL achieved the highest mean areas under the receiver operating characteristic curve (AUROCs) in every cohort, with 0.963 ± 0.009, 0.802 ± 0.009, and 0.910 ± 0.015 in the DELFI/Cristiano 2019, LUCAS/Mathios 2021, and Jiang 2015 cohorts, respectively. End-motif encoding provided the strongest standalone discriminative signal, whereas the optimal fragment pool was cohort-dependent, with open chromatin region fragments performing best for pan-cancer detection and tumor-enriched fragments for disease-specific cohorts. Performance was retained across cancer stages and cancer types. Conclusions: Preserving molecule-level fragmentomic information improves cancer detection, supporting fragment-token learning as a promising framework for next-generation cfDNA fragmentomics.