Early cancer detection via multi-omics cfDNA fragmentation using early-late fusion neural network with sample-modality evaluation.

Abstract

Cell-free DNA (cfDNA) fragmentation patterns reflect epigenetic modifications and are promising biomarkers for early cancer detection. While integrating diverse fragmentomic signals can improve accuracy, high modality dimensionality, and limited samples challenge effective multimodal fusion. We present Early-Late fusion with Sample-Modality evaluation (ELSM), a two-stage neural network integrating 13 fragmentomic feature spaces with sample-wise modality evaluation to capture complementary signals. Across five datasets of 1994 samples from 10 cancer types, ELSM outperforms unimodal and advanced multimodal models for cancer detection and tissue-of-origin prediction, achieving an AUC of 0.972 for pan-cancer diagnosis and 0.922 in an independent gastric cancer cohort, with a median tissue-of-origin accuracy of 0.683. Analysis of key genomic regions identified by ELSM reveals robust interpretability aligned with known oncogenic pathways. ELSM provides a powerful and interpretable framework for integrative multi-omics analysis with strong potential for clinical translation in early cancer detection.