Lung cancer subtyping using cell-free DNA fragmentomes and protein biomarkers

Introduction: Lung cancer is the leading cause of cancer-related mortality worldwide. Accurate histological subtyping to differentiate between lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), and small cell lung cancer (SCLC) is critical for guiding optimal therapeutic strategies. However, up to 20% of patients lack sufficient tissue for conventional histopathological classification. Liquid biopsies using cell-free DNA (cfDNA) fragmentomics offer a promising non-invasive alternative for cancer characterization when tissue is not available. Methods: We examined 761 patients with newly diagnosed, treatment-naive lung cancer of all stages, including lung adenocarcinoma (n=468), squamous cell carcinoma (n=156), small cell carcinoma (n=42), large cell carcinoma (n=15) and other subtypes (n=80) from the prospective Lung Cancer Early Molecular Assessment trial (LEMA, NCT02894853). Low-coverage whole genome sequencing of cfDNA plasma samples was performed to derive genome-wide fragmentation features. Circulating tumor DNA (ctDNA) burden was estimated from fragmentation using the DELFI-TF method. We developed a machine learning classifier trained exclusively on the tissue-based copy number signatures from the Clinical Lung Cancer Genome Project (CLCGP) and applied it to patient cfDNA samples to predict lung cancer subtypes. Results: This tissue-trained subtyping algorithm was evaluated on all available plasma samples, achieving an AUC of 0.99 (95% CI = 0.98-1.00) for distinguishing NSCLC from SCLC and an AUC of 0.91 (95% CI=0.87-0.95) for differentiating LUAD from LUSC. The model correctly classified 88% of SCLC, 80% of LUAD and 87% of LUSC cases where the tumor fraction was ≥0.3% (n=276). Among a subset of 361 NSCLC patients, integration of five blood protein biomarkers resulted in a multimodal model that differentiated LUAD from LUSC across all tumor fractions with high performance (AUC=0.85, 95% CI=0.80-0.90), an improvement over cfDNA (p<0.01; AUC=0.78, 95% CI=0.74-0.82) or protein-only classifiers (p<0.001; AUC=0.70, 95% CI=0.62-0.78). Conclusions: These findings establish cfDNA fragmentation and protein biomarkers as a viable non-invasive approach for lung cancer subtyping when tissue is unavailable, with potential to expedite subtype-specific treatment selection and improve clinical outcomes