Cell-Free DNA Genomic and Fragmentomic Features for Early Outcome Prediction in Diffuse Large B-Cell Lymphoma

Purpose: Diffuse large B-cell lymphoma (DLBCL) patients need an accurate and early risk stratification strategy, as prompt therapy escalation may improve outcomes. Cell-free DNA (cfDNA) is an emerging liquid biopsy biomarker that has demonstrated clinical utility in guiding cancer treatment. Patients and Methods: We evaluated cfDNA genomic and fragmentomic features in 190 patients with large B-cell lymphoma from two multicenter prospective trials using Whole Genome Sequencing (WGS). Patients had a DLBCL or high-grade B-cell lymphoma (HGBL) diagnosis and were treated with R-CHOP or dose-adjusted EPOCH-R (followed by nivolumab consolidaiton for HGBL after achieving remission). Responders were defined as those in radiographic remission at the end-of-treatment (EOT) or those with a negative biopsy following a positive scan. Non-responders were defined as those with EOT PET-CT Deauville 4-5, progression, or death due to lymphoma during treatment. We extracted four cfDNA features: enhanced tumor fraction, proportion of short cfDNA fragments ranging from 20-150bp, the fragment end integrated analysis (FrEIA) score, and the Gini Diversity Index. We trained a Random Forest model with the training cohort (n = 120), performing hyperparameter optimization using 4-fold cross-validation to maximize prediction accuracy. We then evaluated the model’s performance in the validation cohort (n = 61), averaging the ACT scores from the five models for classification predictions. We defined a new metric called the ACT score (Aberrations, Contribution of short fragments, Terminal motif analyses), where a threshold of 0.5 differentiates non-responders (ACT score ≥ 0.5) from responders. We evaluated the prognostic value of the ACT score for progression-free survival (PFS) and overall survival (OS), and in the context of IPI and interim PET-CT. Results: Individual cfDNA features were altered between responders and non-responders after one cycle of treatment, and the ACT score could predict EOT response (AUC 0.70). Patients with a positive ACT score had inferior outcomes compared to ACT score-negative patients [PFS: HR 3.2 (95% CI 1.4-7.2); p < 0.01; OS: HR 4.4 (95% CI 1.7-11.6); p < 0.01]. The 2-year PFS in ACT score-positive and -negative patients was 40% vs. 80%; and OS was 55% vs. 90%, respectively. In the multivariate analysis, the prognostic value of the ACT score was independent of the International Prognostic Index and interim PET-CT. Conclusions: The ACT score, computed from a single plasma sample collected after one cycle of treatment, can predict clinical outcomes. This low-cost and easy-to-interpret tumor-naïve test has the potential to guide treatment in interventional clinical trials and risk-adapted treatment strategies.