Uncovering molecular trajectories in follicular lymphoma using hidden markov models of ctdna dynamics

Abstract

Abstract Circulating tumor DNA (ctDNA) is a valuable prognostic biomarker in follicular lymphoma (FL) frontline treatment, providing a minimally invasive tool to monitor disease. However, its prognostic value in the relapsed or refractory (R/R) setting remains unclear. This study, evaluated ctDNA dynamics across the clinical course of first-line and relapsing FL patients (including CAR-T and bispecific antibodies). To avoid predefined response categories, we applied a longitudinal modeling approach using Hidden Markov Models (HMMs), identifying latent molecular states from serial ctDNA measurements. A total of 391 liquid biopsy samples from 95 FL patients diagnosed between 2018 and 2024 were analyzed. All the patients received treatment (79 first-line immunochemotherapy and 36 treatments in the setting of relapse). MRD was quantified using a personalized NGS DNA assay (Altum-track) with a sensitivity of 10⁻⁴. Evaluation time-points (TPs) were adapted to therapy type: months 4–6 for immunochemotherapy (ICT), and months 6–12 for CAR-T, BsAbs, or Rituximab-Lenalidomide (R²). To evaluate response PET-CT imaging was performed in all cases. Relapsing patients have received a median of 3 lines of therapy (2-6), and 85% of them presented progression of diseases during the first 24 months after therapy (POD24). Median PFS was not reached in first line therapy and was 21 months in the relapsing setting. MRD-positive patients had worse PFS. In the first-line setting, and after a median follow-up of 35 months, 25% of patients relapsed (median 16 months for MRD+ vs NR for MRD-, p<0.001). In the relapse group (n=36), 56% progressed after a median follow-up of 22 months. At evaluation, 19 of 36 were MRD-positive (PFS 7 months vs NR, p<0.001). CtDNA VPP and VPN for PFS were 100% and 90% respectively, compared with 70% and 74% for PET-TC. Additionally, ctDNA and PET/CT combined identified the patients that progressed in less than two years with 88% sensitivity and 100% specificity. HMMs were used to model ctDNA dynamics over time. HMMs were trained using a range of different numbers of hidden states (2-4), with the three-state model demonstrating the best fit (Akaike Information Criterion (AIC)/ Bayesian Information Criterion (BIC)). The Viterbi algorithm assigned each MRD TP to a latent state. As predictive accuracy increased with additional TPs, the analysis focused on patients with ≥3 ctDNA TPs (n = 70). Model validation used case-centric data splitting to prevent information leakage, combined with a 5-fold cross-validation and bootstrap approach with out-of-bag (OOB) testing. HMMs trained on resampled cases predicted states in OOB patients, with clear survival stratification. For clinical interpretation, the final predicted hidden state for each patient (from the full or split-trained model) was used as a grouping factor in a Kaplan-Meier survival analysis. The HMM identified three dynamic states with distinct outcomes: most cases (n = 45) fell into a favorable state (S1), with only two progressions; an intermediate-risk state (S2, n = 10) with median PFS of 42 months (HR vs. S1: 13.5, 95% CI: 2.90–63.9; p < 0.001); And a high-risk state (S3, n = 15) with median PFS of 9 months (HR vs S1: 34.3, 95% CI: 7.7–163.5; p < 0.001). Patients with rising ctDNA levels had worse outcomes, even when MRD remained stable. HMM-inferred trajectories reclassified some cases into higher-risk groups, highlighting the added prognostic value of dynamic trajectory modeling. In summary, dynamic ctDNA modeling with HMM revealed distinct molecular trajectories that predict outcomes beyond static MRD status. This approach provides a robust, data-driven framework tracking disease evolution and refining risk stratification in FL. Cross-validation and bootstrap analyses supported the model's reliability and clinical relevance. Incorporating this method into future trials may enable earlier intervention and more personalized treatment strategies in FL.