Dynamic Noninvasive Genomic Monitoring for Outcome Prediction in Diffuse Large B-Cell Lymphoma

Background: The prognosis for diffuse large B-cell lymphoma (DLBCL) patients who fail initial therapy remains poor. Current prognostic methods to identify patients destined for failure employ baseline molecular profiles or imaging data at fixed milestones, thus sub-optimally capturing functional response dynamics. Noninvasive detection of tumor-specific DNA sequences in the plasma, or circulating tumor DNA (ctDNA), provides a window of opportunity to observe these changes early during therapy. We sought to relate early ctDNA kinetics during therapy to tumor volume, therapeutic responses, and ultimate clinical outcomes.

Methods: Using CAPP-Seq, a next-generation sequencing platform for detection of ctDNA (Newman Nature Medicine 2014), we prospectively profiled patients with DLBCL receiving combination immunochemotherapy at Stanford University. Tumor samples were used to define tumor specific somatic alterations, which were then monitored in plasma. We examined two methods of assessing ctDNA change over time: a simple heuristic model (assessing the change in ctDNA concentration from cycle 1 to cycle 2), and a biologically based mathematical model of ctDNA dynamics to predict tumor volume and patient outcomes.

Results: We sequenced tumor and plasma samples (n=135) from ten patients receiving Rituximab-containing regimens. Plasma samples were collected prior to, during, and immediately after chemotherapy, with a median of 7 samples per patient during the first therapy cycle. Across patients, ctDNA concentrations varied over a 6-log range (Figure 1). The change in ctDNA concentration between cycle 1 and cycle 2 generally tracked with FDG PET/CT response - patients achieving a PR or CR had an average decrease of 2.9±0.8 logs in ctDNA concentration, compared to an increase of 0.3±0.8 logs for those with SD or PD (p<0.001). However, this metric failed to capture some patients who ultimately relapsed after radiographic remission. We therefore developed a multi-compartmental ordinary-differential equation (ODE) model of tumor dynamics capturing tumor volume, ctDNA, and the effect of chemotherapy. We performed nonlinear regression to fit data to this model using serial ctDNA measurements from individual patients, thereby creating continuous, patient-specific models of both ctDNA and tumor volume (Figure 1a-b). This mathematical model significantly fit ctDNA measurements and predicted tumor volumes across patients and samples (Figure 1c). Using ctDNA measurements from the first 2 cycles of therapy, this model accurately predicted clinical outcomes for all ten patients, including relapse after radiographic remission. An additional cohort of patients will be presented at this meeting.

Conclusions: Given its high specificity and large dynamic range, ctDNA provides an opportunity to monitor the dynamics of therapeutic response in patients with DLBCL. Methods capturing these dynamics correlate with radiographic response. Given the complexity of tumor dynamics, heuristic models of ctDNA may less faithfully capture ultimate clinical outcomes. Personalized mathematical models of ctDNA can potentially reflect tumor dynamics and predict clinical outcomes for individual patients.

Figure 1.

View largeDownload slide

Personalized tumor modeling from ctDNA tumor dynamics. a) An example of a model of ctDNA fit to observed data for a single patient (DLBCL010). b) The corresponding tumor volume prediction over time for patient DLBCL010. c) Summary of the mathematical model across ten patients, demonstrating the fit between measured data and the model.

Figure 1.

View largeDownload slide

Personalized tumor modeling from ctDNA tumor dynamics. a) An example of a model of ctDNA fit to observed data for a single patient (DLBCL010). b) The corresponding tumor volume prediction over time for patient DLBCL010. c) Summary of the mathematical model across ten patients, demonstrating the fit between measured data and the model.

Close modal