Treatment-agnostic joint modeling of longitudinal circulating tumor dna predicts survival across first-line regimens in metastatic non-squamous NSCLC

Abstract

Background and Purpose

To develop a treatment-agnostic joint model leveraging longitudinal circulating tumor DNA (ctDNA) dynamics to predict overall survival (OS) across different treatment regimens without incorporating treatment information using data from the IMpower150 trial in treatment-naïve metastatic non-squamous NSCLC patients.

Experimental approach

Patients were randomized 1:1:1 to receive ABCP, ACP, or BCP, with plasma samples collected at baseline and multiple time points to track longitudinal ctDNA changes. The joint model followed a two-step approach by combining two separate submodels: a disease submodel for ctDNA dynamics and a survival submodel for time-to-event analysis. Random effects are derived from the nonlinear mixed effects model for ctDNA dynamics. In the subsequent step, these individual random effects are incorporated as covariates in the survival submodel. A landmark modeling approach was used where ctDNA data from the first 21 weeks posttreatment was used to predicts OS beyond 21 weeks.

Key Results

Among 466 participants, 348 had detectable ctDNA at one or more time points and were stratified into training (n = 181) and test (n = 167) sets. Fourteen ctDNA summary metrics were assessed, with median allele frequency for known/likely mutations emerging as the top-performing metric. Overall, the treatment-agnostic model's predicted survival curves closely matched the observed ones across treatment arms in both training and test sets. Predicted median OS and 2-/3-year OS rates aligned well for ABCP, ACP, and BCP, with discrepancies generally under 20 %. Although the model tended to overpredict OS for ACP—likely due to a small sample size—the final IMpower150 analysis reported median OS values within 10 % of our predictions.

Conclusions & Implications

Early ctDNA monitoring and modeling have the potential to significantly enhance treatment personalization. By enabling the early prediction of patient outcomes across various treatment regimens, this treatment-agnostic model supports more informed clinical decision-making, ultimately improving patient management and outcomes in oncology.