Richard Franzese 1, Joost de Jongh 2, Sebastiaan Goulooze 2, Teun Post 2, Murad Melhem 3, Herbert Struemper 4
1 GSK (Upper Providence, USA), 2 LAP&P Consultants (Leiden, The Netherlands), 3 GSK (Waltham, USA), 4 GSK (Durham, USA)
Objectives:
Tumor size-overall survival (TS-OS) modeling is an established approach for predicting OS in oncology clinical trials and thus supporting oncology development decisions [1,2,3]. More recently, circulating tumor DNA (ctDNA) has been identified as a prognostic biomarker of efficacy and OS; ctDNA metrics, derived from longitudinal data, have been proposed as potential early markers of radiographic response [4], to risk stratify patients [5], and to predict OS [5]. Joint models of ctDNA, TS, and OS have been proposed [6,7]. Here, we evaluated ways in which a non-small cell lung cancer (NSCLC) TS-OS framework may be enhanced by incorporating ctDNA data.
Methods:
This analysis covered first-line NSCLC participants from two Phase 2 randomized controlled trials Study NCT04581824 (dostarlimab + chemotherapy or pembrolizumab + chemotherapy) and Study NCT05565378 (dostarlimab monotherapy or dostarlimab + belrestotug). Longitudinal ctDNA levels were quantified using mean variant allele frequency and TS was quantified as the sum of longest diameters for target lesions. Observations of ctDNA that were included in analysis were planned at Screening and Week 13 for NCT04581824 and at Screening and Weeks 4, 7, and 13 for NCT05565378.
The ctDNA and TS dynamics were characterized via separate Stein models [6,8] with correlated random effects, and an accelerated failure time log-normal survival model was used to describe OS duration. The impact of baseline covariates on OS was identified previously [1] and included as parameters with fixed values. Simultaneous joint estimation of ctDNA, TS, and OS model parameters was performed in NONMEM 7.5. Various combinations of TS and ctDNA Stein model parameters entering as OS link functions were evaluated. Treatment differences in OS were solely driven by differences in any treatment-specific TS and ctDNA Stein model parameters. Changes in the objective function value (OFV), goodness of fit plots, and visual predictive checks were used to evaluate models.
Results:
The ctDNA-TS-OS dataset included 581 patients, with a total of 3072 TS observations and 1080 ctDNA observations.
Using the joint ctDNA-TS-OS model, in a univariate testing of link functions for OS (relative to an OS base model with no ctDNA- or TS-OS link functions) tumor size growth rate (kgrow_TS) was the most influential OS predictor (change in objective function value [∆OFV] = -76.8). Other significant OS predictors during univariate screening were baseline ctDNA (ctDNAb; ∆OFV = -16.1), ctDNA growth rate (kgrow_ctDNA; ∆OFV = -22.6) and baseline tumor size (TSb; ∆OFV = -39.0). The ctDNA clearance rate (kclear_ctDNA) was not significant during univariate testing of OS link functions.
During stepwise model development of the link functions, two additional link functions (on top of kgrow_TS) were identified: ctDNAb (∆OFV = -24.2) and TSb (∆OFV = -6.59). The ctDNA parameters describing post-baseline ctDNA dynamics (kgrow_ctDNA and kclear_ctDNA) were not statistically significant predictors of OS on top of these three link functions. Additionally, all models with only ctDNA-based link functions were statistically inferior to a model with only kgrow_TS as an OS link function. The inclusion of ctDNAb in the model impacted the individual estimates for kgrow_TS, particularly in patients with early OS events (resulting in limited TS information) as high ctDNAb may provide an alternative predictor for the early OS events. A sensitivity analysis showed that the model parameter ctDNAb could be replaced by the observed baseline ctDNA without worsening the statistical fit of the data.
Conclusions:
We jointly characterize ctDNA, TS, and OS data over time, across 4 different treatments for NSCLC. Baseline ctDNA was prognostic of OS on top of TS-based predictors. The TS dynamics (kgrow_TS) were the strongest predictor of OS. It is possible that the lack of significance of ctDNA dynamic parameters to predict OS on top of TS dynamics in the dataset may be caused by the number and timing of ctDNA observations. Our results imply the potential value of using baseline ctDNA as a covariate in TS-OS modeling, but do not show a clear benefit of modeling the dynamics of post-baseline ctDNA for predicting OS when combined with longitudinal modeling of TS dynamics. Future work, using ctDNA data that is more richly sampled and that also includes later time points, is needed to evaluate both the full predictive potential of ctDNA data in combination with TS data and how best to combine these complementary data sources.
References:
[1] Struemper et al. CPT Pharmacometrics Syst Pharmacol (2025) 14(6):1006–1017
[2] Bruno et al. Clin Cancer Res (2020) 26(8):1787–1795
[3] Chan et al. CPT Pharmacometrics Syst Pharmacol (2021) 10(10):1171–1182
[4] Bartolomucci et al. npj Precis Onc (2025) 9(84)
[5] Assaf et al. Nat Med (2023) 16;29(4):859-868
[6] Ribba et al. Front Pharmacol (2023) 13:1058220
[7] Netterberg et al. Clin Cancer Res (2020) 26(18):4892–4900
[8] Stein et al. Clin Cancer Res (2011) 17(4):907–917
Reference: PAGE 34 (2026) Abstr 11910 [www.page-meeting.org/?abstract=11910]
Poster: Drug/Disease Modelling - Oncology