Assil MERLAUD 1,2,3, Marion KERIOUI 1,4, Alyse LIN 5, Phyllis CHAN 5, Benjamin WU 5, Jin Y. JIN 5, Pascal CHANU 2,6, Julie BERTRAND 1
1 Université Paris Cité, IAME, Inserm, F-75018 (Paris, France), 2 Institut Roche (Boulogne-Billancourt, France), 3 Clinical Pharmacology, Genentech-Roche (Paris, France), 4 MRC Biostatistics Unit, University of Cambridge (Cambridge, UK), 5 Clinical Pharmacology, Genentech Inc. South San Francisco (South San Francisco, USA), 6 Clinical Pharmacology, Genentech-Roche (Lyon, France )
Objectives:
Overall survival (OS) is the main endpoint in oncology randomized clinical trials (RCTs), yet its late observation limits the assessment of treatment benefit. Longitudinal tumor size measurements provide earlier indicators of disease progression; however, these conventional RECIST¹-based summaries aggregate lesions at the patient level, masking inter-lesion and inter-organ heterogeneity. Increasing evidence indicates that anatomical location influences tumor dynamics, treatment response, and prognosis²–⁴. Yet systematic quantification of organ-specific tumor dynamics and their prognostic value for survival across cancer types and treatment settings remains limited⁵.
Multilevel joint models (mJM) of tumor dynamics and survival offer a rigorous framework to characterize these associations⁶. Here, we extend this approach using a two-stage Bayesian meta-analytic framework to synthesize evidence across RCTs⁷.
We aimed to: (i) identify phase II/III randomized atezolizumab (ATZ)-containing trials in advanced carcinomas; (ii) characterize organ-specific lesion dynamics; (iii) quantify their time-varying association with OS; and (iv) evaluate consistency and heterogeneity of these effects across cancer types and treatment.
Methods:
A systematic review (PROSPERO CRD420251032620) was conducted in Medline and ClinicalTrials.gov to identify phase II/III RCTs evaluating ATZ (alone or in combination) in stage III/IV carcinomas with OS as primary and secondary endpoint. Transitivity and cross-trial comparability were evaluated using study and baseline patient level characteristics.
In the first stage, each arm of RCT (experimental and control) was analyzed independently using the same Bayesian mJM linking lesion-level tumor dynamics to OS. Lesion size dynamics were described using a Stein model⁸, capturing exponential shrinkage and regrowth, within a nonlinear mixed-effects framework accounting for inter-patient and inter-lesion variability, and organ-specific fixed effects (categorized as lung, lymph nodes, liver, other). OS was modeled using a Weibull baseline hazard incorporating time-varying current lesion size. Both an agnostic association structure and an organ-specific association structure were evaluated. Parameters were estimated in Stan using Hamiltonian Monte Carlo algorithm with weakly informative priors⁹. Model adequacy and predictive performance were assessed using WAIC and posterior predictive checks⁶.
In the second stage, arm-level tumor dynamics-survival association parameter estimates (with their posterior standard deviation) were synthesized using meta-analysis at the study level. A univariate random-effects model was applied to the organ-agnostic association parameter estimate, while multivariate random-effects models were fitted to organ-specific tumor dynamics-survival association parameter estimates. Potential effect modifiers (arm, treatment line, regimen, cancer type) were explored using a stepwise covariate modeling approach.
Results:
Of 23 eligible RCTs, lesion-level data were retrieved for 22 trials encompassing 9 primary solid cancer types. In total, 12,291 patients contributed 28,832 target lesions and 153,583 longitudinal measurements.
Organ-level dynamic patterns were consistent across cancer types. Liver lesions exhibited slower shrinkage and faster post-nadir regrowth than lung lesions, reflecting less favorable dynamics. Differences across cancer types were also evident: small cell lung cancer showed the most treatment-sensitive yet most aggressive tumor profile, whereas renal cell carcinoma displayed the least sensitive and least aggressive dynamics. Although tumor shrinkage was greater in control and chemotherapy-containing regimens than with immunotherapy monotherapy, post-nadir regrowth rates were lower in immunotherapy arms (monotherapy or combination), consistent with a delayed but sustained treatment effect.
Across studies, increases in lesion size were positively associated with higher instantaneous risk of death. In the organ-agnostic meta-analysis, a 10-mm increase in tumor size corresponded to a pooled 2.61% [1.75%; 3.52%] (95% Credibility Interval) increase in hazard, with between-study heterogeneity (CV = 75.26% [47.38%; 128.81%]). Organ-specific meta-analysis reduced unexplained heterogeneity and revealed clear site-dependent prognostic effects in addition to arm effects. In the immunotherapy arms, liver lesions showed the strongest prognostic impact (8.62% [5.97%; 11.45%] hazard increase per 10-mm growth), whereas lung lesions exhibited the smallest impact (2.95% [2.15%; 3.85%]). In the control arms, corresponding hazard increases were 6.79% [4.12%; 9.60%] and 1.80% [1.03%; 2.70%], respectively. Lung and liver effects were comparatively stable across trials under immunotherapy (CV = 50.91% [26.43%; 89.10%] and 70.78% [45.36%; 119.66%]), whereas greater heterogeneity was observed in the control arms, reaching 143.24% [71.06%; 288.10%] for “other” organs.
Conclusions:
The consistent and stable prognostic impact of organ-specific lesion progression across cancers supports the use of organ-level dynamic metrics as mechanistically informed biomarkers. This framework provides a robust statistical foundation for cancer type-agnostic development strategies and cross-indication evidence synthesis. A future one-stage hierarchical extension should help predict the survival of future ongoing studies¹⁰.
References:
References
1. Eisenhauer, E. A. et al. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). Eur. J. Cancer Oxf. Engl. 1990 45, 228–247 (2009).
2. Kerioui, M. et al. Assessing the impact of organ-specific lesion dynamics on survival in patients with recurrent urothelial carcinoma treated with atezolizumab or chemotherapy. ESMO Open 7, 100346 (2022).
3. Zhou, J. et al. Mapping lesion-specific response and progression dynamics and inter-organ variability in metastatic colorectal cancer. Nat. Commun. 14, 417 (2023).
4. Chen, C. et al. Organ-specific tumor dynamics predict survival of patients with metastatic colorectal cancer. Eur. J. Cancer Oxf. Engl. 1990 207, 114147 (2024).
5. Ian, D., Jiawei, Z., Quefeng, L. & Yanguang, C. A Pan-Cancer Analysis of Lesion-Level Treatment Response to Extend the ‘Seed and Soil’ Paradigm. MedRxiv Prepr. Serv. Health Sci. 2025.09.16.25335893 (2025) doi:10.1101/2025.09.16.25335893.
6. Kerioui, M. et al. Nonlinear Multilevel Joint Model for Individual Lesion Kinetics and Survival to Characterize Intra-Individual Heterogeneity in Patients with Advanced Cancer. Biometrics 79, 3752–3763 (2023).
7. Sudell, M., Tudur Smith, C., Gueyffier, F. & Kolamunnage-Dona, R. Investigation of 2-stage meta-analysis methods for joint longitudinal and time-to-event data through simulation and real data application. Stat. Med. 37, 1227–1244 (2018).
8. Stein, W. D. et al. Tumor growth rates derived from data for patients in a clinical trial correlate strongly with patient survival: a novel strategy for evaluation of clinical trial data. The Oncologist 13, 1046–1054 (2008).
9. Carpenter, B. et al. Stan: A Probabilistic Programming Language. J. Stat. Softw. 76, 1 (2017).
10. Sudell, M., Kolamunnage-Dona, R., Gueyffier, F. & Tudur Smith, C. Investigation of one-stage meta-analysis methods for joint longitudinal and time-to-event data through simulation and real data application. Stat. Med. 38, 247–268 (2019).
Reference: PAGE 34 (2026) Abstr 12002 [www.page-meeting.org/?abstract=12002]
Poster: Oral: Drug/Disease Modelling - Oncology