Semi-mechanistic modeling of tumor size and overall survival including machine learning-derived tumor heterogeneity
Diego Vera-Yunca(1, 2), Zinnia P. Parra-Guillen(1, 2), Pascal Girard(3), Iñaki F. Trocóniz(1,2), Nadia Terranova(3)
(1) Pharmacometrics & Systems Pharmacology, Department of Pharmaceutical Technology and Chemistry, School of Pharmacy and Nutrition, University of Navarra, Pamplona, Spain. (2) IdiSNA, Navarra Institute for Health Research, Pamplona, Spain. (3) Merck Institute for Pharmacometrics, Merck Serono S.A., Switzerland, an affiliate of Merck KGaA, Darmstadt, Germany.
Objectives: Total tumor size (TS), expressed as the sum of longest diameters (SLD) of the individual tumor lesions (iTLs), is often used in oncology drug development and treatment individualization as prognostic factor of overall survival (OS) 1. The use of SLD is not exempted from limitations as it represents a naïve metric compared to the iTLs dynamics whose granularity could provide insights into different drug disposition/response across a variety of locations due to tumor heterogeneity. In this work we have adopted an intermediate approach coupling a machine learning (ML)-based analysis of iTLs dynamics with a population semi-mechanistic modeling of SLD and parametric time to event (TTE) analyses, aiming to assess the impact of tumor heterogeneity on tumor growth inhibition (TGI) and OS2,3. As a demonstration of this methodology, data from patients with metastatic colorectal cancer (mCRC) receiving either FOLFIRI or FOLFOX alone or in combination with cetuximab were included in the analysis.
Methods: Data from four clinical studies (1781 patients, 6369 iTLs) were analyzed using the ClassIfication Clustering of Individual Lesions (CICIL) methodology4 to find the degree of iTLs heterogeneity within a tissue (intra-class) or across different tissues (inter-class). iTL dynamics were assessed across different patient groups depending on their treatment or genotype based on their cross-correlation coefficients (CCs)5. After assessing iTLs dynamics, the SLD was computed in order to build a TGI model using NONMEM 7.36. Drug concentrations over time were unavailable, and tumor exposure was inferred through a kinetic-pharmacodynamic (K-PD) model7. The probability of patient drop-out (DO) was taken into account by including it as a parametric time-to-event (TTE) model. OS was also modeled in the same way. The effects of covariates measured during the first 8 weeks on model parameters were explored both on the TGI and OS models. In addition to the standard covariates typically used during oncology modeling, new covariates such as median CC, KRAS mutation status [wild-type (wt) or mutated] and primary tumor lesion location (either right-sided or left-sided CRC) were included in the analysis.
Results: Tumor heterogeneity within the same tissue (intra-class analysis) was found to be lower (71-88% of CCs with similar dynamics) than across different tissues (inter-class analysis, 65%) for those patients receiving cetuximab. This heterogeneity was further increased in patients presenting KRAS mutations (52%) when compared to KRAS wt patients (72%). Tumor heterogeneity across tissues was expressed by computing the median of the CCs for a given patient. The TGI model included both chemotherapy and cetuximab effects in an additive form Resistance to drug treatment developed quicker for chemotherapy (t1/2 = 8.12 weeks) than for cetuximab (t1/2 = 11.87 weeks). Model parameters were precisely estimated [relative standard error (RSE) < 30%]. The DO was characterized by a log-logistic model. The most influencing covariates in the TGI model were (i) the KRAS wt status increasing 3.5 times the cetuximab-induced tumor shrinkage parameter and (ii) the location of the primary tumor lesion on the cetuximab resistance, being 4.7 times higher for those having right-sided CRC. Finally, OS was best described using a Weibull model. The observed appearance of a new lesion during the first 8 weeks was the most significant predictor of OS (HR = 4.14, 95%CI: 3.07-5.57), followed by a large predicted TS at baseline and the median CC (HR = 1.36, 95%CI: 1.07-1.72), our marker of tumor heterogeneity derived from the CICIL methodology.
Conclusions: This work represents the first time that mCRC patients receiving FOLFIRI/FOLFOX plus a targeted therapy are characterized with a population pharmacokinetic/pharmacodynamic (PKPD) modeling of TGI. First, an assessment of iTL dynamics within and across tissues in TS data from a large population of mCRC patients was performed2. This approach allowed us to derive a simple metric that summarizes the heterogeneity of individual lesions to be further integrated into semi-mechanistic TGI and OS models3. Other novel covariates, not previously identified for CRC tumor size models to our knowledge, are KRAS mutation status and primary tumor lesion location. Next steps will involve the development of a tissue-agnostic model in order to continue studying the impact of tumor heterogeneity on patient outcomes.
References:
[1] Ribba B, Holford N, Magni P, et al. A Review of Mixed-Effects Models of Tumor Growth and Effects of Anticancer Drug Treatment Used in Population Analysis. CPT Pharmacometrics Syst Pharmacol. 2014;3(5):113. doi:10.1038/psp.2014.12
[2] Vera-Yunca D, Girard P, Parra-Guillen ZP, Munafo A, Trocóniz IF, Terranova N. Machine Learning Analysis of Individual Tumor Lesions in Four Metastatic Colorectal Cancer Clinical Studies: Linking Tumor Heterogeneity to Overall Survival. AAPS J. 2020;22(3):58. doi:10.1208/s12248-020-0434-7
[3] Vera‐Yunca D, Parra‐Guillen ZP, Girard P, Trocóniz IF, Terranova N. Relevance of primary lesion location, tumour heterogeneity and genetic mutation demonstrated through tumour growth inhibition and overall survival modelling in metastatic colorectal cancer. Br J Clin Pharmacol. June 2021. doi:10.1111/bcp.14937
[4] Terranova N, Girard P, Ioannou K, Klinkhardt U, Munafo A. Assessing Similarity Among Individual Tumor Size Lesion Dynamics: The CICIL Methodology. CPT Pharmacometrics Syst Pharmacol. 2018;7(4):228-236. doi:10.1002/psp4.12284
[5] Wei WWS. Time Series Analysis - Univariate and Multivariate Methods. Addison-Wesley; 1989.
[6] Beal S, Sheiner LB, Boeckmann A, Bauer RJ. NONMEM User’s Guides. Ellicott City, MD, USA Icon Dev Solut. 2009.
[7] Jacqmin P, Snoeck E, Van Schaick EA, et al. Modelling response time profiles in the absence of drug concentrations: Definition and performance evaluation of the K-PD model. J Pharmacokinet Pharmacodyn. 2007;34(1):57-85. doi:10.1007/s10928-006-9035-z