ONCO Subgroup Contributed Talks

Daniel Glazar

Moffitt Cancer Center & Research Institute

"A simulation-based sample size analysis of a joint model of longitudinal and survival data for patients with glioma"

Introduction Patients with recurrent high-grade glioma (rHGG) have poor prognosis with median progression-free survival (PFS) <6 months, and median overall survival <12 months [1]. The Response Assessment in NeuroOncology (RANO) defines radiographic progression as 25% increase in the sum of products of longest diameters of individual lesions (SPD) delineated from MRIs relative to minimum observed SPD [2]. However, there is a wide heterogeneity in response to treatment in these patients with some experiencing disease progression within weeks and others surviving for years. Predicting which patients will progress early or late on different therapeutic regimens may aid the clinician in deciding which regimen with which to treat the patient. Thus, there is an urgent clinical need for a predictive biomarker for patient-specific PFS. To predict PFS and OS, baseline disease characteristics, such molecular markers have been investigated [3]. However, repeated measures of such biomarkers are infeasible due to the invasive procedures involved. To remedy this limitation, predictive mathematical models of patient-specific tumor dynamics in glioma based on non-invasive imaging data have been developed [4–8]. However, most of these do not evaluate PFS by RANO, and those that do only predict PFS as a binary outcome over discrete time horizons [6]. Considering a joint model of longitudinal tumor volume and PFS from a Bayesian perspective has the advantage of predicting PFS as a continuous outcome over continuous time horizons for individual patients while taking into account population-level response dynamics [9–11]. However, before applying such a model to clinical data, we note that due to low incidence and accrual rates of early phase clinical trials for patients with rHGG [12], there is a clinical need to determine the minimum sample size necessary to predict patientspecific TTP using longitudinal tumor volumes. As such, we perform a simulation-based sample size analysis of a joint model of longitudinal tumor volume and PFS on an in silico clinical trial for patients with rHGG. Objectives 1. To develop a joint model of longitudinal tumor volume and PFS for patients with rHGG. This joint model will use tumor volume dynamics as an accurate and precise predictive biomarker of PFS. 2. To perform a sample size analysis on an in silico clinical for patients with rHGG to determine the minimum number of patients needed to make accurate and precise individual Bayesian dynamic predictions of PFS based on longitudinal tumor volumes in order to inform clinical trial design. Methods 1. We developed a joint model of longitudinal tumor volume and TTP for patients with rHGG. Tumor volumes were modeled using a tumor growth inhibition (TGI) model with mixed effects. In it, tumors grow exponentially to replicate the fact that patients inevitably progress before any inflection point in their tumor growth. Tumor response rate declines exponentially due to the inevitable development of therapeutic resistance. TTP was then modeled discretely and defined as the time when tumor volume reached 40% above from nadir, extrapolating from the bi-dimensional 25% threshold in RANO [2]. Due to technical limitations, the hazard function was approximated using a scaled skew normal distribution curve. Population parameter estimates of the developed joint model were estimated using quantile information reported in the literature [13] by maximizing the likelihood of joint order statistics [14]. 2. An in silico clinical trial was conducted to study the effects of sample size on the predictive performance of the developed joint model to dynamically predict patient-specific PFS. We selected sample sizes of 40, 60, 80,..., 200 as most clinically feasible due to the low incidence of rHGG. In silico training and test sets were generated by sampling from model parameter distributions and simulating longitudinal tumor volume and TTP every 6 weeks from treatment initiation with baseline 2 weeks prior to treatment initiation to conform with rHGG clinical trial protocols. Population parameters were estimated using the stochastic approximation of expectation-maximization (SAEM) algorithm [15] and then taken to parameterize a prior distribution to dynamically predict patient-specific TTP for the test patients across landmark times and time horizons [9–11]. Predictive performance was then evaluated using time-dependent Brier score (BS) and area under the receiver operating characteristic curve (AUC) [9,16]. Simulations were performed using the Monolix suite software [17]. Results 1. In estimating population parameters, simulated median PFS was between 24 and 30 weeks, which agrees with clinical observations [6]. Also, the majority of simulated patients had between 3 and 7 observations, also in agreement with clinical observations [6]. 2. For the largest sample size considered (N=200), we evaluated the developed model’s predictive performance for set time horizons of 6, 12, 18, and 24 weeks across all landmark times. The median AUC ranged from 0.55 to 0.63 with median BS ranging from 0.19 to 0.25. We then evaluated the developed model’s performance to predict progression around the median PFS (between 24 and 30 weeks) across landmark times 0, 6, 12, and 18 weeks after treatment initiation. The median AUC ranged from 0.50 to 0.65 with median BS ranging from 0.21 to 0.25. 3. Across all sample sizes tested, there was statistically significant albeit small correlation between sample size and AUC (Pearson’s r=0.20, p<1e-15) across all landmark times and time horizons. However, no statistic significance was reached for the correlation between sample and BS (Pearson’s r=-0.0017, p=0.95). When controlling for landmark time and time horizon, the median Pearson’s correlation between sample size and either AUC or BS were r=0.28, 0.07, respectively. Conclusions We developed a joint model of longitudinal tumor volume and PFS for patients with rHGG and parameterized according to quantile information reported in the literature. The model was able to capture the dynamics and survival profiles of the patient population. The predictive performance of the model was robust across the sample sizes tested. However, the overall predictive performance of the model was only marginally better than chance as measured by AUC and BS. In future studies, we will explore a larger range of sample sizes to investigate how many patients are necessary to meet various performance benchmarks as measured by AUC or BS. To improve model predictive performance, we will include covariates, such as sex, molecular markers, and different treatment arms, which are known to affect rHGG volume dynamics and survival endpoints. We will also consider competing risks in the form of a smooth, continuous hazard function as many patients with rHGG experience progression due to clinical deterioration or new lesion even while exhibiting radiographic response. References: [1] Birzu C, French P, Caccese M, et al. Recurrent Glioblastoma: From Molecular Landscape to New Treatment Perspectives. Cancers. 2021; 13(1):47. https://doi.org/10.3390/cancers13010047. [2] Wen PY, Chang SM, Van den Bent MJ, et al. Response Assessment in Neuro-Oncology Clinical Trials. J Clin Oncol. 2017; 35(21):2439-2449. https://doi:10.1200/JCO.2017.72.7511. [3] Phillips HS, Kharbanda S, Chen R, et al. Molecular subclasses of high-grade glioma predict prognosis, delineate a pattern of disease progression, and resemble stages in neurogenesis. Cancer Cell. 2006; 9(3):157-173. https://doi.org/10.1016/j.ccr.2006.02.019. [4] Swanson KR, Bridge C, Murray JD, et al. Virtual and real brain tumors: using mathematical modeling to quantify glioma growth and invasion. Journal of the Neurological Sciences. 2003; 216(1):1-10. https://doi.org/10.1016/j.jns.2003.06.001. [5] Br¨uningk SC, Peacock J, Whelan CJ, et al. Intermittent radiotherapy as alternative treatment for recurrent high grade glioma: a modeling study based on longitudinal tumor measurements. Sci Rep. 2021; 11:20219. https://doi.org/10.1038/s41598-021-99507-2. [6] Glazar DJ, Grass GD, Arrington JA, et al. Tumor Volume Dynamics as an Early Biomarker for Patient-Specific Evolution of Resistance and Progression in Recurrent High-Grade Glioma. J Clin Med. 2020; 9(7):2019. https://doi: 10.3390/jcm9072019. [7] Hormuth DA., Al Feghali KA, Elliott AM. et al. Image-based personalization of computational models for predicting response of high-grade glioma to chemoradiation. Sci Rep 2021; 11:8520. https://doi.org/10.1038/s41598-021-87887-4. [8] Dean JA, Tanguturi SK, Cagney D, et al. Phase I study of a novel glioblastoma radiation therapy schedule exploiting cell-state plasticity. Neuro-Oncology. 2022. https://doi.org/10.1093/neuonc/noac253. [9] Rizopoulos D. Dynamic predictions and prospective accuracy in joint models for longitudinal and time-to-event data. Biometrics. 2011; 67(3):819–29. [10] Mbogning C, Bleakley K, Lavielle M. Joint modelling of longitudinal and repeated time-to-event data using nonlinear mixed-effects models and the stochastic approximation expectation maximization algorithm. J Stat Comput Simul. 2015; 85(8):1512–28. [11] Desm´ee S, Mentr´e F, Veyrat-Follet C, et al. Nonlinear joint models for individual dynamic prediction of risk of death using Hamiltonian Monte Carlo: application to metastatic prostate cancer. BMC Med Res Methodol. 2017;17:105. https://doi.org/10.1186/s12874-017-0382-9. [12] Central Brain Tumour Registry of the United States. CBTRUS Statistical Report: Primary brain and central nervous system tumours diagnosed in the United States in 2004–2006. 2010. [13] Stensjøen AL, Solheim O, Kvistad KA, et al. Growth dynamics of untreated glioblastomas in vivo. Neuro-Oncology. 2015; 17(10):1402–11. https://doi.org/10.1093/neuonc/nov029. [14] Reiss RD. Approximate Distributions of Order Statistics With Applications to Nonparametric Statistics. Springer New York. 2012. [15] Delyon B, Lavielle M, Moulines E. Convergence of a stochastic approximation version of the EM algorithm. Ann Stat. 1999; 27:94–128. [16] Schoop R, Graf E, Schumacher M. Quantifying the predictive performance of prognostic models for censored survival data with time-dependent covariates. Biometrics. 2008; 64(2):603–10. [17] Monolix Suite 2023R1, Lixoft SAS, a Simulations Plus company.
Additional authors: Solmaz Sahebjam, Neuro-Oncology, National Institutes of Health; Michael Yu, Radiology, Moffitt Cancer Center & Research Institute; Dung-Tsa Chen, Biostatistics, Moffitt Cancer Center & Research Institute; Heiko Enderling, Integrated Mathematical Oncology, Moffitt Cancer Center & Research Institute

---

Guillermo Lorenzo

University of Pavia, Italy

"Personalized MRI-informed predictions of prostate cancer growth during active surveillance"

Active surveillance (AS) an established clinical management option for low to intermediate-risk prostate cancer (PCa), which represents almost 70% of newly-diagnosed cases. During AS, patients have their tumor monitored via multiparametric magnetic resonance imaging (mpMRI), serum prostate-specific antigen (PSA), and biopsies. If any of these data reveal tumor progression towards an increased clinical risk, the patient is prescribed a curative treatment. However, clinical decision-making in AS is usually guided by observational and population-based protocols that do not account for the unique, heterogenous nature of each patient’s tumor. This limitation complicates the personalization of monitoring plans and the early detection of tumor progression, which constitute two critical, unresolved problems in AS. To address these issues, we propose to forecast PCa growth using personalized simulations of an mpMRI-informed mechanistic model solved over the 3D anatomy of the patient's prostate. We describe PCa growth via the dynamics of tumor cell density with a diffusion operator, representing tumor cell mobility, and a logistic reaction term, which accounts for tumor cell net proliferation. Model calibration and validation rely on assessing the mismatch between model predictions of the tumor cell density map with respect to corresponding mpMRI-based estimates. Here, we present a preliminary study of our predictive technology in a small cohort of newly-diagnosed PCa patients (n=11) who enrolled in AS and had three mpMRI scans over a period of 2.6 to 5.6 years. Our results show a median concordance correlation coefficient (CCC) and Dice score (DSC) of 0.59 and 0.79, respectively, for the spatial fit of tumor cell density during model calibration using two mpMRI datasets. Then, model validation at the date of a third mpMRI scan resulted in median CCC and DSC of 0.58 and 0.76, respectively. Additionally, the global CCCs for tumor volume and total tumor cell count were 0.87 and 0.95 during model calibration and of 0.95 and 0.88 at forecasting horizon, respectively. Thus, while further improvement and testing in larger cohorts are required, we believe that our results are promising for the potential use of our methods to personalize AS protocols for PCa and predict tumor progression.
Additional authors: Chengyue Wu, Oden Institute for Computational Engineering and Sciences, The University of Texas at Austin, USA; Joshua P. Yung, Department of Imaging Physics, MD Anderson Cancer Center, USA; John F. Ward, Department of Urology, MD Anderson Cancer Center, USA; Hector Gomez, School of Mechanical Engineering, Purdue University, USA; Alessandro Reali, Department of Civil Engineering and Architecture, University of Pavia, Italy; Thomas E. Yankeelov, Oden Institute for Computational Engineering and Sciences, The University of Texas at Austin, USA; Aradhana M. Venkatesan, Department of Abdominal Imaging, MD Anderson Cancer Center, USA; Thomas J. R. Hughes, Oden Institute for Computational Engineering and Sciences, The University of Texas at Austin, USA

---

Jason T. George

Texas A&M University

"Physical Modeling of the T Cell-Peptide Interaction: Toward Predictive T Cell Immunotherapy"

Despite substantial research activity, reliable and computationally efficient prediction of therapeutically relevant T cell receptor (TCR)-antigen pairs remains elusive owing to limited availability of training data and the formidable dimensionality of TCR and antigen sequence space. Successful prediction, if possible, would open new fields of research, from a systematic understanding of the adaptive immune response to viral adaptation to the rapid and optimal identification of tumor antigen-specific T cells. This talk will detail our recent probabilistic modeling efforts that characterize a post-selection T cell repertoire’s ability to identify foreign antigenic signatures with a high degree of specificity. This modeling framework is then applied to construct a data-driven inferential statistical model trained on primary TCR and peptide primary sequences along with crystal structures of known strong binding TCR-peptide pairs. When restricted to a common major histocompatibility complex allele variant, we demonstrate that this approach successfully identifies therapeutically relevant TCR-peptide pairs with a high degree of sensitivity and specificity. Lastly, the trained model is applied to TCRs derived from the peripheral blood of AML patients with the ultimate goal of rapid in silico prediction of tumor antigen-specific T cells.
Additional authors: Zahra Ghoreishi, Department of Biomedical Engineering, Texas A&M University ; Pujan Shrestha, Department of Biomedical Engineering, Texas A&M University ; Priya J. Koppikar, Department of Hematopoietic Biology and Malignancy, The University of Texas MD Anderson Cancer Center ; Noah Tubo, Department of Hematopoietic Biology and Malignancy, The University of Texas MD Anderson Cancer Center ; Jeffrey J. Molldrem, Department of Hematopoietic Biology and Malignancy, The University of Texas MD Anderson Cancer Center.

---

Jeffrey West

Moffitt Cancer Center

"Markov models predict minimal residual disease in Adult B-Lymphoblastic Leukemia"

Cancer stem cells (CSCs) are hypothesized to promote tumor progression through innate chemoresistance and self-renewal. While ostensible CSCs were first identified via CD34+/CD38- immunophenotyping in acute myeloid leukemia, the temporal variation of CD34 and CD38 expression in B-lymphoblastic leukemia (B-ALL) complicates the search for CSCs in this setting. We present a Markovian mathematical model which combines the concept of CSCs with pattern of B-ALL subpopulation at diagnosis to demonstrate dynamic phenotypic alteration and predict minimal residual disease (MRD). Qualitative flow cytometry performed on diagnostic bone marrow was used to determine the proportions of CD34+/CD38+, CD34+/CD38-, CD34-/CD38+, and CD34-/CD38- cells in 44 patients with B-ALL. An iterative numerical search procedure was used to derive patient-specific Markov matrices, describing the stochastic cell state transitions. Then, these patients were divided into MRD positive (n=9) and MRD negative (n=35) cohorts to compare transition matrix features. Among all patients with adequate (> 3 years) follow-up, all MRD positive patients experienced relapse within 3 years, whereas only 31.3% of MRD negative patients (11/35) experienced relapsed B-ALL. A higher transition probability to CD34+CD38- from CD34+CD38+ was associated with positive MRD . In contrast, higher transition probabilities toward a CD34-CD38+ phenotype strongly favor negative MRD status, irrespective of the starting phenotype of other three patterns (p=0.00286, 0.00112, 0.000494, 0.000915, respectively). Combining these parameters into a simple predictive model achieves a sensitivity of 44.4% and specificity of 97.1% for MRD in this setting. Thus, Markov modeling proves useful in assessment of cell state dynamics in patients with B-ALL, especially to predict MRD and a higher relapse rate of disease.
Additional authors: Curtis Gravenmier MD, Ling Zhang MD, Lynn Moscinski MD (Department of Hematopathology and Laboratory Medicine, Moffitt Cancer Center)

---