Bridging tumor growth dynamics and survival in preclinical oncology drug development – a Tumor Growth Inhibition and Time-to-Event (TGI-TTE) modeling framework Authors: Kumar Kulldeep Niloy, Jamie Horn, Nazmul Hasan Bhuiyan, Suhas S Bhosale, Khaled A Shaaban, Thomas E Prisinzano, Jon S Thorson, Jurgen Rohr, Markos Leggas Published: 2026 ABSTRACT AbstractPurposeTo develop a tumor growth inhibition and time-to-event (TGI-TTE) modeling framework linking drug exposure, tumor dynamics, and survival. Modeling used preclinical efficacy studies with MTMSA-Trp, a preclinical-stage anti-tumor agent for Ewing sarcoma.MethodsTumor volume and survival data from Ewing sarcoma mouse xenografts treated with MTMSA-Trp (0.3–2.85 mg/kg) were analyzed. A Simeoni TGI model was used to estimate individual exponential and linear tumor growth rates. A parametric log-logistic TTE model was used to describe mouse survival across treatment groups by incorporating post hoc TGI metrics and regimen-specific average concentrations derived from the PK model as covariates associated with the model-derived event-time distribution. Models were evaluated using the precision of parameter estimates, goodness-of-fit plots, and bootstraps.ResultsThe TGI model accurately described individual tumor growth dynamics, yielding precise parameter estimates (RSEs < 15%). The final TTE model successfully captured the observed Kaplan-Meier curves across treatment groups (RSEs < 20%). As expected from the endpoint definition, higher regimen-specific average concentration corresponded to longer model-derived survival (coefficient = 0.63), whereas higher exponential and linear tumor growth rates corresponded to shorter model-derived survival (coefficients = − 0.75 and − 0.46, respectively).ConclusionThis framework quantitatively links tumor dynamics, drug exposure, and survival and may support the design, analysis, and simulation-based evaluation of dose regimens in preclinical oncology studies.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s00280-026-04947-0.Keywords: Tumor growth inhibition modeling, Time-to-Event Modeling, Ewing sarcoma, Mithramycin analogues Introduction To characterize the time course of tumor growth and quantitatively evaluate the antitumor efficacy of an anticancer agent in mouse xenograft studies, tumor growth inhibition (TGI) modeling has become a common and valuable approach [1–6]. More broadly, mathematical modeling approaches have been used to integrate tumor, host, and treatment dynamics and to support decision-making in oncology drug development [7, 8]. TGI models describe the tumor growth dynamics of the treated animals by accounting for both natural tumor progression and drug-induced tumor cell kill [9, 10]. However, the potential of an anticancer agent requires extending beyond tumor regression to understand its effect on survival [6, 11, 12]. Although TGI modeling is common in preclinical efficacy assessment, integrated TGI-TTE frameworks remain underused for preclinical xenograft survival data; a notable example by Martin et al. demonstrated their value for optimizing preclinical study designs [13]. To bridge this gap, the present work developed a TGI-TTE modeling framework to quantitatively link drug exposure, tumor growth dynamics, and survival endpoints as a case study in preclinical oncology drug development. The modeling used data from preclinical studies in Ewing sarcoma (ES) models. ES is a pediatric bone and soft tissue cancer characterized by chromosomal translocation leading to the expression of EWS-ETS fusion proteins, which serve as oncogenic transcription factors driving tumor progression [14, 15]. MTMSA-Trp is an EWS-ETS transcription factor inhibitor that is currently under preclinical development for the treatment of ES [16–18]. In ES mouse xenograft studies, this compound demonstrated significant tumor regression and improved survival, as measured by predefined tumor volume [17, 18]. In these xenograft studies, survival was defined as time to reach the prespecified tumor-volume threshold, consistent with standard preclinical efficacy practice [11]. Using tumor volume and survival data from ES mouse xenograft studies treated with different MTMSA-Trp dosing regimens, this study employed a two-stage sequential modeling approach. First, a TGI model was developed to estimate individual TGI metrics. Subsequently, a parametric log-logistic TTE model was developed to test these individual post-hoc TGI metrics as well as the drug exposure metric as covariates associated with the model-derived event time. This framework establishes a quantitative bridge between tumor dynamics and survival, making it more generally applicable to support the design, analysis, and simulation-based exploration of dosing regimens in preclinical oncology drug development. Methods Animal experiments Tumor growth data were collected from in vivo tumor growth inhibition experiments conducted in female athymic nude mice bearing TC-32 tumors [17, 18]. Briefly, female athymic nude mice were implanted on the rear left flank by subcutaneous injection of 0.5 × 106 TC-32 cells in 0.1 mL 50:50 RPMI: Matrigel®. Mice were monitored for tumor growth using slide calipers and, once tumors reached a predefined volume, were randomized to receive either MTMSA-Trp or vehicle (5%:95% HS-15: PBS, pH 8) via bolus tail-vein injection. Randomization was performed in Microsoft Excel by assigning each mouse a random number using the RAND() function and allocating mice to treatment groups accordingly. The initial tumor volumes ranged from 37 to 251 mm3 across three experiments assessing tumor growth inhibition. In experiment 1, mice were dosed at 0.3, 0.6, and 0.9 mg/kg once daily for 5 days (Q1dx5). In experiment 2, mice were dosed at 2.40 mg/kg every 3 days for a total of 6 doses (Q3dx6). In experiment 3, mice were dosed at 2.40 mg/kg for a total of 6 doses (Q3dx6) and at 2.85 mg/kg for a total of 6 doses twice per week (BIWx6). Tumor volumes and animal body weights were measured daily throughout the dosing period, then twice weekly thereafter. Tumor growth data were collected over 100 days across all three experiments. To support estimation of unperturbed tumor growth, vehicle-control data from experiments 1 and 2 were pooled with historical TC-32 control data generated under comparable study conditions. The survival criteria were varied across experiments as follows: tumor dimension > 15 mm for Experiment 1; tumor dimension > 15 mm and tumor volume > 1,500 mm³ for Experiments 2 and 3; and tumor volume > 1,500 mm³ for historical controls. All animal experiments were performed under an institutionally approved protocol (St. Jude Institutional Animal Care and Use Committee protocol 3164). Female athymic nude mice were purchased from Envigo RMS (Athymic Nude-Foxn1nu; Indianapolis, IN) and Charles River (NU(NCr)-Foxn1nu; Wilmington, MA) for pharmacokinetic and efficacy experiments, respectively. The animals were housed in AAALAC-certified, temperature- and humidity-controlled (64 °F to 84 °F and 30% to 70%, respectively) facilities with a 12 h light/dark cycle. Mice were provided with standard commercial diets (rodent 5R53 TestDiet, Richmond, IN) and chlorinated, reverse-osmosis-treated water ad libitum. TGI modeling The TGI modeling dataset comprised 1,377 tumor volume observations from 104 mice. The TGI model was developed in a stepwise manner. First, a natural tumor growth model was identified using data from the control group. Several structural models describing different growth patterns (e.g., exponential, quadratic, Simeoni, and Koch) were evaluated. The best-fitting natural growth model was then extended to include a drug effect term representing MTMSA-Trp–induced tumor cell kill. The drug effect term included concentrations simulated from a previously published one-compartment PK model [19]. Goodness-of-fit plots and predictive performance evaluations for this PK model are provided in the original report. This PK model incorporated an empirical power relationship between dose and clearance to account for dose-dependent disposition. Therefore, the typical population clearance was calculated for each dose level used in the xenograft studies based on this dose-clearance relationship. These dose-specific clearance values were subsequently used to simulate the concentration-time profiles for each regimen. TGI model parameters were estimated by simultaneously fitting tumor volume data from both control and treatment groups. Interindividual variability (IIV) on model parameters was assumed to follow a log-normal distribution. Additive, proportional, and combined residual error models were evaluated to characterize residual unexplained variability. TTE modeling A parametric TTE model was used to describe the distribution of model-derived event times. A sequential modeling approach was employed where individual post-hoc TGI parameters estimated in the first stage were evaluated as covariates in the subsequent TTE model. The TTE dataset included data from experiments 1, 2, and 3 (Table S1). Because these experiments applied different survival criteria, the survival endpoint was standardized to avoid bias from inconsistent endpoint definitions. For uniformity, the survival time for each mouse was derived from its individual tumor-volume profile predicted by the TGI model, and the survival event was defined as the model-predicted time for the tumor volume to reach 1,500 mm³. Mice whose tumors did not reach this threshold were right-censored. In these animals, the tumor was not progressing toward the threshold: the treated animals were complete responders whose tumors had regressed (in some cases leaving a small, non-proliferative residual mass), and the control animal that did not reach the threshold had a tumor that did not grow over the course of the experiment. Several parametric TTE models, including log-logistic, Weibull, Gompertz, and exponential, were fitted to the survival data from all mice. Upon selection of the base TTE model, covariate modeling was conducted to link TGI and PK parameters to the event time. For this, individual TGI parameters were derived from the TGI model, and exposure was summarized as regimen-specific average concentration (). PK model-derived parameters were used to simulate the concentration-time profile of each regimen, and non-compartmental analysis was used to calculate from the simulated concentration-time profiles [20]. Parameter estimation and model evaluation TGI and TTE modeling and parameter estimation were performed in Monolix version 2024R1 (Lixoft). Model selection was based on the Akaike Information Criterion (AIC), and the model with the lowest AIC was selected as the best fit. Model evaluation was based on the precision of parameter estimates, the visual predictive check (VPC) of 1000 simulations, and the goodness-of-fit (GOF) plots. Model stability was assessed by comparing parameter estimates with those from 1000 bootstrap runs. The data were plotted using GraphPad Prism 10.4.0. Results TGI modeling The Simeoni natural tumor growth model yielded the lowest AIC value (Table S2) and was therefore selected as the base model. A linear drug effect was incorporated to describe the drug-induced tumor kill because a nonlinear drug effect led to unstable parameter estimates (Table S3). While a Simeoni transit compartment structure to describe delayed drug-induced cell death was evaluated, this parameterization led to unidentifiability with the current dataset; thus, a direct drug effect without a transit compartment was retained. Given the pooled nature of the dataset, the source experiment was evaluated as a categorical covariate on the tumor growth parameters. Although including experiment effects decreased the − 2 log-likelihood, several experiment-effect and variance parameters were estimated imprecisely, with relative standard errors much greater than 50% in most cases. This indicated that the additional effects were not reliably identifiable (Table S4). Further, re-estimation of the downstream TTE model using individual post hoc parameters from the alternative TGI model produced similar TTE covariate coefficients (Table S8). Therefore, the parsimonious TGI model was retained. The structure of the Simeoni TGI model is shown in Fig. 1 and in Eq. 1. Simeoni TGI model structure where t is time (hours), TV is tumor volume (mm3), is TV at t = 0, and represent the exponential and linear tumor growth rate parameters, is the drug-induced tumor cell kill rate constant, and is the simulated plasma concentration based on the PK model. The was included in the model as a regressor rather than estimated as a population parameter because initial tumor volume measurements were available for each mouse, which precluded the necessity to estimate an additional parameter. The final estimates of the population parameters are summarized in Table 1. All TGI parameters were precisely estimated, as indicated by low relative standard errors (RSE < 15%). Table 1TGI model parametersParameterDefinitionEstimate (%RSE)IIV (%CV)[shrinkage]Bootstrap-derived median (95% CI) Exponential growth rate(1/h)0.0063 (8.56)0.85 (102.82) [6.12%]0.0066(0.0056–0.0076) Linear growth rate(mm3/h)5.93 (11.8)0.73 (83.39) [28.8%]5.88(4.85–8.36) k Tumor kill rate(ml/ng/h)0.000024 (13.4)0.87 (106.68) [30%]0.000024(0.000018–0.000032) a Additive residual error(mm3)11.22-- b Proportional residual error (%)0.22-- TGI model parameters Exponential growth rate (1/h) 0.0066 (0.0056–0.0076) Linear growth rate (mm3/h) 5.88 (4.85–8.36) Tumor kill rate (ml/ng/h) 0.000024 (0.000018–0.000032) Additive residual error (mm3) GOF plots indicated adequate description of the data (Fig. 2). The population predictions showed a wider scatter around the line of unity due to high IIV for the natural growth parameters (Fig. 2a). However, the individual predictions showed strong agreement with the observed tumor volumes (Fig. 2b). Individual weighted residuals (IWRES) were randomly distributed around the line originating at y=0 with no apparent bias over time or individual predicted tumor volumes (Fig. 2c and d). Goodness-of-fit plots of TGI model. Panel a and b show the tumor volume observations versus population and individual predictions, respectively. Panel c and d show individual weighted residuals versus individual predictions and time, respectively Furthermore, the individual model fits demonstrated that the model successfully captured the variability in tumor growth profiles of individual mice across treatment groups (Fig. 3). Overall, the individual post-hoc estimates of TGI parameters were robust and precisely estimated with acceptable shrinkage (Table 1). Individual tumor volume fits using post hoc TGI estimates across treatment groups. Solid lines represent individual fits and symbols represent observed tumor volumes TTE modeling The log-logistic TTE model best described the survival dataset, as it had the lowest AIC value (Table S5) and was therefore selected as the base model. The hazard function of the log-logistic TTE model is shown in Eq. 2. where, h is hazard, t is time (hours), is scale parameter representing median survival time, and s is shape parameter (fixed to 4). As shown in Eq. 3, , and were identified as significant covariates on . Where is the typical population-scale parameter, represents the coefficient for the respective covariate, normalized by their respective median values, and is the random effect representing the residual inter-individual variability in the scale parameter. The final TTE model parameter estimates are summarized in Table 2. All parameters were estimated with high precision, as indicated by low relative standard errors (RSE < 20%). Consistent with the endpoint definition, higher tumor growth rates ( and ) were associated with shorter survival ( = -0.75 and − 0.46, respectively), whereas higher drug exposure () was associated with longer survival ( = 0.63). Table 2TTE model parametersParameterDefinitionEstimate (%RSE)IIV [%CV](%RSE)Bootstrap-derived median(95% CI) Scale703.17(4.72)0.043 [4.26](38.4)724.59(602.79–890.34) Coefficient of -0.75(8.3)--0.75(-1.02 – -0.49) Coefficient of -0.46(17.9)--0.39(-0.71 – -0.04) Coefficient of 0.63(10.4)-0.62(0.37–0.85) s Shape4(fixed)-- TTE model parameters 703.17 (4.72) 0.043 [4.26] (38.4) 724.59 (602.79–890.34) -0.75 (8.3) -0.75 (-1.02 – -0.49) -0.46 (17.9) -0.39 (-0.71 – -0.04) 0.63 (10.4) 0.62 (0.37–0.85) 4 (fixed) VPCs of the base TTE model without covariates captured the overall Kaplan-Meier survival curve; however, they failed to describe the Kaplan-Meier survival curves for individual treatment groups (Fig. 4a). In contrast, the addition of covariates in the scale parameter of the TTE model successfully captured the observed Kaplan-Meier survival curves across all treatment groups (Fig. 4b; Table 2). Moreover, observed survival probabilities were within the 90% prediction intervals derived from the model simulations. VPC plots of TTE model. Panel a and b show dose-stratified VPC plots of base and final TTE models, respectively. Black solid lines represent observed Kaplan-Meier curves. Blue dashed lines and shaded areas represent model predicted median Kaplan-Meier curves and 90% prediction interval Discussion A quantitative understanding of the relationship between tumor growth dynamics and survival is critical for optimizing dosing strategies for anti-cancer agents [21]. Although TGI–TTE modeling frameworks have been extensively applied in clinical settings to describe the relationship between longitudinal tumor size and survival outcomes [22–25], their application to preclinical outcomes remains limited. In this study, using tumor volume and survival data collected from in vivo tumor growth inhibition experiments conducted in TC-32 tumor-bearing female athymic nude mice, we developed a TGI-TTE modeling framework to quantitatively link drug exposure, tumor growth dynamics, and survival endpoints for MTMSA-Trp. To characterize natural tumor growth dynamics, several models were explored, and a Simeoni natural tumor growth model was selected as the base model because it provided the best fit, as indicated by the lowest AIC, consistent with previous reports for a similar tumor type [2]. A linear drug effect was subsequently incorporated to describe the MTMSA-Trp-induced tumor cell kill. Incorporation of a linear drug-effect term into the final TGI model implied that MTMSA-Trp exerts a concentration-dependent tumor kill proportional to its plasma concentration. Non-linear drug effect (i.e., Michaelis-Menten) was also explored. However, it led to unstable parameter estimates and was therefore not retained. The linear drug effect successfully captured variability in tumor growth profiles across treatment groups in individual mice. The individual post-hoc estimates of the TGI parameters were robust and precisely estimated with acceptable shrinkage, suggesting the model was suitable for its intended purpose of estimating individual TGI parameters to support subsequent TTE modeling. Individual event times were obtained from TGI model-predicted tumor trajectories; thus, the TTE analysis should be regarded as a second-stage characterization of a model-derived time-to-threshold endpoint rather than as an independent survival analysis. This distinction has two important implications. First, because the event time was derived directly from each mouse’s predicted tumor growth, we expect tumor-growth parameters (λ ₀ and λ ₁) and the regimen-level exposure metric (Cavg) to be related to the event-time distribution. These relationships result from how the endpoint was defined and should not be interpreted as independent evidence that these variables predict survival. Instead, the TTE model provides a population-level description of the relationship between tumor growth, drug exposure, and the event-time endpoint that can be used to simulate survival under different dosing schedules. Second, and for the same reason, the residual variability captured by the scale- parameter random effect does not represent an additional biological source of between-animal survival variability; it reflects the fact that the TTE model summarizes each tumor-growth profile using a reduced set of covariates (λ₀, λ₁, and regimen-level Cavg) together with a parametric description of the resulting event-time distribution. The primary value of the TTE stage is to provide a compact parametric representation of the event-time distribution for population-level simulation rather than to establish independent survival predictors. This representation enables population-level simulation of survival outcomes, with prediction intervals, under dosing regimens not tested experimentally, and handles right- censoring. Censoring is essential here because animals that did not reach the 1,500 mm³ threshold were not progressing toward it: the treated animals were complete responders whose tumors had regressed, and the control animal that did not reach the threshold had a tumor that did not grow. Because the TGI model does not include a tumor-eradication state, propagating these trajectories beyond the observation period would inevitably predict eventual regrowth, generating progression events that are biologically inconsistent with the observed complete responders. Therefore, right-censoring these animals, as done for complete responders and long-term survivors in clinical time-to-event analysis, is appropriate. A framework that jointly models tumor eradication and regrowth would remove the need for this censoring and represents a useful direction for future work. For the survival analysis, a parametric log-logistic TTE model best described the survival dataset and was selected as the base model. However, the base TTE model failed to capture the distinct Kaplan-Meier curves across treatment groups, as shown in the dose-stratified diagnostic plots. Building on evidence from clinical TGI–TTE analyses, covariate modeling was conducted to test TGI and PK parameters as covariates in the TTE model. In covariate modeling, tumor growth rates and regimen-specific average concentration () were identified as significant covariates on the scale parameter of the TTE model. The addition of covariates significantly improved the TTE model, as evidenced by a large drop in the objective function value (∆-2LL = -87.75). Parameter estimates in the final TTE model were precise, with relative standard errors below 20%. The directions and magnitudes of the covariate effects follow directly from the endpoint definition: faster tumor growth (higher and ) shortens the time to reach the survival endpoint, whereas greater average drug exposure prolongs it. Because was defined at the regimen level, the observed exposure-survival relationship should be interpreted as a between-regimen association rather than an individual-level exposure-response relationship. Model evaluation illustrates the effect of adding covariates on the final TTE model fit. The base model did not reproduce the distinct dose-stratified Kaplan-Meier curves, whereas adding covariates substantially improved agreement across treatment groups. Some discrepancies remained in the smallest groups, particularly the 2.40 mg/kg Q3dx6 (n = 4) and 2.85 mg/kg BIWx6 (n = 7) groups; however, the observed survival probabilities remained within the 90% prediction intervals. Bootstrap resampling also supported the stability of the parameter estimates. In conclusion, this work establishes a TGI–TTE modeling framework that quantitatively links exposure, tumor dynamics, and survival in mouse xenograft studies. This framework complements conventional nonparametric survival analyses by integrating tumor growth dynamics and drug exposure into a parametric survival modeling framework and supports future study design by enabling simulation-based evaluation of dose regimens in preclinical oncology drug development. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements This research included select experiments conducted by the Animal Resources Center at St. Jude Children’s Research Hospital, which is supported by the American Lebanese Syrian Associated Charities. We acknowledge Karen Jackson and Scott Kinison for technical assistance with select experiments. Author contributions K.K.N. and M.L. conceptualized the study. N.H.B., K.A.S., and S.S.B. synthesized and purified MTMSA-Trp. K.K.N. and J.H. performed animal experiments and sample analysis. M.L., T.E.P., J.S.T., and J.R. provided resources and funding for the experiment. K.K.N. performed data analysis and modeling. K.K.N. and M.L. prepared the manuscript. M.L., T.E.P., J.S.T., and J.R. reviewed and edited the manuscript. All authors have read and agreed to the published version of the manuscript. Funding This research was funded by the National Cancer Institute (R01 CA243529), the Kentucky Medical Services Foundation Chair in Pharmacy, the National Institutes of Health (R37 AI052218), the Center of Biomedical Research Excellence (COBRE) for Translational Chemical Biology (CTCB, NIH P20 GM130456), the National Institute of Food and Agriculture (USDA-NIFACBGP, Grant No. 2023-38821-39584), the University of Kentucky College of Pharmacy, the University of Kentucky Markey Cancer Center, the National Center for Advancing Translational Sciences (UL1TR000117 and UL1TR001998), and the American Lebanese Syrian Associated Charities. Data availability Inquiries regarding the datasets should be directed to the corresponding author. Declarations Conflict of interest M.L., J.S.T, K.A.S., and J.R. are inventors on patents and/or patent applications related to mithramycin derivatives with anticancer activity. These intellectual property rights are assigned to the University of Kentucky. The other authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement Inquiries regarding the datasets should be directed to the corresponding author.