Mechanistic data-informed multiscale quantitative systems pharmacology modeling framework enables the clinical translation and efficacy assessment of CAR-T therapy in solid tumors

Abstract

Background

Chimeric antigen receptor (CAR)-T cell therapy represents an innovative and potentially revolutionary modality in cancer treatment. Despite their great success in treating blood cancers, CAR-T therapies exhibit significantly lower effectiveness in treating solid tumors. Moreover, the preclinical-to-clinical translation of CAR-T therapies targeting solid tumors is still a challenging task because of their unique “live cell” nature and the substantial variability in patients’ pathophysiology.

Methods

We have developed a multiscale quantitative systems pharmacology (QSP) model to facilitate the clinical translation of CAR-T therapies in solid tumors. Our mechanistic modeling framework integrates the essential biological features that impact CAR-T cell fate and antitumor cytotoxicity, from cell-level CAR-antigen interaction and activation, to in vivo CAR-T biodistribution, proliferation and phenotype transition, and finally to clinical-level patient tumor heterogeneity and response variability. This modeling framework has been calibrated and validated by multimodal experimental data including published preclinical and clinical data of various CAR-T products and original preclinical data of a novel claudin18.2-targeted CAR-T product LB1908.

Results

We demonstrated the general utility of this framework in facilitating clinical translation and characterizing the paired cellular kinetics-cytotoxicity response of different antigen-targeting solid tumor CAR-T cell therapies. As an example, we generated model-based virtual patients and prospectively simulated the response to claudin18.2-targeted CAR-T therapies under different dosing strategies, including step-fractionated dosing and convenient flat dose-based regimens, to inform future clinical trial implementation.

Conclusions

Our translational QSP platform offers an innovative pathway to integrate multiscale knowledge and inform clinical decision-making of novel solid tumor-targeting CAR-T therapies.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Preclinical-to-clinical translation of chimeric antigen receptor (CAR)-T therapies in solid tumor treatment is a nascent field with significant challenges due to the “live cell” nature of CAR-T cells and the substantial variability within patients with solid tumor. New systems-level methodologies are needed to help researchers better integrate CAR-T pharmacology and multiscale experimental data to derive the optimal CAR-T clinical treatment strategies in solid tumor settings.

WHAT THIS STUDY ADDS

• We have here developed a mechanism-based multiscale quantitative systems pharmacology model for CAR-T therapies in solid tumors to advance the state-of-the-art paradigm of CAR-T research and development. We also presented original data of a novel claudin18.2-targeted CAR-T product LB1908, and we demonstrated the translational utility of our QSP framework in terms of understanding multiscale CAR-T mechanism of action, deconvoluting complex CAR-T cellular kinetics-response relationships, projecting clinical antitumor efficacy, and optimizing CAR-T dosing regimens.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The developed QSP model framework provides a novel quantitative approach to bridge multiscale knowledge and support clinical decision-making as well as dosing design of emerging solid tumor-targeting CAR-T therapies.

Background

Chimeric antigen receptor (CAR)-T cell therapy is an innovative and potentially groundbreaking approach for cancer treatment that has demonstrated substantial clinical efficacy in various hematologic malignancies, with seven products approved thus far by the US Food and Drug Administration. In addition to its therapeutic success in treating hematologic cancers, extensive efforts have been made to explore its anticancer potential in solid tumors. As of March 2022, there are 229 clinical trials testing CAR-T therapies in solid tumors (and the number is still increasing),¹ and commonly investigated molecular targets include claudin18.2 (CLDN18.2),² human epidermal growth factor receptor 2 (HER2),³ glypican-3 (GPC3),^{4 5} mesothelin (MSLN),⁶ epidermal growth factor receptor (EGFR)7,9 and so on. Among these, CLDN18.2 is recognized as a promising therapeutic target for gastrointestinal cancers because of its high protein expression in gastric tissues and its high epitope visibility during malignant transformation; thus, CAR-T therapies targeting CLDN18.2 have also been actively tested in clinical trials.10,12 However, the clinical efficacy of CAR-T therapies in solid tumors is generally much lower than that in blood cancers; in addition, preclinical-to-clinical translation of CAR-T therapies is still a nascent field with significant challenges due to the unique “live cell” nature of CAR-T cells and the substantial pathophysiological variability of patients with solid tumor.¹³ Therefore, new methodologies are needed to help researchers integrate CAR-T pharmacology and multimodal experimental data in order to derive the optimal CAR-T clinical treatment strategies for patients with solid tumor.

However, it is usually not feasible to apply classical pharmacological concepts such as pharmacokinetics (PK) or corresponding experimental approaches to assess the complex in vivo fate and efficacy of CAR-T therapies. A key challenge is that CAR-T cells typically display intricate patterns of cellular kinetics in the body due to their unique multiphasic distribution and dynamic life cycle, which can involve initial fast diffusion, short-term proliferation and expansion, contraction, and enduring persistence.¹⁴ Additionally, clinical CAR-T products usually contain a number of different CAR+ T cell subsets (eg, CD4+ and CD8+ T cells), and all of these T cells can have complex interactions with tumor cells, stromal cells in the tumor microenvironment, and themselves. Thus, multiscale physiological factors can influence the way in which CAR-T cells behave in the body. Certain product parameters, such as CAR affinity and CAR density, can impact CAR-T potency, as they control CAR-mediated recognition of tumor cells.¹⁵ Tumor types and patient tumor burden have also been reported to influence CAR-T in vivo durability and treatment efficacy in patients.^{16 17} Furthermore, features related to CAR-T manufacturing and clinical implementation, such as the ratio of different CAR+ T cell subsets in the CAR-T product, the clinical dose, and the choice of lymphodepletion regimen, can significantly affect the final clinical performance of CAR-T therapies.^{18 19} In solid tumors, particularly, one additional factor that significantly contributes to the limited effectiveness of CAR-T cells is the insufficient presence of CAR-T cells in the tumor tissue.^{20 21} Moreover, various obstacles in vivo can impede the penetration of CAR-T cells into solid tumors, such as physical barriers formed by fibrosis and tumor-associated fibroblasts, aberrant blood vessels within the tumor, and undesired cytokines/chemokines secreted by tumor tissue.^{22 23} Research has shown that the solid tumor microenvironment can also perform immunosuppressive functions and induce CAR-T inactivation through various routes, including regulatory immune cells (eg, myeloid-derived suppressor cells, tumor-associated macrophages, regulatory T cells), hypoxia, the upregulation of immune checkpoints, and the secretion of immunosuppressive mediators.2024,26 In addition, unlike blood cancers where B/T cell-specific antigens are usually uniformly expressed in high amounts, tumor-associated antigen targets in solid tumors can exhibit significant heterogeneity in terms of intratumoral distribution and expression intensity, and immune evasion resulting from loss of antigen targets in solid tumors can also result in low CAR-T treatment efficacy and resistance.²⁷

Given the multitude of factors that potentially contribute to differential CAR-T cellular dynamics and tumoricidal potency in solid tumor settings, mechanistic quantitative systems pharmacology (QSP) modeling that integrates CAR-T-related pathophysiological and pharmacological mechanisms together with multiscale experimental data has therefore emerged as a promising research framework to advance our understanding of the multiphasic behavior of CAR-T cells and facilitate their clinical translation. Tsai et al recently proposed a mechanism-based framework to characterize the organ-level biodistribution and tumor killing kinetics of CAR-T in preclinical animals, in addition to their series of QSP modeling work of anti-BCMA (B cell maturation antigen) and anti-CD19 CAR-T therapies in blood cancers.28,30 The above work demonstrated the unique benefits of using computational models to identify optimal CAR-T delivery routes in preclinical mouse tumors; however, currently, there is still no comprehensive mechanistic QSP model of CAR-T treatment in solid tumor settings that can integrate all relevant biological perspectives, such as cell-level CAR-T-tumor interactions, in vivo CAR-T cell biodistribution/proliferation/phenotype transition, clinical-level tumor heterogeneity and patient response simulation, as well as calibration/validation with multiscale experimental data (in vitro, in vivo, and clinical) and application generality with respect to different tumor-associated antigens. Thus, here we developed a first-of-its-kind mechanism-based multiscale QSP model for CAR-T therapies in solid tumors that incorporates all the aforementioned translational features to advance the state-of-the-art paradigm of CAR-T research and clinical development. Using the tumor target CLDN18.2 as an example, the model was quantitatively calibrated and validated against multimodal in vitro cell culture and in vivo animal datasets for an innovative CLDN18.2-targeted CAR-T cell therapy (LB1908). The translational utility and mechanistic adaptability of our QSP model framework were further demonstrated using multiscale CAR-T data (in vitro, in vivo, clinical) for another four solid tumor targets: HER2, EGFR, GPC3, and MSLN. By employing sensitivity analyses and kernel density estimation, we assessed the primary physiological characteristics that may influence the antitumor response in patients. Then, to assess the translational value and potential clinical dosing design of LB1908 in patients with CLDN18.2-positive advanced gastric cancer, we generated model-based virtual patients and simulated the clinical-level response of different CAR-T dosing strategies. Through integrated in silico analyses, we demonstrated that individual patients can exhibit highly different clinical responses to increasing CAR-T doses, and that flat dose-based regimens as well as step-fractionated dosing were also proposed and prospectively evaluated. Our mechanistic QSP model-based translational platform can serve as a new route to efficiently integrate multiscale knowledge and quantitatively guide the clinical decision-making of novel CAR-T therapies in solid tumor treatment.

Materials and methods

Cell lines

Human gastric cancer cell lines KATOIII.luc, NUGC4.luc and SNU16.luc were developed by lentiviral transduction in house. Human pancreatic cancer cell line PANC-1-luc was developed by lentiviral transduction in house, and different amounts of linearized CLDN18.2 messenger RNA (mRNA) were delivered to CLDN18.2-negative PANC-1-luc cells by electroporation.

Quantitative evaluation of human CLDN18.2 expression on target cells by flow cytometry

CLDN18.2 protein expression level was evaluated by flow cytometry on KATOIII.luc, NUGC4.luc, SNU16.luc and PANC-1-luc cells. Cells were harvested and re-suspended with 100 µL Dulbecco's Phosphate-Buffered Saline (DPBS) containing human IgG1 Fc protein isotype (AcroBiosystems), and incubated for 1 hour at 4°C. After centrifugation at 200×g for 5 min, the cell pellets were re-suspended with 100 µL DPBS containing PE anti-human IgG Fc secondary antibody (BioLegend) and 0.5 µL LIVE/DEAD Fixable Violet Dead Cell Stain (Invitrogen) and incubated for 30 min at 4°C. After this staining reaction, the cells were washed with 200 µL DPBS and CLDN18.2 expression was detected by flow cytometry (ACEA Biosciences).

In vitro cytotoxicity assay

CLDN18.2-targeted CAR-T cells were prepared from two healthy donors. In vitro cytotoxicity of CAR-T cells was evaluated for KATOIII.luc, NUGC4.luc, SNU16.luc and PANC-1-luc cells. For KATOIII.luc, NUGC4.luc and SNU16.luc cells, the target cells were harvested and co-cultured with effector cells (CAR-T cells and untransduced T cells as control) at the effector:tumor (E:T) ratio of 8:1, 4:1 and 2:1, respectively. For PANC-1-luc cells, the target cells were harvested and co-cultured with effector cells (CAR-T cells and untransduced T cells as control) at the E:T ratio of 8:1. After 20 hours incubation in co-culture, ONE-Glo Luciferase Assay System solutions (Promega) was added to each assay well, mixed gently and incubated for 5 min at room temperature in the dark. The assay plates were then read in a microplate reader (Perkin Elmer Envision) to calculate in vitro cytotoxicity readouts.

In vitro cytokine release assay

To detect the release of interferon (IFN)-γ and tumor necrosis factor (TNF)-α, 100 µL supernatant from the co-culture system was collected. Detection of human IFN-γ and TNF-α in the co-culture supernatants was performed using HTRF kit (Cisbio) following kit protocols. The assay plates were then read in a microplate reader (Perkin Elmer Envision 2105).

Measurement of plasma expansion and antitumor effect of CLDN18.2-targeted CAR-T cells in NCG xenograft mice

NCG mice were injected with 3.0×10⁶ human gastric cancer cells (NUGC4) in the right flank under aseptic conditions. 90 animals were randomly divided into nine groups (10 animals per group): vehicle, UnT (untransduced T cells) from Donor 1, UnT from Donor 2, CLDN18.2-targeted CAR-T from Donor 1 low dose (0.3×10⁶ CAR-T cells/animal), medium dose (1.0×10⁶ CAR-T cells/animal) and high dose (3.0×10⁶ CAR-T cells/animal), and CLDN18.2-targeted CAR-T from Donor 2 low dose (0.3×10⁶ CAR-T cells/animal), medium dose (1.0×10⁶ CAR-T cells/animal) and high dose (3.0×10⁶ CAR-T cells/animal). All mice were dosed with UnT, CAR-T or Hanks' Balanced Salt Solution (HBSS) once by intravenous infusion. The first day of dosing was indicated as D1. Tumor volumes were measured two times before grouping and twice a week after dosing on D4, D7, D11, D14, D18, D21, D25 and D28, respectively. The tumor volume was calculated using the formula: V=(1/2)×major axis×minor axis².

For quantification of CAR-T expansion, 100 µL of peripheral blood per mouse was collected in anticoagulant (EDTA) tubes from all study animals on days 1, 7, 14, 21 and 28. The genomic DNA was extracted using the QIAamp DNA Mini Kit, and then analyzed by Droplet Digital PCR to derive quantitative readouts for plasma CAR-T cell expansion and concentration.

Multiscale QSP model formulation and virtual patient simulations

The multiscale QSP model was constructed based on ordinary differential equations using mass-action laws and Hill-type functions. The model integrated experimental data from multiple scales as described above (in vitro assays, preclinical animal experiments, clinical CAR-T expansion and efficacy measurements). In the minimal physiologically-based pharmacokinetic (mPBPK) module of our QSP model, the amount of CAR-T that can directly interact with tumor is quantitatively characterized by the CAR-T concentration in the extravascular component (representing tissue) of tumor compartment, and we assumed that when the tumor mass volume is less than or equal to 0.01 mL (as per Response Evaluation Criteria in Solid Tumors (RECIST) V.1.1 criteria, mimicking an almost complete response state),³¹ the tumor will cease to grow and the tumor antigens can no longer stimulate CAR-T cell proliferation and cytotoxicity. In the model, the tumor compartment volume in individual patients (which is independent of the actual dynamic tumor mass volume, see online supplemental table S1 for detailed definitions) which represents the volume of space that was initially taken up by the tumor within the gastrointestinal tissues was determined by the initial tumor burden and remained unchanged regardless of CAR-T killing. In addition, the “other” compartment in the mPBPK module was essentially the agglomeration of redundant organs such as the kidney, spleen, and liver (with their Q, L and V rates quantitatively considered and recalculated for the “other” compartment) when we performed model reduction of the original full PBPK.

The calibration of the in vitro pharmacodynamic (PD) module in the case of CLDN18.2-targeted CAR-T relied primarily on experimental data obtained from SNU16.luc, KATOIII.luc, and NUGC4.luc cell lines. The CLDN18.2 antigen density on SNU16.luc cells was first calculated to be 100,000 per cell from quantitative literature sources.³² Since we have quantitatively measured the mean fluorescence intensity (MFI) of CLDN18.2 in SNU16.luc, KATOIII.luc, and NUGC4.luc cells, we proportionally calculated the amount of CLDN18.2 on the other two cell lines using SNU-16.luc as a reference (antigen density for KATOIII.luc and NUGC4.luc was calculated to be 538,200 and 1,255,100, respectively). To demonstrate the accuracy and feasibility of this method, we tested our calculations using the experimental data from the PANC-1 cell line, which was transfected with different amounts (0.000128, 0.00064, 0.0032, 0.016, 0.08, 0.4, 2, 10 µg) of CLDN18.2 mRNA. The resulting MFIs were again measured for each transfection scenario, and the corresponding CLDN18.2 antigen densities were calculated using the same reference method and MFI ratios. By correspondingly plugging in the calculated CLDN18.2 densities, the in vitro module can still accurately reproduce the cytotoxicity of CLDN18.2-targeted CAR-T cells when co-cultured together with transfected PANC-1 cells (with different RNA amounts), which validates our proposed method. The initial conditions and parameter values associated with the several cancer cell lines used in our model formulation, together with their corresponding descriptions, are also summarized in onlinesupplemental tables S1S4.

We used experimental datasets obtained from tumor-bearing mice for in vivo model calibration. Our QSP model accurately captured all the variations in tumor growth dynamics and CAR-T cell kinetics (manufactured from two different donors) in mice inoculated with NUGC4-luc cancer cells, by adjusting only the parameters related to tumor proliferation (batch-specific) and antigen-induced CAR-T cell proliferation and killing (donor-specific). We also used data from all individual mice to establish the parameter ranges for our combined model at the preclinical in vivo scenario, especially during generation of virtual mice and simulation of preclinical dose-response relationships. When translating the model from mouse to human (single patient), we first adjusted the parameters relating to PBPK considering the physiological differences between mouse and human (blood and organ volumes, organ volumetric flow rates, lymph flow rates were reset to human values),^{29 33 34} while maintaining the overall model structure and reactions. Certain physiological parameters were also adjusted for human simulations during calibration against clinical data (considering potentially slower tumor growth and more drastic CAR-T proliferation in patient vs preclinical settings),^{28 29 35} including tumor growth and killing rates (decreased in number), tumor-mediated inhibition of CAR-T proliferation (increased in number), antigen-mediated CAR-T proliferation (increased in number), and memory CAR-T cell proliferation (increased in number). Other major parameters such as CAR density, tumor antigen density, CAR binding rates and transitional activation/deactivation rates were unchanged. Further, at the clinical population level, based on information derived during in vitro and preclinical calibration, we specifically varied a small set of parameters representative of different patient characteristics to create virtual patients to be treated by CLDN18.2-targeted CAR-T therapy. For patients with CLDN18.2-positive gastric cancer as an example, we varied 15 physiological parameters and resampled them within predetermined parameter ranges to first formulate a large population of plausible patients (n=1,000) for further analyses. Individual patient response was calculated based on RECIST V.1.1 criteria every 45 days (to resemble the clinical follow-up protocol).

All model reactions were implemented in MATLAB Simbiology Toolbox (MathWorks, Natick, Massachusetts, USA), and model simulations were performed using the ode15s solver. For parameter estimation and optimization, the function pattern search in global optimization toolbox was used. A summary of model species, parameters, reactions, and their descriptions was provided in onlinesupplemental tables S1S4.

Global sensitivity analyses

We used the partial rank correlation coefficient algorithm to assess global parameter sensitivities of our model.³⁶ The algorithm setup included 5,000 simulation iterations for each run with parameter values ranging from half to two times of their baseline, and the major output of interest in all sensitivity analyses is the tumor mass volume change on day 29 (relative to day 0) after CAR-T administration. Sensitivity analyses encompassed cell line-level scenarios (using only the in vitro PD module), preclinical animal-level scenarios (using the integrative QSP model for mouse), and clinical-level scenarios (for patients with CLDN18.2-positive cancer). Parameters suggested to have strong and statistically significant correlations with the model output in each condition were selected, verified and presented.

Random forest analyses and kernel density estimation of model parameters

To identify the most influential parameters on population-level treatment response, we used the TreeBagger function in MATLAB to construct a random forest model. For model optimization, we adjusted the hyperparameters of the random forest model, including the number of trees, maximum splits, and minimum leaf node size, to enhance the model predictive performance. By evaluating the out-of-bag classification error rate under different parameter settings, we established an efficient random forest model that quantitatively ranked important parameters in terms of their influence on population-level clinical outcomes and selected a smaller set for further kernel density analyses. For kernel density estimation of how different parameter values contribute to patient-level treatment response, we used the ksdensity function in MATLAB to visualize the influence of heterogeneity in the generated 15,000 different parameter sets on individual therapeutic response in terms of the four response types (progressive disease [PD], stable disease [SD], partial response [PR] and complete response [CR]).

Results

Overview of QSP model formulation and experimental data integration

The QSP model we developed comprises three submodules (figure 1). Module A includes an mPBPK model (mechanistically reduced from a whole PBPK model) to depict the in vivo biodistribution of CAR-T cells in the blood, lung, gastrointestinal system, tumor, and lymph nodes. Each organ compartment is separated into an extravascular/tissue component (green) and a vascular component (red). On intravenous administration, CAR-T cells first enter the lungs from the blood and then move to vascular components in different organs. The majority of CAR-T cells will return to the blood while a portion can also migrate from the vascular components into the tissue in different organs, and these CAR-T cells eventually return to the blood through lymphatic flow, as assumed by prior studies.^{29 30} Module B is an in vitro PD model that describes the physical interaction between tumor cells and CAR-T cells. CAR-T cells with a defined CAR density and affinity can bind to tumor-associated antigens on tumor cells, triggering a cascade of signals to activate CAR-T cells. This activation is triggered by CAR-T-antigen complex formation in the model and leads to CAR-T cell proliferation, cytokine release, and direct tumor killing. Module C describes the in vivo PD effects in addition to the aforementioned in vitro proliferation and cytotoxicity effects. In both mice and patients, we also investigated the formation of memory-type CAR-T cells, tumor-mediated CAR-T cell exhaustion, and progressive tumor death. Thus, modules A and B can be independently calibrated and validated by corresponding preclinical data, and the full QSP model combining all three modules together can be directly compared with CAR-T cell kinetics and antitumor efficacy data in mice and patients. Details of the model mechanisms and reactions are available in onlinesupplemental tables S1S4.

Figure 1. Workflow of mechanistic QSP modeling analyses and general model structure. The basic building blocks of the overall QSP model of CAR-T treatment in solid tumors are the following two parts. (A) Minimal physiologically-based pharmacokinetic (mPBPK) models that were mechanistically reduced from whole PBPK models were used here to illustrate the in vivo biodistribution of CAR-T cells by dividing the whole body into different key organ compartments (eg, blood, lung, gastrointestinal tract, lymph nodes, tumor, and other tissues and organs). Each organ compartment is then divided into a vascular component and an extravascular component (representing tissue). (B) A cell-level pharmacodynamic model of CAR-T cell activity was used to mechanistically describe CAR-T-mediated tumor cell killing. CAR-T cells can physically recognize tumor antigens and trigger their activation, which results in tumor cell elimination, CAR-T cell proliferation and cytokine production. (C) When the in vitro PD module was integrated into the mPBPK module, additional mechanisms such as stepwise signal transduction downstream of CAR-T recognition, transitional tumor eradication, and formation of memory CAR-T cells were further included to formulate the final in vivo QSP model for mouse and human simulations. (D) During QSP model building and analyses, stepwise model/module calibration using corresponding experimental data was enforced for each physiological scale (cell, animal, human). Finally, at the clinical level, virtual patients were generated based on the QSP model and virtual clinical trials were conducted to predict efficacy and dose-response relationships. CAR, chimeric antigen receptor; CR, complete response; PR, partial response; PD, progressive disease; SD, stable disease; IV, intravenous; GI, gastrointestinal tract; LN, lymph node; PD, pharmacodynamic; QSP, quantitative systems pharmacology.

To ensure a rigorous model calibration and validation process, we used multimodal in-house experimental data (see Materials and methods for details) together with literature data in a stepwise manner throughout model formulation. For module A, published data on the in vivo organ biodistribution of CAR-T cells were used during formulation and calibration of the full PBPK and mPBPK models. For module B, in-house in vitro assays of LB1908 CAR-T cells co-cultured with tumor cells at different E(CAR-T):T(tumor) ratios were conducted, and antitumor cytotoxicity and proinflammatory cytokine secretion profiles were measured quantitatively from three different gastric cancer cell lines (with varying CLDN18.2 expression) and used as calibration data for module B. Next, PANC-1 cells transfected with different amounts of CLDN18.2 RNA were co-cultured with LB1908 CAR-T cells from different donors and cytotoxicity as well as cytokine secretion data were then collected and used as validation data for module B. After the modules were combined, in vivo in-house data on mouse tumor growth inhibition and plasma CAR-T cellular kinetics on treatment with different doses of LB1908 CAR-T infusion were used together to validate the model. As LB1908 is in an early-phase clinical trial and patient response data are still immature at the time of model development, only preclinical data of LB1908 were used for the current work. Additionally, we used multiscale CAR-T experimental data (in vitro, in vivo, clinical) for another four solid tumor targets (HER2, EGFR, GPC3, and MSLN) in similar stepwise calibration-validation cycles to further strengthen the validity and adaptive utility of our QSP model framework. To our knowledge, our model is by far the most extensively calibrated QSP model in terms of the amount of calibrated mechanistic targets, quantitative experimental data usage, and multimodal clinical data consistency.

Stepwise model calibration and validation using in vitro LB1908 cell experimental datasets

We first used LB1908 in vitro experimental data obtained from three gastric cancer cell lines, KATOIII.luc, NUGC4.luc, and SNU16.luc, for the calibration of our primary model. Flow cytometry data suggested that NUGC4.luc cells express the highest amount of CLDN18.2, whereas KATOIII.luc cells express medium levels of CLDN18.2, and SNU16.luc cells express minimal levels of CLDN18.2. In vitro co-culture data from two different T-cell donors revealed a clear association between increased E(CAR-T cell):T(tumor cell) culture ratio and tumor cell killing across all three cell lines, and overall the killing efficacy was greater in cells expressing higher amounts of CLDN18.2; both phenomena were quantitatively captured by module B of our QSP model (figure 2A–F). Furthermore, PANC-1 cells were transfected with different amounts of CLDN18.2 RNA, and in vitro CAR-T cell cytotoxicity was measured experimentally to validate our model. As CLDN18.2 antigen expression increases in PANC-1 cells, cytotoxicity generally increases (figure 2G.H); however, CAR-T cells manufactured from different donors can have different killing strength, and this variation is also reproduced by our QSP model. In addition, we measured the trend of IFN-γ and TNF-α secretion after co-culturing CAR-T cells with the above four different cancer cell lines. Generally, the observed trend is very similar to the in vitro cytotoxicity observed in cells expressing higher amounts of CLDN18.2, and under conditions with higher E:T culture ratios, the secretion of IFN-γ and TNF-α generally increases and becomes more significant. We incorporated this mechanism into the QSP model, and the resulting simulations can simultaneously match all the IFN-γ and TNF-α secretion profiles with high accuracy (figure 2I–V), along with the aforementioned in vitro cytotoxicity data. As a control, untransduced T cells (donor T cells without CAR) were cultured together with PANC-1 cells expressing different amounts of CLDN18.2, and the resulting proinflammatory cytokine secretion remained at baseline as expected (figure 2T,X).

Figure 2. Model calibration and validation using in vitro CAR-T/tumor co-culture data. The in vitro pharmacodynamic module of the quantitative systems pharmacology model was quantitatively calibrated then validated against numerous cytotoxicity and cytokine release datasets collected for LB1908 (anti-CLDN18.2 CAR-T cells). (A–F) Results from in vitro cytotoxicity experiments of CAR-T cells from two different donors (donor 1 and donor 2) co-cultured with two CLDN18.2-positive cell lines (KATOIII.luc, NUGC4.luc) and one CLDN18.2-negative cell line (SNU16.luc) at different effector/tumor (E/T) ratios. Cytotoxicity (expressed as percentages) was assessed at 20 hours. (G–H) PANC-1 cells were transfected with varying amounts of CLDN18.2 messenger RNA (control, 0.000128, 0.00064, 0.0032, 0.016, 0.08, 0.4, 2, 10 µg) and co-cultured with CAR-T cells from two donors at an 8:1 E/T ratio, with cytotoxicity measured at 20 hours and used as model validation. (I–X) CAR-T cells from two donors were co-cultured with the above four cancer cell lines at different E/T ratios or with different CLDN18.2 transfection (with E/T=8:1), and in vitro cytokine release (IFN-γ and TNF-α) was measured at 20 hours. In (T) and (X), TNF-α release from the co-culturing of untransduced T cells (UnT) with PANC-1-luc cells was measured. CAR, chimeric antigen receptor; CLDN18.2, claudin18.2; D, experimental data; IFN, interferon; S, model simulation; TNF, tumor necrosis factor.

The combined QSP model enables accurate translation and prediction of CLDN18.2 CAR-T cellular kinetics and efficacy in vivo

To characterize the biodistribution of CAR-T cells in vivo, we initially used the PBPK modeling approach, and the developed model can quantitatively characterize the dynamic distribution of exogenously infused T cells in the mouse liver, lungs, blood, kidneys, spleen, lymph nodes, gastrointestinal tract, and tumor (figure 3A–D).3334 37,39 To reduce complexity, we then simplified the full PBPK model to an mPBPK model that can still accurately characterize the biodistribution of exogenous T cells in key organ compartments (including the blood, lungs, gastrointestinal tract, and tumor) (figure 3E–H) using published data.³⁰ We integrated the mPBPK model (module A) with the in vitro/in vivo PD model (modules B/C) to generate the full QSP model and depict the behavior of CAR-T cells in tumor-bearing mice (figure 1C). A total of 120 NCG mice were inoculated with NUGC4 cells (CLDN18.2 high expression) and treated with LB1908 CAR-T cells at three doses: 0.3/1/3 M (million cells). Figure 3I–L showed the mean plasma CAR-T cell kinetics and corresponding tumor growth inhibition data summarized from the mouse experiments, together with our QSP model simulations. Data from LB1908 CAR-T cells generated from two different donors revealed strong dose-dependent increases in plasma CAR-T cell expansion and in vivo tumor killing effects in tumor-bearing mice, and these phenomena were captured well by our QSP model. In addition, on different parameterization, our QSP model can reproduce all the dynamic individual mouse data in terms of plasma CAR-T expansion dynamics and tumor growth inhibition (at all three CAR-T doses and for different donors), suggesting that the QSP model can enable accurate translation and prediction of in vivo CAR-T treatment response (figure 3M–T and online supplemental figures S1 and S2).

Figure 3. Model calibration using biodistribution and efficacy data of CAR-T cells in mice. (A–H) PBPK model prediction (line) and experimentally measured time-course profiles (circles) of CAR-T cells in mouse organs. Cells were injected intravenously (20 million per mouse), and biodistribution was quantified in (A) liver, (B) lung, (C) kidney, and (D) spleen. The reduced mPBPK model again reproduced CAR-T cell biodistribution in (E) blood, (F) lung, (G) GI, and (H) tumor in mice following intravenous injection. Experimental data used here and model simulations were expressed in % of injected dose per gram of tissue (%ID/g). (I–T) In NUGC4 (CLDN18.2-high) xenograft mice, two donors’ LB1908 CAR-T cells were infused and corresponding tumor growth inhibition and CAR-T cell expansion were measured at three different doses: 0.3 million cells/animal (low), 1 million cells/animal (medium), and 3 million cells/animal (high). As a calibration step, these quantitative in vivo data were reproduced by the model. (I–L) Averaged profiles (mean) and (M–R) representative individual mouse profiles of CAR-T biodistribution and in vivo efficacy were shown and compared with model simulations. (S–T) Model-simulated and experimentally measured tumor growth dynamics under the control condition (vehicle). (I–T) CAR-T expansion profiles were expressed in units of copies/μg gDNA, and tumor growth/volume was expressed in units of mm³. Dots represent experimental data and lines represent model simulations. CAR, chimeric antigen receptor; D, data; gDNA, genomic DNA; GI, gastrointestinal tract; mPBPK, minimal physiologically-based pharmacokinetic; PD, pharmacodynamics; PK, pharmacokinetics; S, simulation; TV, tumor volume.

The multiscale QSP modeling framework supports the general clinical translation of solid tumor-targeting CAR-T therapies

To demonstrate the general utility of our QSP model in CAR-T research and translational development, we again validated our modeling framework by integrating published preclinical and clinical CAR-T datasets for four additional solid tumor targets: HER2, EGFR, GPC3, and MSLN.48 9 40,46 Using HER2 and EGFR as examples, the same in vitro module of our QSP model (upon corresponding parameter value adjustments specific to the different antigens and CAR-T products) can replicate multimodal in vitro cytotoxicity and cell expansion data of antigen-specific CAR-T cells simultaneously (including scenarios of varying CAR affinities, co-culturing with cancer cells with different HER2/EGFR antigen densities and under different E:T ratios) (figure 4A,B).^{47 48} Similarly, the combined multiscale QSP model can quantitatively predict in vivo time-course tumor growth inhibition in response to different scenarios of CAR-T treatment (different CAR affinities and infusion doses) in mice bearing antigen-expressing cancer cell-derived xenografts (figure 4C,D).^{44 49} At the clinical level, our QSP model was used to generate virtual patients that can again quantitatively match the observed patient response data (including plasma CAR-T expansion and tumor size changes over time) from published anti-HER2 CAR-T and anti-EGFR CAR-T trials (figure 4E,F).^{9 39} The overall sequence of model calibration and validation follows the same order as presented previously with CLDN18.2, and the general utility of our multiscale QSP model was also demonstrated using experimental datasets (including preclinical and clinical data) for GPC3 and MSLN (online supplemental figures S3 and S4).

Figure 4. Multiscale experimental datasets validate and reinforce the translational applications of the QSP model framework. Preclinical and clinical data for additional solid tumor-targeting CAR-T cells (HER2 and EGFR) were used to reinforce the QSP model’s predictive utility. (A–B) At the in vitro level, the QSP model can quantitatively reproduce their respective in vitro cytotoxicity and CAR-T expansion data. (A) For HER2, this includes in vitro cytotoxicity data (in circles) of NALM-6-CBG cells transfected with 0.1 (red), 1 (green) and 10 (blue) μg of HER2 messenger RNA, respectively, and co-cultured with anti-HER2 CAR-T cells with differential affinity (4D5-7:Kd=3.2 nM, 4D5-3:Kd=3.9 µM) at varying E:T ratios (0.5:1-16:1); proliferation (in fold expansion) of the above anti-HER2 CAR-T cells as a function of antigen densities on HER2-expressing cells (1:1 E:T co-culture for 7 days). (B) For EGFR, this includes in vitro cell viability data of EGFR cell lines (with different antigen densities) co-cultured together with anti-EGFR CAR-T cells at different E:T ratios for 4 hours. (C–D) For anti-HER2 and anti-EGFR CAR-T, at the preclinical animal level, in vivo tumor growth kinetics in (C) mice inoculated with two HER2 cell lines (SKOV3 for HER2-high and PC3 for HER2-low) then treated with anti-HER2 CAR-T cells (4D5-7 and 4D5-3), and (D) mice inoculated with EGFR-expressing MDA-MB-231-fluc cells then treated with low doses (2.5e6 cells) and high doses (5e6 cells) anti-EGFR CAR-T. (E–F) The QSP model enables translational prediction of clinical biodistribution and efficacy of anti-HER2 and anti-EGFR CAR-T. (E) For HER2, paired plasma CAR-T expansion data as well as antitumor efficacy data were used (from four patients receiving anti-HER2 CAR-T infusion at 100 M CAR-T cells). (F) For EGFR, paired CAR-T expansion and antitumor efficacy data were used from another four patients receiving anti-EGFR CAR-T at 150 M CAR-T cells. The dashed orange lines indicate 20% and −30% changes in target lesion diameter. Clinical data were extracted from NCT01935843 and NCT01869166 and patient labels were taken directly from the corresponding publications. Dots represent experimental data and lines represent model simulations. For the trial results from Feng et al, it was assumed that the reported best tumor SLD% response occurred at 4 weeks according to their publication, so the simulated SLD% responses at 4 weeks were evaluated against individual patient data; for the results from Liu et al, the durations of actual individual patient response were reported, so the simulated dynamic SLD% responses at different time points were evaluated against individual patient data. CAR, chimeric antigen receptor; EGFR, epidermal growth factor receptor; E:T, effector:tumor; gDNA, genomic DNA; HER2, human epidermal growth factor receptor 2; QSP, quantitative systems pharmacology; SLD%, relative percent change in sum of lesion diameters; TV, tumor volume.

Global sensitivity analyses and virtual clinical trials enable identification of factors that can affect patient-level treatment response

We performed global sensitivity analyses on our QSP model to look for parameters that can potentially affect model output (eg, tumor volume) at the clinical level for patients with CLDN18.2-positive cancer, including both patients with gastric and pancreatic cancer as two model subversions, by referencing relevant information from the literature such as cancer type-specific CLDN18.2 expression (figure 5A,B, online supplemental figure S5a and S6). All virtual patients used in the sensitivity analyses were administered a sample dose of 150 M anti-CLDN18.2 CAR-T cells (by referencing the doses used in published solid tumor CAR-T studies). As expected, larger parameter values for CAR-T activation rate (cmplxa), CAR density on CAR-T cells (agCAR), and CAR-T dose (adddose) led to reduced tumor volume, whereas larger parameter values for tumor inhibitory potential (inhibeff, representing the negative regulatory effect of the tumor microenvironment on CAR-T function and persistence), tumor growth rate (tumor growth), and CAR-antigen dissociation rate (koff) resulted in increased tumor volume. Interestingly, parameters for CAR-T cell proliferation potential (prolifmax) and tumor antigen density (agtumor) were determined to be more influential in patients with pancreatic cancer than in patients with gastric cancer, whereas the CAR-T killing potency (killmax) was suggested to be much more influential in patients with gastric cancer. This can be mechanistically explained by the fact that pancreatic cancer cells and tissues generally express much lower levels of CLDN18.2 than gastric cancer⁵⁰; thus, patients with pancreatic cancer with higher tumor CLDN18.2 expression are much more likely to achieve greater CAR-T activation and tumor killing potency. On the other hand, in gastric cancer, most patients already have high enough CLDN18.2 expression so that antigen density is no longer a critical factor in determining antitumor efficacy; instead, the innate cytotoxic potential of patients’ autologous T cells (as identified from our analyses) could play a decisive role.

Figure 5. Sensitivity analyses and virtual clinical trials to assess factors influencing patient response. (A) Partial rank correlation coefficient (PRCC) indices of parameters that can significantly impact tumor volume in CLDN18.2-positive gastric cancer simulations. The positive or negative signs of PRCC values represent a positive or negative effect on the model output (tumor volume). (B) Simulated spider plots that depict the tumor size change profiles over time for a sample simulation run of 200 virtual patients with gastric cancer receiving CLDN18.2-targeted CAR T cells (150 million cells). (C) Heatmap illustrating the relative importance of different parameters on clinical efficacy as analyzed by the random forest algorithm. Parameters are listed from left to right in the order of increasing contribution to clinical efficacy. Detailed meaning of all listed parameters can be found in onlinesupplemental tables S1S4. (D–I) Kernel density estimation analyses of the probability density distribution of different model parameters including (D) agCAR (number of CARs per CAR-T cell), (E) killmax (killing rate of bound CAR-T), (F) prolifmax (rate of CAR-T proliferation), (G) tumor burden (initial tumor size), (H) tumor growth (rate of natural tumor growth), and (I) inhibieff (inhibitory effect of tumor microenvironment on CAR-T) in terms of their impact on patient treatment response. CAR, chimeric antigen receptor; CLDN18.2, claudin18.2; CR, complete response; PD, progressive disease; PR, partial response; SD, stable disease.

From another perspective, we also used the random forest algorithm to evaluate the importance of different key parameters on the clinical-level efficacy of anti-CLDN18.2 CAR-T therapies (figure 5C). We selected 6 variables out of the 15 most significant parameters from the random forest analysis and conducted kernel density estimation to graphically visualize how changes in parameter values can influence patient response (figure 5D–I and online supplemental figure S7). Our analyses suggest that patients with lower CAR expression (figure 5D), weaker CAR-T killing (figure 5E) or less potent CAR-T proliferation (figure 5F) generally have non-responder-type outcomes (PD, SD). Patients with smaller initial tumor burdens (figure 5G), lower tumor growth (figure 5H) or less immunosuppressive tumor microenvironments (figure 5I) were more likely to achieve a complete response or partial response.

Model-based virtual clinical trials provide quantitative insights for optimal CLDN18.2-targeted CAR-T dose selection and dosing regimen design

To assess the potential dose-response relationships of CLDN18.2-targeted CAR-T therapies in mice and humans and to propose plausible dosing strategies for clinical trials, we initially created three hundred virtual mice on the basis of the preclinical QSP model and simulated three different doses of CAR-T cells. Pooled analyses indicated that CAR-T-mediated tumor growth inhibition in mice increased significantly with increasing dose (online supplemental figure S8). We then generated different virtual patients treated with three representative clinical CAR-T dose levels to explore interindividual variability (50 M – low dose, 150 M – medium dose, and 250 M – high dose) (figure 6A and online supplemental figure S9). Model-generated virtual patients P1-P3 (representing SD, PR and PD) did not show significant changes in efficacy with increasing doses, whereas the response of patient P4 to 50 M CAR-T cells was quite different from that to the 150 M and 250 M doses, indicating substantial interindividual variability in terms of dose-response.

Figure 6. Quantitative systems pharmacology model-based virtual patient simulations provided insights for claudin18.2-targeted CAR-T dosing strategy design. (A) The simulated time-course percentage changes in tumor lesion diameter (compared with baseline) for virtual patients P1 through P4, each treated with varying CAR-T doses of 50 M, 150 M, and 250 M cells. (B) The absolute percentage differences in clinical level response rate (in terms of simulated best objective response rate) when comparing weight-based dosing (3 M/kg) to flat dosing (blue:150 M CAR-T cells; orange:250 M CAR-T cells) in light-weight (35–50 kg) and heavy-weight (50–80 kg) patients. (C) The simulated percentage change in tumor diameter over time for one light-weight virtual patient (P5, 40 kg) and three heavy-weight virtual patients (P6, P7, P8, 60 kg) under the 3 M/kg or 150 M CAR-T cells regimens. The dashed orange lines indicate 20% and −30% changes in target lesion diameter. CAR, chimeric antigen receptor.

We further used our model to assess the potential differences in the efficacy of weight-based dosing and flat dosing of CAR-T cells in patients with CLDN18.2-positive gastric cancer (figure 6B), given that flat dosing could be preferable in CAR-T treatment settings for reasons such as the convenience of preparation and administration. In a model-based virtual patient population with lighter weights (35–50 kg), the administration of CAR-T cells at 3 M/kg compared with a flat dose of 150 M cells was suggested to achieve comparable efficacy (~2% responder percentage difference). In contrast, in the model-based virtual patient population with heavier weights (50–80 kg), there was a notable clinical response difference (in terms of responder percentage) between the two dosing methods, with 150 M flat dosing being inferior to 3 M/kg dosing (~9% responder percentage difference). Model simulations further suggested that a higher flat dose of 250 M cells for the heavier-weight population can achieve a similar clinical-level response compared with the 3 M/kg group (~1% difference). Similar findings were also observed in representative single-patient simulations (figure 6C and online supplemental figure S9). We also investigated the potential impact of dose fractionation on CAR-T efficacy in solid tumors, using CLDN18.2-positive gastric cancer and targeted CAR-T as an example. We tested two dose fractionation strategies that can presumably mitigate adverse events and compared the results with the classical single-dose strategy. Simulated patients P9 and P10 received a single dose of CLDN18.2-targeted CAR-T cells on day 0 (150 M and 250 M, respectively) and their differential responses (including both responders and non-responders) were recorded (figure 7). Then, we simulated and compared the two fractionated regimens, 3-day averaged regimen (days 0/1/2 equal dose) and 3-day step-up regimen (10% on day 0, 30% on day 1, 60% on day 2), with the response of the classical single-dose regimen in P9 and P10, and the results indicated that the patient responses were overall highly comparable (figure 7A,B). In another two virtual patients P11 and P12, we compared one-time CAR-T dosing with similar step-up regimens but over the span of 7 days (the fractionated doses were now administered on days 0/3/6). The simulations again suggested that the antitumor responses were very similar among the three different dosing strategies (figure 7C,D). We further tested the hypotheses using model-based virtual patient populations, and the results again revealed that fractionated dosing strategies can achieve quantitatively similar efficacy compared with the classical regimen at the population level (~1–3% in responder percentage difference) (figure 7E). Overall, these in silico analyses at the clinical level suggested overall similar efficacy of fractionated CAR-T dosing compared with one-time dosing in solid tumor settings (given the same total dose).

Figure 7. Analyzing the impact of step-fractionated dosing of CAR-T therapies in virtual patients. (A–D) The simulated percentage change in tumor diameter over time for virtual patients P9-P12 under the single dosing and fractionated dosing regimens. The individual patients were assumed to have relatively large initial tumor burdens. (A) One-time 150 M CAR-T cells, 3-day 50 M cells/per day, and 3-day step-up regimen (15 M, 45 M, 90 M) were simulated in P9. (B) One-time 250 M cells, 3-day 83.3 M cells/per day, and 3-day step-up regimen (25 M, 75 M, 150 M) were simulated in P10. (C) One-time 150 M cells, 50 M cells on days 0/3/6, and step-up regimen (dosing on days 0/3/6) were simulated in P11. (D) One-time 250 M cells, 83.3 M cells on days 0/3/6, and step-up regimen (dosing on days 0/3/6) were simulated in P12. (E) The absolute percentage differences in clinical level response rate (in terms of simulated best objective response rate) when comparing fractionated dosing regimens (blue: 3-day regimen; orange: 7-day regimen) to the classical single-dose strategy (assuming a total dose of 250 M cells). The dashed orange lines indicate 20% and −30% changes in target lesion diameter. CAR, chimeric antigen receptor.

Discussion

Given the growing interest in the application of CAR-T therapies in the treatment of solid tumors and increasing expenditures of clinical trials, it is highly valuable to develop novel techniques that can quantitatively assess the mechanism of action and in vivo fate of CAR-T cells and therefore guide optimal clinical trial design. This study presents a novel translational QSP model that mechanistically incorporates various features of solid tumor-targeting CAR-T therapies, including whole-body biodistribution, solid tumor antigen recognition and CAR binding, CAR-T cell proliferation and persistence, cytokine release, and antitumor cytotoxicity. During model calibration and validation, we used a substantial amount of multiscale quantitative data in a stepwise manner that included scenarios of in vitro CAR-T-tumor co-culture, in vivo CAR-T cell biodistribution and tumor growth inhibition in mice, and patient-level paired CAR-T expansion and antitumor response, thus advancing the state-of-the-art translational modeling of solid tumor-targeting CAR-T therapies. The current model framework primarily considers general effector and memory-type CAR-T cells with different functions. Interestingly, increasing evidence suggests that variations in the cell type composition of CAR-T products can impact their efficacy. In neuroblastoma, researchers have reported that the numbers of CD4+T cells and central memory T cells in infused GD2-targeted CAR-T products are positively associated with increased CAR-T persistence in patients.⁵¹ For blood cancers, studies have also shown that higher percentages of CD8+/CD45RA+CCR7+ (memory stem cell-like) CD19 CAR-T cells in the infused products are correlated with greater expansion in patients, which may lead to enhanced antitumor activity, and that the composition of CD8+ versus CD4+ CAR T cells in the infused products can profoundly impact overall CAR-T function and therapeutic efficacy.^{52 53} On the other hand, one study revealed that higher numbers of CAR+ regulatory T cells (targeting CD19) were strong markers for progressive disease and treatment failure in patients with blood cancer.⁵⁴ Such emerging features of CAR-T cell product composition have also been reflected in models simulating CAR-T expansion in blood cancers: a recent work by Salem et al considered a total of eight CAR-T cell phenotypes (CD8+ and CD4+ subsets, with each subset further divided into stem-cell memory, central memory, effector memory, and effector cells).²⁸ Although only a small set of patient CAR-T expansion data were used for their model calibration, it still demonstrated exciting potential of using mechanistic models in guiding CAR-T product optimization in order to achieve the best possible clinical response and maximize the clinical success rate. Thus, our next step is to expand the model into a comprehensive platform incorporating multiple CAR-T cell subsets in solid tumor settings that can quantitatively connect CAR-T product manufacturing information to in vivo CAR-T cellular kinetics and individual-level/population-level clinical response. The resulting computational platform can be further calibrated by patient-level cell expansion (of all different cell subsets) together with cytokine secretion, host T-cell expansion, and tumor size data, and this will enable a systems-level multidimensional description of the in vivo interconnected dynamic regulation between all different CAR-T cell subsets as well as their collective impact on overall CAR-T efficacy. In this way, virtual clinical trials can be conducted to prospectively screen and guide the selection of the optimal cell composition and dosage of CAR-T products for future CAR-T clinical trials targeting patients with solid tumor.

CLDN18.2 is significantly expressed in pancreatic cancer and gastric cancer tissues and is associated with the onset and prognosis of cancer, making it a promising target for therapeutic intervention. Our QSP model simulations suggested that in both cancer types, the administration of CLDN18.2-targeted CAR-T can yield promising efficacy. Recently, in multicenter open-label phase I/Ib trials (NCT03874897, NCT04581473) conducted using a CLDN18.2-targeted CAR-T therapy CT041, the objective response rate in patients with gastric/gastroesophageal junction cancer was reported to be 54.9%,^{10 55} which is generally higher than that reported in patients with pancreatic cancer.¹² Our model sensitivity analyses revealed that the CLDN18.2 antigen density in patients with pancreatic cancer may be particularly influential in determining the anticancer CAR-T response. According to the Human Protein Atlas database,⁵⁰ the expression of claudin family proteins (including CLDN18) in the human pancreas is generally lower than that in the gastrointestinal tract, which partly explains the relatively reduced clinical efficacy of CLDN18.2-targeted CAR-T therapy in patients with pancreatic cancer. Another potential explanation is that, compared with gastric cancer cells, pancreatic cancer cells may exhibit more rapid proliferation and migration. For example, the experimentally measured doubling time of PANC-1 (a widely used pancreatic cancer cell line) is approximately 21 hours, while the doubling time for the gastric cancer cell line SNU-16 is significantly longer (29.9 hours).^{56 57} Other factors, such as the barrier-like dense stroma resulting from the desmoplastic reactions commonly observed in pancreatic cancer, may lead to reduced CAR-T tissue penetration and thus decreased efficacy in patients with pancreatic cancer.⁵⁸

Optimal dose selection for CAR-T therapy is always a challenge in clinical settings. Research has indicated that the occurrence of severe adverse events can be strongly associated with CAR-T dose.^{59 60} For example, studies have reported that severe cytokine release syndrome and immune cell-associated neurotoxicity syndrome would more likely occur in patients with blood cancer treated with high CAR-T doses.61,63 For CLDN18.2-targeted CAR-T therapies, the common adverse events in patients are cytokine release syndrome, central nervous system toxicity, tumor lysis syndrome and hemocytopenia.⁶⁴ It was reported that one patient who was administered with CT041 experienced grade 4 gastrointestinal hemorrhage at a very high dose of 500 M, and with this concern, the researchers later reduced the dose to lower levels (eg, 250 M) in their trial expansion phase.¹⁰ On the other hand, our model suggested that CAR-T cells administered at doses less than 100 M would result in significantly reduced antitumor efficacy in patients with CLDN18.2-positive cancer (internal simulations). This predicted feature is qualitatively supported by prior findings of other CAR-T products that showed significantly decreased efficacy at low CAR-T doses in patients (eg, below 50 M), although the overall pattern of CAR-T dose-response in general (especially at high doses) is still under debate and may likely vary in a case-by-case manner.^{65 66} Furthermore, the recently released clinical results of six patients treated with 0.5×10⁶ cells/kg LB1908 showed that two patients have achieved maximal target lesion shrinkage of more than 30%, and this efficacy response is within our model-predicted objective response rate (ORR) ranges for that given dose (internal simulations); however, the clinical sample size so far (n=6) is relatively small so the efficacy results may change as sample size increases.⁶⁷ In response to the above safety and efficacy considerations, an area of active research in clinical CAR-T dose selection is the potential of applying dose fractionation.⁶⁸ Studies by Frey et al and Xu et al examined the comparative efficacy of single-dose CAR-T and fractionated doses in patients with cancer and demonstrated that the response rates in the fractionated dose groups were very close to the overall historical response rates observed in the single-dose groups, which is also consistent with the simulation results of our model.69,72 Thus, dose fractionation may significantly reduce clinical adverse events for patients, enabling flexible titration of CAR-T in step-up manners as well as temporary halts in treatment in the event of severe adverse events after receiving fractionated doses.^{70 73}

As mentioned earlier, quantitative modeling has recently emerged as a new tool for understanding CAR-T proliferation and cytotoxic mechanisms in vivo, deconvoluting complex cellular kinetics-response relationships, and designing better clinical regimens for patients.^{74 75} In addition to the models used to study CAR-T efficacy, Hanson et al built a theoretical model for leukemia that aimed to investigate the general interplay between CAR-T cells, endogenous T cells and B cells, tumor burden as well as the treatment-related inflammatory toxicity in patients. Model simulations of the inflammatory cytokine storm resulting from different CAR-T treatment regimens and initial patient tumor burdens can match published population-level data well, which opens up a new avenue for model-informed forecasting of clinical adverse events.⁷⁶ Another model by Hardiansyah and Ng incorporated the production and turnover processes of specific inflammatory cytokines, such as interleukin-6 and IFN-γ, and the authors demonstrated that the model can enable individual-level quantitative prediction of dynamic cytokine profiles on CAR-T treatment in blood cancers.⁷⁷ Overall, the field of quantitative CAR-T modeling (especially in solid tumors) is still in a nascent stage with increasing interest and expanding translational applications, since similar multiscale modeling strategies have already demonstrated significant research applications and drug development impacts in many other disease areas and therapeutic modalities.78,80 Given the physiological complexity of CAR-T therapies, we envision that mechanistic QSP-type modeling, as we presented here, has unique advantages over empirical models in terms of understanding CAR-T pharmacology and toxicology, which can be further leveraged to improve CAR-T therapeutic design, optimize clinical dosing and enhance patient outcomes.

Limitations of the study

While our modeling results showed exciting potential of using high-throughput simulations in guiding CAR-T translational research against solid tumors, there are still limitations regarding model applications that should raise attention. First, the innate features of preclinical experimental systems (in vitro co-culture, CDX and PDX mouse) used in CAR-T research are different from the actual human tissue microenvironment; for example, there is no host immune system in CDX or PDX mice, while in humans, there is an intact immune system that could exert beneficial or suppressive effects on CAR-T function. We envision that in future extensions of the model, researchers can add dynamics of host T cells and essential proliferation-related cytokines (as well as the regimens and effects of lymphodepletion) to more systematically describe the potential competition between host immune system and exogenously infused CAR-T cells. Second, considerations of CAR-T-induced adverse events (of known or unknown driving mechanisms) are also important for model-informed clinical translation. Although CAR-T-related CRS in solid tumors is usually not as serious as in blood cancers, it was reported (eg, in the clinical studies of CT041) that significant portions of patients can experience high-grade hematological toxicities (due to both preconditioning and also possibly CAR-T). Therefore, incorporating the driving mechanisms of treatment-induced hematotoxicity through mechanistic modeling (eg, proliferation and differentiation of various hematopoietic cell lineages and the killing effects of drugs) in future modeling efforts is also of great value in deriving optimal dosing regimens of CAR-T in solid tumor settings by balancing both efficacy and toxicity. In addition, since CAR-T cells would be dynamically educated by the complex immune environment and heterogenous cancer tissues in solid tumor settings, extending the current model to include more reactions describing intratumoral immune regulation and tumor tissue features (eg, antigen heterogeneity, extracellular matrix, vasculature) would further advance the predictive capacity of model simulations. As a number of cytokines secreted by activated CAR-T cells were also known to regulate CAR-T function and cell fate, future extensions of the QSP model can further consider cytokine secretion/regulation in more sophisticated ways, together with emerging data from preclinical and clinical measurements, to deliver more mechanistic insights for clinical decision-making. Then, by generating virtual patient populations that can physiologically represent interpatient variability in the above features and running virtual clinical trials through repetitively sampling, such model-based investigations can provide a more confident and comprehensive projection of all possible patient response scenarios to better inform the clinical translation of solid tumor-targeting CAR-T products.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information.