key: cord-0427969-fenbmufm authors: Creemers, J. H. A.; Roes, K. C. B.; Mehra, N.; Figdor, C. G.; de Vries, I. J. M.; Textor, J. title: In silico cancer immunotherapy trials uncover the consequences of therapy-specific response patterns for clinical trial design and outcome date: 2021-09-13 journal: nan DOI: 10.1101/2021.09.09.21263319 sha: 99a5d7d0179c38221683ce78bf4eccfb96e167d8 doc_id: 427969 cord_uid: fenbmufm Background Late-stage cancer immunotherapy trials strive to demonstrate the clinical efficacy of novel immunotherapies, which is leading to exceptional responses and long-term survival in subsets of patients. To establish the clinical efficacy of an immunotherapy, it is critical to adjust the trial's design to the expected immunotherapy-specific response patterns. Methods In silico cancer immunotherapy trials are virtual clinical trials that simulate the kinetics and outcome of immunotherapy depending on the type and treatment schedule. We used an ordinary differential equation model to simulate (1) cellular interactions within the tumor microenvironment, (2) translates these into disease courses in patients, and (3) assemble populations of virtual patients to simulate in silico late-stage immunotherapy, chemotherapy, or combination trials. We predict trial outcomes and investigate how therapy-specific response patterns affect the probability of their success. Results In silico cancer immunotherapy trials reveal that immunotherapy-derived survival kinetics -- such as delayed curve separation and plateauing curve of the treatment arm -- arise naturally due to biological interactions in the tumor microenvironment. In silico clinical trials are capable of translating these biological interactions into survival kinetics. Considering four aspects of clinical trial design -- sample size calculations, endpoint and randomization rate selection, and interim analysis planning -- we illustrate that failing to consider such distinctive response patterns can significantly reduce the power of novel immunotherapy trials. Conclusion In silico trials have three significant implications for immuno-oncology. First, they provide an economical approach to verify the robustness of biological assumptions underlying an immunotherapy trial and help to scrutinize its design. Second, the biological basis of these trials facilitates and encourages communication between biomedical researchers, doctors, and trialists. Third, its application as an educational tool can illustrate design principles to scientists in training, contributing to improved designs and higher success rates of future immunotherapy trials. Since a constant time-independent tumor growth rate would unlikely be observed in a clinical 124 setting, we have added two additional parameters to the model that influence the growth rate of 125 tumors in a time-dependent manner, which are: 126 • The tumor growth rate decline (Dr): a parameter that describes to which extent the 127 proliferation rate of tumor cells gradually declines over time. 128 • The decay rate of the tumor growth rate decline (rd; hereafter referred to as 'decline decay 129 rate'): a parameter that indicates at which pace the tumor growth rate decline decreases. 130 The model parameters and their values are listed in Table 1 ; their rationale is described previously 27 . 131 Moreover, the values of the tumor growth rate decline and its decay rate were set to augment the 132 interpatient variability in tumor development and allow for more extensive disease trajectories. At 133 baseline, only one tumor cell and a pool of 10 6 naive T cells are presumed to be present, while activated 134 T cells are absent, yielding the following initial conditions for the simulations: T(0) = 1, I(0) = 0, S(0) = 135 0, and N(0) = 10 6 . 136 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 137 Simulating untreated disease, chemotherapy, and immunotherapy in individual patients 138 Using this ODE model, we simulated cancer development and disease trajectories in patients. We 139 varied the tumor properties (i.e., the tumor growth rate, the growth rate decline, and the decline decay 140 rate) between patients extensively to guarantee interpatient variation in disease courses. Unless 141 otherwise specified, the remaining model parameters are set to the same values for all patients (Table 142 1). 143 Each patient is simulated from cancer onset (i.e., malignant transformation of the first cell) for 144 up to more than two years (i.e., 800 days). A simulated time step corresponds to one day. The diagnosis 145 threshold of a tumor mass was set to 65 * 10 8 cells, corresponding to the size at which common 146 malignancies are diagnosed. The lethal tumor burden is set to 10 12 tumor cells (a tumor volume of 147 approximately 10.6 dm 3 ). 148 Disease trajectories of patients with cancer can be steered with therapy. Given their prominent 149 roles in the oncological treatment landscape, we included immune checkpoint inhibitors (ICI) and 150 chemotherapy in the model. Both treatments function through their primary modes of action. ICI are 151 implemented as follows: once a cancer reaches a diagnosis threshold, immune checkpoint inhibitors 152 increase the killing rate of cytotoxic T cells (multiplication factor: 0-7), enabling them to eradicate 153 tumor cells. The duration and potency of the ICI treatment eventually determines patient outcome. 154 In patients treated with chemotherapy, the immune system is still present; however, it is not 155 boosted (as is the case during ICI treatment), hence the T cells are not potent enough to curb the tumor 156 growth. Once a patient is diagnosed with cancer, chemotherapy can reduce the tumor growth rate 157 with its cytotoxic capacity (multiplication factor: 0-1). Again, the duration and potency determine 158 patient outcome. By default, the treatment duration for ICI and chemotherapy are two years and six 159 months, respectively. 160 Simulating untreated cohorts in an in silico trial: model fitting 162 To expand our modeling approach from a single patient into a trial cohort, we simulated multiple 163 patients with individualized disease courses based on unique tumor properties. As an illustrative 164 example, we took a publicly available dataset of patients with advanced lung cancer from the North 165 Central Cancer Treatment Group (NCCTG) and regarded the survival times of these patients as if they 166 were untreated 30 . We fitted our model to the NCCTG dataset to show that our trial model can 167 reproduce authentic survival kinetics as observed in clinical trials. Specifically, we searched for 168 parameter combinations reflecting realistic survival times for each patient in our model. Since a single-169 parameter fitting approach could not generate a sufficiently wide range of survival times, we used a 170 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint multidimensional fitting approach. An overview of the fitting approach is depicted in Supplementary 171 Figure 1 . 172 Specifically, we fitted the tumor growth rate (r), the decline in tumor growth rate over time 173 (∆r) , and the decline's decay rate (rd) of patients to OS. The fitting approach comprised two processes: 174 1) sampling non-censored survival values and 2) translating these survival values to model parameters. 175 We fitted a (parametric) Weibull distribution to the NCCTG lung cancer dataset (shape k: 1.32, scale l: 176 417.76; Supplementary figure 1A). The choice for a Weibull distribution is based on the Akaike 177 information criterion and the fact that a Weibull model is a survival model from which the parameters 178 (i.e., scale and shape) contain a mechanistic meaning and can, therefore, be interpreted. Late-stage (i.e., phase III) clinical trials traditionally contain two arms: a control arm and a treatment 191 arm. The control arm can be a placebo (i.e., untreated) or a standard of care therapy. To construct 192 phase III in silico immunotherapy trials, we extended the simulations with treatment cohorts (mono-193 chemotherapy, mono-immunotherapy, chemoimmunotherapy, or induction chemotherapy followed 194 by immunotherapy). These cohorts facilitate the comparison between various treatment regimens. A 195 treatment cohort uses the same baseline parameter distribution as a control cohort. It differs in one 196 critical aspect, though: once patients in the treatment arm reach a tumor burden that corresponds to 197 the diagnosis threshold, patients can be treated with chemotherapy, ICI, or combination therapy, as 198 described above. The distribution of survival parameters is, unless otherwise specified, derived from 199 the most mature, digitized data from the CA184-024 trial, as shown below 31 . At inclusion into the trial, 200 patients are randomly assigned to a study arm (randomization). 201 The primary endpoint of the trials is the 2-year OS. Given the absence of accrual times in in 202 silico trials, the trial duration equals two years, which provides each patient in the trial with 24 months 203 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint We used in silico cancer immunotherapy trials -a mechanism-based simulation platform of cancer-227 immune dynamics -to investigate the consequences of immunotherapy-specific response patterns on 228 trial design principles 27 . Patients within these trials are simulated with an ODE model, which describes 229 cancer development in a patient by modeling the interaction between tumor cells and the immune 230 patients are randomized into two cohorts to resemble conventional phase III trials: a control group 244 (either placebo or chemotherapy) and a treatment group (immunotherapy, chemoimmunotherapy, or 245 induction chemotherapy followed by immunotherapy; Figure 1D ). Since the cellular dynamics (e.g., 246 tumor burden over time or the efficacy of T cell killing) and survival outcomes of these patients are 247 known and can be modified, in silico clinical trials are suited to answer questions like: "Assuming that 248 a novel treatment X increases T cell killing by 5%, how does this translate to a survival benefit in 249 patients? Moreover, how many patients are needed to establish this benefit in a clinical trial? When 250 should one analyze the results?" ( Figure 1E ). 251 252 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. Subsequently, cohorts of patients were constructed based on the fitted parameters to simulate actual immunotherapy trials. (E) Applications of these trials include predicting survival outcomes of trials, estimating appropriate sample sizes, selecting 263 endpoints and randomization ratios, and investigating the timing of interim analyses. In silico late-stage immunotherapy trials yield realistic survival outcomes 266 To illustrate that this in silico clinical trial approach can generate realistic survival kinetics as observed 267 in late-stage immunotherapy trials, we fitted the simulation to three different datasets: (1) The choice for these trials is based on the size of the trials and the maturity of the data. The follow-up 272 of the CA184-024 trial and the CheckMate 066 trial were five and three years, respectively. As the last 273 two datasets were not publicly available, we extracted the data using image digitization (see Methods). 274 As a reference for the in silico trials, we visualized the Kaplan-Meier estimators of these datasets 275 ( Figure 2A ). Both trials were digitized correctly, as reflected by the nearly identical risk tables compared 276 to the original manuscripts 31, 32 . Next, we fitted our trial simulation model on the NCCTG dataset and 277 the control arms of the CA184-024 and CheckMate 066 trials ( Figure 2B ; black lines). Given the limited 278 response rates of dacarbazine for metastatic melanoma (15%), the patients in the control arm were 279 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint regarded as untreated, and the model was fitted as such. For simplicity, we did not simulate dropout 280 or censoring in the trials shown in this paper, although it could be added to the simulation. On average, 281 the simulations capture the survival kinetics of the trials accurately, which is reflected by similar 282 median overall survival values and reasonably corresponding risk tables. The final step to fully 283 resemble late-stage immunotherapy trials in a simulation setting is replicating the treatment arms of 284 the CA184-024 and CheckMate 066 trials. These simulated patients were treated with ICI upon 285 diagnosis. ICI increased their T cell killing rate seven-fold and prolonged their survival, leading to OS 286 benefit in the in silico trial that matched the original trial. Hence, these in silico trials couple the disease 287 mechanism and mechanistic treatment effect to a realistic clinical trial outcome. Interestingly In silico immunotherapy trials enable a priori prediction of trial outcomes and uncover 302 immunotherapy-specific response patterns 303 304 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint The design and the success rate of any clinical trial depends, among others, on an accurate a priori 305 prediction of the survival kinetics -i.e., the shape of the survival curves and the trial outcome. For late-306 stage immunotherapy trials, commonly observed immunotherapy-induced response patterns are a 307 delayed curve separation and a plateauing tail of the survival curve of the treatment arm (Figure 2) . 308 These characteristic survival curve shapes reveal a violation of a vital premise at the basis of many 309 clinical trials: the proportional hazard assumption (PHA). The PHA states that the 'instantaneous death 310 rate' of a patient (i.e., the hazard rate) in both arms of the trial should be proportional, resulting in a 311 constant hazard ratio. Many traditional design methods, ranging from sample size calculations to 312 outcome analyses, depend on this theory. For late-stage immunotherapy trials, this induces two 313 problems: (1) while a violation of the PHA needs to be addressed during trial planning, the hazard rates 314 -and thereby the fact if the trial adheres to or violates the PHA -become available after analysis of 315 the trial, and (2) if a trial does not adhere to a PHA, what will be the shape of the survival kinetics? 316 Especially in an era where treatment and control arm regimens are becoming increasingly complex, 317 adjusting the design and analysis methods to unknown survival kinetics is challenging. 318 In silico clinical trials can provide principled estimates of the shape of the survival curve, 319 including the underlying hazard rates and hazard ratios before trial execution. The most traditional 320 scenario would be a trial in which patients are randomized 1:1 to mono-chemotherapy or placebo. 321 Given the direct chemotherapy effect, the PHA is generally assumed to hold for these trials. An in silico 322 trial in which chemotherapy reduces the tumor growth rate to 70% for the duration of the trial indeed 323 replicates these assumptions (Supplementary Figure 2) : the survival curves separate from the start of 324 the trial, and the hazard ratio is constant over time. However, what happens if the chemotherapy 325 effect does not last for the entire trial but for -maybe more realistically -6 months? Initial 326 proportional separation of the survival curves is followed by a nearly parallel decay of both curves, 327 leading to an early survival benefit for the chemotherapy arm ( Figure 3A ). Consequently, for any 328 therapy with a non-constant treatment effect -even for chemotherapy trials -deviations from the 329 PHA might be observed. When we switch to immunotherapy in the treatment arm, a violated PHA 330 becomes immediately apparent. Recall that in our model, immunotherapy exerts its mechanistic effect 331 indirectly on the tumor via an increase in the killing rate of T cells. Through approximately the first six 332 months, the hazard rates remain constant over time, but after that, they start to decline in the 333 immunotherapy group (red line), yielding a non-constant hazard ratio over time ( Figure 3B) . 334 The flexibility of in silico trials lies in their ability to incorporate complex treatment regimens. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint 3E). Mechanism-based immunotherapy trials provide the means to translate biological assumptions 340 regarding the disease and treatment effects into survival kinetics (including its hazard rate/ratio 341 estimates). These survival kinetics, such as crossing survival curves ( Figure 3D ) or a temporary curve 342 separation ( Figure 3E ), may be hard to predict otherwise and can be detrimental for the trial outcome 343 if not dealt with appropriately. 344 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint (T = treatment, C = control). The red dot in column three indicates the overall hazard ratio. For an accurate prediction of the 355 hazard rates and hazard ratios, saturated survival curves (n=100.000 patients per arm) were used, and the data was smoothed 356 before plotting. Ignoring immunotherapy-specific response patterns might cause an overestimation of an 359 immunotherapy trial's power 360 To investigate the consequences of violating the PHA on the power of a clinical trial, we compared the 361 power calculated using a PHA-dependent method (the Log Rank test) with a non-PHA-dependent 362 method (Pearson's Chi-squared test) for different clinical scenarios. An essential difference between 363 both methods is that the Log Rank test considers the entire survival curve, while Pearson's Chi-squared 364 test only compares the number of events in both arms at 24 months. Logically, in a scenario that 365 approximates the PHA the closest (such as a chemotherapy vs. placebo trial), a PHA-dependent 366 method is superior ( Figure 4A ). However, a traditional immunotherapy trial violates the PHA, leading 367 to a vast underestimation of the power of the trial when PHA-dependent methods are used for its 368 planning ( Figure is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. In silico trials can validate endpoints and randomization ratios before trial execution 391 Clearly, the success rate of novel immunotherapy trials depends on more than its sample size alone. 392 To establish an OS benefit of the treatment arm, it is crucial to analyze the trial once the data have 393 reached a certain maturity -i.e., the treatment needs to be granted sufficient time to induce a survival 394 benefit. We assumed that a delayed curve separation in immunotherapy trials would prolong the 395 follow-up needed to establish an OS benefit of immunotherapy and thereby defer reaching maturity 396 of the trial data. If the therapy is effective, data maturity can be regarded as the time point when a 397 treatment effect can be observed. Hence, an optimal trial endpoint would be the earliest time at which 398 this treatment effect can be detected with sufficient power. Therefore, we analyzed the power of 399 differently sized trials with respect to their OS endpoint. Herein, we distinguished trials that were 400 subject or were not subject to a delayed curve separation (immunotherapy and chemotherapy, 401 respectively). In a classic chemotherapy trial, the treatment effect translates directly to a survival 402 benefit in the treatment arm -the survival curves separate from the start. Therefore, the highest 403 power will be obtained after the total duration of the treatment effect ( Figure 5A, panel 1) . In this case, 404 the treatment effect lasts for six months, leading to the 6-months OS as the endpoint with the highest 405 power. The delayed curve separation in immunotherapy trials renders it futile to analyze OS data early 406 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint on in the trial ( Figure 5B, panel 1) . A practical ramification is that in the presence of a delayed curve 407 separation, the trial requires a sufficiently long follow-up and an adequate size to gain power and 408 detect immunotherapy-specific treatment effects. Mechanism-and simulation-based power 409 calculations with in silico trials can consider these specific kinetics when determining the sample size 410 for upcoming trials. 411 Given the observation that both the size of an immunotherapy trial and its endpoint heavily 412 influence the probability of finding the survival benefit of interest, we presumed that increasing the 413 size of the treatment arm -i.e., an unequal randomization scheme -would similarly affect the power. 414 Instead of varying the study size, we now varied the randomization ratio (second panel of Figure 5A /B). 415 Interestingly, while the power logically depended on the OS endpoint, the randomization ratio did not 416 greatly affect the power ( Figure 5B ). Considering that an unequal treatment allocation may provide 417 ethical benefits, we confirm that the randomization ratio in immunotherapy trials is of secondary 418 importance compared to its size or primary OS endpoint. In summary, our in silico immunotherapy Since the survival curves in classical chemotherapy trials separate from the trial onset, the highest power -and most optimal 428 endpoint -is obtained at the end of the treatment interval (i.e., after six months in this example; see Figure 3A ). Although 429 less influential, a similar observation can be made for randomization ratios (study size panel 2: 300 patients). (B) Delayed 430 curve separation in immunotherapy trials emphasizes that a premature final analysis of the primary OS endpoint is 431 detrimental to the trial outcome. These trials permit validating the pre-specified survival outcomes of novel trials a priori. Commonly selected randomization ratios do not seem to be heavily influenced by immunotherapy-specific response patterns 433 (study size panel 2: 1200 patient). Trial characteristics are similar to Figure 3A is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint 435 We have observed a clear tradeoff between the power of an immunotherapy trial on the one hand, 437 and the primary OS endpoint, and correspondingly the data maturity, on the other. Luckily, the two 438 are not entirely mutually exclusive: interim analyses have been developed for ethical purposes to 439 establish positive or harmful treatment effects early. However, there is a catch: the necessity to control 440 for multiple testing at each interim analysis lowers the significance threshold on the final analysis to 441 maintain the same overall type I error rate. This begs the question: "How many interim analyses should 442 you plan, and when should you plan them?" Again, principled answers to such questions can be 443 obtained with the help of in silico immunotherapy trials. To illustrate this, we simulated 1000 444 immunotherapy trials with 1200 patients per trial, randomized 1:1 over immunotherapy with a strong 445 treatment effect or a placebo ( Figure 6A ). In the absence of interim analyses, the vast majority of the 446 trials are predicted to end up positive. Adding interim analyses (O'Brien-Flemming approach) to the 447 equation induces a tradeoff. On the one hand, increasing the number of equally-spaced interim 448 analyses increases the probability of early detecting a positive treatment effect (e.g., approximately 449 60% of the trials are positive after 18 months in the case of three interim analyses; Figure 6A ). On the 450 other hand, the overall probability of ending up with a negative trial due to more stringent analyses 451 (i.e., less power) also increases, especially in the case of immunotherapies with a weaker treatment 452 effect (±57% without an interim analysis vs. ±63% with three interim analyses; Figure 6B ). In an actual 453 trial, the latter needs to be corrected by including additional patients to maintain the pre-planned 454 power. Furthermore, we observe that the timing of the interim analysis is crucial. Whereas an interim 455 analysis at 18 months provides additional value to the trial, interim analyses before 16 months are 456 predicted to be wasteful: both due to the presence of non-proportional hazards and less mature data. 457 As a control, we simulated trials without any treatment effect. By design, approximately 95% of the 458 trials should end up negative irrespective of the number of interim analyses, which seemed to be the 459 case ( Figure 6C ). Logically, the weaker the treatment effect, the higher the probability of erroneously 460 finding a harmful treatment effect -a characteristic that the simulation also exhibits ( Figure is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. In this study, we used mechanism-based in silico cancer immunotherapy trials to predict survival 476 kinetics and response profiles of novel immunotherapy trials. Complementary to conventional design 477 methods, in silico trials provide the ability to investigate the implications of a researcher's biological 478 (as opposed to statistical) hypotheses of a drug's mechanism of action for the design, conduct, analysis, 479 and outcome of clinical trials. When comparing the simulated outcomes to actual immunotherapy trial 480 outcomes, we showed that in silico trials are suited to translate complex biological mechanisms (such 481 as observed during the treatment of patients with ICI) into realistic trial outcomes. Crucially, the 482 survival kinetics that arose from these mechanism-based simulations reflected two pivotal 483 components often found in immunotherapy trials: a delayed curve separation and a plateauing tail of 484 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint the survival curve at later stages of the trial. In line with genuine immunotherapy trials, we find that 485 these immunotherapy-specific response patterns differ considerably from chemotherapy-based 486 kinetics. Our findings confirm that diversity in survival kinetics profoundly impacts the outcomes of 487 immunotherapy trials 33 . Consequently, these features need to be considered when deciding on the 488 sample size, endpoint, randomization ratio, and the number and timing of interim analyses of a novel 489 immunotherapy trial. 490 Over the past two decades, in silico clinical trials are gaining in popularity. These trials enable 491 investigating, among others, how novel drugs, treatment schedules, dosing regimens, and interpatient 492 heterogeneity affect the outcome of a clinical trial 34 . In silico clinical studies have a wide range of 493 applicability from pediatric infectious 35 and orphan diseases 36 to diabetes 37 , inflammatory 494 autoimmune diseases 38 , traumatic injury 39 , psychiatric illness 40 , and cancer. In oncology, several in 495 silico clinical trials involving chemotherapy and tyrosine kinase inhibitors have been performed 41, 42 . 496 Moreover, with the onset of checkpoint inhibitors, in silico immunotherapy trials have gained interest, 497 leading to trials with anti-CTLA-4-antibodies and anti-PD-(L)1 antibodies 43, 44, 45 . The common 498 denominator in these trials is that they primarily center around the therapies' dosing regimens and 499 treatment schedules. Herein lies the main difference with our simulation approach: although the 'key 500 ingredients' of these approaches are similar -they are based on a mathematical abstraction of a 501 disease mechanism -our trials do not aim to optimize treatment schedules. Instead, we complement 502 traditional design methodology by adding the means to predict trial outcomes and elucidate trial 503 kinetics a priori to steer design decisions of novel immunotherapy trials. These trials differ from 504 traditional trial design research in that these, often statistically-grounded, approaches simulate clinical 505 trials based on population-level assumptions (e.g., with particular distributions of survival times, study 506 durations, or with a specific censoring mechanism). Examples of these high-level simulation 507 approaches include, but are certainly not limited to, studies aiming to calculate the sample size and 508 power of clinical trials 46, 47, 48 . Since these methods lack a direct link to the underlying biological disease 509 mechanism, interpreting their parameters for individual trial participants is difficult or even 510 impossible. In contrast, in silico trials are founded on biological assumptions but then translate these 511 assumptions into statistical concepts such as hazard ratio kinetics. In this manner, simulated trials 512 encourage an interdisciplinary discussion about the design of an upcoming trial. 513 In silico clinical trials are applicable in several settings. First, they provide the means to verify 514 clinical trial and treatment assumptions before investing extensive amounts of work and funds into 515 the development and execution of a clinical trial and can, thereby, function as a proof of principle of 516 the soundness of the hypotheses for an upcoming trial. Scrutinizing each aspect of the trial supports 517 optimal design decisions and might reduce unanticipated outcomes. Moreover, this mechanism-based 518 approach does not necessitate a deep understanding of complex mathematical theorems; instead, it 519 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint requires a biological understanding of a disease. This mechanistic basis is intuitive, which benefits the 520 communication between clinical doctors and biomedical researchers on the one hand and statisticians 521 and clinical trialists on the other. Additionally, in silico trials might serve as excellent educational tools. 522 The ability to simulate a wide range -from basic to highly advanced -research questions can be 523 exploited in teaching activities for entry-level clinicians to experienced trialists. A final implication, 524 which holds for any trial simulation, is that they may provide insight when conventional clinical trials 525 are unfeasible due to practical or ethical constraints (e.g., clinical trials in rare diseases, pediatrics, or 526 critical care medicine). 527 Nonetheless, in silico clinical trials have to be considered in light of some limitations. The most 528 critical limitation is universal to any -either in vitro, in vivo, or computational -scientific model: the 529 immunotherapy trial outcomes depend heavily (if not entirely) on the biological assumptions of the 530 model, meaning that incorrect interactions or erroneous parametrization of the model might induce 531 inaccurate outcomes. The parameterization, in particular, might pose a problem: given the often novel 532 treatment mechanisms, data to fine-tune the parameters of the model accurately might be scarce. In 533 these cases, the simulation itself can be used as a sensitivity analysis to assess to what extent a certain 534 parameter range influences the robustness of the predictions. In addition, while the model itself is 535 intuitive to understand, translating biological principles into an ODE model and implementing it into a 536 simulation requires thorough knowledge of computational methods, limiting its widespread 537 applicability. 538 In summary, in silico cancer immunotherapy trials offer a versatile approach to simulate 539 immunotherapy trials based on biological assumptions. Furthermore, as a simulation tool, they 540 facilitate the verification of trial design decisions to optimize the probability of a successful 541 immunotherapy trial and contribute to high-quality research for cancer patients. 542 543 544 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 13, 2021. ; https://doi.org/10.1101/2021.09.09.21263319 doi: medRxiv preprint Evaluation of Overall Response Rate and Progression-Free 597 Survival as Potential Surrogate Endpoints for Overall Survival in Immunotherapy Trials Milestone Survival: A Potential Intermediate Endpoint for Immune Checkpoint 601 Inhibitors Durable response rate as an endpoint in cancer immunotherapy: insights 604 from oncolytic virus clinical trials Development of tumor mutation burden as an immunotherapy biomarker: 607 utility for the oncology clinic Microsatellite Instability as a Biomarker for PD-1 PD-L1 Expression as a Predictive Biomarker in Cancer Immunotherapy Designing Late-Stage Randomized Clinical Trials with Cancer Immunotherapy: Can 619 We Make It Simpler? Restricted mean survival time: an alternative to the hazard ratio for 622 the design and analysis of randomized trials with a time-to-event outcome Designing therapeutic cancer vaccine trials with delayed treatment 626 effect A tipping point in cancer-immune dynamics leads to divergent 629 immunotherapy responses and hampers biomarker discovery Extending the quasi-steady state approximation by changing 632 variables A general functional response of cytotoxic 635 T lymphocyte-mediated killing of target cells Prospective evaluation of prognostic variables from patient-completed 638 questionnaires. North Central Cancer Treatment Group Five-year survival rates for treatment-naive patients with advanced melanoma 641 who received ipilimumab plus dacarbazine in a phase III trial Survival Outcomes in Patients With Previously Untreated BRAF Wild-Type 644 Advanced Melanoma Treated With Nivolumab Therapy: Three-Year Follow-up of a 645 Randomized Phase 3 Trial Predicting analysis times in randomized clinical trials with cancer immunotherapy. 648 Translational approaches to treating dynamical diseases through 651 in silico clinical trials Novel model-based dosing guidelines for gentamicin and tobramycin in 654 preterm and term neonates In silico clinical trials for pediatric orphan 657 diseases Integrating epidemiological data into a mechanistic model of type 2 diabetes: 660 validating the prevalence of virtual patients Alternate virtual populations elucidate the type I 663 interferon signature predictive of the response to rituximab in rheumatoid arthritis Trauma in silico: Individual-specific mathematical models and virtual clinical 667 populations Dosing and Switching Strategies for Paliperidone Palmitate 3-Month 670 Formulation in Patients with Schizophrenia Based on Population Pharmacokinetic Modeling 671 and Simulation, and Clinical Trial Data Computational design of improved standardized chemotherapy 674 protocols for grade II oligodendrogliomas Reduced tyrosine kinase inhibitor dose is predicted 677 to be as effective as standard dose in chronic myeloid leukemia: a simulation study based on 678 phase III trial data A Computational Model of Neoadjuvant PD-1 Inhibition in Non-Small Cell 681 Lung Cancer A QSP Model for Predicting Clinical Responses to Monotherapy, Combination 684 and Sequential Therapy Following CTLA-4, PD-1, and PD-L1 Checkpoint Blockade In silico simulation of a clinical trial with anti-CTLA-4 and anti-PD-L1 688 immunotherapies in metastatic breast cancer using a systems pharmacology model Sample size calculation for simulation-based multiple-testing 692 procedures Power and Sample Size Calculations in Clinical 695 Trials with Patient-Reported Outcomes under Equal and Unequal Group Sizes Based on Graded 696 Response Model: A Simulation Study Efficient and flexible simulation-based 699 sample size determination for clinical trials with multiple design parameters ICI: immune checkpoint inhibitor; NCCTG: North Central Cancer Treatment Group; ODE: ordinary 705 differential equation; OS: overall survival COVID-19. Nat Rev Drug Discov, (2020