Mathematical modeling framework enhances clinical trial design for maintenance treatment in oncology Authors: Emilia Kozłowska, Ulla-Maija Haltia, Krzysztof Puszynski, Anniina Färkkilä Published: 2024 ABSTRACT AbstractClinical trials are costly and time-intensive endeavors, with a high rate of drug candidate failures. Moreover, the standard approaches often evaluate drugs under a limited number of protocols. In oncology, where multiple treatment protocols can yield divergent outcomes, addressing this issue is crucial. Here, we present a computational framework that simulates clinical trials through a combination of mathematical and statistical models. This approach offers a means to explore diverse treatment protocols efficiently and identify optimal strategies for oncological drug administration. We developed a computational framework with a stochastic mathematical model as its core, capable of simulating virtual clinical trials closely recapitulating the clinical scenarios. Testing our framework on the landmark SOLO-1 clinical trial investigating Poly-ADP-Ribose Polymerase maintenance treatment in high-grade serous ovarian cancer, we demonstrate that managing toxicity through treatment interruptions or dose reductions does not compromise treatment’s clinical benefits. Additionally, we provide evidence suggesting that further reduction of hematological toxicity could significantly improve the clinical outcomes. The value of this computational framework lies in its ability to expedite the exploration of new treatment protocols, delivering critical insights pivotal to shaping the landscape of upcoming clinical trials.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-024-80768-6.Keywords: Branching process model, Virtual clinical trials, High-grade serous ovarian cancer, PARP inhibitorsSubject terms: Ovarian cancer, Computational models, Predictive medicine, Software, Virtual drug screening, Mathematics and computing Introduction High-grade serous ovarian cancer (HGSC) is the most common and lethal subtype of ovarian cancer that is typically diagnosed at an advanced stage1,2. The cornerstone of therapy for newly diagnosed HGSC is cytoreductive surgery combined with platinum-based chemotherapy3. However, nearly all patients experience relapse with a tumor that develops resistance to conventional chemotherapeutic approaches4–6, resulting in poor overall survival4. Over the past decade, Poly-ADP Ribose Polymerase (PARP) inhibitors have revolutionized HGSC treatment, particularly in patients with germline or somatic BRCA1/2 mutations, or tumors with homologous recombination deficiency (HRD)5,6. PARP inhibitors are a group of pharmacological inhibitors of PARP enzymes, which play roles in transcription regulation, apoptosis, and DNA damage response. In HRD cells, the inhibition of PARP prevents the DNA repair process, disrupting cellular homeostasis and ultimately leading to cell death7. Approximately 50% of ovarian cancer patients with HRD and/or sensitivity to platinum-based chemotherapy are especially sensitive to PARP inhibitors8. Thus, HRD and platinum sensitivity are the primary biomarkers for PARPi response8,9. Olaparib was among the first PARPi that had favorable outcomes in the landmark SOLO-1 clinical trial as a first-line maintenance treatment after debulking surgery and platinum-based chemotherapy in patients with germline BRCA1/2 mutations10. Currently, three PARP inhibitors have been approved by the U.S. Food and Drug Administration (FDA) and European Medicines Agency (EMA) for the treatment of ovarian cancer: olaparib (in 2014)11, rucaparib (in 2016)12, and niraparib (in 2017)13. Despite recommendations for schedules and doses based on results from phase I/II–III clinical trials, not all potential treatment protocols can be clinically tested due to practical and ethical constraints. Therefore, different dosing schemes, treatment durations, and the effects of toxicities on clinical outcomes remain unknown. In addition, not all patients respond similarly to the treatment, and thus acquired PARPi resistance and patient population heterogeneity are emerging clinical problems. Thus, computational biologists have employed mathematical modeling to perform clinical trials in silico14. So-called virtual clinical trials are emerging for example in drug design to estimate the relative benefit of one dosing protocol over another in fast and non-invasive manner. The emergence of therapy resistance remains the biggest obstacle in improving long-term outcomes of cancer patients. Previous preclinical and translational work has suggested that PARPi resistance is associated with significant clonal, genomic, and functional heterogeneity15,16. However, because clinical biopsies are not typically performed upon PARPi resistance after first-line maintenance, the clonal dynamics and their clinical relevance are unknown. Thus, virtual trials based on computer simulations can provide valuable supporting evidence for the clinical testing of various hypotheses regarding novel therapeutic agents17,18. We have designed an advanced computational framework, which has been crafted for the purpose of exploring a range of drug treatment protocols that are not feasible to assess through traditional clinical trials. Here, we applied the computational framework to optimize the maintenance treatment with olaparib PARP inhibitor in HGSC. We calibrated the mathematical model, which is the core of the platform, on the SOLO-1 olaparib maintenance clinical trial19,20 and a phase II clinical trial of platinum-based chemotherapy21, and show that our model faithfully reproduces these clinical trials. Subsequently, we investigated the effects of treatment duration, optimal dosing, and management of hematological toxicity on long-term clinical outcomes. Our findings indicate that modification of dosing protocols, including dose reductions and interruptions for toxicity management, does not compromise patient responses. Notably, we provide evidence suggesting that further reduction of hematological toxicity could significantly improve clinical outcomes in first-line PARPi maintenance treatment for ovarian cancer. Developed within the MATLAB environment, our framework is freely available as open-source software. Its modular design enables easy adaptation for evaluating other targeted therapies. The inherent utility of this framework lies in its ability to streamline the exploration of novel treatment regimens, thereby offering valuable insights to inform the development of future clinical trials. Results An integrative computational framework for virtual clinical trials of 1st-line maintenance treatment of cancer patients We developed an integrated computational framework for conducting clinical trials on cancer patients, considering the following treatment modalities: surgery, chemotherapy, and maintenance treatment (Fig. 1). The platform includes a mechanistic model of cancer cell dynamics that is simulated numerically and a statistical model represented by a Kaplan–Meier estimator. Mathematical modeling framework for virtual clinical trials of 1st-line maintenance treatment of ovarian cancer patients. (A) Cancer cells in the model differ in the level of drug resistance to platinum-based chemotherapy (Pt-based CX) and targeted treatment with PARPi (olaparib). Resistance mechanisms can exist in a cancer cell either at diagnosis or develop during cancer treatment, with variably resistant clones formed through asymmetric division. (B) The model includes three types of treatment: surgery, chemotherapy (with platinum compounds), and targeted treatment with PARPi (olaparib). (C) Pharmacokinetics of chemotherapy and targeted treatment are modeled using a one-compartmental model. (D) To evaluate hematological toxicity, white blood cell count is included as a one-type branching process model. (E) The developed model was applied to simulate the SOLO-1 clinical trial by integrating the mathematical model with the Kaplan–Meier estimator. (F) We utilized the model to assess the most suitable method for managing adverse events with olaparib. (G) Tumor evolutionary dynamics of drug resistance to chemotherapy and PARPi were extracted to reveal features associated with response and resistance. (H) The first progression-free survival, and (I) the second progression-free survival from the SOLO-1 clinical trial and the model simulations. To evaluate the goodness of fit between our model and real-life PFS plots, RMSE was used as a statistical method. Our model shows an excellent fit to the real-life data. (J) The lowest value of white blood cells (WBCs) after each chemotherapy cycle (from clinical and virtual patient cohorts). (K) The tumor burden and the level of WBCs over time in a single exemplary virtual patient. The mechanistic model is a modified branching process model that is a discrete-time and discrete-state model. The model also includes a discrete-time model of olaparib and platinum-based chemotherapy pharmacokinetics and WBC dynamics with a logistic model. In the model, the tumor is initiated from a single malignant cell, normal WBC levels, without treatment perturbations. Cancer cells divide and mutate, resulting in a heterogeneous, well-mixed mass of cancer cells that can acquire resistance to chemotherapy (platinum-based) and PARP inhibitor both during the initial progression or during the treatments via the accumulation of (epi)genetic aberrations22 (Fig. 1A). Resistance to PARP inhibitors and platinum have been shown to exhibit significant patterns of cross resistance23,24, occurring commonly as a single step, by a reversion mutation, where a new mutation restores the open reading frame the original mutation25. The cancer treatment includes debulking surgery (modelled with removal of fraction of tumor cells at one time point) and 1st-line chemotherapy, followed by maintenance treatment with a PARP inhibitor (Fig. 1B). The pharmacokinetics of platinum-based chemotherapy and olaparib are represented by a one-compartmental model describing exponential drug elimination (Fig. 1C). Because the time interval from drug administration to the highest drug concentration in the peripheral blood (Tmax), is in hours, and simulations are on a day scale, we assume that Tmax = 0. Thus, we modeled intravenous bolus or peroral administration for carboplatin and olaparib, respectively. In the third layer, we included the hematological toxicity of platinum-based chemotherapy and olaparib, measured as the number of WBCs in the peripheral blood (Fig. 1D), as an indicator of myelosuppression, including thrombocytopenia and anemia, which are common side effects of chemotherapy21 and PARP inhibitors13,22,26. Thus, we modeled the dynamics of WBCs using a discrete-time logistic model. To conclude, the model is composed of three layers such as (1) cancer dynamics, (2) drugs pharmacokinetics, and (3) WBC dynamics that are simulated concurrently. The mathematical model developed herein serves as the foundational framework for conducting virtual clinical trials in cancer. Through simulations, we acquire data on the long-term response of virtual patients, represented by the first and second progression-free survival (PFS). This data is then utilized by a statistical model known as the Kaplan–Meier estimator (Fig. 1E). Moreover, in addition to insights into patient long-term response, the model provides information about tumor evolutionary dynamics during PARPi maintenance treatment (Fig. 1G). This information can be further investigated using Cox regression (and other statistical or machine learning models), which utilize survival data to predict the risk of events such as patient relapse (as demonstrated here) or death highlighting the generalizability of our model. Mathematical modeling framework faithfully reproduces the clinical outcomes of the SOLO-1 trial First, we calibrated the model to match the clinical outcomes (the first and the second PFS) and toxicities observed in the clinical trials applied for model calibration. The effects of surgery and chemotherapy on the cancer cells were adjusted27 to the pharmacokinetics of administered drugs and clinical outcomes. The toxicity parameter during chemotherapy was adjusted to the lowest level of WBCs after each cycle and the time when the lowest WBC value was observed (nadir), as adapted from a clinical trial21. Olaparib toxicity was adjusted according to the reported rate of 40% of patients experiencing grade 3 hematologic toxicity that led to treatment interruption or dose reduction in the SOLO-1 trial13,26. Our model showed an excellent fit to the first PFS and the second PFS (Fig. 1H and I) of the SOLO-1 clinical trial, with a low root mean squared error (RMSE) serving as a metric for model fitness. It is the first model to our best knowledge that is able to recapitulate not only the response to the first-line line treatment but also the second line. In addition, the hematological toxicity in form of the WBC level is also calibrated and shows excellent fit to the reported toxicities (Fig. 1J). Interestingly, real-world data suggest that the responses to platinum-based therapies in the second line have been remarkably poor in previously PARPi-treated patients28. We then explored how the second PFS to platinum-based therapies of the model fit real-world data. Our observations indicated that our model reproduces the inferior responses to chemotherapy after PARPi maintenance observed in real-world data (Fig. 1I). By correctly predicting the results it was not fitted for, our model proofs its usefulness as a tool for prediction of the therapy outcome. Consequently, we successfully reproduced the clinical course, hematological toxicity, and tumor dynamics of virtual ovarian cancer patients undergoing 1st- and 2nd-line treatment with surgery and platinum-based chemotherapy followed by PARPi maintenance (Fig. 1K). Thus, the herein-developed mathematical framework can be used to statistically evaluate different treatment schemes and tumor dynamics to optimize the treatment. Cancer cell dynamics and heterogeneous patient populations suggest fitness, selection advantage, and chemosensitivity as determinants of PARPi response To investigate cancer cell dynamics in a large patient population, we simulated 10,000 virtual ovarian cancer patients undergoing primary debulking surgery and adjuvant chemotherapy followed by PARP inhibitor maintenance therapy. Interestingly, we observed four distinct patterns of tumor dynamics in the patient groups (Fig. 2A–E). The first and largest patient group (Fig. 2A, “eradicated”) is characterized by a rapid decrease in sensitive and resistant cell populations during cytoreductive surgery and chemotherapy. In this patient population, only sensitive cells remained at the beginning of PARPi treatment, ultimately leading to the complete eradication of cancer cells. The second-largest patient population (Fig. 2B, “heterogeneous”) is represented by partial responses to 1st-line treatment and rapid progression of the tumor during the PARPi maintenance treatment, resulting in a heterogeneous tumor with fully or partially resistant cells at relapse. The third-largest population (Fig. 2C, “fully resistant”) was characterized by a fast emergence of a fully resistant population right after finishing chemotherapy. Moreover, we observed a relative overgrowth of resistant cells during PARPi treatment, but still, the presence of a sizeable fraction of sensitive cells at relapse after 2 years of PARPi maintenance treatment. The fourth patient population (Fig. 2D, “sensitive”) is characterized by a slow progression of sensitive cells leading to relapse after a relatively long PFS > 4 years. Even though there is a minority of resistant cells at relapse, these patients, at least partially, would likely respond to a second-line platinum-based chemotherapy. PARPi maintenance treatment induces heterogeneous cancer cell dynamics and patient populations with distinct clinical outcomes. (A) eradicated, (B) heterogenous, (C) fully resistant, and (D) sensitive. (E) Proportions of each distinct clinical outcome in a virtual patient cohort. (F) Kaplan–Meier estimator for the first progression-free survival of each of the four patient types. (G) Model parameters associated with the response and resistance to PARPi. Univariate Cox regression for features related to tumor dynamics, with the c-index as a metric of model performance. Features within the rectangle show good predictive performance. Interestingly, the Kaplan–Meier plot of the different patient populations revealed distinct responses to PARPi maintenance treatment among the patient groups (Fig. 2F). As expected, the “eradicated”, and “sensitive” had an extremely favorable prognosis, whereas the “fully resistant” had an intermediate , and the “heterogeneous” patient group had the shortest PFS to 1st-line olaparib maintenance. For the fully resistant patients, this can be potentially explained by the initial presence of sensitive cancer cells and their gradual replacement by the fully resistant cells at relapse, whereas different partially resistant cell populations arise earlier in the “heterogeneous” patient group. To further investigate which model features are most significantly associated with PFS on PARP inhibitors, we extracted the outputs of the model simulations that describe tumor growth and heterogeneity (Fig. 2G) and applied the simulations with each feature, followed by the Cox regression model with tenfold cross-validation. Interestingly, seven parameters were significantly associated with PFS, indicating that PFS is multifactorial, and tumor intrinsic resistance mechanisms do not define PFS as a single feature. The fitness of the cancer cells (i.e., selection advantage, which defines how fast the cancer cells are growing in comparison with other cells in the tumor) and selection under olaparib were the two most significant features associated with a shorter PFS. In addition, the WBC levels at diagnosis, time of maintenance treatment, and discontinuation of PARPi treatment due to toxicity were also associated with shorter PFS. By contrast, the parameters significantly associated with prolonged PFS were the response to first-line chemotherapy (measured as the fraction of cancer cells killed by chemotherapy), and reduced lowering of the white blood cell levels as an indicator of lower hematological toxicity. Drug dosing and toxicity management affect clinical outcomes of PARPi maintenance treatment In the design of clinical trials, the maximum tolerated dose of the drug from phase I–II is typically utilized in phase III to investigate the efficacy of clinical outcomes. However, the impact of different drug doses and dose reductions on clinical outcomes is not well understood. Therefore, we set out to investigate the effect of different starting doses of olaparib as maintenance treatment using our integrative computational platform. We performed a simulation of 10,000 virtual HGSC patients treated with a wide range of olaparib doses between 100 to 400 mg twice a day, and the median PFS was computed (Fig. 3A). We observed significant increases in PFS times with olaparib starting dose increases between 200 to 250 and 250 to 300 mg and a drop in doses higher than 300 mg twice a day (Fig. 3A), suggesting that the toxicities included in the model can result in inferior clinical outcomes with doses over 300 mg bi-daily. Effects of PARPi dosing, duration, and toxicity management on PFS. (A) The first PFS as a function of (A) olaparib dose and (B) treatment duration. (C) Effect of dose reduction on the first PFS. (D) Effect of the duration of treatment interruption on the first PFS. (E) The first PFS, as a function of the toxicity level, measured as a percentage of patients requiring toxicity management. (F) The Kaplan–Meier estimator for patients with higher-than-normal levels of PARPi toxicity (60%), olaparib (40%), and the second-generation PARPi (30%). Asterisks indicate statistical significance compared to the previous dose with p < 0.05 (t-test). Next, we investigated the optimal duration of maintenance treatment and observed that the optimal treatment duration with olaparib was 27 months (Fig. 3B). Notably, no additional benefit was obtained from prolonged olaparib maintenance beyond 27 months. We then examined the dose reductions from the starting dose of 300 mg twice a day after toxicity and noted that no differences were observed in PFS, indicating that dose reduction due to hematological toxicity, even down to a 50 mg bi-daily dose, does not impair the efficacy of PARPi maintenance (Fig. 3C). To investigate the effect of treatment interruptions, we included treatment breaks from 1 to 10 weeks in the model. In our analyses, the interruption of treatment in that range does not decrease or increase the first PFS (Fig. 3D). Recently, more specific 2nd generation PARP inhibitors have been introduced with lower toxicity profiles and potentially better-tolerated selective targeting of the PARP-1 enzyme29. Lastly, we investigated how the administration of PARPis with similar potency but different toxicity levels affects PFS. We observed a significant shortening of PFS with toxicity levels higher than 40% (Fig. 3E). Importantly, the reduction of toxicity to a level where 30% of patients undergo at least one form of toxicity management, such as with the recently developed 2nd generation PARP1 specific inhibitors, led to a significantly prolonged PFS (log-rank test p-value < 0.0001) (Fig. 3F). Thus, our results suggest significant clinical benefits of the novel, less toxic PARP inhibitors, and highlight the importance of the inclusion of systems-level toxicities in mathematical cancer therapy models. Discussion PARP inhibitors have yielded significant clinical benefits in the first-line maintenance treatment of high-grade serous ovarian cancer. However, the widespread clinical adoption of PARPi maintenance has given rise to several unanswered questions regarding this therapy and patient management. These questions encompass the optimal duration of olaparib administration, the appropriate dosage scale of the medication, and whether the measures taken to manage toxicity impact the clinical benefits derived from the treatment. Additionally, the dynamics of tumors during PARP inhibitor maintenance remain unknown, posing challenges post the development of therapy resistance. To address these uncertainties, we have developed a computational platform comprising both a mechanistic model and statistical analyses. By employing a virtual clinical trials approach, we demonstrate that heterogeneous tumor dynamics and toxicity management significantly influence clinical outcomes. Our findings underscore the efficacy of virtual clinical trials in optimizing treatment strategies and provide valuable insights for guiding the design of future clinical trials. Even though olaparib has shown efficacy in patients with BRCA1/2 mutated/HRD-positive tumors, not all patients respond uniformly well. This heterogeneity is apparent even in clinical trials, where close to 30% of the patients relapse during PARPi maintenance due to emerging PARP inhibitor resistance. Numerous resistance mechanisms to PARP inhibitors have been identified31. Importantly, resistance to platinum-based chemotherapies is a strong predictor of PARPi resistance, suggesting shared underlying mechanisms30. In general, a secondary mutation in BRCA1/2 restoring the protein function is the best clinically characterized mechanism of platinum agent and PARPi resistance in HGSC31–35. In addition, several non-genetic mechanisms have been identified, including the downregulation of DNA repair proteins16,38,39 and enhanced replication fork protection36–39. Interestingly, previous work has suggested that PARPi resistance is associated with significant clonal, genomic, and functional heterogeneity16. In alignment with the clinical observations regarding the inferior 2nd PFS in the clinical setting, we observed inferior responses to subsequent therapy after relapse on PARPi treatment, confirming that our mathematical model adeptly reproduces the cancer evolution biology and clinical behavior of overlapping PARPi and platinum therapy resistance mechanisms. Our study presents cancer cell fitness, chemotherapy response, and selection pressure as predictive markers of the response to olaparib, forming a robust foundation for the development of predictive combination biomarkers. Furthermore, our model emphasizes the role of hematological toxicity parameters as determinants of clinical benefit in first-line maintenance treatment with PARP inhibitors. Olaparib is generally well-tolerated, however a substantial number of patients report mild or moderate side effects, including fatigue, nausea, and other gastrointestinal symptoms. Many of these issues can be effectively managed with supportive therapy or dose modifications and do not require pausing the treatment. The most prevalent grade 3–4 adverse effects include anemia and neutropenia36, frequently necessitating treatment interruption and dose reduction. In the SOLO-1 trial, adverse effects led to treatment interruption in 51.9% of olaparib patients, dose reduction in 28.5%, and treatment discontinuation in 11.5% of patients19. Our findings show that treatment interruptions or dose reductions did not compromise the maintenance treatment outcomes and that dose reductions, even to low doses, should be prioritized over treatment discontinuation. This aligns with previous studies concerning olaparib interruptions in the recurrent setting40. Our data further support that reductions and interruptions do not compromise the clinical outcomes, providing valuable insights into clinical decision-making and patient counseling concerning toxicity management. Importantly, we show using our computational platform that the use of PARP inhibitors with reduced toxicity results in improved progression-free survival, thereby supporting the development of next-generation molecules with reduced toxicity profiles for maintenance treatments. Noteworthy, toxicity management data from SOLO1-trial is not publicly available and therefore we were not able to mimic it in our trial. In our model, the proportion of patients who had to discontinue the treatment is apparently higher than in real-life setting, due to the fact that we could not perform repeated dose reductions. The use of mathematical models and virtual patient concepts has been used many times in the past. For example, Gatenby et al. in41 presents an alternative approach to anti-cancer therapy in which treatment is so modulated to achieve desired tumor population i.e. to keep sensitive cells population big enough to suppress the growth of the resistant ones instead of completely remove the cancer cells. Chmielecki et al.42 uses evolutionary mathematical cancer models to find therapeutic strategies which delay the development of resistance. Beckman et al. in43 uses a mathematical model of personalized therapy to find augmented, nonstandard strategies of therapy which may lead to superior patient outcomes compared with standard methods. Leder et al.44 optimized delivery schedules of radiotherapy to overcome tumor heterogeneity and its radioresistance. Yeang and Beckman45 showed how the virtual patients may be used to predict the possible outcomes of multistep therapies and thus allows for optimal strategy choice. Our novel computational platform considers not only the response to drugs but also drug resistance and toxicity. The platform is generic and freely available, and thus, with minimal changes, it could be applied to test other anticancer drugs, facilitating the design of new clinical trials. However, the platform is not without limitations. The branching process model does not include the interaction or spatial relationships between cancer cells and all other cells in the tumor due to computational constrains and the inherent challenges in quantifying these interactions at the patient scale. Furthermore, the simulator operates under the assumption that oncogenic cells are synchronized, meaning they concurrently undergo apoptosis and mitosis. This assumption is pivotal in enabling the simulation to deliver results within a practical timeframe. Addressing these limitations constitutes another aspect of our ongoing research and development agenda. We acknowledge that the p-value given in the results may not be the best indicator of the statistical significance of the results. This is due to the results being based on computer simulations, while p-value tends to decrease in value with increasing number of samples. Thus, by increasing the number of simulated patients, we decrease the p-value. For this reason, use other indicators of statistical significance should be preferred46. One have to remember also that the model parameters marked as “calibrated” may not reflected the actual rate of the described processes. For example the mutation rate in the presence of Olaparib follows the general knowledge that PARP inhibitors increase chromosomal instability and mutagenesis in BRCA deficient cells47 but the range of the change of mutation rate of the particular genes responsible for cis-platinum and Olaparib resistance is not known, thus the real rate may be different that provided by us. The model sensitivity to the change of the “calibrated” parameters is provided in the supplementary materials Text S8. Altogether, our findings underscore the efficacy of virtual clinical trials that include modeling of resistance mechanisms, tumor dynamics, and patient toxicities in optimizing treatment strategies and provide valuable insights to guide clinical treatment and future clinical trial design. Materials and methods SOLO-1 clinical trial data SOLO-1 is a Phase III randomized, double-blinded, placebo-controlled clinical trial designed to evaluate the efficacy and safety of olaparib tablets (300 mg twice a day) as a maintenance monotherapy compared with a placebo. The admission criteria for the trial were as follows: (1) newly diagnosed advanced BRCA1/2 mutated ovarian cancer, (2) the patient must be in complete or partial response following platinum-based chemotherapy, and (3) the patient must have normal bone marrow function. In total, 391 patients were enrolled in this trial. Patients were randomized to receive olaparib or a placebo for 2 years or, in the case of relapse, until disease progression. The primary endpoint of the trial was progression-free survival, and the toxicity profile was also assessed. In our study, the mathematical model was fitted to data on progression-free survival rates (depicted as a Kaplan–Meier plot) from the SOLO-1 clinical trial. In addition, the hematological toxicity profiles of olaparib (from the SOLO-1 trial) and platinum-based chemotherapy21 were extracted to fit the model parameters to include hematological toxicity in the model. Stochastic mathematical model We developed a mechanistic model consisting of three layers. The first layer is the cancer cell dynamics. The second is the pharmacokinetics of the drugs, and the last one is the hematological toxicity of the drug measured by the number of white blood cells (WBCs). We included cancer cells as a heterogeneous, well-mixed mass differing in the level of resistance to platinum-based chemotherapy and olaparib. During tumor growth, both with and without treatment, the accumulation of (epi)genetic aberrations results in cancer cells gaining new resistance mechanisms. Each cell is defined by two numbers, where the first and second ones denote the number of resistance mechanisms to carboplatin and olaparib, respectively. For example, a cell defined as “1,3” has one and three resistance mechanisms to carboplatin and olaparib, respectively. The number of resistance mechanisms that could accumulate in cancer cells is not restricted and can go to infinity. The only limitation is the tumor burden reaching a level where symptoms become clinically evident, and the patient relapses. It is worth to notice that in our model the drug resistance is not a direct result of the exposure to the drug but it is a result of stochastic events of the emerging mutations during cell division which make drug resistance parameters not dependent on the pharmacokinetic ones. The second layer, describing the pharmacokinetics of carboplatin and olaparib, is included as a one-compartmental model depicting exponential drug decay. Because the time interval from drug administration to the highest drug concentration in the peripheral blood (Tmax), is in hours, and simulations occur on a daily scale, we assume that Tmax = 0. In such a way, we modeled intravenous bolus intravenous or peroral administration for carboplatin cisplatin and olaparib, respectively. Subsequently, we included the hematological toxicity of platinum-based chemotherapy and olaparib in the model. Because hematological toxicity is associated with a low level of WBCs, we modeled the dynamics of WBCs throughout treatment and treatment intervals. In the absence of treatment, the number of WBCs is at dynamic equilibrium. During treatment, however, the level of WBCs decreases. We modeled WBC growth as a density-limiting growth using a logistic model. This allows WBCs to oscillate around the equilibrium in the absence of treatment and the decay of WBC during treatment intervention, as explained in detail in Supplementary Text 5. A computer simulator of HGSC patient course We developed a simulator for the course of HGSC patients from cancer initiation to the second recurrence. The computer simulations include the following phases: (1) pre-diagnosis, (2) primary treatment phase, (3) maintenance treatment phase, (4) post-treatment phase, (5) second-line treatment, and (6) the second post-treatment phase. During the pre-diagnosis phase, the tumor grows from a single sensitive cell (“0,0”) without any perturbations. During tumor growth, there is a probability that a cell will gain one resistance mechanism to one of two drugs during cell division. During this simulation phase, the tumor evolves according to the neutral evolution law. This phase concludes when the tumor reaches cancer cells, which reflects the time of diagnosis. We assume that treatment starts at the time of diagnosis without any loss of generality. The primary treatment phase starts with upfront cytoreductive surgery (primary debulking surgery—PDS). Surgery is modeled by the removal of a fraction of cancer cells at a one-time point, followed by a 3-week recovery period and platinum-based chemotherapy. Chemotherapy is administered as an intravenous injection (bolus IV). Each patient in the virtual cohort undergoes 6 cycles of platinum-based chemotherapy, with the median interval between them equal to 3 weeks. Maintenance treatment with olaparib starts after the WBC level is within the normal range after chemotherapy. Patients then receive olaparib as peroral administration twice a day. This phase is simulated for 2 years, until relapse or if toxicity becomes too high, depending on which time is shorter. Subsequently, the patient is simulated in a follow-up until relapse, if relapse does not occur during maintenance treatment. During this period, the tumor grows without perturbation as the patient is asymptomatic, and the tumor remains undetectable. The phase concludes with the first relapse. The subsequent phase is the second-line platinum-based chemotherapy. Similar to the primary treatment, patients receive 6 cycles of chemotherapy at a 3-week interval. The final phase of the simulation is a follow-up after the second-line treatment. Here, cancer cells can grow without perturbation, reflecting the time when the patient is in clinical remission or receives palliative treatment, depending on the chemotherapy response. This phase concludes with the second relapse. The main output of the simulation is a response rate after the first- and second-line treatments, measured as progression-free survival and toxicity level to platinum-based chemotherapy and olaparib. The simulator is freely available on GitHub at https://github.com/EmiliaKo/VCT_olaparib. Mathematical model calibration using SOLO-1 clinical trial and toxicity profile data The developed mathematical model was fitted to two main types of data: SOLO-1 clinical trial data (comprising treatment protocol and response rate in the form of progression-free survival) and the toxicity profile of carboplatin and PARPi. A detailed description of model calibration and a list of all parameters along with their values are provided in Supplementary Text 7 and Supplementary Table S4, respectively. In brief, we initially set all model parameters that could be directly extracted from clinical data as constant values. Subsequently, we fitted toxicity parameters to align with the toxicity profile of the two drugs included in the model (carboplatin and olaparib). In the final step, we fitted the remaining parameters to ensure that the Kaplan–Meier plot produced from the model simulations agrees with the one obtained from the SOLO-1 clinical trial. Because of the stochastic nature of the model pure automatic fitting was not possible. Thus, manual adjustment was made with least-squares as a calibration criteria. A detailed description of the model fitting process is presented in Supplementary Text 7. After the model fitting, sensitivity analysis was performed. It is results show that model in general is not sensitive on the change of the parameters related to drug-resistance gain/loose parameters but it is sensitive on the change of the parameters related to drug’s effectiveness and toxicity. Detailed description of the sensitivity analysis performed and its results is presented in Supplementary Text 8. Tumor feature extraction and Cox regression We extracted the outputs of the model simulations that describe tumor dynamics and heterogeneity. Subsequently, we performed Cox regression on a cohort of 10,000 virtual patients, utilizing all extracted features (see Supplementary Text 11 and Supplementary Table S6). As a metric of Cox regression performance, we applied the concordance index (c-index). We performed the analysis only in the univariate scenario using tenfold cross-validation. Treatment duration optimization In the optimization of treatment duration, we simulated 10,000 virtual HGSC patients under a given condition and calculated the median PFS. For maintenance duration, we initially conducted a simulation without any maintenance treatment to establish a baseline. Subsequently, we incrementally increased the duration of olaparib maintenance by 1 month at a time. We assessed the impact of the extended duration on PFS and continued extending the treatment period in the event of a significant increase in PFS until no further difference was observed. Similarly, we optimized the olaparib dose. Toxicity management We also investigated the performance of a new generation of PARPi that is less toxic by reducing the parameter responsible for decreasing WBC level during olaparib administration. The toxicity level was assessed by calculating the percentage of virtual patients who developed hematological toxicity to olaparib. Importantly, we incorporated three methods for dealing with olaparib toxicity: dose reduction, dose interruption, and treatment discontinuation. From real-world data, we estimated that a total of 40% of the patients had grade 2–3 hematological toxicity, resulting in one or several forms of toxicity management in SOLO-126. However, real-world data were not available on the proportions of these patients who underwent more than one toxicity management procedure. Thus, in the model, upon hematological toxicity, each patient would undergo only one toxicity management, and after repeated toxicity, the treatment was discontinued. A detailed description of the methods to overcome olaparib toxicity is provided in Supplementary Text 9. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Calculations were performed on the Ziemowit computer cluster in the Laboratory of Bioinformatics and Computational Biology, created in the EU Innovative Economy Programme POIG.02.01.00-00-166/08 and expanded in the POIG.02.03.01-00-040/13 project. This work was carried out in part by the Silesian University of Technology internal research funding. The work was supported by grants from the Sigrid Jusélius Foundation (A.F.), Cancer Society of Finland (A.F.), Academy of Finland (grant number 339805, 350396 to A.F.), The Finnish Medical Foundation (A.F.), University of Helsinki (A.F.), and Helsinki University Hospital state-level funding (A.F.). This study was co-funded by the European Union (ERC, SPACE, 101076096). Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. The data generated in this study are available within the article and its supplementary data files. Author contributions Conceptualization was done by A.F. (biological part) and K.P. (software and simulation part); Methodology developed by A.F. (biology) and K.P with E.K (software and simulation); Software was writen by E.K. with the help of K.P. Formal analysis and investigation was done by A.F. and U.H. in biological part and K.P. and E.K. in software and simulation part. E.K. was responsible for resource gathering and data curation. A.F. (biology) and K.P. (software and simulation) validated the results and supervise the project. The writing of the main manuscript, supplement and figure preparation was done by all authors who also reviewed the final manuscript. Funding acquisition was done by A.F. Data availability The framework is implemented in MATLAB environment and available as open-source at: https://github.com/EmiliaKo/VCT_olaparib. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Krzysztof Puszynski, Email: krzysztof.puszynski@polsl.pl. Anniina Färkkilä, Email: anniina.farkkila@helsinki.fi. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The framework is implemented in MATLAB environment and available as open-source at: https://github.com/EmiliaKo/VCT_olaparib.