Background
The haematopoietic growth factor G-CSF is routinely used in cancer therapy to prevent or ameliorate leukopenic conditions. Its effectiveness has been shown in several studies (Kosaka et al.
2015; Lee et al.
2013; Vogel et al.
2005; Altwairgi et al.
2013; Dale
2002,
2003; Kuderer et al.
2007; Crawford et al.
1991; Bohlius et al.
2008; Sung et al.
2007; Cooper et al.
2011; Mhaskar et al.
2014; Clark et al.
2003,
2005). Although G-CSF is expensive, its application often results in an overall cost-reduction due to the reduced number of severe events (Zagonel et al.
1994; Wang et al.
2016).
With the introduction of G-CSF support, more intense chemotherapies became feasible in order to improve outcome of patients (Trumper et al.
2008; Untch et al.
2011a,
b; Pettengell et al.
1992; Pfreundschuh et al.
2004a,
b,
2008; Diehl et al.
2003; Sieber et al.
2003; Blayney et al.
2003,
2005). A number of G-CSF pharmaceuticals are in use differing in both, pharmacokinetic and pharmacodynamic properties (Kuwabara et al.
1994,
1996a,
b; Yang et al.
2004; Zamboni
2003; Molineux
2002; Houston et al.
1999). Various generics are available or under development.
Several in vivo modes of action of G-CSF are known, namely increased proliferation, accelerated maturation and improved release of mature bone marrow granulopoietic cells (Lord et al.
1989; Schmitz et al.
1993). In combination with the relatively short half-life of blood granulocytes and the bone-marrow suppressive effects of cytotoxic chemotherapy, application of G-CSF results in complex dynamics of blood granulocytes which cannot easily be predicted. As a consequence, optimal G-CSF support for a given chemotherapy and patient population is a non-trivial task. It depends on a large number of variable therapy parameters such as the type of cytotoxic drugs, granulotoxic risk factors of patients, type of G-CSF derivative applied and its dosing and timing (Bennett et al.
2013).
In clinical trials, it is practically impossible to control for each of these factors. Therefore, only limited attempts were made to compare the efficacy of different G-CSF schedules in the context of clinical trials (Danova et al.
2009; Holmes et al.
2002; Loibl et al.
2011; Lyman et al.
2009; Vose et al.
2003; Zwick et al.
2011; Faber et al.
2006; Crawford et al.
1997; Leonard et al.
2015). However, available clinical trials showed that considerable improvements can be expected by optimized G-CSF schedules. Since such trials are both, cost and time-intensive, there is relevant need to predict the outcome of alternative G-CSF schedules prior to clinical application. On the basis of large clinical and experimental data sets, we developed a comprehensive biomathematical model of human granulopoiesis including detailed information on injection, pharmacokinetics and pharmacodynamics of both, chemotherapeutic drugs and three G-CSF derivatives namely filgrastim, pegfilgrastim and the experimental drug Maxy-G34 (Scholz et al.
2005,
2009a,
2009b,
2012; Chua et al.
2014; Engel et al.
2004; Schirm et al.
2013,
2014b). The model was validated in several settings and is now ready to make clinically relevant predictions regarding G-CSF schedules optimized for given chemotherapeutic regimens.
In this paper, we present our approach for developing optimized dosing and timing schedules of G-CSF for a variety of applications, i.e. for different chemotherapy schedules, risk groups of patients and usage of filgrastim or pegfilgrastim. Different measures of treatment outcome are considered. We also show examples of model predictions validated in the context of clinical trials.
Discussion
Although the haematopoietic growth factor G-CSF is routinely applied in clinical practice since many years, its optimal use in a given clinical situation is often unknown or not well investigated. The reason is that the performance of alternative G-CSF schedules is difficult to predict in view of the strong interaction of chemotherapy-induced leukopenia, pharmacokinetic properties of G-CSF and the resulting effects on bone marrow leukopoiesis.
In view of the large number of variable therapy options (dosing and scheduling of cytotoxic drugs and G-CSF, different G-CSF pharmaceuticals, individual risk factors of patients), it is practically impossible to study this problem solely on the basis of clinical trials. Thus, there is a strong need for predictive modelling of G-CSF applications. Pastor et al. (
2015) proposed a statistical model, while Quartino et al. (
2014) proposed a semi-mechanistic model for this purpose. Craig et al. (
2015) used their granulopoiesis model to explore alternative filgrastim schedules for general 14-day chemotherapy cycles. Here we propose to use our recently established biomathematical model of human granulopoiesis under G-CSF and chemotherapy treatments to address this task. Our model is based on biological assumptions on bone marrow haematopoiesis, PK and PD effects of G-CSF injections and the cytotoxic effects of chemotherapy.
The model was developed on the basis of large clinical and literature data sets (Schirm et al.
2013,
2014a,
b). To apply the model, it is necessary to estimate the bone marrow toxicity of an applied cytotoxic drug or drug combination, which can be achieved by studying time-series data of patients treated under this condition. By this approach, we were able to quantify bone marrow toxicities of a total of 10 drugs and 33 schedules (Schirm et al.
2014b). After quantifying the toxicity, the performance of alternative G-CSF schedules can be simulated by the model. We already applied this method in the planning phase of a number of clinical trials. Data collected under the newly proposed schedules showed that our predictions are in good agreement with the data. In view of these encouraging results, we propose additional optimized G-CSF schedules here. The proposed schedules are intended to be verified in clinical trials.
Different G-CSF derivatives are in practical use. Here we focused on filgrastim and pegfilgrastim which are generally considered as equally potent to prevent leukopenia if properly applied. A few studies and meta-analyses indicate advantages for pegfilgrastim (Clark et al.
2003,
2005; Cooper et al.
2011; Mhaskar et al.
2014; Lambertini et al.
2015). Indeed, pegfilgrastim can be applied more easily increasing compliance. But filgrastim can be dosed more precisely allowing individual adaptations. This is especially relevant for risk-adapted G-CSF treatments. Moreover, it is supposed that the amount of pegfilgrastim injected by a single standard syringe might be too high for some patients (Ishiguro et al.
2008; Djulbegovic et al.
2013; Masuda et al.
2015). Therefore, going beyond pure variation of starting time of pegfilgrastim, we also considered scenarios with reduced dosage of pegfilgrastim.
We studied different outcomes to assess the resulting cytotoxic outcome of a schedule, namely WBCAOC, DoL and MLC. Pros and cons of these outcomes are discussed elsewhere (Scholz et al.
2006) and we propose WBCAOC as the most reasonable choice. This allows us to compare different G-CSF schedules with respect to their expected cytotoxic outcome, and finally, to optimize the schedules. The relationship between the degree of chemotherapy-induced leukopenia and resulting risk for infections is well-established (Colotta et al.
1992; Bennett et al.
2013; Li et al.
2016).
As practical applications of our model, we considered for example different starting times of pegfilgrastim for the adjuvant breast cancer chemotherapy ETC in the patient population studied in Moebus et al. (
2010). According to our simulations, we predict that the application at d4 after chemotherapy is superior to d2 and that d6 is optimal. However, the differences are small. Moreover, the strongest leukopenic risk is expected for the cycles with cyclophosphamide applications. Here, the nadir occurs in a narrow time interval which might be difficult to capture in a clinical trial. This could explain the results of Loibl et al. (
2011) who observed a (non-significant) trend towards better performance of the d4 schedule compared to d2.
For the BEACOPP escalated regimen to treat advance stage Hodgkin’s lymphoma (Diehl et al.
2003), we predict that pegfilgrastim is optimally applied at d6–7 after chemotherapy. However, this would still fall into the period of procarbacine treatment.
We also propose optimized filgrastim treatment for three scenarios: for BEACOPP escalated, we predict that starting 1 day earlier and increasing the number of G-CSF injections would result in improved leukopenia prophylaxis. For the ETC chemotherapy mentioned above, we predict that filgrastim d7–14 after chemotherapy is clearly superior to the current standard d4–11. But since the haematotoxic risk clearly depends on the applied drugs, we also considered different filgrastim schedules for cycles 1–3 (epirubicin), 4–6 (paclitaxel) and 7–9 (cyclophosphamide), respectively. However, only small improvements were predicted compared to the d7–14 schedule.
Since time-intensified CHOP is advantageous for the treatment of high-grade non-Hodgkin’s lymphoma in elderly patients (Pfreundschuh et al.
2004a; Roesch et al.
2014; Rosch et al.
2016), we designed a regimen with six cycles of CHOP repeated every 12 days (CHOP-12). We predict that with optimal filgrastim support at d7–12, the toxicity might be tolerable but slightly cumulates over six cycles.
Another application of the model is the development of risk-adapted G-CSF schedules as recommended (Kuderer et al.
2006; Georgala and Klastersky
2015). This is achieved under the assumption that risk groups differ in sensitivity to chemotherapeutic drugs rather than response to G-CSF treatment (Chatta et al.
1994). We established a statistical model of leukopenia risk, depending on pre-therapeutical (i.e. age, sex) and intra-therapeutical (observed toxicity in first cycle) risk factors for patients of high-grade non-Hodgkin’s lymphoma in the past (Ziepert et al.
2008). However, the risk score did not result in recommendations regarding individualized G-CSF regimen so far. We addressed this issue in our paper by dividing patients into tertiles for which we propose specific G-CSF schedules. Indeed, we could detect some potential for risk-dependent filgrastim treatment: For the optimal schedules, number of filgrastim injections differed between four for the low-risk group, six for the medium-risk group and eight for the high-risk group. No optimization potential was detected for single pegfilgrastim injections (optimum d6–7 after chemotherapy for all risk groups). This approach can be generalized to other therapy schedules for which a leukopenia risk score is available.
As a general recommendation observed throughout our scenarios, we conclude that filgrastim and pegfilgrastim treatment should not be started too early after chemotherapy. The major reason is that G-CSF releases the bone marrow reserve of granulocytes which should be avoided if the number of granulocytes is still sufficiently high. However, this might be applicable only for intense chemotherapies with a high risk of leukopenia (Whitworth et al.
2009; Cheng et al.
2014). Moreover, filgrastim should not be stopped too early. Even if granulocytes are recovered, we predict a benefit of maintained G-CSF treatment in the subsequent chemotherapy cycle. According to our model simulations, we also expect that there is some potential to reduce the dose of single pegfilgrastim injections without loss of efficacy. However, this prediction must be considered with caution since it is based on extrapolation of absorption kinetics.
A limitation of our method is that we only consider the number of leukocytes and not the clinically more relevant outcome of infection. Although there are strong relationships between leukocyte counts and risk for infection (Bennett et al.
2013; Colotta et al.
1992), our method does not account for leukocyte function or other measures to prevent infections such as prophylactic antibiotic treatment or hospitalization. Another limitation is that we optimized G-CSF therapy for the medians of patient populations or risk groups while patient extremes are most relevant. We aim at addressing this issue by modelling individual time courses in the future.