Development of an Agent‐Based Model to Investigate Durability of Factor IX Activity in Hemophilia B Patients Treated With Etranacogene Dezaparvovec

ABSTRACT

Many currently approved gene therapies use adeno‐associated virus (AAV) to deliver DNA sequences encoding protein(s)‐of‐interest into cells. The AAV viral genome forms stable, circular DNA structures called episomes after entering the nuclei. Therapeutic proteins are then generated in vivo from transcription and translation of these episomes, and long‐term durability thus depends on episome stability. Prior modeling work has utilized differential equation‐based models to characterize AAV uptake and subsequent protein production. However, episome loss associated with target cell turnover is poorly described with these models. Here, an agent‐based model (ABM) to overcome this shortcoming is developed. The liver was used as the example organ as it is known to be self‐renewing and has been a common target for gene therapies. In this model, each hepatocyte is an agent, capable of division and death. When transduced, these agents acquire and carry episomes. During cell division, episomes are passed from mother to daughter cells. All episomes are presumed lost when transduced cells die. The ABM was applied to etranacogene dezaparvovec (formerly AMT‐061 or CSL222), a liver‐targeting gene therapy for hemophilia B consisting of AAV serotype 5 particles encoding a transgene for the Padua variant (R338L) of coagulation factor IX (FIX). ABM‐simulated FIX activity in patients receiving this therapy was consistent with clinical observations over more than 2 years following treatment. This ABM was then used to generate 20‐year predictions and explore biological mechanisms underlying long‐term durability. The ABM framework and approach should be applicable to investigating durability of other AAV‐based gene therapies.

Study Highlights

• What is the current knowledge on the topic?

  • ○
    There have been many attempts to treat hereditary diseases with gene therapies. Despite positive clinical data, long‐term durability of nonintegrating gene therapy targeting self‐renewing organs remains hotly disputed, with no published model to address this knowledge gap.
• What question did this study address?

  • ○
    Here, we developed an agent‐based model (ABM) to investigate the long‐term profile of coagulation factor IX (FIX) activity in hemophilia B patients treated with etranacogene dezaparvovec, which delivers a transgene encoding the Padua variant of FIX.
• What does this study add to our knowledge?

  • ○
    The ABM effectively predicted the median FIX activity plateau observed in patients following gene therapy. Predictions revealed that transduced polyploid hepatocytes, which have slower turnover, may sustain the durability of FIX expression at therapeutic levels out to 20 years.
• How might this change drug discovery, development, and/or therapeutics?

  • ○
    This model provided a novel way to quantitatively investigate the long‐term durability of nonintegrating gene therapies targeting self‐renewing organs. The findings instill confidence in the long‐term efficacy of liver‐targeting gene therapies.

1. Introduction

Gene therapy involves the administration of genetic material (e.g., DNA or RNA) via a carrier, known as a “vector,” for delivery of genetic material into target cells. Adeno‐associated virus (AAV) is a widely used somatic cell‐targeting gene therapy vector. Currently approved AAV‐based gene therapies include onasemnogene abeparvovec (ZOLGENSMA; Novartis, Bannockburn, IL) [1, 2], to treat spinal muscular atrophy; etranacogene dezaparvovec (HEMGENIX; CSL Behring, Lexington, MA) [3, 4] and fidanacogene elaparvovec (BEQVEZ; Pfizer, New York, NY) [5], to treat hemophilia B; and valoctocogene roxaparvovec (ROCTAVIAN; BioMarin Pharmaceutical Inc., Novato, CA) [6, 7], to treat hemophilia A. Many additional AAV‐based gene therapies are under clinical development [8].

The liver has been an organ of choice for AAV‐based gene therapy because the liver is directly or indirectly involved in many essential processes and is directly affected by numerous inherited diseases [9]. After a viral particle enters a hepatocyte, its genome can undergo a sequence of events before forming a stable, circular DNA structure called an episome. Expression of episomes in hepatocytes then results in the production of protein‐of‐interest.

While proffering curative treatment with limited toxicity [10], the durability of somatic gene therapy has been questioned [11]. The liver continuously self‐renews, resulting in a hepatocyte lifespan averaging less than 2 years [12, 13]. Assuming episomes in a transduced hepatocyte are lost after cell death, the rapid turnover of liver tissue as compared with skeletal muscle or neural cells [14], has led to speculation that sustained transgene expression would not be achieved in liver‐targeting gene therapies [15], suggesting that the durability of effect for these AAV‐based gene therapies may decline over time.

To assess the potential for a reduction in effect over time, an agent‐based model (ABM) was implemented to facilitate understanding of the relationships between hepatocyte turnover, episome dynamics, and systemic protein production following treatment with liver‐targeted gene therapies. ABMs are rule‐based, probabilistic models composed of many autonomous agents, with discrete events taking place in discrete time steps [16, 17]. This methodology was chosen for its ability to represent a series of complex biological processes in a heterogeneous population intuitively and without statistical assumptions, enabling the investigation of episome loss associated with target cell turnover and subsequent protein production. ABM has been used for chimeric antigen receptor (CAR)‐T‐related modeling [18]. The application of ABM to AAV‐gene therapy has been discussed [19], but to our knowledge this is the first implementation.

Individual hepatocytes are represented as agents, and agents follow a consistent set of “rules,” specifying cell division and death, episome expression and loss. The ABM was parameterized using biological information and assumptions from literature and was then exercised to assess etranacogene dezaparvovec (formerly AMT‐061 or CSL222), an AAV serotype 5 (AAV5)‐based gene therapy for hemophilia B that delivers a transgene encoding the Padua variant of coagulation factor IX (FIX), administered via intravenous (IV) infusion. Etranacogene dezaparvovec was the first available gene therapy for hemophilia B, approved by the Food and Drug Administration (FDA) in November 2022 [20]. As long‐term clinical data on FIX durability is currently not available, the prediction of FIX activity levels following a treatment that is administered once to a given patient is of relevance to clinicians, patients, and payors. The ABM framework enables exploration of the biological and clinical processes that manifest in long‐term treatment outcomes, and the insights and methodology may be applicable to other AAV‐based gene therapies.

2. Materials and Methods

The ABM [17] was developed to represent the livers of virtual patients, wherein each agent represents a hepatocyte with specified ploidy (diploid vs. polyploid) and transduction state (Figure 1). A virtual liver was composed of 1000 agents (hepatocytes), intended to generate statistically reproducible simulations rather than represent the physical size of an actual liver. Larger numbers of agents were tested but found not to provide statistical benefit (unpublished data). Individual hepatocytes could be transduced by AAV, proliferate, die, and/or change ploidy with time (Figure 1). During ABM simulations, the states of all agents were updated simultaneously. Multiple time steps between 0.01 years (or 3.6 days) and 0.25 years (or 90 days) were tested, and a time step of 0.05 years (or 18 days) was deemed sufficiently short to capture episome dynamics while being computationally efficient. Henceforward, this process was repeated with a time step of 0.05 years (or 18 days), which effectively converted the probability of an agent's change of state (informed by Heinke et al. [13]; Table S1) into a rate of change over time.

FIGURE 1.

Schematic of the agent‐based model. β₂ and β_p are the division probabilities of diploid and polyploid hepatocytes, respectively. δ₂ and δ_p are the death probabilities of diploid and polyploid hepatocytes, respectively. K_2p is the probability of diploid hepatocytes changing ploidy to polyploid hepatocytes. K_p2 is the probability of polyploid hepatocytes changing ploidy to diploid hepatocytes. k_prodp is the protein‐of‐interest production rate per episome per transduced cell. k_deg ^protein is the degradation rate for the protein‐of‐interest. Parameter values can be found in Table S1.

The following assumptions were made about hepatocytes and viral genomes (AAV). All hepatocyte loss was caused by natural cell turnover (i.e., the liver remained healthy and suffered no injury or immune response). The viral genome existed only in the form of episomes in transduced hepatocytes. Each episome was formed by a viral genome. Episomes were stable, did not degrade unless their host cells died, and did not replicate. During cell division, all episomes of the mother cell were distributed evenly between daughter cells (i.e., all episomes were maintained during mitosis and lost during cell death). Other than in episome quantities, daughter cells were identical to their mother cells. Diploid or polyploid hepatocytes carried a comparable quantity of episomes after transduction. Transduced hepatocytes were identical to other hepatocytes except in episome quantities.

The FIX production rate per episome was estimated based on data from nonhuman primates (NHPs; unpublished data) and was assumed constant over time for all hepatocytes, even during cell division. It was assumed that no more than one to five episomes could be expressed per cell (Figure 1, Table S2) and all FIX production originated from these episomes (i.e., in highly transduced cells, only a small subset of episomes were expressed). This assumption was based on the limited transcription and translation machinery in a cell. The FIX production rate per expressing episome in polyploid hepatocytes was assumed twice that of diploid cells, as polyploid hepatocytes are larger and have at least twice the number of nuclei [21]. Once produced, FIX rapidly entered the bloodstream and degraded over time (Table S1, half‐life ~7 h) [22]. The final model contained nine parameters, seven of which were fixed from literature, one was estimated as described, and one varied in the virtual population simulations (Tables S1 and S2).

Treatment was simulated by initializing the ABM (parameter ranges summarized in Table S2). Initial states of the virtual livers reflected male adult livers between the ages of 20 and 80 years [23]. A single IV dose of etranacogene dezaparvovec (2 × 10¹³ genome copies per kilogram of body weight) resulted in successful transduction of 10%–70% of all virtual hepatocytes (unpublished data) [24]. The episome quantity per transduced hepatocyte was assumed to be between 10 and 1000, calculated from the AAV copy number per genomic DNA in a preclinical study [24]. The initial fraction of diploid versus polyploid hepatocytes was specified as a function of patient age (i.e., $\mathit{\text{diploid fraction}}=\frac{92-0.5125\times \mathit{\text{human}}\mathit{age}}{100}$) [13]. Each simulation represented a population of 160,000 virtual livers, a number empirically chosen for statistical robustness. All parameters were generated by random sampling across uniform distributions, with the exception of episome number, which was sampled from a log‐uniform distribution using a Latin hypercube scheme [25].

The FIX molecule count as projected by the ABM simulation was normalized and scaled to FIX activity based on plasma FIX concentration in an average person (87.8 nmol/L) [26], plasma volume (3.1 L) [27], male liver weight (1561 g) [28], and hepatocellularity or number of hepatocytes per gram of liver (139 × 10⁶/g) [29]. In addition, the etranacogene dezaparvovec FIX Padua variant was assumed to have activity of six times that of wild‐type FIX [30]. That is,

$$\begin{array}{c}\mathit{\text{normalized}}\mathit{FIX}\mathit{\text{activity}}=\hfill \\ 6\times \frac{\mathit{\text{simulated number of}}\mathit{FIX}\mathit{\text{molecules}}}{\mathit{\text{hepatocyte count in virtual liver}}}∕\frac{\mathit{\text{normal number of}}\mathit{FIX}\mathit{\text{molecules}}}{\mathit{\text{normal liver hepatocyte count}}}.\hfill \end{array}$$

The resulting simulated FIX activity levels did not reach a relatively stable trajectory until 6 months post treatment, once transduction was complete and FIX expression kinetics were fully established. Only simulation results describing times later than 6 months post treatment were used for further analysis.

The study is conducted in accordance with International Council for Harmonisation Good Clinical Practice guidelines and ethical principles originating in the Declaration of Helsinki. The protocol was approved by institutional review boards and independent ethics committees at each study site. All simulations were performed in Julia 1.10.4, using the Distributions and QuasiMonteCarlo packages. ABM simulation results were compared with patients' FIX activity observed in Phase 2b (NCT03489291) and Phase 3 (NCT03569891) clinical studies of etranacogene dezaparvovec [30, 31].

3. Results

3.1. Comparison Between Model Simulations and Clinical Dataset

Model‐predicted FIX activity for patients receiving etranacogene dezaparvovec was compared with clinical observations for 3 years following treatment. The clinical observations were divided into training and validation sets based on availability at model development. The training dataset (2‐year data) was available when the model was built. The validation dataset (out to 3 years [31]) matured after model parameterization was completed. Predicted median FIX activity at Years 1 and 2 after receiving gene therapy was lower but tracked the median FIX activity level observed in the clinical studies (Figure 2). The predicted median FIX activity 1 year after receiving gene therapy was 35.3% (equivalent to 35.3 IU/dL; 95% confidence interval [CI] 7.82%, 99.1%), close to the observed median value of 40.8% (range 5.9%–113%), computed from one‐stage assay measurement in Phase 2b and Phase 3 studies. The predicted median FIX activity 2 years and 3 years after treatment was 31.8% (95% CI 7.01%, 91.7%) and 28.2% (95% CI 6.12%, 83.5%), respectively, again close to the median FIX activity of 34.6% (range 4.7%–99.2%) and 35.95% (range 4.8%–80.3%), respectively, computed from clinical observations.

FIGURE 2.

ABM‐simulated and observed FIX activity as a percent of normal versus time for the first 4 years post treatment. The light gray points represent individual clinical observations in the training data, the blue triangles represent observations in the validation data. The light gray dash lines connect observations from the same subject. The green solid line represents the simulated median FIX activity, and the shaded areas represent the 50%, 80%, and 95% CIs (the darker shading corresponds to the tighter CI). The dark gray dash line represents the median observed FIX activity of the training data, the blue dash line represents the median observed FIX activity of the validation data. The dash‐dot lines outline the 95% CI of the observed FIX activity. ABM, agent‐based model; CI, confidence interval; FIX, factor IX.

3.2. Long‐Term Predictions and Drivers of Durability

At the time of writing, additional data became available. The predicted mean FIX trajectory tracked the clinical dataset released by CSL 5 years into the HOPE‐B trial, with mild but noticeable underprediction of normalized FIX activity during the fourth and fifth year—and thus treatment durability (Figure 3). The model‐predicted mean FIX activity at Year 5 was 23.5% (95% CI 4.92%, 72.2%), overlapping with the reported range of 31.6% ± 15.7% (mean ± standard deviation) [31].

FIGURE 3.

ABM‐simulated and observed FIX activity from the most recent data cut of the HOPE‐B study as a percent of normal versus time for the first 4 years post treatment. The light gray points represent individual clinical observations in the training data, the blue triangles represent observations in the validation data. The light gray dash lines connect observations from the same subject. The green solid line represents the simulated mean FIX activity, and the shaded areas represent the 50%, 80%, and 95% CIs (the darker shading corresponds to the tighter CI). The solid black line represents summarized clinical data of mean FIX activity observed in 47 patients enrolled in the HOPE‐B trial over 4 years, with black bars representing the standard error. ABM, agent‐based model; CI, confidence interval; FIX, factor IX.

While the ABM predicted a later plateau on median FIX activity level (Figure 4), the fact that it mirrors the observed data from the clinical studies is remarkable and allows us to explore the possibilities for longer‐term durability. Following the predicted decrease in FIX activity during the first 5 years, the ABM indicates that FIX activity would stabilize thereafter (Figure 4) with the median predicted FIX activity at 22.6% (95% CI 4.58%, 71%) by Year 20. This exceeds the FIX activity threshold (2% FIX activity) reported to allow patients to “lead normal lives,” with a decrease in bleed number and long‐term preservation of musculoskeletal function [32, 33].

FIGURE 4.

ABM‐predicted FIX activity versus time over 20 years post treatment. The solid blue line represents the predicted median FIX activity, and the shaded area represents the 95% CI. Red and blue dashed lines represent the 2% and 5% thresholds of normal FIX activity, respectively. ABM, agent‐based model; CI, confidence interval; FIX, factor IX.

Disentanglement of FIX expression by diploid versus polyploid hepatocytes revealed that the elevated FIX activity was primarily sustained by FIX production in the latter cell population (Figure S1). This was driven by the slow turnover of polyploid hepatocytes, which have an average lifespan exceeding 5 years. In addition, polyploid hepatocytes were more likely to divide than die as compared with diploid cells. Thus, the fraction of transduced polyploid hepatocytes increased over time. In contrast, the rapid turnover of diploid hepatocytes (average lifespan < 1 year) led to the decrease in their contribution to FIX activity over the first 5 years.

3.3. Sensitivity Analysis of Long‐Term Predictions

To understand the biological processes underlying long‐term FIX activity, sensitivity analyses on the number of episomes per transduced cell, episome passing efficiency (i.e., the fraction of episomes that could be passed on from mother cell to daughter cells during cell division, or the dilution of episomes between mother and daughter cells), and polyploid hepatocyte fraction and dynamics were carried out.

The sensitivity analysis of transduction efficiency (the number of episomes per transduced cell at treatment time) revealed that a decrease in initial episome transduction efficiency resulted in decreased long‐term FIX activity (Figure 5A). When the range of episomes per transduced cell decreased 10‐fold, from a range of (10–1000) to a range of (1–100), the predicted median FIX activity at 1 year decreased disproportionately less, from 38.4% to 34.1%. By Year 5, the predicted median FIX activity decreased from 26.5% to 22.1% and by Year 20, from 24.7% to 19.7%.

FIGURE 5.

Effect of a 10‐fold decrease in transduction efficacy (the episome count per transduced hepatocyte described by model parameter Init_Episome) (A) and decreased episome passing efficiency (B) on ABM‐predicted FIX activity for the 20 years post treatment. In (A), solid lines represent the median normalized FIX activity levels predicted for two separate scenarios. The green line represents the default scenario where the initial episome count per transduced hepatocyte ranged from 10–1000. The red line represents a reduced transduction scenario where the initial episome count per transduced hepatocyte ranged from 1–100. Shaded areas represent the respective 95% CIs. The red and blue dashed lines represent the 2% and 5% thresholds of normal FIX activity, respectively. In (B), solid lines represent the median predicted normalized FIX activity levels for five individual scenarios, representing a range of episome passing efficiencies between 0 and 1. When the episome passing efficiency equals 1, all episomes of a transduced hepatocyte are split equally between daughter cells during cell division, an assumption made during model development. With decreasing passing efficiency, proportionally fewer or no episomes are passed from mother to daughter cells during proliferation. Shaded areas represent the 95% CIs, with boundaries outlined in dashes. Red and blue dashed lines represent the 2% and 5% thresholds of normal FIX activity, respectively. ABM, agent‐based model; CI, confidence interval; FIX, factor IX.

Reduced episome passing efficiency (i.e., increase in episome dilution) during cell division was predicted to have a more pronounced effect on long‐term FIX activity (Figure 5B). With the baseline of episome passing efficiency = 1 in all hepatocytes (i.e., no episome loss), the predicted median FIX activity decreased from 38.4% at Year 1 to 19.1% by Year 20. At the other extreme, when the baseline of episome passing efficiency = 0 (i.e., no episomes are passed from mother to daughter cells), the predicted median FIX activity decreased from 19.5% to 6.35% over the same time period.

Additional sensitivity analyses focused on polyploid hepatocyte biology, given their predicted importance in maintaining FIX activity over extended periods of time. Specifically, the sensitivity analyses were carried out on the fraction of polyploid cells in virtual livers at the time of treatment, their proliferation and death probabilities, and the FIX production rate per episome.

At baseline, the initial percentage of polyploid hepatocytes in a virtual liver was determined by the age of the patient [13], where the fraction of diploid hepatocytes was observed to decrease approximately linearly with age (Figure S2). To test how the long‐term FIX activity would be impacted if this relationship was invalid, the initial percentage of polyploid hepatocytes was set randomly to between 20% and 60% of the total hepatocyte count, independent of the virtual patient's age. Limited changes were predicted in the FIX activity (Figure 6A): the predicted median FIX activity 20 years after treatment was 25.3% (95% CI 5.13%, 77.5%), close to the default scenario of 22.6% (95% CI 4.58%, 71%).

FIGURE 6.

Effects of age‐varying or age‐independent diploid hepatocyte fractions in virtual livers at treatment time (A), different proliferation and death rates of polyploid hepatocytes (B), and different FIX synthesis rates in polyploid compared with diploid hepatocytes (C) on ABM‐predicted FIX activity versus time for 20 years post treatment. In (A), solid lines represent the median normalized FIX activity levels predicted for two separate scenarios. The green line represents the default scenario where the fraction of diploid hepatocytes in a virtual liver depends on the age of the virtual patient at the time of treatment (“Varying with age”). The red line represents an alternative scenario where the fraction of diploid hepatocytes in a virtual liver does not vary with age of the virtual patient at the time of treatment (“Independent of age”). Shaded areas represent the respective 95% CIs. Red and blue dashed lines represent the 2% and 5% thresholds of normal FIX activity, respectively. In (B), seven different scenarios were simulated in which the proliferation and death rates of polyploid hepatocytes were varied simultaneously by the same ratio, compared with their default parameterizations. The two (relatively faster) rates for diploid hepatocytes' proliferation and death were not varied in this analysis. A ratio of 2 meant that the polyploid hepatocytes proliferated and died at twice their default rates (shown in Table S1), resulting in a polyploid cell turn over at twice the default rate, while maintaining the same polyploid hepatocyte population size in a given virtual liver. Solid lines represent the median predicted normalized FIX activity levels for the individual scenarios, shaded areas represent their 95% CIs. Red and blue dashed lines represent the 2% and 5% thresholds of normal FIX activity, respectively. In (C), the baseline ratio of FIX synthesis rate between polyploid and diploid hepatocytes was assumed to be two (as shown in Table S1), with four additional ratios ranging from 0.2 to 5 being simulated for comparison. Solid lines represent the median predicted normalized FIX activity levels for each individual scenario, shaded areas represent their 95% CIs. Red and blue dashed lines represent the 2% and 5% thresholds of normal FIX activity, respectively. ABM, agent‐based model; CI, confidence interval; FIX, factor IX.

The impact of polyploid hepatocyte turnover was analyzed by varying proliferation and death probabilities simultaneously over a span of 0.2‐ to 5‐fold compared with the baseline rates (Figure 6B). Increased cellular turnover of polyploid hepatocytes led to a decrease in predicted FIX activity over time. This was most significant when the probabilities for proliferation and death were increased by a factor of 5: the median predicted FIX activity dropped to 8.18% (95% CI 1.05%, 37.6%) 20 years after treatment, down from the default projection of 22.6%. This outcome was expected; higher turnover means that more episomes are lost to cell death, and the episome content of proliferating cells will be diluted over time. This effect was less pronounced when the change in probability of polyploid hepatocyte turnover was smaller. When the probabilities of proliferation and death remained scaled between 0.2 and 1.1 of their original value, FIX activity was predicted to change very slightly.

An increase in the FIX production rate in polyploid hepatocytes resulted in higher predicted FIX activity over time (Figure 6C). When this increase was tuned up to 5‐fold, there was very little decline in FIX activity over time, and simulations were not consistent with clinical observations (Figure 2). This suggests the initial assumption of a 2‐fold higher FIX production rate in polyploid hepatocytes compared with diploid cells was reasonable.

In summary, the sensitivity analyses reveal that changes in hepatocyte and episome biology affect long‐term FIX expression in line with intuition. The median of all local sensitivity coefficients (calculated as described by Chen et al. [34]) was below 1 (Figure S3), meaning durability of transgene function is relatively robust to each parameter value (shown in Table S1 with ranges listed in Table S2). However, the local sensitivity coefficients for episome passing efficiency and polyploid hepatocyte turnover had wider ranges beyond 1 (Figure S3), indicating the durability of transgene expression could still be influenced by their values. In addition, the ratio of FIX synthesis rate between polyploid and diploid hepatocytes was predicted to have a local sensitivity coefficient of 1 at Year 20. This was consistent with the prediction that all FIX expressions in the long term were from polyploid hepatocytes. Across a range of biologically plausible assumptions, the 20‐year FIX activity was predicted to remain above the threshold requiring prophylaxis [32, 33].

4. Discussion

An ABM was developed to predict the durability of FIX activity levels following gene therapy targeting a self‐renewing organ. To our knowledge, this is the first ABM developed for an AAV‐based gene therapy, despite the potential of this approach being discussed in literature [19]. The model was applied to etranacogene dezaparvovec, a liver‐targeting gene therapy for hemophilia B that consists of an AAV vector encoding the Padua variant of FIX. Simulations of FIX activity in patients receiving this therapy effectively predicted a plateau in the median FIX activity level following treatment, although this was observed to be occurring slightly earlier in the clinical data. The model provided an interesting projection of the anticipated longer‐term durability of effect, including sensitivity analyses around the biological processes modulating therapeutic durability. The exploration of median FIX activity levels under different hypothetical conditions suggested that virtual patients experienced FIX activity well above normal levels (2%) at 20 years post‐therapy.

A key feature of the ABM is self‐renewal of the liver. Homeostasis is driven by the balance between hepatocyte proliferation and death, a model proposed and assessed by Heinke et al. [13]. Alternative mechanisms for liver regeneration, such as diploid hepatocyte generation from hepatic progenitor cells or biliary epithelial cells [35, 36], were explored by Heinke et al. [13], but were estimated to contribute only minimally to the diploid hepatocyte pool (< 0.1% per year). The simplest model proposed by Heinke et al. was therefore implemented here, as it reflected the most relevant physiological target for etranacogene dezaparvovec treatment.

Sustained FIX activity was predicted to be driven by polyploid hepatocytes (Figure S1) due to the slower turnover as compared with diploid cells, and consequential maintenance of FIX‐expressing episomes. Diploid hepatocytes were predicted to minimally contribute to the exogenous FIX production after the first 6 years of gene transfer due to the rapid turnover (Figure S1). Notably, durability is driven by long‐lived polyploid subpopulations and diploid–polyploid dynamics, rather than total average hepatocyte turnover. These predictions indicate that the concerns associated with AAV‐based gene therapy targeting self‐renewing organs are valid, as transgene expression is lost during cell turnover. Yet, for AAV‐based liver‐targeting gene therapies, the polyploid hepatocytes are predicted to sustain therapeutically meaningful transgene expression for at least 20 years under a range of biologically plausible scenarios, and this seems to be supported by the observed clinical results for etranacogene dezaparvovec, as well as other gene therapies.

Model‐predicted FIX activity showed limited sensitivity to changes in polyploid turnover probabilities and the initial diploid to polyploid ratio, despite the importance of polyploid hepatocytes in sustaining transgene expression beyond 6 years following treatment. This robustness to input assumptions instills confidence in the long‐term, population‐level predictions. Sensitivity analysis revealed that the long‐term FIX activity decreases as the number of episomes per transduced hepatocyte decreases (i.e., transduction efficiency; Figure 5A). This is consistent with studies reporting a positive correlation between AAV vector dose and transgene persistence [37, 38], though the model does not quantitatively link administered dose with the number of transduced hepatocytes. This is because many other factors may impact the relationship, such as the interindividual variability in AAV uptake and episome formation, as well as tier of pre‐existing neutralizing antibodies.

Furthermore, sensitivity analysis also indicated that the long‐term FIX activity was sensitive to how effectively episomes pass from mother to daughter cells (Figure 5B). This model assumed that episomes were maintained during mitosis and were only lost due to cell death (i.e., apoptosis or necrosis). However, literature reports vary and have, at times, contradictory views on the stability of episomes in dividing cells, which may differ by context [38, 39, 40, 41]. The model herein constrained hepatocyte turnover to homeostatic conditions, as opposed to liver regeneration after injury (e.g., due to hepatotoxins), immune response, or hepatectomy.

The differences between the virtual patient population of 160,000 livers being simulated with the ABM and the FIX activity observations from actual patients may reflect the assumptions relating to the compatibility of the NHP data to humans and/or the constant FIX production rate per episome, as well as our assumption of no replication of the episomes. However, the purpose of virtual population simulation is to conscientiously include potential ‘edge’ cases; assumptions that may not describe the average patient but may occur in a smaller subset of the population and may affect their treatment outcomes. To see the virtual population simulations arrive at a realistic, but slightly lower, average FIX activity than the data supports illustrates this conservative approach.

While there are no data available for independent, external validation, a recent study that analyzed liver biopsies collected from patients treated with valoctocogene roxaparvovec (AAV5‐hFVIII‐SQ), an AAV‐based gene therapy for hemophilia A, reported that 30%–50% of hepatocytes were transduced at 2.6–4.1 years post treatment [42]. Despite the results not being directly comparable due to differences in vector design and dosing (2–3‐fold higher than etranacogene dezaparvovec), our simulations fall within a similar range (medians of 31% and 23% at Years 2.5 and 4, respectively, with 95% CI intervals ranging from 6.12% to 60.51%).

In contrast to a published differential equations‐based framework that predicted early (< 6 months) FIX dynamics following treatment [43], our work focused solely on the long‐term durability of FIX activity, with a special focus on the impact from episome loss. The difference between these two types of models was mainly driven by the different questions they aimed to answer; the ABM model focused on the drivers of durability over 20 years, while the differential equations‐based models focused on other aspects, such as preclinical‐to‐clinical translation and dose optimization. Consequently, the ABM simulations were initiated at 6 months post treatment, once FIX expression dynamics had settled. The omission of episome loss in previously published differential equations‐based frameworks likely leads to an overestimation of actual FIX activity in the long term. Without accounting for the loss of episomes during target cell turnover, such models cannot adequately address existing concerns regarding the long‐term durability of nonintegrating gene therapies in self‐renewing organs like the liver.

Despite the model's biological intuition, some limitations and caveats should be noted. First and foremost, many initial conditions (i.e., the range of episome count per hepatocytes and the initial fraction of transduced hepatocytes) in the simulation were informed by NHP studies (Table S2). It is not clear how translatable these conditions are between NHPs and humans. Second, this model does not account for the impact of pre‐existing neutralizing antibodies, which may affect gene therapy transduction efficiency and long‐term outcomes [44]. Third, this model assumed transgenes exist stably in episomal form. This is a simplification, as episomes may be degraded or integrated into host genomes [38, 45]. In addition, the model assumed that the maximum episomes that can be expressed in a transduced hepatocyte was between 1 and 5. This was based on experts' assessment, and was not backed up by experimental observations. Further sensitivity analyses need to be conducted to test the model's robustness against this assumption. Moreover, this model assumes a consistent FIX production rate per episome per cell over time that could partly explain the divergence in FIX levels between the predicted and observed median activity levels at Year 3. Episomes may also be silenced [14, 46], potentially leading to decreased FIX production rates. However, there may also be an interim increase in the FIX production rate which is not reflected in the assumed constant rate of production, which also may account for the divergence in FIX levels between the predicted and observed median activity levels at Year 3. Furthermore, this model assumes a healthy, homeostatic virtual liver, while patients with hemophilia B are prone to liver comorbidities, and disease states and injury may affect episome propagation [47]. These factors could all impact long‐term transgene expression but are beyond the scope of this current research. Despite these uncertainties, model simulations effectively predicted the median FIX activity plateau seen in clinical observations and enabled the predictive exploration of physiological factors underlying long‐term efficacy of etranacogene dezaparvovec. The novel modeling framework presented herein is well positioned to serve as a platform for investigating the durability of other nonintegrating gene therapies targeting self‐renewal organs.

Author Contributions

Y.L., A.K.W., D.C.K., P.N., and S.N. wrote the manuscript. Y.L., E.J., K.P., P.N., and S.N. designed the research. Y.L., E.J., and T.K. performed the research and analyzed the data.

Funding

This work was sponsored by CSL Behring.

Conflicts of Interest

P.N., K.P., and S.N. are (or were, at the time this work was conducted) employees and shareholders of CSL Behring. Y.L., E.J., T.K., D.C.K., and A.K.W. are (or were, at the time this work was conducted) employees of Metrum Research Group. All authors declared no other competing interests for this work.

Supporting information

Table S1: ABM parameters describing hepatocyte dynamics and FIX production after treatment with etranacogene dezaparvovec.

Table S2: Parameter ranges used to set up virtual livers for patients treated with etranacogene dezaparvovec.

Figure S1: ABM‐predicted FIX activity segmented based on contributions from diploid and polyploid transduced hepatocyte populations versus time over 20 years post treatment.

Figure S2: Calibration of model parameters describing the proliferation and death probabilities of polyploid hepatocytes to capture the reduced diploid hepatocyte fraction as a function of age.

Figure S3: Local sensitivity coefficients of episome passing efficiency, initial episome count per transduced cell, the fold‐change of polyploid hepatocyte turnover rates, and the ratio of FIX synthesis rate between polyploid and diploid hepatocytes.

Data S1: Supporting information.

Li Y., Nandy P., Jordie E., et al., “Development of an Agent‐Based Model to Investigate Durability of Factor IX Activity in Hemophilia B Patients Treated With Etranacogene Dezaparvovec,” CPT: Pharmacometrics & Systems Pharmacology 15, no. 7 (2026): e70286, 10.1002/psp4.70286.