Integrating MBMA and QSP to Identify Key Covariates and Predict Treatment Outcomes in Relapsed/Refractory Multiple Myeloma

ABSTRACT

This study demonstrates the application of a model based meta analysis (MBMA) framework to characterize the safety and efficacy profiles of therapies in relapsed and refractory multiple myeloma (RRMM). Published clinical trial data were analyzed to evaluate the incidence of Grade ≥ 3 neutropenia and overall response rate (ORR), providing a quantitative foundation for model‐informed drug development. The final model incorporated trial‐ and treatment‐level covariates and was evaluated using visual predictive checks and predictive simulations. Results revealed increased neutropenia risk associated with alkylating agents and higher ORR in regimens with background corticosteroids and in patients with only one prior line of therapy. MBMA‐derived estimates facilitated systematic comparisons across regimens, accounting for heterogeneity in trial design and populations. The MBMA estimates can also support benchmarking of internal regimens against current standards. A quantitative systems pharmacology (QSP) model, developed in parallel, was also used to simulate patient responses across a broad array of RRMM treatments, including novel combinations involving T‐cell engagers (TCEs) and CELMoD agents. Trial‐calibrated virtual patients and a classifier for prior therapy exposure enabled the prediction of regimen‐specific ORR across different treatment histories. Together, the MBMA‐informed and QSP‐supported modeling strategy enabled a comprehensive benefit–risk assessment by combining statistical estimation with mechanistic simulation. This coordinated approach enhances clinical decision‐making by enabling comparison of novel or investigational therapies to the evolving treatment landscape, particularly in the absence of head‐to‐head trials.

Study Highlights.

• What is the current knowledge on the topic?

  • ○
    Relapsed/refractory multiple myeloma (RRMM) is a heterogeneous disease. Model based meta analysis (MBMA) can be used to identify the most impactful covariate, and quantitative systems pharmacology (QSP) models are often used to predict novel treatment outcomes.
• What question did this study address?

  • ○
    How can MBMA and QSP work together to predict novel combinations that consider the impact of important covariates?
• What does this study add to our knowledge?

  • ○
    MBMA identified prior treatment history, background steroid use, and alkylating agent use as significant covariates of response and Grade ≥ 3 neutropenia, respectively. QSP predicted the outcomes of a novel treatment combination (immunomodulatory agents and T‐cell engagers).
• How might this change drug discovery, development, and/or therapeutics?

  • ○
    MBMA can be used early to identify and justify the most important mechanisms for platform QSP models to focus on. Platform QSP models can semi‐mechanistically account for significant covariates, broadening their scope in drug development. This dual approach enables prediction and benchmarking without head‐to‐head trials.

1. Introduction

Relapsed and refractory multiple myeloma (RRMM) remains a significant clinical challenge, with patients often progressing through multiple lines of therapy. Over the last decade, the treatment landscape for RRMM has diversified rapidly, now including proteasome inhibitors, immunomodulatory drugs (IMiDs), monoclonal antibodies (mAbs), antibody‐drug conjugates, bispecific T‐cell engagers, and CAR‐T cell therapies. While the integration of novel agents into treatment regimens offers the possibility of long‐term survival and improved quality of life, the constantly changing landscape means that it may be challenging to choose optimal therapy [1]. Most contemporary RRMM regimens are used in combination formats, incorporating backbone therapies such as steroids, alkylating agents, or chemotherapy, all of which can impact both efficacy and safety outcomes. As a result, making informed treatment decisions has become increasingly complex, especially in the absence of direct head‐to‐head clinical trials.

To address these challenges, model based meta analysis (MBMA) applies nonlinear mixed‐effects (NLME) modeling to integrate published summary data with internal data, enabling the assessment of safety and efficacy endpoints across heterogeneous studies and facilitating indirect comparison between treatments. MBMA extends traditional meta‐analysis methods by incorporating pharmacological principles, such as dose–response relationships, and accounting for between‐trial and between‐treatment‐arm variability [2, 3]. As a result, MBMA offers a flexible framework for interpreting aggregated data from historical reference studies and is a standard tool within model‐informed drug development (MIDD) [4].

Treatment decisions in the clinic are affected by a wide range of patient‐ and disease‐specific characteristics, such as prior lines of therapy, cytogenetic risk, renal function, age, and performance status [5]. These variabilities further complicate direct comparisons between regimens across diverse patient populations. MBMA addresses this complexity by quantifying the effects of covariates and specific drugs on treatment outcomes. By accounting for both prognostic and predictive factors, MBMA enables a more clinically relevant assessment of treatment efficacy and safety. Moreover, MBMA integrates internal and external data sources to perform benefit–risk assessments and benchmark new treatments against historical outcomes [3]. This understanding is essential for guiding clinical development decisions (e.g., go/no‐go) and reducing the risk of late‐stage trial failure.

In this study, an MBMA framework was applied to evaluate the safety (Grade ≥ 3 neutropenia) and efficacy (ORR) profiles of RRMM therapies. By integrating data across diverse studies, the analysis enabled treatment benchmarking, supported benefit–risk assessment, and informed future drug development strategies. Quantitative systems pharmacology (QSP) was used as a complementary approach, offering mechanistic insight into covariate effects and extrapolating to untested treatments. Both approaches can integrate large datasets, but they operate at different levels of abstraction. While MBMA operates at a statistical level, ideal for screening impactful covariates, it may be limited in interpreting biological mechanisms or predicting novel combination outcomes. In contrast, QSP models represent biological systems mechanistically, capturing interactions among cells, cytokines, and molecular targets. Together, these methods enabled a dual modeling workflow: MBMA identified key covariates, which were mechanistically integrated into a QSP model to enhance predictive performance. MBMA contributed to both safety and efficacy characterization, while QSP supported interpretation and extrapolation.

2. Methods

2.1. MBMA Approach

A comprehensive RRMM clinical trials database from Certara's Clinical Outcomes Database (CODEx) [6] platform was utilized for the MBMA analysis. The database comprises 352 randomized clinical trials with 512 treatment arms and 177,692 RRMM patients. The database was refined to include only treatment arms with more than 10 patients and treatments supported by at least three independent studies. This ensured that treatment effects could be reliably quantified. MBMA analysis incorporated a broad spectrum of both single‐agent and combination therapies commonly used in the treatment of RRMM. Key therapeutic classes included proteasome inhibitors (e.g., bortezomib, carfilzomib), IMiDs (e.g., pomalidomide, lenalidomide), mAbs (e.g., daratumumab, isatuximab), and B‐cell maturation antigen (BCMA) directed therapies (e.g., teclistamab, elranatamab). Examples of commonly used combinations incorporated in the analysis include proteasome inhibitors with IMiDs (e.g., bortezomib + lenalidomide, carfilzomib + pomalidomide) and mAbs with either proteasome inhibitors or IMiDs and a corticosteroid medication (e.g., daratumumab + bortezomib + dexamethasone, daratumumab + pomalidomide + dexamethasone) (Table 1).

TABLE 1.

Summary of data for ORR and Grade ≥ 3 neutropenia by treatment class and regimen.

Treatment class	Treatment	ORR			Grade ≥ 3 neutropenia
		No. of studies	No. of arms	No. of patients	No. of studies	No. of arms	No. of patients
Proteasome inhibitor	Bortezomib	61	75	7165	41	49	5972
Immunomodulators (IMiD)	Pomalidomide	35	38	3276	32	35	3328
IMiD	Lenalidomide	24	25	3048	20	21	2805
Proteasome inhibitor	Carfilzomib	22	25	2966	19	21	2288
Anti‐CD38 mAb	Daratumumab	9	12	1252	6	7	1091
Proteasome inhibitor	Ixazomib	6	7	344	5	6	324
Anti‐CD38 mAb	Isatuximab	5	9	355	—	—	—
Anti‐CD38 mAb + IMiD	Isatuximab + Pomalidomide	4	7	288	3	3	229
Anti‐CD38 mAb + IMiD	Daratumumab + Pomalidomide	4	4	411	3	3	364
Proteasome inhibitor + IMiD	Ixazomib + Pomalidomide	4	4	118	—	—	—
Proteasome inhibitor + IMiD	Carfilzomib + Lenalidomide	3	3	494	3	3	490
Anti‐CD38 mAb + Proteasome inhibitor	Daratumumab + Carfilzomib	3	3	460	—	—	—
Anti‐CD38 mAb + Proteasome inhibitor	Daratumumab + Bortezomib	3	3	403	—	—	—
Anti‐CD38 mAb + IMiD	Daratumumab + Lenalidomide	3	3	378	—	—	—
Anti‐BCMA TCE	Elranatamab	3	3	242	—	—	—
Proteasome inhibitor + IMiD	Bortezomib + Thalidomide	3	3	240	—	—	—
Anti‐CD38 mAb + Proteasome inhibitor	Isatuximab + Carfilzomib	3	3	235	—	—	—
Anti‐BCMA TCE	Teclistamab	3	3	230	3	3	243
Proteasome inhibitor + IMiD	Bortezomib + Lenalidomide	3	3	127	—	—	—
Anti‐CD38 mAb + recombinant human hyaluronidase antibody (rHuPH20)	Daratumumab + rHuPH20	3	3	112	3	3	112
Exportin‐1 (XPO1) inhibitor + Proteasome inhibitor	Selinexor + carfilzomib	3	3	83	—	—	—

Note: This table presents the number of studies and treatment arms included in the MBMA dataset for each treatment regimen in RRMM. Counts are provided separately for the efficacy endpoint (ORR) and the safety endpoint (Grade ≥ 3 neutropenia), stratified by treatment class and specific agents. Table 1 reflects regimen‐level data; studies contributing to multiple regimens may be counted more than once. Unique study counts are provided in the text.

An NLME model was employed to quantify probabilities of events (i.e., Grade ≥ 3 neutropenia and overall response rate) using binomial regression, incorporating between‐study variability (BSV) and between‐treatment‐arm variability (BTAV). The treatment effect for study i and arm j (EFF _i,j) is defined in (1), where EFF _θ represents the typical treatment effect specific to each unique treatment, η _i ^study denotes the BSV for study i, η _i,j ^arm is BTAV for study i and arm j, and n _i,j is the total patient number in study i, arm j. The BTAV was normalized using the square root of the patient number per treatment arm to ensure appropriate weighting. To account for overdispersion beyond binomial variability, an arm‐specific error term (BTAV) was included in the model and normalized by √n to reflect the inverse relationship between sample size and standard error for binary outcomes. This approach, consistent with established methods in binary meta‐analyses, avoids the instability of weighting by estimated standard errors in sparse data [7]. The BSV components representing across‐trial heterogeneity remained unaffected by arm‐level sample sizes [8]. An inverse‐logit function was applied to transform the normally distributed parameter EFF _i,j to the probability of events p _i,j (2). A probability mass function describing the probability of observing k events (e.g., 0, 1, 2…) among n patients is shown in (3).

$${\mathit{EFF}}_{i,j}={\mathit{EFF}}_{\theta }+{\eta }_i^{\mathit{\text{study}}}+\frac{{\eta }_{i,j}^{\mathit{arm}}}{\sqrt{n_{i,j}}}$$ (1)

$$p_{i,j}=\frac{\mathit{exp}\left({\mathit{EFF}}_{i,j}\right)}{1+\mathit{exp}\left({\mathit{EFF}}_{i,j}\right)}$$ (2)

$$P\left(Y=k\right)=\frac{n!}{k!\times \left(n-k\right)!}\times p^k\times {\left(1-p\right)}^{n-k}$$ (3)

The model was developed in Monolix (MonolixSuite2023R1, Simulations Plus Inc., Research Triangle Park, NC), while simulations and all plots were prepared in R (version 4.0.3). Covariate effects were evaluated to identify potential predictors of ORR and Grade ≥ 3 neutropenia across trials using a stepwise covariate model approach, where model selection was guided by the Bayesian Information Criterion (BIC), a metric that balances model fit with complexity by penalizing excessive parameters. Covariate analysis was restricted to variables with ≤ 40% missing data to ensure robustness and interpretability of model estimates. Missing covariate values were imputed using the median, a method chosen for its robustness to outliers and its suitability for handling summary‐level data. Dose–response effects for monotherapies were tested to improve model fit, and the dose intensities were normalized to the FDA‐approved dosing frequency for consistency across treatments. Model evaluation included visual predictive checks (VPCs), which were generated by simulating the actual studies from the analysis dataset, using 500 replicates per study to assess the model adequacy and capture variability.

2.2. QSP Approach

The QSP model was developed to represent the disease biology, including disease progression, and the mechanisms of action of various therapies for RRMM. The QSP model incorporated clinical endpoints defined by the International Myeloma Working Group (IMWG) [9], including ORR, best overall response (BOR), time to response (TTR), and progression‐free survival (PFS) [9]. The QSP model was calibrated using a similar dataset to the MBMA analysis, with added data for the cereblon E3 ligase modulatory drugs (CELMoD) drug class. Additional proprietary data were included to constrain baseline characteristics (e.g., serum B‐cell maturation antigen, soluble free light chain, urine and kidney monoclonal protein, bone marrow myeloma cell fraction) and on‐treatment cytokine release dynamics and immune cell response (unpublished data). Model simulations used drug‐specific pharmacokinetic (PK) parameters to recreate the trial dosing protocols. However, all T cell engagers (TCE) used the same PK and PD implementation parameterized using internal datasets on alnuctamab (unpublished data). TCE simulations were done using a single escalating subcutaneous dosing regimen, regardless of the TCE molecule or route of administration, to pool data from different TCE molecules. Pooling TCE outcomes eliminates the model's ability to distinguish between TCE molecules but is not expected to introduce much bias because TCE efficacy parameters were fit. The QSP model was validated against a set of hold‐out clinical trials, which included CELMoD‐based triplet therapies, bispecific TCE combination therapies, and daratumumab‐based combinations. Trial arms were split into fitting and validation by drug, corresponding to a within‐pathway validation scheme [9], and monotherapy versus combination therapy.

The QSP model was developed in the Thales platform (Simulations Plus Inc., Research Triangle Park, NC), which provides an end‐to‐end platform for model building, simulation, optimization, and analysis of QSP models. Thales can simulate intricate treatment sequences; however, the breadth of possible treatment histories for RRMM patients makes this intractable. Instead, a more efficient, approximate approach was employed. This approach involved simulating RRMM virtual patients and subsequently classifying them as representative of real patients who had received a specific number of prior therapies. Specifically, the QSP model simulated the clinical dosing protocols for each clinical trial in the calibration dataset, recorded clinical response in each treatment scenario ($\mathit{ORR}$), calculated the fraction of scenarios in which a patient responded f _resp, (4), and probabilistically assigned each virtual patient to a prior line category (N _prior). This classification was performed by creating a linear predictor z, (5), scale parameters b _i and location parameter (δ) for the unnormalized probability of each N _prior category using f _resp, then normalizing the linear combinations z into proper probabilities (p) using the softmax function (6):

$$f_{\mathit{\text{resp}},j}=\frac{\sum \mathit{\text{OR}}{R}_{j,k}}{K}$$ (4)

$$z_{i,j}=b_i\left(f_{\mathit{\text{resp}},j}-\delta \right)$$ (5)

$$p_{i,j}=p\left(N_{\mathit{\text{prior}},i,j}|f_{\mathit{\text{resp}},j}\right)=\mathit{\text{softmax}}\left(z_{i,j}\right)=\frac{e^{z_{i,j}}}{{\sum }_{i=1}^I{e}^{z_{i,j}}}$$ (6)

for K simulations per patient, prior line categories i ∈ {1, 2–3, 4+} and the j ^th patient. Real trial results were stratified by the median number of prior lines. The prior line binning scheme {1, 2–3, 4+} was chosen to minimize the number of classifier parameters to fit while still providing an informative dataset (not shown). When a clinical trial reported ORR stratified by N _prior (alongside the ORR for the overall arm), these ORRs were included alongside the arm‐level ORR and BOR.

Except for prediction intervals, virtual population statistics stratified by N _prior were weighted averages ŷ, (7) of virtual patient responses (y) using weights ŵ, (8) defined both by prevalence weights (w) and the N _prior probability assignments:

$${\widehat{y}}_i={\sum }_{j=1}^J{\widehat{w}}_{i,j}{y}_j$$ (7)

$${\widehat{w}}_{i,j}=\frac{w_j{p}_{i,j}}{{\sum }_{j=1}^J{w}_j{p}_{i,j}}.$$ (8)

Model calibration used an iterative loop of sampling virtual patients, simulating their outcomes, assigning a prior line probability, optimizing their prevalence weights, and proposing new patients using a genetic algorithm (Figure 1). Using this approach, parameters of the QSP model and the classifier were simultaneously calibrated to trial data, stratified by the median number of prior lines of each trial result. The model predictions were then compared to the hold‐out validation data. Finally, the QSP model was used to make blind predictions for CELMoD and TCE combo therapy, reporting ORR stratified by N _prior.

FIGURE 1.

Workflow for QSP model. (1) Each virtual patient is simulated across multiple RRMM treatments in parallel; clinical responses are calculated using IMWG criteria. (2) A patient's general responsiveness to treatment is calculated by dividing the number of scenarios in which they responded by the total number of treatments they received. The histogram shows the final weighted distribution of f _resp in the virtual population. (3) Each virtual patient is associated with a prior treatment history in terms of the number of prior therapies the patient has received. Unnormalized probability for each prior line category is calculated as a linear predictor of the patient's average response rate, whose probabilities are then normalized using the softmax function. Plot shows the final N _prior assignment function. (4) Population‐level statistics are compared between real trials, stratified by the median number of prior lines, and the virtual population. Virtual population statistics are derived using a prevalence‐ and N _prior‐weighted sum of virtual patient outcomes. (5) Optimize the population prevalence weights and classifier parameters and propose new virtual patients. (1–5) repeat until convergence. (6) Validate model against hold‐out clinical response data. (7) Blind prediction of TCE + mezigdomide (1 mg days 1–21/28 starting cycle 2 day 1).

Once a virtual population is established, (degenerate) prediction intervals (PIs) for population‐level statistics are generated by sampling from this population to generate virtual trials. N patients were sampled according to the patient prevalence weights assigned during calibration, and the result of each “virtual trial” was recorded. Prediction intervals were used to compare the model's expected range of trial outcomes with actual clinical trial results, such that N matched the actual trial size of each observation. One thousand virtual trials were sampled for each comparison to data, and the inner 90% of these trial results defined the model PIs for a given output. These intervals are termed degenerate (or “plug in”) prediction intervals rather than confidence intervals because the latter require explicit estimates of parameter uncertainty, whereas the QSP model parameter distributions do not separately consider uncertainty and variability [10, 11]. Visualizations were done using the Python library Seaborn (0.13.02) [12].

3. Results

3.1. Datasets for MBMA and QSP Analyses

For the MBMA analysis of the safety endpoint Grade ≥ 3 neutropenia, 130 trials (154 treatment arms) representing 17,246 patients were included in the analysis (Table 1). For analysis of the efficacy endpoint ORR, 192 trials (239 treatment arms) were included, covering 22,227 patients (Table 1). For the QSP model, the calibration and validation datasets amounted to 123 distinct trial arms (103 for fitting, 20 for validation) across 36 distinct drug combinations (24 for fitting, 12 for validation) (Table S1). The MBMA and QSP datasets overlapped, especially for highly relevant drug combinations (Table S2). The QSP model included CELMoDs, anti‐SLAMF7 elotuzumab, and data from certain TCE (alnuctamab, talquetamab), but omitted some less common drugs that the MBMA did include (hyaluronidase enzyme rHuPH20, exportin‐1 inhibitor selinexor, anti‐CD38 mAb isatuximab).

Treatment‐specific fixed effect (EFF) parameters were estimated for 11 and 21 regimens in the Grade ≥ 3 neutropenia and ORR models, respectively. Parameter estimates (Tables S3 and S4) and Monolix model code are provided in the Supporting Information.

3.2. MBMA Identifies Combination With Alkylating Agent as Significant Covariate for Grade ≥ 3 Neutropenia

The covariates tested in the MBMA analysis of Grade ≥ 3 neutropenia are presented in Table S5. Mean age ranged from 54 to 76 with an average of 65 years. The percentage of males ranged from 36.8% to 77.1% with a mean percentage of 55.8%. The mean percentage of white races across treatment arms was 67.6%. The mean percentages of patients in ISS Stage 1, Stage 2, and Stage 3 [13] of the disease were 36.7%, 32.6%, and 27.2%, respectively. In this MBMA analysis, the effect of prior line was represented using the percent of a trial's patients who had received 1 prior line of therapy. The mean percentage of patients with only 1 prior line of therapy was 29.1%.

The final model for Grade ≥ 3 neutropenia adequately described the observed data as demonstrated by the visual predictive check shown in Figure 2A. The blue bar represents the predicted 90% interval for each treatment, and the red dot is the observed median from the data. The majority of observed medians fall well within the prediction interval, thus supporting the use of this model for simulation and benchmarking purposes. Exploratory analysis suggested dose–response trends for bortezomib and lenalidomide (not shown); however, only the inclusion of a linear dose–response relationship for lenalidomide significantly improved model fit (i.e., ∆BIC < 0). Among the covariates evaluated, the presence of a background alkylating agent was found to be a significant predictor of Grade ≥ 3 neutropenia, as shown in Figure S1A.

FIGURE 2.

Visual predictive checks (VPCs) for the final MBMA model predictions of Grade ≥ 3 neutropenia (A) and ORR (B). Red points with red horizontal lines represent observed median values across studies, with empirical 95% percentile intervals reflecting between‐study uncertainty, blue dots indicate model‐predicted median values, and blue bars indicate 90% prediction intervals across treatment regimens incorporating uncertainty due to covariate effects and random effects.

A benchmark plot was generated to compare the model‐derived incidence of Grade ≥ 3 neutropenia across key RRMM regimens (Figure 3A). Each regimen is represented by a point estimate with an associated 95% confidence interval (CI), calculated using the standard error of the estimate. The CI reflects the uncertainty in the model's fixed effect parameters. This visualization enabled direct comparison of hematologic toxicity across regimens, identifying teclistamab and isatuximab + pomalidomide with the highest predicted incidence of Grade ≥ 3 neutropenia. The combination of pomalidomide with various agents (e.g., daratumumab or isatuximab) consistently resulted in higher neutropenia risk, supporting the known hematologic toxicity of IMiD‐containing regimens [14]. Carfilzomib monotherapy had the lowest predicted neutropenia incidence, consistent with its nonmyelosuppressive profile [15]. Wider confidence intervals observed for some regimens reflect greater uncertainty due to limited trial data (Table 1).

FIGURE 3.

Benchmarking of Treatment Regimens for (A) Grade ≥ 3 neutropenia and (B) ORR using MBMA‐Derived Estimates. Estimated treatment effects with 95% CI for Grade ≥ 3 neutropenia and ORR across RRMM regimes, based on the final MBMA model. Estimates shown here are model‐derived probabilities of Grade ≥ 3 neutropenia and ORR and can be used for benchmarking treatments across safety and efficacy outcomes.

To explore the influence of background alkylating agents, a simulation‐based analysis was performed (Figure 4A). Using the MBMA‐derived parameter estimates, 1000 virtual trials were simulated for each regimen under conditions with and without alkylating agents. The predicted incidence of Grade ≥ 3 neutropenia was summarized using median values and 90% prediction intervals (PIs). All simulated doses reflected FDA‐approved schedules. Simulations showed increased Grade ≥ 3 neutropenia risk in the presence of alkylating agents but did not affect relative rankings.

FIGURE 4.

Impact of key covariates on Predicted (A) Grade ≥ 3 neutropenia and (B and C) ORR across RRMM regimens based on trial simulations. (A) illustrates the impact of background alkylating agent use, with blue and orange representing regimens without and with background alkylating agent, respectively. (B) shows the effect of prior lines of therapy: Orange indicates regimens with a higher percentage of patients in the second line of therapy (i.e., 1 prior line), while blue corresponds to regimens with patients in more than or equal to the third line of therapy (i.e., ≥ 2 prior lines). (C) displays the effect of background corticosteroid use, where blue and orange denote regimens without and with steroid use, respectively. For ORR simulations (B and C), the reference values for all other covariates were fixed at their most observed values across trials: Ptherapy1% = 12% and background steroid = 1 (representing steroid use). Each point represents the median predicted value with 90% prediction intervals (PI), based on 1000 simulations per regimen. Doses shown in parentheses represent FDA‐approved dosing regimens for each drug. All simulations were performed using schedules aligned with approved dosing.

3.3. MBMA Identifies Prior Lines and Background Steroid as Significant Covariates for ORR

The covariates evaluated in the MBMA analysis for ORR are presented in Table S6. Mean age ranged from 48 to 76 years, with an average of 64.7 years. The percentage of males ranged from 23.6% to 77.1% with a mean percentage of 55.7%. The mean percentage of white races across treatment arms was 66.0%. The mean percentages of patients in ISS Stage 1, Stage 2, and Stage 3 [13] of the disease were 35%, 31.7%, and 29.1%, respectively. The mean percentage of patients with only one prior line of therapy was 29.2%.

The final model for ORR adequately captured the observed data, as demonstrated by the VPC shown in Figure 2B. A linear relationship best described the dose–response for carfilzomib on ORR. Among the covariates evaluated, the percentage of patients with only one prior line of therapy and the use of combination therapy with steroids were identified as significant positive predictors of ORR (Figure S1B).

Figure 3B presents a benchmark plot showing point estimates and 95% CI for ORR, enabling direct comparison of efficacy across treatments. Daratumumab + lenalidomide, bortezomib + thalidomide, and isatuximab + carfilzomib combinations demonstrated the highest ORR estimates. Notably, the higher ORR for daratumumab + pomalidomide versus daratumumab + lenalidomide likely reflects differences in patient populations, as all patients in the former group had prior lenalidomide exposure, indicating a more refractory population. Regimens containing daratumumab in combination with other agents generally clustered toward the higher end of the ORR spectrum, suggesting a consistent additive benefit across backbones [16]. The wide confidence intervals for many combinations reflect greater uncertainty, likely due to limited data (Table 1).

Figure 4 (B, C) illustrates model‐predicted ORR (%) across various RRMM treatment regimens based on trial simulations (n = 1000) using median predicted ORR and associated 90% prediction intervals for each regimen. Regimens were ranked by median predicted ORR, with predictions stratified by key covariates—percentage of patients with 1 prior line of therapy (B) and background corticosteroid use (C). Regimens with a greater proportion of patients in earlier lines of therapy or with background steroid use demonstrated higher predicted ORR (rightward shift), while the relative treatment ranking remained consistent across stratifications.

3.4. QSP Corroborates MBMA Findings for ORR and Predicts CELMoD+TCE

Across both fitting and validation datasets, the fraction of trial outcomes that fell within the QSP model PIs was roughly consistent across drug combinations. Overall, 77.2% of the clinical response data (ORR, BOR, TTR, PFS, DOR, n = 1308 distinct mean targets) used to calibrate the QSP model fell within model confidence intervals, and 78.9% of the hold‐out data (n = 256 distinct mean targets) fell within the model confidence intervals (Figure 5). The real validation performance of the QSP model may be overstated because the median trial size within the validation dataset was lower than within the calibration dataset.

FIGURE 5.

Visual predictive checks for the QSP model calibration and validation. (A) Predicted versus observed outcomes for ORR, BOR, TTR, and PFS targets. Markers are distinct trial results; marker size correlates with trial size, and marker style denotes whether the model captured that trial outcome within the model's prediction intervals (O: Data in model PI, X: Data outside of model PI). (B) The fraction (x axis) of trial data that falls within model prediction intervals, grouped by drug combination (y axis). The table lists the number of targets for each drug combination and the corresponding fraction of those data that fall within model PIs. Color denotes the purpose of the data; blue: Used to fit the model, lilac: Used to validate the model predictions.

The model captured the negative relationship between the median number of prior lines and ORR (Figure 6). This result is intuitive and accords with the MBMA analysis showing that patients with fewer previous lines of therapy have a higher chance of responding to the treatment. Model validation showed accurate CELMoD combination therapy outcomes, but slightly underpredicted daratumumab and TCE combination outcomes. Running the calibration algorithm (Figure 1) several times with different starting patient samples showed low inter‐run variability (not shown), suggesting a well‐constrained system. The model showed a similar decline in disease control with increasing N_prior for other clinical outcomes (PFS) included in the calibration and validation datasets (Figure S2). Finally, the model predicted a large increase in ORR for CELMoD+TCE combination therapy for N_prior of 1 and 2–3 compared to 4+.

FIGURE 6.

ORR vs. Nprior for QSP model fitting, validation, and prediction. Each panel is a distinct drug combination and shows the %ORR (y axis) versus the prior line group (x axis). Panels contain model results for a given dose/schedule, shown as a horizontal line connected by lines when the same dose/schedule combination has data across N_prior categories. Markers are trial results; marker size correlates with trial size, and marker style denotes whether the model captured that trial outcome within the model's prediction intervals (O: Data in model PI, X: Data outside of model PI). Color denotes the purpose of the data: blue: Used to fit the model, lilac: Used to validate the model predictions, red: Model prediction with no clinical data for comparison.

4. Discussion

This study demonstrates the application of a robust MBMA framework to characterize the safety and efficacy profiles of therapies in RRMM. By systematically integrating data across hundreds of clinical trials and accounting for treatment‐ and trial‐level variability, the analysis provides quantitative insights into a highly heterogeneous therapeutic landscape where direct head‐to‐head comparisons between regimens are often unavailable.

The identification of key covariates such as prior lines of therapy and background steroid or alkylating agent use is consistent with biological and clinical expectations [17, 18, 19, 20]. For instance, patients with fewer prior lines of therapy or those receiving steroids (like dexamethasone) generally exhibited higher response rates, consistent with dexamethasone's pro‐apoptotic effects in MM cells [18], which were included as its primary MoA in the QSP model. Moreover, alkylating agents such as melphalan and cyclophosphamide are known to be myelotoxic. Their combined use with other myelosuppressive drugs increases granulopoiesis suppression [19], explaining the higher Grade ≥ 3 neutropenia risk seen with background alkylator use [20]. This mechanistic alignment supports the biological plausibility and translational relevance of our MBMA.

MBMA also offers a powerful framework for benchmarking. While this paper does not explicitly compare internal regimens, the framework enables model‐based comparisons against the standard of care with quantified uncertainty, facilitating future benchmarking and model‐informed decision making.

One of the strengths of this work lies in the dual focus on both efficacy (ORR) and safety (Grade ≥ 3 neutropenia), enabling a comprehensive benefit–risk characterization across treatment classes—particularly important in RRMM, where high response rates are often achieved at the cost of hematologic toxicity. One limitation of this analysis is the lack of consistent reporting on prophylactic granulocyte colony‐stimulating factor (G‐CSF or GCSF) [21] use across studies, which prevented its inclusion as a covariate in the model. This may bias neutropenia risk estimates in regimens where prophylactic care was administered, making them appear safer than they might be without such intervention. Future work can incorporate data on prophylactic interventions to enhance model precision and applicability.

The identification of prior lines of therapy as a significant predictor of ORR informed the QSP model design. QSP models typically rely on detailed mechanistic relationships, limiting their ability to explain known covariates that do not have clear mechanistic hypotheses. Further, explicitly simulating patient treatment history can be complex. Adding a statistical layer to the QSP model to associate a virtual patient with a “prior line” overcomes these limitations by considering the model's wide, multidrug calibration as a sufficient proxy for the behavior of real patients. This is not the first study to add a statistical layer onto a QSP model [22]; however, it differs from other approaches because the input variable represents a patient characteristic (f_resp) that holistically describes the virtual patient, rather than a 1:1 conversion between model variables. Other studies have also demonstrated utility from parallel MBMA and QSP analyses, but with limited information flow between modeling approaches [23].

The QSP model corroborated the MBMA finding that increasing numbers of prior lines of therapy result in worse outcomes. It extended the MBMA by including other important clinical endpoints (BOR, PFS) in model calibration and validation and by predicting efficacy for a novel treatment combination (TCE + CELMoD) as a function of prior line. The N_prior classifier for the QSP model had several limitations. First, the f_resp metric was derived by a simple average of response status across all simulated scenarios, instead of being weighted by the number of patients or trials corresponding to each simulation. This could lead to issues if the model is used to simulate many scenarios of one drug, but the data for that drug has a minor contribution to the overall fit due to a small trial size or few trials. Second, only simulated ORR (via f_resp) was used to determine N_prior, even though all clinical data were stratified by N_prior. In future work, including more simulated patient features could make the N_prior classifier richer. For instance, including simulated BOR and PFS in the N_prior classifier would align the set of N_prior inputs with the data types used for calibration. Similarly, including both simulated and actual biomarker data (e.g., baseline cell or tumor concentrations) would increase the totality of evidence used to define patient types. Third, it ignores the potential ways that exposure to a given drug class can affect treatment outcomes for subsequent treatments, such as exposure to one proteasome inhibitor affecting a patient's treatment outcomes to either a different proteasome inhibitor or a different drug class altogether. Stratifying clinical responses by specific treatment histories in this way would constrain the distribution of f_resp and allow the model to distinguish more mechanistic causes of nonresponse or disease progression.

In summary, this study demonstrates the utility of integrating MBMA and QSP to identify key covariates and predict treatment outcomes across a broad RRMM landscape. MBMA can be used to identify and justify the most influential clinical drivers for inclusion in platform QSP models. In turn, QSP models can semi‐mechanistically account for these significant covariates, broadening their scope and applicability in drug development. This coordinated approach enables a dynamic and data‐driven framework for exploring treatments, understanding patient heterogeneity, and simulating response under diverse clinical conditions. Together, MBMA and QSP enable more informed decisions around dose, regimen, and patient population, particularly in settings where head‐to‐head clinical trials are not feasible.

Author Contributions

Z.S., C.M.A., and J.Z. wrote the manuscript. Z.S., C.M.A., J.Z, C.H., A.V.R., and A.G.K. designed the research; Z.S., C.M.A, and J.Z. performed the research. Z.S., C.M.A., J.Z., K.D.M., C.V., and W.D. analyzed data. W.D. contributed new analytical tools.

Conflicts of Interest

Zeel Shah, Chuanpu Hu, Alexander Ratushny, and Anna G. Kondic are employees of Bristol Myers Squibb (BMS). Clifton Anderson, Kevin McCormick, Celeste Vallejo, and William Duncan are employees of Simulations Plus Inc. Jian Zhou was employed at BMS at the time of writing.

Supporting information

Data S1: psp470145‐sup‐0001‐DataS1.docx.