Comparing Multiple Virtual Population Generation Approaches using as a base, a QSP Model of Rheumatoid Arthritis.

Dinesh Bedathuru (1), Tanvi Joshi (1), Goutam Nair (1), Aijaz Shaliban (1), Prakash Packrisamy (1), Rukmini Kumar (1)

(1) Vantage Research, Delaware, USA

Introduction: Virtual cohort and Virtual Population (VPop) generation are important outputs in the QSP process that attempt to capture the statistics of clinical trial data, including inter-individual variability across the clinical population. Virtual populations are calibrated to match the mean, standard deviation & other statistics of clinical outcomes (baseline and post-therapy) identified from clinical trials relevant to model design. Virtual Populations that are a good fit build confidence in the model’s predictions. 

Different approaches have been proposed in the QSP literature to generate virtual populations. We evaluate and compare these approaches on relevant metrics, using the same model as base. Finally we use VPops generated by the different approaches to predict clinical outcomes.

Objectives: 

1. To generate VPops for a QSP model of RA, using the following approaches.
1. Prevalence weighting (via gQSPsim [1])
2. Selection using a Genetic Algorithm (via VQM Tools [2])
3. Allen’s Method [3]
4. MAPEL algorithm (via QSP Toolbox) [4]
2. To compare these approaches across relevant Virtual population quality metrics as described in the methods section.

Methods: 

1. Model Development: An ODE based QSP model was developed to include 9 cell types, 17 cytokines and 42 ODEs(including PK, physiology etc.). The model includes DAS28_CRP and ACR as the clinical scores. The model was developed in MATLAB Simbiology.
2. Virtual Cohort generation: A virtual cohort of size ~20k is generated by introducing variability to select parameters and filtered such that in the untreated (baseline) conditions, they are 
1. At steady-state in the untreated condition 
2. Within  variability in cell, cytokines & clinical score  observed in public literature. 

This step is common for all approaches & the VPop is selected from this cohort of plausible Virtual Patients. 

1. Virtual Population generation: Using the above mentioned approaches, VPops were generated to be consistent with the observed phase 3 trial outcomes for Methotrexate [5], Adalimumab [6] and Tocilizumab [7] therapies. The analysis was done using both tools shared by the authors and/or scripts developed by us
2. VPopMetrics: VPops were compared using the following metrics
1. Goodness of fit: This considers how efficient the approach is in minimizing the objective function.
2. Computational efficiency: This considers the time taken for the algorithm to produce the virtual population. 
3. Diversity within a Vpop: This considers the parameter and outcome distributions within the generated virtual population.
4. Ability to provide diverse Vpops: Generation of diverse Vpops helps evaluate the confidence in the predictions made. The capability of the approaches in providing these, is evaluated by capturing the parameter variability & correlations. 
3. Prediction of a predetermined scenario: The goal here was to assess if the VPops generated from these approaches can lead to a similar predictive outcome. For this, we compared the predicted efficacy of Adalimumab on a Tocilizumab Non-responder subpopulation and vice-versa.

Results and Conclusions: This exercise provides a comparison of VPop approaches available in public space, and provides an analysis of the benefits/drawbacks of each. Approaches 2 and 3, provide the ability to generate multiple VPops, while requiring similar computational resources, and are therefore important in providing confidence intervals to the predicted population level outcomes (e.g. % remission). 

An important conclusion was that outcomes predicted by the different VPop generation methods were within an acceptable range  based on several key characteristics, improving confidence in the publicly available VPop methods.

References:
[1] PMID: 31957304 PMCID: PMC7080534 DOI: 10.1002/psp4.12494
[2] https://www.mathworks.com/matlabcentral/fileexchange/71942-vqmtools
[3] PMID: 27069777 PMCID: PMC4809626 DOI: 10.1002/psp4.12063
[4] PMID: 28540623 DOI: 10.1208/s12248-017-0100-x
[5] PMID: 10573044 DOI: 10.1001/archinte.159.21.2542
[6] PMID: 22562973 PMCID: PMC3551224 DOI: 10.1136/annrheumdis-2011-201247
[7] PMID: 21949007 DOI: 10.1136/ard.2010.148700

Reference: PAGE 31 (2023) Abstr 10638 [www.page-meeting.org/?abstract=10638]

Poster: Methodology - Other topics