<p>QSP models are increasingly being applied to drug development. Calibrating virtual populations (VPops) to clinical data helps explore patient variability, study populations of interest, and predict clinical outcomes for new therapies. In previous work, we have developed the QSP Toolbox, a library of tools and workflows for developing VPops. The aim is to create a VPop that fits clinical data from a variety of sources and with a variety of data types, allowing for differences in the quality of the data. The goodness-of-fit (GOF) of the VPop is given by a scalar function <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12248_2025_1134_Article_IEq1.gif" Format="GIF" Height="12" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">$p$</EquationSource> <EquationSource Format="MATHML"><math> <mi>p</mi> </math></EquationSource> </InlineEquation> whose form is inspired by the p-value of Fisher’s combined hypothesis test. However, our VPop development workflow has been hampered by the need to perform repeated optimization of “prevalence weights” to maximize <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12248_2025_1134_Article_IEq1.gif" Format="GIF" Height="12" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">$p$</EquationSource> <EquationSource Format="MATHML"><math> <mi>p</mi> </math></EquationSource> </InlineEquation>. In this article, we describe how we can speed up our workflow by using proxy GOF functions whose maximization can be reduced to the relatively easy problem of minimizing a convex quadratic objective function. A successful run of this “proxy-guided” workflow generates a VPop with very good <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12248_2025_1134_Article_IEq1.gif" Format="GIF" Height="12" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">$p$</EquationSource> <EquationSource Format="MATHML"><math> <mi>p</mi> </math></EquationSource> </InlineEquation>-fit, but without having to continually maximize <InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12248_2025_1134_Article_IEq1.gif" Format="GIF" Height="12" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">$p$</EquationSource> <EquationSource Format="MATHML"><math> <mi>p</mi> </math></EquationSource> </InlineEquation> throughout the development process. We find that our proxy-guided workflow can calibrate VPops much faster than the original (<InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12248_2025_1134_Article_IEq1.gif" Format="GIF" Height="12" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">$p$</EquationSource> <EquationSource Format="MATHML"><math> <mi>p</mi> </math></EquationSource> </InlineEquation>-guided) workflow.</p> Graphic Abstract <p></p>

A Proxy-guided Workflow for Virtual Population Development

期刊论文
出版模式：

开放获取
发表日期：

2025年10月13日

查看全文

Lu Huang,
Yinbo Chen,
Satyendra Suryawanshi,
Amir Molavi,
Eric Sison,
Brian J. Schmidt

摘要

QSP models are increasingly being applied to drug development. Calibrating virtual populations (VPops) to clinical data helps explore patient variability, study populations of interest, and predict clinical outcomes for new therapies. In previous work, we have developed the QSP Toolbox, a library of tools and workflows for developing VPops. The aim is to create a VPop that fits clinical data from a variety of sources and with a variety of data types, allowing for differences in the quality of the data. The goodness-of-fit (GOF) of the VPop is given by a scalar function $p$ $p$ whose form is inspired by the p-value of Fisher’s combined hypothesis test. However, our VPop development workflow has been hampered by the need to perform repeated optimization of “prevalence weights” to maximize $p$ $p$ . In this article, we describe how we can speed up our workflow by using proxy GOF functions whose maximization can be reduced to the relatively easy problem of minimizing a convex quadratic objective function. A successful run of this “proxy-guided” workflow generates a VPop with very good $p$ $p$ -fit, but without having to continually maximize $p$ $p$ throughout the development process. We find that our proxy-guided workflow can calibrate VPops much faster than the original ( $p$ $p$ -guided) workflow.

Graphic Abstract