From virtual patients to digital twins in immuno-oncology: lessons learned from mechanistic quantitative systems pharmacology modeling

Hanwen Wang, Theinmozhi Arulraj, Alberto Ippolito, Aleksander S. PopelNpj Digit. Medicine 2024 · DOI

TL;DR

This paper explores the challenges and opportunities in creating "virtual patients" and "digital twins" for immuno-oncology research. By leveraging mechanistic modeling, these computational tools aim to accelerate drug development and improve patient outcomes, with the authors sharing their experiences and insights from developing these sophisticated models.

What Problem Does This Paper Address?

Cancer drug development, particularly in immuno-oncology, faces a significant hurdle: low success rates in clinical trials. This is exacerbated by the complexity of the immune system's interaction with cancer and the vast number of potential drug combinations, doses, and biomarkers to test. Traditional clinical trials are expensive, time-consuming, and often fail to identify effective treatments. To address this, researchers are turning to computational approaches like "virtual patients" and "digital twins." Virtual patients are simulated populations that mimic real patient characteristics, allowing for in silico testing of therapies. Digital twins are even more personalized, aiming to represent an individual patient's biological state to optimize treatment. However, generating these models, especially for the intricate field of immuno-oncology, is challenging due to data limitations, complex biological mechanisms, and inter-patient variability. This paper aims to discuss these challenges and propose ways forward.

Methods

The authors utilize Quantitative Systems Pharmacology (QSP) models, which are mechanistic mathematical models that integrate pharmacokinetic, pharmacodynamic, and systems biology information. Specifically, they discuss the development of Quantitative Systems Pharmacology for Immuno-Oncology (QSP-IO) models.

Key methodological aspects include:

  • Mechanistic Modeling: Building models based on known biological mechanisms of cancer and the immune system, rather than purely data-driven approaches.
  • Virtual Patient Generation:
    • Parameterization: Defining physiologically plausible ranges for model parameters, often informed by experimental and clinical data.
  • Latin Hypercube Sampling: A method for randomly sampling parameter values to generate a diverse virtual patient population.
  • Probability of Inclusion: A technique to select virtual patients whose characteristics statistically match real patient data, using information from public databases like iAtlas.
  • Integration of Multi-omics Data: Incorporating data from genomics, transcriptomics, and digital pathology to better represent tumor microenvironment heterogeneity and patient-specific features.
  • Agent-Based Modeling (spQSP-IO): Integrating spatial aspects into QSP models to capture the spatial heterogeneity within tumors and their microenvironment.
  • Model Calibration: Iteratively adjusting model parameters by comparing simulation outputs to clinical measurements, a process that can be time-consuming due to the non-linear nature of the models.

Datasets

The paper references the use of several public datasets for model calibration and validation:

  • iAtlas database: Used for immune cell subset ratios in real patient data (lung adenocarcinoma and squamous cell carcinoma).
  • RNA-seq data on breast cancer metastases: Generated by Siegel et al. and used to inform virtual patient generation for metastatic breast cancer.
  • Multiplex digital pathology and genomic analysis data: Used to inform QSP-IO models.

The paper does not introduce new benchmark datasets but rather leverages existing public resources.

Key Insights and Ideas

  • Immuno-oncology's Complexity is a Major Hurdle: The intricate interplay of the immune system with cancer, including the mechanism of action of new therapeutics, is often not fully understood, making mechanistic modeling particularly challenging.
  • Heterogeneity is King (and a Pain): Inter-patient, inter-tumoral, and intra-tumoral heterogeneities in cancer require extensive clinical datasets to accurately represent virtual patients.
  • Multi-omics Data is Crucial: Emerging multi-omics data holds promise for resolving the challenge of representing patient heterogeneity by providing detailed molecular characterization of the tumor microenvironment.
  • Virtual Patients vs. Digital Twins: While related, virtual patients are often more generic and used for population-level predictions (e.g., drug efficacy), whereas digital twins are highly personalized, aiming to optimize individual patient care.
  • Public Databases are Invaluable: Resources like iAtlas, which contain correlated patient-level data, are essential for generating physiologically plausible virtual patients and validating models.
  • The "Probability of Inclusion" Method: This approach offers a statistically sound way to select virtual patients that closely resemble real patient populations, improving model predictive power.
  • Bridging the Gap: The authors highlight the need for greater collaboration and research to apply methods proven in other fields (like cardiovascular diseases) to immuno-oncology, and vice versa, to advance both virtual patient and digital twin development.

Key Results

  • The paper demonstrates the successful application of QSP-IO models to analyze anti-PD-L1 treatment in non-small cell lung cancer and predict the effectiveness of masked antibodies in triple-negative breast cancer.
  • By using the "Probability of Inclusion" method with iAtlas data, the authors showed that the distributions of immune cell subset ratios in their virtual patient cohort were statistically similar to those in real patient data.
  • The study highlights that while correlations between immune cell subset ratios in virtual patients can be similar to real data, differences can arise due to outliers and selection biases, underscoring the importance of careful patient selection and model validation.
  • The integration of RNA-seq data and deconvolution methods allowed for the characterization of immune cell proportions in metastatic lesions, informing virtual patient generation for metastatic breast cancer.

Limitations and Future Work

  • Data Scarcity: A significant limitation for both virtual patient and digital twin development is the lack of sufficient, high-quality patient data, especially longitudinal time-course data.
  • Mechanistic Understanding Gaps: In immuno-oncology, the precise mechanisms of action for many novel therapeutics are still being elucidated, which limits the completeness of mechanistic models.
  • Computational Cost: Calibrating QSP models and generating large virtual patient populations can be computationally intensive.
  • Future Work:
    • Continued research into applying and adapting methods from other therapeutic areas to immuno-oncology.
  • Developing more robust methods for generating realistic virtual patients from digital twins and vice versa, fostering collaboration between these fields.
  • Leveraging new methods to generate virtual patients and digital twins, particularly in fields where data is sparse.
  • Improving the applicability and robustness of existing methods in immuno-oncology.

Summary updated

BibTeX
@article{wang2024from,
  author = {Hanwen Wang and Theinmozhi Arulraj and Alberto Ippolito and Aleksander S. Popel},
  title = {{From virtual patients to digital twins in immuno-oncology: lessons learned from mechanistic quantitative systems pharmacology modeling}},
  journal = {Npj Digit. Medicine},
  year = {2024},
  doi = {10.1038/s41746-024-01188-4},
}
Figure