Oncology Digital Twin

Updated 29 November 2025

Oncology Digital Twin is a computational virtual replica that integrates multimodal patient data to capture tumor heterogeneity and dynamic treatment responses.
It features a three-layer framework combining data input, mechanistic and AI-driven modeling, and decision-support outputs for precision oncology.
The approach employs multiscale models—including PBPK, dosimetry, and radiobiological simulations—to provide personalized therapy planning with quantified uncertainty.

An oncology digital twin (ODT) is a computational, patient-specific virtual replica of a cancer system—encompassing tumor, host, therapeutic interventions, and dynamic physiological responses—designed to synthesize multimodal data, mechanistic modeling, and data-driven prediction in a closed-loop clinical workflow. Unlike generic digital twins, ODTs must accommodate hallmark complexities of cancer such as cellular heterogeneity, evolving genetics, spatiotemporal tumor microenvironment, and therapy adaptation. They integrate patient-specific imaging, histopathology, omics, dosimetric, and biomarker data via multiscale mathematical models and provide decision-support for prognostic and treatment-planning tasks with quantified uncertainty (Ryhiner et al., 4 Nov 2025, Zhang et al., 24 Nov 2025).

1. Architectural Principles and Data Integration

Oncology digital twin frameworks are fundamentally layered. The architecture includes three primary layers (Ryhiner et al., 4 Nov 2025, Zhang et al., 24 Nov 2025):

Data-Input Layer: Incorporates multi-parametric imaging (diagnostic PET/CT, SPECT/CT for functional and post-therapy biodistribution), histology, molecular/genomic profiles (ctDNA, mutation status), and dosimetric inputs (voxelwise time-activity curves, microdosimetric S-values).
Mathematical/Computational Core: Implements mechanistic multiscale models for pharmacokinetics (PBPK), dosimetry (Monte Carlo, S-value computation), pharmacodynamics/radiobiology (LQ and repair-kinetic models), and optimization/AI for parameter inference and fast predictions.
Decision-Support Outputs: Delivers patient-specific probabilities for tumor control (TCP) and toxicity (NTCP), optimized therapy regimens (timing, activity, isotope choice), combination therapy suggestions, and posterior confidence intervals for uncertainty quantification.

A rigorous preprocessing pipeline standardizes imaging (DICOM), clinical (HL7/FHIR), and omic data; feature-level fusion in ODTs leverages machine-learning encoders, graph neural networks, and multi-modal integrators to enable coherent model initialization (Zhang et al., 24 Nov 2025, Pandey et al., 26 Sep 2024).

2. Mathematical and Computational Modeling

ODTs employ a suite of interlocking mathematical models calibrated to individual patients (Ryhiner et al., 4 Nov 2025, Pash et al., 13 May 2025, Chaudhuri et al., 2023):

PBPK Compartmental Models: Define temporal activity profiles across organs and tumor compartments (e.g., plasma, liver, tumor interstitium, membrane-bound, internalized cytoplasm). Compartmental ODE systems govern activity transfer, receptor binding, internalization, and nuclear decay.

$\frac{dC_p}{dt} = -\sum_{i}k_{p\to i}\,C_p + \sum_{i}k_{i\to p}\,C_i - \mu\,C_p$

Dosimetry: Absorbed dose rates computed via:

$\dot D_j(t) = \sum_i A_i(t) S_{i\to j}$

For voxelwise accuracy, Monte Carlo transport simulations (Geant4-DNA, GATE) or deep-learning surrogates generate spatial dose kernels.

Radiobiological Response: Cell survival after radioisotope or external radiation is modeled with LQ or extended LQ-Lea–Catcheside models, integrating repair kinetics:

$S = \exp[-\alpha D - \beta G D^2]$

where $G$ incorporates the temporal repair kernel.

Mechanistic Multi-scale Extensions: Agent-based models (ABMs) enable representation of spatial cell–cell dynamics, mutational evolution, and immune–tumor interactions. Hybrid PDE-ABM frameworks couple reaction–diffusion tumor evolution, immune recruitment, and local agent behaviors (Zhang et al., 24 Nov 2025, Wang et al., 5 Mar 2024).
Optimization and Decision Support: Bayesian parameter estimation (MCMC, profile likelihood), uncertainty propagation, and Pareto-optimal multi-objective criteria produce personalized therapy recommendations balancing efficacy against risk (Chaudhuri et al., 2023, Pash et al., 13 May 2025).

3. Clinical Workflows and Adaptive Loop

A prototypical ODT-driven workflow consists of (Ryhiner et al., 4 Nov 2025, Banerjee et al., 30 Sep 2025):

Data Acquisition: Pre-therapy imaging (e.g., PET/CT), liquid biopsy for ctDNA, biopsy-based immunohistochemistry.
Model Calibration: Fit PK/dosimetry/radiobiology model parameters to individual multi-timepoint images and biomarker dynamics; update structural uncertainties via Bayesian assimilation.
Simulation and Plan Ranking: Forward-simulate the digital twin under alternative dosing/fractionation/timing scenarios; compute TCP, NTCP, and rank plans.
Treatment Delivery: Select and implement protocol based on multi-objective recommendation.
Monitoring and Twin Update: Iterative update of model state post-treatment via new response data (imaging, ctDNA, PFS, OS), restarting calibration and optimizing subsequent cycles.

This closed-loop adaptive paradigm underpins ODT personalization and supports real-time therapy adjustment.

4. Illustrative Applications and Case Studies

ODTs have demonstrated clinical utility across radiopharmaceutical therapy, external beam radiotherapy, and multimodal care pathways:

Theranostic Digital Twins (TDTs) for RPT: Enable individualized injected activity and isotope selection (e.g., $^{177}$ Lu-PSMA, $^{225}$ Ac) to maximize tumor ablation and minimize toxicity (Ryhiner et al., 4 Nov 2025). Multi-isotope scheduling exploits DNA-damage synergy, as shown in models coupling DSB induction with PARP inhibitor pharmacodynamics.
Risk-Aware Radiotherapy in High-Grade Glioma: Predictive digital twins integrating tumor growth ODEs with Bayesian calibration enable personalized dose schedules that outperform standard-of-care in modeled time-to-progression, supporting reductions in unnecessary toxicity (Chaudhuri et al., 2023).
Adaptive Proton Therapy: CBCT-guided digital twins reduce target margins and improve dose conformity in prostate SBRT; leveraging deep-learning DIR and uncertainty quantification supports rapid online adaptation (Chang et al., 17 Jun 2025, Chang et al., 16 May 2024).
Cognitive Digital Twins in Neuro-Oncology: IoT-integrated frameworks combine real-time EEG and MRI streaming with advanced transformer-based models to deliver interpretable, continuous monitoring and volumetric tumor kinetics prediction (Banerjee et al., 30 Sep 2025).

5. Validation, Performance Metrics, and Uncertainty

Systematic validation measures DT fidelity, predictive accuracy, and decision support reliability:

Quantitative Metrics: Dice similarity coefficients for segmentation ( $>0.85$ in DT-based reconstructions), prediction errors for tumor volume ( $<10\%$ at 3-month follow-up), and AUC for survival predictions ($0.8$–$0.9$) (Zhang et al., 24 Nov 2025, Pash et al., 13 May 2025).
Uncertainty Quantification: Bayesian and Laplace-approximation approaches provide credible intervals for predicted outcomes, supporting risk-calibrated clinical decisions and optimal experimental design (Chaudhuri et al., 2023, Pash et al., 13 May 2025).
External Evaluation: Cross-validation with retrospective/prospective clinical cohorts, simulation-based virtual trials, and in vivo measurement agreement are employed (e.g., dosimetric errors $<5\%$ ).

Advanced surrogate modeling enables high-throughput uncertainty propagation for real-time optimization without loss of model transparency (Bhattacharya et al., 29 Sep 2025, Panneerselvam et al., 30 Aug 2025).

6. Challenges and Future Directions

Outstanding technical, biological, and regulatory barriers remain:

Data Heterogeneity and Interoperability: Lack of standardized dynamic imaging/genomic data formats and ontologies complicates data integration. Unified interoperability standards are required (Ryhiner et al., 4 Nov 2025).
Model Generalizability and Biological Complexity: Tumor and patient heterogeneity, limited alpha-emitters for multi-isotope RPT, and sparse validation datasets constrain widespread deployment. Multi-center, in silico trials and VVUQ (Verification, Validation, Uncertainty Quantification) are priorities (Ryhiner et al., 4 Nov 2025).
Computational Barriers: Whole-organ Monte Carlo and large-scale ABMs remain resource-intensive; GPU-acceleration and deep-learning surrogates mitigate but require careful V&V.
Explainability and Trust: Clinician-facing transparency via explainable AI layers and explicit decision pathway reporting are essential for adoption (Pandey et al., 26 Sep 2024, Zhang et al., 24 Nov 2025).
Regulation and Ethics: Data privacy (federated learning, encryption), auditable model chains, and clear digital governance frameworks will underpin safe clinical transition.

Future research converges on "biologically informed" DTs integrating real-time microenvironmental, immunological, and repair data, together with federated learning platforms able to balance privacy, explainability, and generalization (Ryhiner et al., 4 Nov 2025, Wang et al., 5 Mar 2024).

7. Summary Table: Principal Components of Radiopharmaceutical Oncology Digital Twins

Layer	Example Components	Methods/Tools
Data Input	PET/SPECT/CT, histology, ctDNA, time–activity curves	DICOM, HL7/FHIR, S-values
Pharmacokinetics (PK)	PBPK, TMDD compartmental ODE systems	Stiff ODE solvers, Bayesian estimation
Dosimetry	Monte Carlo, voxel S-value computation	Geant4-DNA, GATE, GPU kernels
Radiobiology	LQ, LQ+Lea–Catcheside, MEDRAS repair models	Mechanistic DNA damage ODEs
Optimization & AI	Bayesian/MCMC calibration, machine-learning surrogates	Gaussian process, deep learning
Decision Support Outputs	TCP, NTCP, optimal schedule and isotope selection	Multi-objective ranking, confidence bounds

Successful clinical implementation of ODTs will continue to depend on robust integration of these components, rigorous validation, and cross-disciplinary collaboration spanning nuclear medicine, computational modeling, systems pharmacology, and regulatory science (Ryhiner et al., 4 Nov 2025, Zhang et al., 24 Nov 2025).