Digital Twin–Enabled Prediction of Toxicity and Tumor Control in Prostate Cancer Radiotherapy
Principal Investigator
Name
Samantha Rincon Sabatino
Institution
InfotechSoft
Position Title
Bioinformatics Scientist
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCOI-2075
Initial CDAS Request Approval
Aug 24, 2026
Title
Digital Twin–Enabled Prediction of Toxicity and Tumor Control in Prostate Cancer Radiotherapy
Summary
Despite excellent cure rates, the growing population of prostate cancer survivors faces a significant burden of long-term treatment-related morbidity, making prostate cancer the leading contributor to cancer treatment-related years lived with disability worldwide. More than 20% of men experience persistent late gastrointestinal, genitourinary, or sexual toxicities beginning months after radiotherapy, often resulting in lifelong reductions in quality of life. Yet clinicians currently select among competing radiation treatment plans using primarily dosimetric metrics, with little ability to incorporate individualized biological susceptibility or accurately predict patient-specific tumor control and toxicity before treatment. To address these limitations, we propose the Multi-scale Adaptive Simulation for Clinical Oncology and Treatment (MASCOT) Radiotherapy Digital Twin (DT), a high-fidelity, multimodal modeling framework that generates individualized tumor control probability (TCP) and normal tissue complication probability (NTCP) predictions and integrates these estimates into treatment planning workflows to guide plan comparison, optimization, and selection. Building on our prior work using graph neural networks to model disease progression, MASCOT integrates longitudinal clinical data, multimodal imaging, radiotherapy dosimetry, genomics, and transcriptomics within a temporally resolved knowledge representation. The project will engage experts in radiation oncology, genomics, radiomics, and prostate cancer clinical trials to assemble curated datasets, develop predictive models, define workflow-driven use cases, and demonstrate feasibility in RT planning. Phase I of this project aims to establish technical feasibility and single-institution validation. We will define system architecture and clinical use cases through multidisciplinary expert engagement and design workflow-integrated user interfaces supporting personalized toxicity forecasting, followed by formal usability testing. An automated, ontology-driven bioinformatics pipeline will harmonize multimodal prostate cancer data into an extensible temporal knowledge graph by extending mCODE to incorporate radiomics, RT dosimetry, and toxicity. Using the single-institution GenePARE cohort, we will validate semantic coverage, data fidelity, and interoperability. We will develop modality-specific feature representation models, including autoencoders for genomic/transcriptomic data and CNN-based architectures for imaging, radiotherapy, and histopathology, to learn biologically meaningful embeddings, which will be pre-trained on a subset of PCa datasets in Phase I and evaluated across multiple outcomes including toxicity, recurrence, and survival. Phase II will fuse these pre-trained models to develop a graph neural network–based DT prototype to model longitudinal patient trajectories and predict TCP and NTCP, with performance evaluated on a blinded REQUITE test set. Phase II will also advance the validated prototype into a production-ready, multi-institutional clinical decision support platform. We will refine workflow integration, expand data harmonization to additional datasets, and enhance the DT architecture through improved feature engineering, uncertainty modeling, and robustness testing, with external validation using the REQUITE dataset. A structured pilot study will benchmark MASCOT against conventional NTCP/TCP models and evaluate clinical plausibility, plan quality, expert acceptance, and usability to establish deployment readiness.
Aims
Phase I Aims: Demonstrate that MASCOT’s multimodal data integration framework and predictive modeling architecture are technically feasible and ensure clinical workflows are validated by experts.
I.1: Build and verify a standards-based multimodal data integration and knowledge modeling pipeline using a single-institution cohort. Develop an ontology-driven bioinformatics pipeline to harmonize GenePARE multimodal data, perform feature extraction and uncertainty estimation, and instantiate a temporal knowledge graph storing annotated features with provenance and traceability. Success Criteria: Interoperability (TBox compatible with Protégé); Audited GenePARE sample achieves: >95% semantic coverage; >98% data fidelity; >95% feature provenance.
I.2. Develop multimodal transformer frameworks to quantify predictive value of integrated radiogenomic representations. Develop transformer-based multimodal models that learn representations from radiomic, genomic, histopathology, and radiotherapy data. Evaluate toxicity, recurrence, and survival prediction and compare integrated radiogenomic models with radiomics- and genomics-only models using a blinded TCGA-PRAD test set. Success Criteria: Integrated radiogenomic model demonstrates improved (ΔAUC) for prediction of late Grade ≥2 GI/GU toxicity compared with unimodal models.
I.3. Prototype UI for three high-value radiotherapy workflows and establish usability. Engage a multidisciplinary expert panel to refine RT workflows, develop workflow-integrated user interfaces supporting personalized RT outcome forecasting, and evaluate usability and feasibility. Success Criteria: Finalized use cases reviewed by the team; Mean System Usability Scale (SUS) of ≥70 with representative users.
Phase II Aims: Transform the Phase I prototype into a production-ready MASCOT CDS platform with demonstrated multi-institution generalizability, validation, and evidence supporting deployment readiness.
II.1: Scale the multimodal data integration and knowledge modeling pipeline to multi-institution cohorts. Expand the ingestion and harmonization pipeline to the REQUITE multi-institutional cohort, addressing inter-site variability, quality, and batch effects across sites. Success Criteria: Audited sample of REQUITE achieves: >95% semantic coverage; >98% data fidelity; >95% feature provenance.
II.2: Enhance the DT with advanced feature engineering, uncertainty modeling, and multi-institution validation. Expand the DT architecture to support the REQUITE cohort, integrate RT workflow interfaces, and evaluate predictive performance on a blinded REQUITE test set. Benchmark MASCOT against published NTCP models (LKB, logistic regression) and Poisson TCP models for late rectal toxicity, urinary toxicity, sexual dysfunction, and local tumor control. Success Criteria: MASCOT significantly outperforms conventional dosimetric and clinical NTCP/TCP models on blinded validation of one primary endoint.
II.3: Conduct a pilot study benchmarking clinical performance and validating RT workflows for deployment readiness. Conduct a structured pilot study evaluating DT-guided RT planning through expert review and OpenTPS plan assessment. Success Criteria: a) ≥80% agreement between DT-supported and expert plan selection; Mean SUS score of ≥70; improved clinician confidence without increased review time.
Collaborators
Samantha Rincon Sabatino InfotechSoft
Thomas Taylor InfotechSoft
Ray Bradley InfotechSoft