OmniBioTwin: A System-of-Twinned-Systems Framework for Health Digital Twins

Technical Summary: OmniBioTwin: A System-of-Twinned-Systems Framework for Health Digital Twins

Problem Statement
Health Digital Twins (HDTs) offer a paradigm for patient-specific modeling and decision support in precision medicine. However, current implementations are largely monolithic, designed around single organs, specific biological mechanisms, or isolated clinical endpoints. While effective for narrow tasks, these frameworks struggle to capture the complexity of systemic diseases like Alzheimer's Disease (AD), which involve inter-organ communication, cross-scale dependencies (molecular to organismal), and heterogeneous biological processes evolving on disparate timescales. Existing engineering approaches, such as System-of-Systems (SoS) digital twins, have not been adequately adapted to biomedical modeling, which requires handling biological scale heterogeneity, sparse asynchronous observations, and the need for mechanistic interpretability.

Methodology: The OmniBioTwin Framework
The authors propose OmniBioTwin, a System-of-Twinned-Systems (SoTS) framework that reimagines HDTs as modular computational entities coupled through explicit interaction operators within a multi-layer network architecture.

1. Architectural Abstraction (SoTS):

   • Instead of a single end-to-end model, the system treats each digital twin as an autonomous subsystem representing a specific biological component (e.g., an organ, cell type) or process.
   • The framework utilizes a multi-layer network abstraction where nodes represent entities/processes and edges represent interactions.
   • Intra-layer edges capture interactions at similar scales, while inter-layer edges encode vertical cross-scale coupling.
   • Hypernodes are introduced to model collective behaviors involving many-to-many interactions (e.g., inflammatory signaling involving multiple cell types), grouping multiple twins into composite interaction units.

2. Seven-Layer Architecture:
   The framework is organized into seven coordinated layers to integrate heterogeneous data, multi-paradigm models, and decision processes:

   • Data Layer: Aggregates heterogeneous measurements (omics, imaging, EHRs, wearables) while preserving scale, temporal structure, and provenance. It outputs twin-specific input bundles containing processed observations, temporal metadata, uncertainty, and intervention records.
   • Twin Layer: Contains the core computational units. Each twin ($T_i$) is an autonomous unit with its own state ($S_i$), transition function ($f_i$), and uncertainty representation ($U_i$). Twins can utilize different modeling paradigms (mechanistic, stochastic, ML-based) appropriate to their subsystem.
   • Coupling Layer: Formalizes how twins influence one another. It applies coupling operators ($C_{j \to i}$) to transform source twin states into representations admissible for target twins. These operators handle cross-scale translation, semantic alignment, unit conversion, and uncertainty propagation, generating aggregate coupling messages ($\Gamma_i$).
   • Synchronization Layer: Provides temporal orchestration for sparse and asynchronous biomedical data. It determines the scheduling of updates ($\delta_i$) based on dependency signals and temporal metadata, coordinating processes that evolve on unequal timescales.
   • Decision Layer: Acts as the system-level inference engine. It transforms synchronized global states and uncertainties into actionable outputs, including risk estimates, treatment-response assessments, and candidate interventions.
   • Interaction Layer: Serves as the human-in-the-loop interface, rendering model behavior interpretable for clinicians and researchers. It allows users to specify constraints, review candidate actions, and approve interventions.
   • Audit Layer: Ensures traceability and governance by aggregating provenance records, state trajectories, coupling traces, and decision logs into structured audit reports for validation and regulatory assessment.

Case Study: GLP-1 Pathways in Alzheimer's Disease
To demonstrate the framework, the authors instantiated a multiscale twin for Glucagon-like peptide-1 (GLP-1) signaling pathways in AD. This case study connects four distinct twins:

• Peripheral Twin: Constructed from treatment records and peripheral biomarkers (glucose, insulin, HbA1c, BMI) to characterize systemic drug exposure and metabolic response.
• Molecular Twin: Integrates brain-directed signaling drives with molecular observations (CSF/plasma biomarkers, PET imaging) to reconstruct latent molecular pathology (Aβ burden, tau phosphorylation, neuroinflammation, brain insulin resistance).
• Cellular Twin: Models the state changes and interactions of brain cell populations (neurons, astrocytes, microglia) based on molecular pressures, outputting functional integrity and cellular vulnerability.
• Organ Twin: Combines cell-level summaries with neuroimaging and cognitive testing data to model macroscopic disease progression, including network dysfunction and cognitive decline trajectories.

These twins are coupled to simulate how GLP-1 receptor agonists (GLP-1 RAs) propagate from systemic metabolism to molecular, cellular, and organ-level effects, closing the loop with clinical decision support.

Key Contributions

• Conceptual Shift: Proposes the SoTS paradigm for biomedical HDTs, moving away from monolithic models toward modular, coupled systems.
• Structural Framework: Introduces a seven-layer architecture that explicitly separates data integration, autonomous modeling, cross-scale coupling, temporal synchronization, and human-in-the-loop decision support.
• Mechanism for Integration: Defines coupling operators and hypernodes to manage cross-scale translation and collective biological interactions while preserving the autonomy and interpretability of individual components.
• Proof of Concept: Demonstrates the instantiation of a multiscale AD twin, illustrating how heterogeneous data sources (omics, imaging, EHRs) and modeling strategies can be unified within a single system.

Results and Claims
The paper does not present empirical results comparing OmniBioTwin against monolithic models in a clinical trial or retrospective study. Instead, the "results" are the successful instantiation of the framework in the AD/GLP-1 context. The authors claim that this architecture:

• Supports modularity, allowing new twins to be incorporated without redesigning the entire system.
• Enhances interpretability by maintaining distinct modeling strategies for different subsystems and explicit coupling mechanisms.
• Provides a scalable foundation for integrating heterogeneous data and multi-paradigm models.
• Enables closed-loop clinical management where interventions are proposed, approved, applied, and re-evaluated based on updated observations.

Significance
The paper argues that OmniBioTwin addresses the critical gap between isolated organ models and the complex, multiscale reality of systemic diseases like AD. By adapting the SoS paradigm to biomedical modeling, the framework offers a path toward generalizable, patient-specific system representations that can handle the heterogeneity of biological scales and data types. The authors position this work as a necessary step toward future HDTs that can support precision medicine through faithful computational representations of distributed human disease processes, while maintaining the rigor required for biomedical validation and regulatory assessment.