A Multi-Scale, Evidence-Orchestrated Digital Twin Architecture for Therapeutic Development: A Systems-Theoretic Framework

Ian Jenkins¹, Krista Casazza², Vaishnavi Narayan¹, Waldemar Lernhardt¹, Valentina Savich¹, Eric Mathur¹, Robert Tinder¹, Jayson Uffens¹ and Jonathan RT Lakey^1,2*

Citation: Jenkins I, Casazza K, Narayan V, Lernhardt W, Savich V, et al. A Multi-Scale, Evidence-Orchestrated Digital Twin Architecture for Therapeutic Development: A Systems-Theoretic Framework. Adv Clin Med Res. 7(2):1-10.

Received: April 22, 2026 | Published:  May 10, 2026

Copyright^© 2026 Genesis Pub by Jenkins I, et al. This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International License (CC BY 4.0). This license permits unrestricted use, distribution, and reproduction in any medium, provided the original author(s) and source are properly credited.

DOI: https://doi.org/10.52793/ACMR.2026.7(2)-118

Abstract

Digital twins in biomedicine aim to construct computational counterparts of biological systems capable of simulating perturbations, forecasting outcomes, and informing intervention strategies. Despite substantial advances, many current implementations remain constrained by narrow modality integration, limited mechanistic grounding, or insufficient multi-scale coupling. Here, we present a systems-theoretic digital twin architecture grounded in multiomic integration, distributed probabilistic modeling, constraint-based reasoning, reinforcement learning, and neural inference, as instantiated within the Operon™ computational platform. We formalize a multi-scale biological twin spanning molecular, cellular, tissue, and organism-level representations, integrated through structured normalization layers and scenario-based simulation. The framework emphasizes (i) systems-level inference over single-target correlation models, (ii) explicit uncertainty propagation, (iii) traceable evidence synthesis, and (iv) validation under blinded evaluation conditions. We describe architectural principles, mathematical underpinnings, model orchestration strategies, and validation methodology. Independent blinded evaluation demonstrated high concordance under controlled conditions. We discuss implications for digital twin design in therapeutic discovery and outline future research directions for adaptive, patient-specific instantiation.

Keywords

Multi-Scale; Evidence-Orchestrated; Digital Twin Architecture; Therapeutic; Systems-Theoretic Framework.

Introduction

Digital twins are computational counterparts of physical systems that are continuously (or periodically) synchronized with empirical observations [1]. To date, digital twins have become a unifying paradigm in engineering and manufacturing because they enable prospective reasoning [2]. That is, digital twins facilitate prediction under interventions, optimization under constraints, and uncertainty-aware decision support [3]. In medicine, conceptual translation is compelling but substantially more difficult. The “physical system” is a multi-scale, adaptive, partially observed biological organism whose dynamics are driven by nonlinear feedback, context dependence, and substantial inter-individual variability [4]. Recent perspectives and reviews emphasize that medical digital twins (MDTs) should be understood not as a single model, but as a system that integrates evolving data, mechanistic and statistical models, and a defined interface to support prediction and decision-making.

Precision (i.e., personalized) medicine seeks to tailor prevention and therapy by integrating molecular profiles, clinical phenotypes, environment, and longitudinal response [5,6]. Digital twins are increasingly positioned as a technical construct capable of operationalizing this ambition because they can represent patient state trajectories and evaluate “what-if” interventions in silico [7]. As such, MDT link mechanistic plausibility with patient-specific data assimilation. In contrast to purely associative risk scores, a digital twin aspires to simulate downstream consequences of perturbations (e.g., drug exposure, dosing regimens, combination therapies, or eligibility criteria) and to quantify uncertainty around those predictions [8]. The field has expanded rapidly. Contemporaneously, mapping and scoping reviews document accelerating publications, increasing patent activity, and a diversification of use cases spanning cardiovascular disease, oncology, diabetes, nephrology, and beyond. Despite this momentum, there remains limited consensus on (i) the minimal components of an MDT, (ii) how to integrate heterogeneous model classes without collapsing into a black box, and (iii) what constitutes adequate validation for clinical or development-facing decisions. Recent policy and framework articles propose definitional decompositions. For example, the patient, data connection, patient in silico, interface, and synchronization highlight that successful MDTs must be engineered as closed-loop systems rather than static predictive models.

Therapeutic discovery and development present an extreme testbed for MDT methodology. The core inferential problem is to predict how molecular perturbations propagate through biological hierarchies to generate organism-level phenotypes, including both intended efficacy and unintended off-target and safety liabilities [9]. This is not merely a data sparsity issue. Rather, the issue resides in structural identifiability. Multiple mechanistic routes can yield similar observed phenotypes, and most clinical measurements are indirect, noisy, and temporally sparse [10]. Consequently, “target-centric” or single-modality approaches often fail to capture emergent properties that arise from network interactions, compensatory pathways, tissue microenvironments, and patient heterogeneity. These limitations are repeatedly underscored in contemporary discussions of MDTs for precision medicine.

In addition, drug development decisions must be made under high uncertainty and high stakes. Candidate selection, dosing, trial design, enrichment strategies, endpoint selection, and risk–benefit tradeoffs each require integrating diverse evidence streams, including (but not limited to) preclinical, -omics, clinical, and real-world [11]. This is one reason why the MDT concept has become tightly coupled with the resurgence of in silico trials and model-informed drug development. Computational “twins” can be used to explore counterfactual strategies and quantify sensitivity to assumptions before committing to expensive and irreversible experimental pathways [12]. Notwithstanding the foregoing, a rigorous therapeutic-development digital twin must assimilate heterogeneous data across scales and contexts: genomics, transcriptomics, proteomics, metabolomics, epigenetics, curated biological knowledge, preclinical models, clinical trial data, and real-world evidence. The scientific rationale is that disease and drug response are distributed phenomena. Thus, causality is often mediated by regulatory and signaling networks, metabolic and immune states, tissue-level organization, and systemic physiology. The practical challenge is that these data are heterogeneous in structure, resolution, measurement error, batch effects, missingness, and temporal sampling. Reviews emphasize that MDTs will only be clinically useful if they can robustly integrate multi-modal data while preserving biological plausibility and avoiding spurious correlations. This yields a central architectural requirement. That is, a normalization and representation layer that transforms disparate data modalities into coherent internal representations suitable for cross-domain inference [13]. Without this layer, MDT systems devolve into pipelines that are brittle, non-transferable across contexts, and difficult to validate because the mapping between raw data and model state is ill-defined. The need for principled representation is also heightened by increasing interest in generative and foundation-model approaches for MDT which promise flexible synthesis and imputation but amplify concerns about traceability, causal validity, and failure modes.  A key theme in recent MDT scholarship is that no single modeling paradigm is sufficient across the full therapeutic-development stack. Mechanistic models offer interpretability and extrapolation under interventions but can be parameter-hungry and incomplete; machine learning models can capture complex patterns but may fail under distribution shift and may lack mechanistic interpretability; probabilistic models naturally represent uncertainty but can be computationally expensive at scale. Accordingly, the emerging consensus is that high-impact MDTs will require hybridization, i.e., ensembles or orchestrations of complementary model classes, paired with governance mechanisms to ensure consistency and interpretability [14].

Within this scientific context, GATC has developed Operon™ designed explicitly as an internal scientific computing platform that orchestrates multiple computational approaches (e.g., probabilistic inference, constraint-based modeling, reinforcement learning strategies, and neural pattern recognition) rather than relying on a single model class. Equally important, Operon is multi-scale by design, supporting modeling from molecular interactions through pathway and tissue dynamics to whole-organism phenotypic outcomes, with the aim of assessing how therapeutic perturbations propagate through complex biological systems. This theoretical posture aligns with the direction of MDT thought leadership: the “twin” is a structured inference system that must reconcile heterogeneous signals and remain robust under uncertainty. MDTs become materially useful when they shift from passive prediction (“what will happen?”) to counterfactual reasoning (“what would happen if we intervened differently?”). In drug development, this includes simulating dose–response, inclusion/exclusion criteria, endpoint selection, and patient subgroup enrichment. Recent work on in silico trials and medical digital twins suggests that scenario-based simulation can enhance the efficiency of clinical innovation by allowing investigators to predefine and stress-test therapeutic strategies, reduce design-related risks in clinical trials, and systematically evaluate the biological and contextual conditions under which an intervention is likely to succeed or fail [15].

Operon explicitly supports scenario-based simulation to evaluate how variations in biological assumptions, patient characteristics, and intervention strategies influence outcomes, with the stated objective of informing expert-led study design decisions upstream of protocol authoring and regulatory documentation. A recurring barrier to MDT adoption is that increasingly complex models can reduce transparency, weaken trust and limit clinical or development uptake. Contemporary reviews emphasize interpretability, auditability, and the ability to communicate uncertainty as foundational (not optional) characteristics for medical digital twins. This creates a strong rationale for explicit evidence synthesis layers that aggregate heterogeneous model outputs into structured decision frameworks. Operon incorporates structured evidence synthesis workflows that aggregate signals across modeling outputs into interpretable scoring and grading frameworks designed for comparative assessment and expert review.  In the MDT framing, such a layer functions as a “decision interface” that bridges complex internal inference with human-in-the-loop governance, helping ensure that outputs remain traceable to underlying evidence rather than being treated as unanalyzed algorithmic verdicts. Finally, the impact potential of MDTs depends on validation. Recent scoping and challenge-focused papers highlight the need to distinguish (i) technical validation (does the system compute as intended?), (ii) predictive validation (does it forecast outcomes in held-out or prospect; iii) clinical/development utility (does it improve decisions compared to standard practice?). The performance of Operon was assessed using a held-out test subset from the original dataset, reserved exclusively for evaluation while the remaining data were used for model training, and further examined through independent blinded academic review; under these controlled conditions, the framework demonstrated high concordance (true positive rate 86%, true negative rate 91%), while appropriately acknowledging that such validation metrics reflect performance within the evaluation dataset and do not guarantee prospective clinical success. These details motivate a methodological contribution. Contemporaneously, MDT architecture is designed so that validation is not an afterthought, but a first-class requirement supported by traceable workflows and controlled evaluation designs.

Systems-Theoretic Foundations of the Biological Digital Twin Biological System Representation

A biological digital twin can be understood as a dynamic computer model that mirrors the biological state of a person or disease process closely enough to test treatment strategies before they are used in real patients. In contrast to traditional prediction tools that focus on statistical associations, digital twins are designed to simulate responses to interventions, such as drugs or therapy adjustments, and to provide probabilistic forecasts of outcomes [16-18]. This approach has emerged partly because many drug development programs fail late. The failure is often not due to lack of efficacy in ideal conditions, but because early-stage models fail to capture how biological systems respond in complex, real-world contexts [19].

Instead of viewing biology as a single static process, digital twins organize information across multiple biological scales. At the lowest level are molecular activities such as gene expression and protein interactions. Above this are cellular pathways, then tissue or organ systems, and finally whole-organism outcomes such as symptom progression or changes in laboratory values. These layers are interconnected. That is, a change at the molecular level (for example, a drug binding its target) can cascade through cells and tissues to produce measurable effects on whole-body physiology [20]. A twin that fails to account for cross-scale interactions may make predictions that seem plausible at one level but are biologically inconsistent overall. One of the biggest practical challenges is that biomedical data come in many different forms (e.g., genomic sequences, gene expression profiles, protein measurements, metabolic readouts, imaging, clinical lab tests, and patient-reported outcomes) [21]. Without a systematic way to bring these diverse datasets together, models can become “siloed,” making it difficult to draw meaningful inferences across domains [22]. To address this, a digital twin must include a data harmonization layer that standardizes diverse datasets into coherent internal representations. This harmonized input allows the system to reason across different types of evidence rather than making decisions based on disconnected pieces of information.

Moreover, because medical decisions involve uncertainty, a digital twin must be capable of expressing uncertainty and evidence weighting in ways that experts can interpret [23]. Rather than simply providing a single prediction, it should present the evidence supporting the prediction, highlight which pieces of data were most influential, and indicate how confident the system is in its conclusions. This kind of interpretability is essential for clinical users and researchers, who must understand the limits as well as the strengths of the model. For these reasons, modern digital twins in medicine are increasingly built as orchestrated ensembles of complementary modeling approaches, combining mechanistic understanding with statistical learning and probabilistic inference to leverage the strengths of each while compensating for their individual limitations [24]. This hybrid strategy enables more robust, biologically grounded predictions that can inform both research and clinical decisions.

Evidence Synthesis and Scoring Frameworks

In complex biological modeling, individual computational components often produce multiple intermediate outputs [25]. These include (but are not limited to) probabilities, effect size estimates, risk signals, mechanistic plausibility scores, pathway activation metrics, or simulated outcome distributions. While each signal may be informative, decision-makers rarely act on isolated model outputs. Instead, these signals must be integrated into a coherent, interpretable framework that reflects the totality of evidence. Within a digital twin architecture, this integration step functions as a structured evidence synthesis layer. Conceptually, each modeling module contributes a signal that captures a distinct dimension of biological or clinical inference, such as predicted efficacy, off-target liability, pharmacokinetic feasibility, mechanistic robustness, or variability across simulated patient subgroups [26]. These signals may differ in scale, uncertainty structure, and evidentiary strength. A formal aggregation operator (Ψ) combines them into a composite score or grading output that supports comparative assessment and prioritization. Importantly, this aggregation process is not a simple arithmetic average. Evidence synthesis must account for:

• Signal reliability (e.g., strength and quality of supporting data),
• Model concordance (agreement or divergence across independent modeling approaches),
• Biological plausibility, and
• Uncertainty propagation from upstream simulations.

Thus, weighting schemes may vary depending on the specific decision context. For example, early discovery triage versus late-stage risk assessment, while remaining fully traceable and reproducible represent such contexts [27]. This context sensitivity reflects a core principle in decision science. The relative importance of different evidence domains changes with the stakes and objectives of the decision. Equally critical is transparency. The scoring framework preserving visibility into its components, enabling experts to examine how individual signals contributed to the final assessment is essential. In this sense, the scoring layer translates high-dimensional computational inference into structured decision-theoretic objects without obscuring the underlying uncertainty or mechanistic rationale [28]. By converting complex, multi-scale model outputs into interpretable and auditable summaries, the evidence synthesis layer bridges computational modeling and expert judgment. It does not replace human oversight; rather, it provides a systematic, quantitatively grounded structure within which expert interpretation and domain knowledge can operate. 

Validation and Performance Characterization

Robust validation is foundational to the credibility of any biomedical digital twin, particularly when outputs may influence high-stakes research or development decisions. Contemporary guidance in artificial intelligence (AI) for healthcare emphasizes that methodological validation must extend beyond internal accuracy testing to include independent evaluation, pre-specified performance metrics, and transparency regarding uncertainty and limitations [29-31]. In this context, validation serves not only to quantify predictive performance but also to establish reproducibility, generalizability, and resistance to bias. Methodological validation of the present digital twin framework includes both structured internal testing and independent blinded academic evaluation. In blinded studies, outcome labels were withheld from evaluators until predictions were finalized, reducing the risk of information leakage and confirmation bias, an approach consistent with recommended best practices for AI validation in biomedical research [32].

These metrics indicate strong discriminative performance across positive and negative cases within the defined evaluation dataset. Sensitivity and specificity remain core performance indicators in translational predictive modeling because they reflect the model’s ability to detect meaningful signals while limiting false classifications [29]. Importantly, balanced reporting of both measures mitigates misleading conclusions that can arise when only a single metric (e.g., accuracy or AUC) is presented [33]. However, as emphasized in contemporary reporting standards for prediction models and AI systems, performance metrics obtained under controlled evaluation conditions should not be interpreted as guarantees of prospective clinical success. Model performance may shift under distributional changes, varying data quality, or evolving biological contexts. Therefore, observed concordance reflects robust signal detection within the boundaries of the evaluation framework, not deterministic prediction of real-world therapeutic outcomes.

Recent frameworks for AI evaluation in medicine underscore several principles relevant here: (1) separation of development and validation datasets, (2) independent or external testing where feasible, (3) reporting of uncertainty intervals, and (4) explicit acknowledgment of performance limits [29,31,32]. The inclusion of blinded independent evaluation strengthens methodological rigor by approximating real-world deployment conditions and reducing optimism bias. In aggregate, the reported performance characteristics demonstrate meaningful predictive discrimination under uncertainty while aligning with current standards for transparent validation of biomedical AI systems. Continued evaluation in expanded and prospective settings remains essential to further characterize generalizability and decision impact.

Distinguishing Characteristics Relative to Existing Digital Twins

Many “digital twin” efforts in biomedicine have advanced along relatively narrow tracks, often optimizing one slice of the problem rather than supporting end-to-end therapeutic reasoning. For example, a large and influential class of twins focuses on organ-level physiology, where patient-specific anatomy and biophysics are used to simulate organ function and test interventions (well-illustrated in precision cardiology digital twins) [34,35]. A second major class emphasizes imaging-derived structural twins, where high-fidelity anatomic replicas (and sometimes device or procedure simulations) are primarily built from radiology and related imaging data streams [36]. A third class is oriented toward target- or mechanism-specific molecular simulation in drug R&D, where the “twin” frequently functions as an organized systems pharmacology or translational modeling environment for evaluating candidate mechanisms and therapeutic hypotheses [37]. In parallel, there is growing momentum for in silico trials and “virtual patient” approaches that use digital twins to augment or partially replace conventional clinical trial evidence generation, but these often prioritize trial emulation and cohort-level inference rather than deeply integrated mechanistic-to-clinical reasoning [17].

Against this backdrop, the distinguishing contribution of the present framework is not a claim of being “more predictive” in the abstract, but a different architectural emphasis: (1) multiomic, cross-domain integration so that molecular, pathway, clinical, and real-world signals can be represented together rather than handled as disconnected analytic pipelines; (2) explicit orchestration of heterogeneous model classes (mechanistic, statistical/ML, causal, and knowledge-driven components) rather than reliance on a single dominant modeling paradigm [38]; (3) structured evidence synthesis that translates diverse model outputs into transparent, auditable grades and confidence characterizations (rather than exposing raw predictions without traceable justification); and (4) scenario-based “risk landscape” mapping, where the twin is used to explore counterfactual intervention strategies, identify sensitivity to assumptions and subpopulation shifts, and explicitly represent uncertainty in the downstream decision object. Finally, the system is intentionally positioned as a scientific computing platform operated in controlled environments, not a general-purpose SaaS interface. This is a methodological choice that supports governance: it enables pre-specified evaluation protocols, versioned model execution, reproducible evidence trails, and expert review of how the score was produced. Such capabilities are consistently highlighted as necessary for translating AI-enabled clinical tools into reliable biomedical practice [38].

Limitations and Future Directions

Despite its structured, multi-scale design, this digital twin framework is subject to several important limitations that are common across advanced biomedical modeling systems. First, performance is inherently dependent on the completeness, representativeness, and quality of input data. Multiomic and clinical datasets frequently contain missing values, batch effects, measurement variability, and population biases. If underlying data are sparse or systematically skewed, model outputs may inherit those distortions. This challenge has been widely recognized in digital twin and AI-for-healthcare research, particularly in relation to distribution shift and real-world generalizability. Second, knowledge base curation may introduce bias. Biological networks, pathway databases, and curated literature resources are shaped by historical research priorities and uneven experimental coverage. Diseases that are well studied may be overrepresented in training and inference pipelines, whereas rare or underfunded conditions may lack sufficient mechanistic depth. As a result, digital twins risk reflecting the structure of available knowledge rather than the true structure of biology. Ongoing curation transparency and version tracking are therefore essential to mitigate epistemic bias. Third, computational complexity increases substantially at full multi-scale resolution. Modeling interactions across molecular, cellular, tissue, and organism-level layers can require significant computational resources. High-resolution simulations may be constrained by processing time, memory, and scalability limits. Strategic abstraction, hierarchical modeling, and adaptive resolution techniques are necessary to balance fidelity with tractability. Fourth, although blinded validation demonstrates strong classification concordance under controlled conditions, prospective validation across diverse therapeutic domains remains essential. Performance characteristics established in retrospective or semi-prospective evaluations may not fully capture real-world variability, evolving standards of care, or emerging biological mechanisms. Broader validation in oncology, immunology, metabolic disease, and rare disorders would strengthen external generalizability.

Looking forward, several research directions are particularly promising. One priority is the development of adaptive, patient-specific twin instantiation, in which models dynamically recalibrate to individual-level data streams rather than operating solely at a program or population level. Such personalization would align more directly with the goals of precision medicine, enabling individualized response simulation and dynamic risk assessment. Another important direction is real-time Bayesian updating using longitudinal clinical inputs. As new laboratory values, imaging findings, or treatment responses are observed, posterior probability distributions could be updated iteratively, allowing the twin to evolve alongside the patient or therapeutic program. This would transform the twin from a static analytical snapshot into a continuously learning system. Further, rigorous uncertainty quantification via probabilistic programming frameworks could strengthen interpretability and decision reliability. Explicit representation of parameter uncertainty, structural uncertainty, and model-form uncertainty would allow downstream decision-makers to understand confidence bounds and sensitivity drivers rather than relying on point estimates. Integration of causal discovery and causal inference frameworks represents another key frontier. Moving beyond associative pattern recognition toward structured causal modeling would enhance counterfactual reasoning and intervention simulation. Finally, as digital twins mature, there is a clear need for regulatory-grade validation standards and certification pathways. Transparent benchmarking, reproducibility audits, performance drift monitoring, and standardized reporting guidelines will be necessary for broader acceptance in clinical and therapeutic development settings.

Conclusion

We describe a rigorously structured, multi-scale digital twin architecture grounded in systems biology, probabilistic inference, constraint-based modeling, and reinforcement learning–based orchestration. The framework integrates heterogeneous data modalities into coherent biological representations, supports counterfactual scenario simulation, and aggregates high-dimensional inference into interpretable scoring systems designed for expert review. Independent blinded validation demonstrates strong classification concordance under controlled evaluation conditions, indicating robust signal detection within defined performance boundaries. At the same time, careful acknowledgment of uncertainty, data limitations, and validation scope remains essential. By emphasizing multi-scale biological integration, uncertainty-aware reasoning, heterogeneous model orchestration, and structured evidence synthesis, this architecture advances the field of biomedical digital twins beyond narrow correlation-driven prediction toward systems-level therapeutic foresight. In doing so, it contributes to the evolving foundation of computationally enabled personalized medicine, where rigorous modeling, transparent validation, and human oversight remain central to responsible innovation.

References

1. National Academies of Sciences, Engineering, and Medicine. Digital Twin Landscape. Published online March 28, 2024.
2. Tao F, Pickl S, Yi L. (2025) Digital twins optimizing the world. Digit Eng. 7.
3. Qin R, Zi X, Ge X. (2026) Digital twin-based intelligent risk assessment and decision support system for university student entrepreneurial projects. Sci Rep. 16(1):5770.
4. Common Themes at the Intersection of Biological and Physical Sciences. Common Themes at the Intersection of Biological and Physical Sciences. Published online 2010.
5. Schork NJ, Goetz LH. (2025) From precision interventions to precision health. Nat Commun. 16(1):5024.
6. Jamalinia M, Weiskirchen R. (2025) Advances in personalized medicine: translating genomic insights into targeted therapies for cancer treatment. Ann Transl Med. 2025;13(2):18.
7. Alharthi S. (2025) AI-powered in silico twins: redefining precision medicine through simulation, personalization, and predictive healthcare. Saudi Pharm J SPJ Off Publ Saudi Pharm Soc. 34(1):1.
8. Akbarialiabad H, Pasdar A, Murrell DF, et al. (2025) Enhancing randomized clinical trials with digital twins. NPJ Syst Biol Appl. 11(1):110.
9. Chen R, Duffy Á, Do R. (2026) Genomics of drug target prioritization for complex diseases. Nat Rev Genet. 27(3):231-245.
10. Liyanage YR, Chowell G, Pogudin G, Tuncer N. (2025) Structural and Practical Identifiability of Phenomenological Growth Models for Epidemic Forecasting. Viruses. 17(4):496.
11. Cheng CH, Shi SS. (2025) Artificial intelligence in cancer: applications, challenges, and future perspectives. Mol Cancer. 24(1):274.
12. Qin R, Zi X, Ge X. (2026) Digital twin-based intelligent risk assessment and decision support system for university student entrepreneurial projects. Sci Rep. 16(1):5770.
13. Yinghua M, Khan A, Heng Y, Gul Khan F, Aldhahri A, et al. (2025) A contrastive adversarial encoder for multi-omics data integration. PloS One. 20(10):e0333134.
14. Ma A, Newing TP, O’Shea R, et al. (2024) Genomic multidisciplinary teams: A model for navigating genetic mainstreaming and precision medicine. J Paediatr Child Health. 60(4-5):118-124.
15. Brown LV, Wagg J, Darley R, et al. (2022) De-risking clinical trial failure through mechanistic simulation. Immunother Adv. 2(1):ltac017.
16. Sadée C, Testa S, Barba T, et al. (2025) Medical digital twins: enabling precision medicine and medical artificial intelligence. Lancet Digit Health. 7(7):100864.
17. Samei E. (2025) The future of in silico trials and digital twins in medicine. PNAS Nexus. 4(5):123.
18. Avraham R, Yarden Y. (2011) Feedback regulation of EGFR Signaling: decision making by early and delayed loops. Nat Rev Mol Cell Biol. 12(2):104-117.
19. Feng X, Wang H, Takata H, Day TJ, Willen J, et al. (2011) Transcription factor Foxp1 exerts essential cell-intrinsic regulation of the quiescence of naive T cells. Nat Immunol. 12(6):544-550.
20. Schuetz AN, Guarner J, Packard MM, Zaki SR, Shehata BM, et al. (2011) Infectious disease immunohistochemistry in placentas from HIV-positive and HIV-negative patients. Pediatr Dev Pathol Off J Soc Pediatr Pathol Paediatr Pathol Soc. 14(3):180-88.
21. Hasin Y, Seldin M, Lusis A. (2017) Multi-omics approaches to disease. Genome Biol. 18(1):83.
22. Subramanian I, Verma S, Kumar S, Jere A, Anamika K. (2020) Multi-omics Data Integration, Interpretation, and Its Application. Bioinforma Biol Insights. 14:1177932219899051.
23. Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. (2019) Causability and explainability of artificial intelligence in medicine. Wiley Interdiscip Rev Data Min Knowl Discov. 9(4):e1312.
24. Ting Sim JZ, Fong QW, Huang W, Tan CH. (2023) Machine learning in medicine: what clinicians should know. Singapore Med J. 64(2):91-97.
25. Ji Z, Yan K, Li W, Hu H, Zhu X. (2017) Mathematical and Computational Modeling in Complex Biological Systems. BioMed Res Int. 2017:5958321.
26. Marques L, Costa B, Pereira M, et al. (2024) Advancing Precision Medicine: A Review of Innovative In Silico Approaches for Drug Development, Clinical Pharmacology and Personalized Healthcare. Pharmaceutics. 16(3):332.
27. Vali Y, Leeflang MMG, Bossuyt PMM. (2021) Application of weighting methods for presenting risk-of-bias assessments in systematic reviews of diagnostic test accuracy studies. Syst Rev. 10(1):191.
28. Chen J, Yang X, Wu H. (2024) A Multibranch Neural Network for Drug-Target Affinity Prediction Using Similarity Information. ACS Omega. 9(33):35978-89.
29. Collins GS, Dhiman P, Andaur Navarro CL, et al. (2021) Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ Open. 11(7):e048008.
30. Cruz Rivera S, Liu X, Chan AW, et al. (2020) Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nat Med. 26(9):1351-63.
31. Wynants L, Van Calster B, Collins GS, et al. (2020) Prediction models for diagnosis and prognosis of covid-19: systematic review and critical appraisal. BMJ. 369:m1328.
32. Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. (2019) Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 17(1):195.
33. Wynants L, Van Calster B, Collins GS, et al. (2020) Prediction models for diagnosis and prognosis of covid-19: systematic review and critical appraisal. BMJ. 369:m1328.
34. Corral-Acero J, Margara F, Marciniak M, et al. (2020) The “Digital Twin” to enable the vision of precision cardiology. Eur Heart J. 41(48):4556-64.
35. Thangaraj PM, Benson SH, Oikonomou EK, Asselbergs FW, Khera R. (2024) Cardiovascular care with digital twin technology in the era of generative artificial intelligence. Eur Heart J. 45(45):4808-21.
36. Pesapane F, Rotili A, Penco S, Nicosia L, Cassano E. (2022) Digital Twins in Radiology. J Clin Med. 11(21):6553.
37. An G, Cockrell C. (2022) Drug Development Digital Twins for Drug Discovery, Testing and Repurposing: A Schema for Requirements and Development. Front Syst Biol. 2:928387.
38. Drummond D, Gonsard A. (2024) Definitions and Characteristics of Patient Digital Twins Being Developed for Clinical Use: Scoping Review. J Med Internet Res. 26:e58504.