In silico models in oncology, neurology, and epidemiology: systems-level and multiscale perspectives Authors: Matteo Italia, José Garcia Otero, Juan Jiménez-Sánchez, Fabio Dercole, Juan Belmonte-Beitia Published: 2026 ABSTRACT Computational medicine uses mathematical modelling, high-performance computing, and the availability of large-scale biomedical data to study multiscale complex diseases. This review synthesises recent developments in in silico oncology, neurology, and epidemiology, highlighting their shared methodological foundations in systems theory and complex networks. We discuss how combining mechanistic approaches with machine learning integrates heterogeneous data. Finally, we identify translation barriers and outline future directions like digital twins for predictive, personalised healthcare. Introduction In recent decades, the intersection of mathematical modelling, computational tools, data analytics, and medical science has engendered a profound transformation in healthcare research and clinical practice. This evolution has culminated in the emergence of computational and in silico medicine, a multi-faceted discipline leveraging the utilisation of sophisticated computational methodologies to simulate, analyse, and forecast the behaviour of biological systems1,2. This paradigm shift has not only expanded our comprehension of intricate disease mechanisms but also revolutionised therapeutic modalities for complex pathological conditions. Within the expansive domain of in silico medicine, oncology3,4, neurology5,6, and epidemiology7,8 have emerged as prominent fields of enquiry, each presenting distinct challenges and opportunities for computational approaches. Despite their disparate focuses, these disciplines share a fundamental underpinning: they all benefit from the application of systems theory and the principles and tools of complexity science to elucidate their intricate nonlinear dynamics9. By adopting a holistic perspective that integrates different time and space scales, in silico methodologies offer a comprehensive framework to investigate the subtleties of disease progression and pathophysiology, and to translate models into clinical practice10. For instance, in silico oncology employs computational models to simulate the biology of cancers, encompassing tumour growth, evolutionary dynamics of arising clones, spatio-temporal intratumour heterogeneity, metastatic dissemination, and therapeutic response profiles, among others. These models facilitate the design of optimal personalised treatment combinations11,12 and schedules13, and support the discovery14 and repurposing15 of effective therapeutic agents and novel biomarkers16. Likewise, computational neurology endeavours to elucidate the intricate network dynamics of the central nervous system17, shedding light on the pathogenesis of neurological afflictions such as Alzheimer’s18 and Parkinson’s19 diseases, and Restless Legs Syndrome20. These insights inform the development of targeted interventions aimed at alleviating disease burden and enhancing patient outcomes. Concurrently, computational epidemiology7 harnesses predictive modelling and data-driven analytics to forecast disease outbreaks21, evaluate intervention strategies22, and optimise public health policies23. Notably, recent endeavours in this domain have been particularly important to mitigate the impact of the COVID-19 pandemic24,25. This review aims to synthesise the wealth of knowledge accrued across these three diverse domains, elucidating the transformative potential of in silico medicine in reshaping contemporary healthcare paradigms. By fostering interdisciplinary collaboration and leveraging state-of-the-art computational methodologies, we envision a future wherein computational models serve as indispensable tools for understanding disease mechanisms, informing clinical decision-making, and advancing the boundaries of precision medicine. In the following sections, we provide a structured overview of recent advances in in silico and computational medicine in the three before-mentioned fields: oncology, neurology, and epidemiology. The literature was surveyed using major bibliographic databases (Web of Science and Scopus), focusing on recent and representative contributions in computational and in silico oncology, neurology, and epidemiology. Rather than a systematic review, this work provides a narrative, concept-driven synthesis emphasising shared modelling paradigms, translational challenges, and emerging directions. Thus, the structure of the article is as follows: Introduction corresponds to the opening of the article, as we have just outlined. The Results section is the core part of this review and focuses on the various in silico results that have been obtained in the literature. The In silico oncology section focuses on in silico results in the field of oncology. Section In silico neurology focuses on the different in silico approaches obtained in the field of neurology, and Section In silico epidemiology reviews the results obtained for epidemiology. The Shared paradigms and bottlenecks: cross-disciplinary synthesis section offers a comprehensive overview bringing these three fields together, by highlighting their common paradigms and their shared bottlenecks. Finally, Conclusions summarises the main conclusions from this overview of the state-of-the-art of in silico computational approaches in the three fields studied: oncology, neurology, and epidemiology. Results In silico oncology Cancer is a major global health challenge, with profound socio-economic impact that affects millions of lives26. Actually, the World Health Organization (WHO) states that cancer is the major cause of death globally27. According to the most recent estimates from the International Agency for Research on Cancer (IARC), there were nearly 20 million new cancer cases and 9.7 million cancer-related deaths worldwide in 2022 26. These numbers indicate that about one in five individuals will develop cancer during their lifetime, while approximately one in nine men and one in twelve women will die from it. Figure 1 shows global cancer incidence and mortality in 2022 by countries and by cancer types, respectively. Thus, understanding the growth and evolution of cancers and how to exploit them in optimal treatments is an important open question28 that could benefit from the contributions of mathematics29–35.Fig. 1Cancer worldwide impact in 2022.Cancer incidence (a) and mortality (b): 20.0 million new cases and 9.7 million deaths. Maps show country-level age-standardised rates (ASR), while pie charts display the distribution by cancer type based on absolute numbers. Lung cancer is the leading contributor in both incidence and mortality (12.4% and 18.7%, respectively), followed by colorectal (9.6% and 9.3%), breast (11.6% and 7.8%), liver (4.3% and 7.8%), and stomach cancer (4.9% and 6.8%). Prostate, thyroid, bladder, and non-Hodgkin lymphoma are among the most frequent cancers in incidence only (a), whereas pancreatic, oesophageal cancers, and leukaemia appear among the leading causes of mortality only (b). Cervical cancer is present in both incidence and mortality (3.3% and 3.6%, respectively). Remaining cancers account for 36.6% of cases and 30.3% of deaths. ASR refers to a weighted average of age-specific rates per 100,000 people, where weights correspond to the population distribution across age groups. Data from the Global Cancer Observatory of the IARC26,311. Figure adapted from ref. 323. Icons used in the figure are provided by Flaticon and are used under their standard attribution license. Cancer worldwide impact in 2022. Cancer incidence (a) and mortality (b): 20.0 million new cases and 9.7 million deaths. Maps show country-level age-standardised rates (ASR), while pie charts display the distribution by cancer type based on absolute numbers. Lung cancer is the leading contributor in both incidence and mortality (12.4% and 18.7%, respectively), followed by colorectal (9.6% and 9.3%), breast (11.6% and 7.8%), liver (4.3% and 7.8%), and stomach cancer (4.9% and 6.8%). Prostate, thyroid, bladder, and non-Hodgkin lymphoma are among the most frequent cancers in incidence only (a), whereas pancreatic, oesophageal cancers, and leukaemia appear among the leading causes of mortality only (b). Cervical cancer is present in both incidence and mortality (3.3% and 3.6%, respectively). Remaining cancers account for 36.6% of cases and 30.3% of deaths. ASR refers to a weighted average of age-specific rates per 100,000 people, where weights correspond to the population distribution across age groups. Data from the Global Cancer Observatory of the IARC26,311. Figure adapted from ref. 323. Icons used in the figure are provided by Flaticon and are used under their standard attribution license. The transformation of healthy cells into tumour cells takes multiple steps, starting from a pre-cancerous lesion up to a malignant tumour27. This evolutionary process, better known as carcinogenesis, has never been explicitly tracked in vivo, and although there is consensus regarding the mechanism whereby it occurs (the somatic mutation theory36), its exact unfolding is still a subject of debate37. It is widely accepted that several key modifications (namely, driver mutations) are required for a healthy cell to promote to an unrestrained growth state, which, if uncontrolled, may eventually give rise to a tumour. These modifications are the result of the combined effect of the subject’s genetic characteristics and external environmental influences, i.e., physical, chemical, and biological carcinogens27. This interplay of factors is one of the reasons why cancer is considered one of the most dynamic and complex groups of diseases. To make matters worse, cancers can invade or spread to other tissues or organs, which are not always nearby (in contrast with benign tumours, which lack this invasive potential. This contributes substantially to cancer-related mortality. This invasiveness is but one of cancer hallmarks38,39. There has been a recent effort to characterize the most relevant cancer trademarks, which encompass replicative immortality (each cell can potentially duplicate infinite times), genome instability (high mutation rate), evasion of growth-suppressor signals, sustained proliferation, resist cellular death, avoid immune destruction, altered metabolism, cancer-promoting inflammation, angiogenesis, and activation of invasion and metastases38 together with emerging enabling features including unlocking phenotypic plasticity, nonmutational epigenetic reprogramming, the accumulation of senescent cells, and polymorphic microbiomes39 (Fig. 2).Fig. 2Conceptual mapping of cancer hallmarks and representative in silico modelling approaches.Cancer hallmarks (inside the grey circle) as reported in the original framework by Hanahan and Weinberg38,39, complemented with representative associations to in silico modelling approaches (outside the grey circle and withi n the same two radii), e.g., hybrid models for the hallmark of activating invasion and metastasis. The associations shown are not intended to be exhaustive, but rather to provide a conceptual mapping of plausible relationships. Icons used in the figure are sourced from Flaticon (https://www.flaticon.com) under their standard attribution license, and graphical elements are adapted from Servier Medical Art (https://smart.servier.com), licensed under CC BY 4.0. Conceptual mapping of cancer hallmarks and representative in silico modelling approaches. Cancer hallmarks (inside the grey circle) as reported in the original framework by Hanahan and Weinberg38,39, complemented with representative associations to in silico modelling approaches (outside the grey circle and withi n the same two radii), e.g., hybrid models for the hallmark of activating invasion and metastasis. The associations shown are not intended to be exhaustive, but rather to provide a conceptual mapping of plausible relationships. Icons used in the figure are sourced from Flaticon (https://www.flaticon.com) under their standard attribution license, and graphical elements are adapted from Servier Medical Art (https://smart.servier.com), licensed under CC BY 4.0. Clinically speaking, the choice of treatment depends on the type and stage of cancer, as well as on individual patient factors. The main types of cancer treatments include surgery40,41, radiation therapy42–46, immunotherapy47–49, chemotherapy50–57, targeted therapy58–63, personalized medicine64–69, nanotherapy70–74, oncolytic virus75–79, and many more. Whenever possible, these treatment options help either remove the malignant tumour cells completely or, at least, control their growth pace. Unfortunately, clinical practice often falls short when it comes to thoroughly exploring the effect of different treatment schemes, due to small sample sizes, high inter-patient heterogeneity, and the lack of untreated controls (which would be ethically dubious). To help overcome these shortcomings, we have in silico oncology. Indeed, in silico oncology applies many different suitable methodologies to different specific research problems. These important studies bring new insight into cancer biology80, evolution12,13,81,82, clinical45 and also both pre-clinical83 and post-clinical84 information, bio-markers identification85, treatment12,13,45,74,82,86,87, and even prevention88. Recently, computational-based patents, tackling some relevant cancer treatment aspects, have been successfully proposed89. Among all the different suitable methodologies that in silico oncology exploits, we can start by mentioning network-based approaches. Network science has been highly applied to model cancer inter- and intra- cellular communications11,90, i.e., pathways and signal transduction, starting from the simpler Boolean networks up to more complex differential equations and Bayesian networks91–93. Moreover, lattices and even more intricate networks are exploited to model interconnections among the agents (sub-populations13,87 or single cells82), adding a spatial or functional structure to the model. Since cancer cells are deeply intertwined and constantly communicating (even with non-cancer cells in their surroundings, like cancer-associated fibroblasts in the tumour microenvironment94) and exchanging signalling molecules, these network-based approaches help us untie the convoluted web of interactions intrinsic to tumours. Virtual screening techniques, such as the one illustrated by Zhu et al.95, accelerate the identification of novel anti-cancer compounds. Moreover, in numerous cancer types, the utilisation of drug repurposing15,63 and drug combinations96,97 has demonstrated enhanced durability and reduced resistance to existing first-line therapies. However, the conventional (in vitro and in vivo) approaches to drug discovery are characterised by prohibitive costs and work98. To address this challenge, there is an urgent need for computational methods that can effectively model the intricate interactions of drug target pathways and identify the underlying mechanisms driving the synergistic effects of drug combinations. An example of this approach is the SynGeNet (Synergy from Gene expression and Network mining) method11, which integrates transcriptomics-based connectivity mapping and network centrality analysis to study disease networks and predict the efficacy of drug combinations. Pharmacophore modelling99,100, a subfield of structural biology and computational chemistry, predicts the 3D arrangement of molecular features necessary for a drug to bind to its target protein, identifying potential drugs that may selectively inhibit cancer-specific targets. Molecular docking predicts the binding interaction between a small molecule (e.g., a drug) and its target protein at the atomic level. This method helps in screening and prioritising potential drugs, evaluating also toxicological effects101. Also, molecular dynamics simulations, whereby protein movements at the atom level are modelled using classical mechanics (and sometimes even quantum effects for local interactions), are of relevance for easing the drug discovery and screening processes. This method is useful for validating binding modes predicted by molecular docking, for mapping target proteins' hot-spots, and for generating protein conformation ensembles for in silico drug design14. Bioinformatics, and more specifically, omics studies, are commonly used for data mining in cancer (and genetic diseases in general) research. Bioinformatics encompasses such a wide field, covering a plethora of different techniques and approaches to biological data. To narrow it down, the main areas of interest within bioinformatics cover the development and analysis of algorithms to process biological information, the design of databases to store, connect and contextualise vast amounts of data (sourcing from DNA sequences, protein structures or gene expression profiles), and the study of statistical methods to distil meaningful insights from biological data. Processing high-throughput biological information and creating adequate databases to store it has resulted in the development of omics (from classical genomics, proteomics, and transcriptomics, to more recent additions such as lipidomics, epigenomics, and metabolomics), which have substantially sped up research in oncology, but have also raised new problems and concerns, mostly about data integration and representation, and dimensionality reduction102. The contextualization of previous data has resulted in useful initiatives aimed at coherently and controllably annotating and labelling biological entities, such as the Gene Ontology Database103, which allows researchers to identify a gene’s/protein’s function and location, enabling fast interpretation of multiplexed data. The integration of bioinformatics with oncology not only lets us build knowledge about the disease, but also personalises treatments and comes closer to a real precision oncology104. Specific goals are to profile cancer cells and their microenvironment, e.g., for target therapy11 and immunotherapeutic discovery105. In the last few years, machine learning (ML) and artificial intelligence (AI) methods have been proposed in the literature with the potential to transform oncology106 (and medicine in general), enabling the prediction of drug response, enhancing theranostics107, aiding in personalised treatment selection, and improving not only nanomedicine74, but also the detection of cancers108. These approaches now increasingly exploit multimodal and multi-omics integration (combining imaging, histopathology, genomics, transcriptomics and clinical data) to derive richer, patient-specific signatures that improve diagnosis, prognostication and therapy selection109. Deep learning models and specialised libraries (for example, frameworks designed for drug-response prediction) have shown promising performance in predicting tumour sensitivity and resistance patterns from genomic and pharmacological datasets, although benchmarking and standardisation remain challenging110,111. In theranostics, AI is being applied to automate image segmentation, perform patient selection and improve dosimetry and treatment planning for radiopharmaceuticals, potentially enabling more precise, individualised theranostic workflows112,113. In nanomedicine, ML accelerates nanoparticle design, predicts biodistribution and toxicity, and helps to optimise formulation and delivery–approaches that are shortening the design-test cycle for nanoscale therapeutics and theranostics17,114. Other active lines of development include federated learning and privacy-preserving training (to share models across hospitals without centralising sensitive data), explainable AI methods (to increase clinical trustworthiness) and physics/knowledge-informed ML that embeds prior mechanistic knowledge to improve generalisability115,116. Nonetheless, important barriers persist–data heterogeneity, limited annotated clinical datasets for many tumour types, issues of model interpretability, prospective clinical validation and regulatory approval are active areas of research that must be addressed before broad clinical deployment117,118. More recently, advances in computational pathology and spatial omics are increasingly redefining how deep representation learning is grounded in biological reality, with explicit modelling of spatial relationships and entropy-based complexity measures emerging as key anchoring mechanisms. Rather than treating tissues as texture-rich but topologically flat images, modern workflows abstract samples into spatial point processes—cells or spots with defined coordinates and phenotypic identities—thereby making cell-cell proximity, neighbourhood structure, and multiscale organisation explicit analytical objects. By encoding who interacts with whom, and at what spatial scale, spatial descriptors shift representation learning away from purely appearance-driven correlations toward features that correspond to known microenvironmental processes, including immune exclusion, invasion fronts, and stromal remodelling119. Within this spatially explicit framework, entropy-based measures provide a principled bridge between tissue organisation and deep learning representations by quantifying heterogeneity, disorder, and complexity across scales120. Entropy metrics capture not only compositional diversity but, critically, the organisation of that diversity in space, distinguishing structured microenvironments from chaotic or transitional states. Since entropy abstracts complex spatial configurations into scale-agnostic measures of order and uncertainty, it enables deep models to encode tissue ecology rather than dataset-specific artefacts. Collectively, the coupling of spatial relationship extraction with entropy-based complexity metrics reframes deep representation learning as an ecological observer of biological microenvironments, ensuring that computational models remain anchored to mechanistically interpretable patterns of tissue organisation while retaining the flexibility and predictive power of modern deep learning. Complementing this, adaptive multimodal fusion mechanisms are emerging as a powerful paradigm for integrating such spatial representations with deep networks. Notably, entropy-guided hybrid architectures that dynamically fuse GNN-based (graph neural networks) local topology with state-space models such as Mamba demonstrate how predictive uncertainty can be used to balance local and global information on a per-sample basis, avoiding rigid fusion assumptions and enabling scalable modelling of complex, heterogeneous tissues121. Together, these developments illustrate a broader shift toward uncertainty-aware, spatially grounded, and dynamically reconfigurable AI systems that parallel and inform advances in multimodal integration across multi-omics research. Multiscale in silico models–with scales spanning from biomolecules and metabolites up to tissues and entire patient bodies–provide a powerful framework for understanding cancer growth and therapy response68,122. In Table 1, we provide an updated and extended taxonomy of evolutionary cancer models, building upon the framework previously established in ref. 82, introducing two significant advancements. First, in addition to population-based (PB) and cell-based (CB) models, we have introduced a specific category for Hybrid (H) models: these couple a CB and PB approach by using a discrete individual-based description for cancer cells with continuous dynamics for the diffusion of oxygen, nutrients, and/or drugs. We detail the three categories hereafter. Consequently, several models previously classified as purely CB in ref. 82 have been repositioned to better reflect their nature. Second, recent state-of-the-art references have been added and discussed.Table 1Literature review of PB, CB, and hybrid (H) evolutionary modelsRef.Model typeResist. phen.Opt. strat.Notes87PBDNIn silico trials for malignant glioma patients under chemotherapy and immunotherapy.123PBDNLeukaemia model linking relapse risk to initial chemotherapy response.124PBDNIn silico trials for low-grade gliomas under temozolomide, overcoming drug resistance.127PBDNPre-existing vs induced resistance in slow-growing tumours.130PBDNIn silico trials for malignant glioma patients under chemotherapy and radiotherapy.320PBDNSpatiotemporal cell growth dynamics and response to doxorubicin in breast cancer.321PBDYIn vitro experiments and game-theory model for the double-bind in breast cancer.125PBDYOptimal control strategies for drug resistance.126PBDYEvolutionary dynamics and optimal control to manage drug resistance in cancer therapy.13PBD/CYMinimisation of neuroblastoma under two drugs (D for genetic resistance, C for plastic).129PBCNDynamics of a structured PB model endowed with constrained plasticity of traits.128PBCYOptimising therapy for tumour reduction under a drug-resistance continuum.134CBDNProstate cancer model studying intra-tumoural competition during adaptive therapy.135CBDNBRAF-mutated melanoma model investigating growth rate and phenotypic adaptation.322CBDYAnalysis of temozolomide schemes for gliomas with extended spacing between doses.82CBCYOptimal two-drug treatment with trade-offs in allocation of drug-specific resistance.139HDNMultiscale model of macrophage polarisation in glioma and therapeutic implications.140HDNAnalysis of angiogenesis-induced resistance in brain tumours with combined treatments.65HDNCell cycle dynamics in a vascularized tumour under chemo-radiotherapy.141HDNInterplay among intracellular, extracellular, and intercellular factors in drug resistance.61HDNRole of the spatial structure on the effectiveness of static and adaptive chemotherapies.142HDNVascular impact on tumour evolution and thermal therapy outcomes.143HDNMulti-scale neuroblastoma model for patient-specific simulations.144HCNMetabolic heterogeneity effect on vascularized tumour growth under different therapies.The table highlights the modelling of the resistance phenotype (D for discrete vs C for continuous) and whether optimal strategies have been investigated (Y for yes vs N for no). This table is adapted and extended from ref. 82, introducing the new Hybrid (H) category and incorporating recent literature. Table 1Literature review of PB, CB, and hybrid (H) evolutionary modelsRef.Model typeResist. phen.Opt. strat.Notes87PBDNIn silico trials for malignant glioma patients under chemotherapy and immunotherapy.123PBDNLeukaemia model linking relapse risk to initial chemotherapy response.124PBDNIn silico trials for low-grade gliomas under temozolomide, overcoming drug resistance.127PBDNPre-existing vs induced resistance in slow-growing tumours.130PBDNIn silico trials for malignant glioma patients under chemotherapy and radiotherapy.320PBDNSpatiotemporal cell growth dynamics and response to doxorubicin in breast cancer.321PBDYIn vitro experiments and game-theory model for the double-bind in breast cancer.125PBDYOptimal control strategies for drug resistance.126PBDYEvolutionary dynamics and optimal control to manage drug resistance in cancer therapy.13PBD/CYMinimisation of neuroblastoma under two drugs (D for genetic resistance, C for plastic).129PBCNDynamics of a structured PB model endowed with constrained plasticity of traits.128PBCYOptimising therapy for tumour reduction under a drug-resistance continuum.134CBDNProstate cancer model studying intra-tumoural competition during adaptive therapy.135CBDNBRAF-mutated melanoma model investigating growth rate and phenotypic adaptation.322CBDYAnalysis of temozolomide schemes for gliomas with extended spacing between doses.82CBCYOptimal two-drug treatment with trade-offs in allocation of drug-specific resistance.139HDNMultiscale model of macrophage polarisation in glioma and therapeutic implications.140HDNAnalysis of angiogenesis-induced resistance in brain tumours with combined treatments.65HDNCell cycle dynamics in a vascularized tumour under chemo-radiotherapy.141HDNInterplay among intracellular, extracellular, and intercellular factors in drug resistance.61HDNRole of the spatial structure on the effectiveness of static and adaptive chemotherapies.142HDNVascular impact on tumour evolution and thermal therapy outcomes.143HDNMulti-scale neuroblastoma model for patient-specific simulations.144HCNMetabolic heterogeneity effect on vascularized tumour growth under different therapies.The table highlights the modelling of the resistance phenotype (D for discrete vs C for continuous) and whether optimal strategies have been investigated (Y for yes vs N for no). This table is adapted and extended from ref. 82, introducing the new Hybrid (H) category and incorporating recent literature. Literature review of PB, CB, and hybrid (H) evolutionary models The table highlights the modelling of the resistance phenotype (D for discrete vs C for continuous) and whether optimal strategies have been investigated (Y for yes vs N for no). This table is adapted and extended from ref. 82, introducing the new Hybrid (H) category and incorporating recent literature. PB models describe cancer cells as homogeneous populations or structured sub-populations through continuous variables13,87,123–130. While they are amenable to formal mathematical analysis and optimal control theory–making them essential for linking macroscopic clinical data to mechanistic parameters131,132–they often lack the resolution to capture fine-grained stochasticity, since they are more fit to capture the averaged behaviour of large ensembles of interacting cells/agents. It is possible to formally derive them from CB models133. CB models, or agent/individual-based models, treat each cell as an independent unit82,134,135. This allows for a high-fidelity representation of genetic heterogeneity and emergent phenomena like clonal selection136,137, though at a higher computational cost138, and by limiting the scale of the simulation, making it difficult to simulate the number of cells present in tumours of clinically relevant size. Hybrid (H) models represent the synthesis of these approaches61,65,139–144. By combining individual cell descriptions with continuum formalisms for the microenvironment, they serve as the backbone for modern digital twin (DT) approaches132,145. As shown in Table 1, these models are increasingly used to manage the trade-off between model complexity and realism versus computational tractability and interpretability of results. In Table 1, we refer to multiscale models that include adaptive resistance to treatments (i.e. phenotypes describing the type and level of cancer resistance to a treatment), reporting the list of references with the model's basic features. All models mentioned in Table 1 have been used to investigate the interplay between different control strategies and the adaptive response of cancer to treatment. In particular, such models provide the theoretical foundation for adaptive and metronomic regimens that aim to preserve drug-sensitive competitors and delay resistant takeover, a strategy supported by both mathematical studies and preclinical/clinical work146–148. Detailed CB and H models are so complex, typically designed in a stochastic framework, that they can be solved only through numerical simulation. However, they have the potential to match in vitro, in vivo, or clinical case studies and provide reliable predictions61,65,138. Indeed, in all CB (except for work82) and H cases, different control strategies have been compared in terms of a target of interest, such as cancer size at the end or during treatment, without facing the heavy computational challenge of optimal control in the CB framework (No optimal strategy, ‘N’ in Table 1). Instead, in some cases of PB, the search for the optimal strategy has been carried out (‘Y’ in Table 1). Among PB evolutionary models, the unstructured framework proposed by Orlando, Gatenby, and Brown51 represents a seminal contribution to the study of treatment-driven cancer evolution. Despite its simplicity, this model introduced the concept of exploiting resistance trade-offs through an evolutionary double bind149, whereby adaptation to one cytotoxic drug increases susceptibility to another. By explicitly modelling drug-specific resistance allocation, the framework demonstrated how evolutionary constraints can be leveraged to design treatment strategies that delay resistance emergence. Subsequent studies have reinforced the potential of double-bind and trade-off-based strategies, showing that sequential or adaptive schedules can outperform classical maximum-tolerated-dose regimens in heterogeneous tumours in silico13,150–152. Importantly, later work highlighted that incorporating pharmacokinetics/pharmacodynamics and spatial structure can qualitatively alter selective pressures and optimal control policies, underscoring the limitations of well-mixed assumptions152–154. Within this evolutionary perspective, adaptive therapy has emerged as a prominent paradigm that shifts treatment goals from eradication to long-term control by dynamically adjusting dosing in response to tumour evolution146. Initially explored in silico148,155, adaptive therapy represents one of the clearest examples of a mathematically motivated strategy that has progressed to clinical testing, including trials in prostate cancer156. Nevertheless, open challenges remain regarding model design, protocol validation, and clinical implementation157. Building on these PB foundations, our work extended the Orlando et al. framework to a CB setting to assess the robustness of optimal control strategies under genetic heterogeneity, spatial competition, and stochasticity82. By retaining sufficient phenotypic complexity while relaxing homogeneity assumptions, this approach demonstrated how candidate control policies identified in reduced PB models can be critically reassessed in more realistic multiscale contexts. In particular, incorporating spatial structure and pharmacokinetic constraints, together with discrete dosing and monitoring schedules reflecting clinical practice, was shown to alter the class of optimal solutions, often favouring adaptive or pulsed protocols153,158. Comparative PB-CB analyses further support the use of multiscale in silico frameworks to evaluate the translational robustness of evolution-aware treatment strategies152,159. Instead, in13, a PB model of neuroblastoma treated with vincristine and cyclophosphamide—two standard agents in the COJEC protocol56—was developed to explore how tumour composition and resistance phenotypes influence optimal treatment strategies. By considering distinct initial tumour states as virtual patients, this framework illustrates how in silico models can support the investigation of personalised therapy design under treatment-induced evolutionary pressure. Indeed, diagnosis and treatment strategies based on personalised or precision medicine are achieving increasing interest thanks to the potential applicability of engineered studies in many medical fields, especially in oncology, where personalised medicine seems to be the most promising approach against treatment failures65,69, also in an evolutionary modelling framework81. Recent developments have reinforced this trend by showing that precision oncology is no longer limited to single-gene markers but increasingly relies on multi-omics integration, liquid biopsies, and comprehensive biomarker signatures that combine genomic, transcriptomic, proteomic, and imaging-derived features to define patient-specific therapeutic windows16,160. These advances have driven new clinical-trial designs (basket, umbrella and platform trials) and molecular tumour boards that operationalise personalised decisions in the clinic, enabling rapid matching of patients to targeted agents or combination strategies161,162. From a modelling perspective, evolutionary and adaptive-therapy frameworks provide a natural bridge between precision diagnostics and treatment scheduling: by explicitly modelling intra-tumour heterogeneity and competitive interactions among subpopulations, evolutionary models can propose adaptive dosing schedules that prolong control and delay resistance compared with fixed maximum-tolerated-dose policies152. Notably, clinical and in silico studies have demonstrated that model-informed adaptive protocols can increase time to progression in some settings, underscoring the translational potential of combining patient-specific data with evolutionary control algorithms156,157. Moreover, recent work highlights two essential modelling ingredients for truly personalised in silico predictions: (i) explicit incorporation of pharmacokinetics/pharmacodynamics (PK/PD) and (ii) calibration to patient-specific longitudinal biomarkers (e.g., ctDNA dynamics, imaging volumetrics). Including PK/PD constrains the timing and magnitude of selective pressures, while longitudinal biomarker calibration enables real-time updating of model parameters for closed-loop adaptive strategies153,163. Together, these ingredients improve the reliability of model-based treatment recommendations and pave the way for hybrid workflows that combine mechanistic models, statistical emulators, and machine-learning surrogates to produce clinically actionable, patient-tailored schedules154,164. Finally, the convergence of precision diagnostics, flexible trial designs, and evolution-aware modelling suggests a practical route to test personalised, model-informed therapies: integrate multi-omic profiling and frequent minimally invasive monitoring (liquid biopsy), use evolutionary models to propose candidate adaptive protocols, and evaluate them in biomarker-driven platform trials that allow rapid iteration between model prediction and clinical observation161,163,165. All these highlight the relevance of mechanistic modelling and mechanistic learning in computational oncology166. The conceptual pipeline for such an integrated multiscale approach is detailed in Fig. 3.Fig. 3Multiscale framework for in silico oncology.The workflow integrates biological knowledge and data from different time- and spatial-scales (from DNA and Omics data, to longitudinal biomarker and pharmacokinetics, up to imaging and patient survival) into a computational framework. The latter is a CB, PB, or hybrid (H) model, see Table 1 for evolutionary models. The Physics-Informed ML layer acts as a surrogate model, trained on high-fidelity simulations to achieve model acceleration for real-time predictive pipelines. Finally, the framework supports closed-loop control, such as adaptive therapy, where clinical outputs and resistance patterns are fed back (dashed line) as input data to recalibrate the multiscale model and adjust subsequent treatments, enabling personalised, actionable decisions. Icons used in the figure are provided by Flaticon and are used under their standard attribution license. Multiscale framework for in silico oncology. The workflow integrates biological knowledge and data from different time- and spatial-scales (from DNA and Omics data, to longitudinal biomarker and pharmacokinetics, up to imaging and patient survival) into a computational framework. The latter is a CB, PB, or hybrid (H) model, see Table 1 for evolutionary models. The Physics-Informed ML layer acts as a surrogate model, trained on high-fidelity simulations to achieve model acceleration for real-time predictive pipelines. Finally, the framework supports closed-loop control, such as adaptive therapy, where clinical outputs and resistance patterns are fed back (dashed line) as input data to recalibrate the multiscale model and adjust subsequent treatments, enabling personalised, actionable decisions. Icons used in the figure are provided by Flaticon and are used under their standard attribution license. In silico neurology Nowadays, due to the ageing of the population, neurological disorders are recognised as the second leading cause of death167. Indeed, just the disorders affecting the human brain account for 13% of the total disease spectrum168. The term “neuro” pertains to nerve cells, while “logia” refers to study; neurological disorders are hence any disease of the nervous system. Indeed, neurology encompasses diagnosing and treating all states and diseases affecting the nervous system. Instead, “degeneration” refers to the deterioration of tissues or organs, leading to a loss of structure or function. Consequently, neurodegeneration can be described as a progressive decline in both the functionality and structure of neurons, eventually culminating in death169. Figure 4a shows the mortality rates related to neurological diseases in each country in 2012. Almost 10 million people died in 2019 because of neurological disorders, plus there were 349 million disability-adjusted life years lost (representing the loss of the equivalent of one year of full health)170, as shown in Fig. 4b.Fig. 4Neurological disease worldwide impact.Mortality (a) and disability-adjusted life years lost (b) globally. a Global distribution of deaths from neurological conditions in 2012. Data grouped by deciles (WHO Global Health Estimates 2012). Map adapted from an image available on Wikimedia Commons (file: neurological conditions world map-Deaths per million persons-WHO2012.svg). Licensed under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0). b Age-standardised disability-adjusted life year rates (ASR) attributed to neurological disorders for all ages and both sexes across 204 countries and territories in 2019. Figure adapted from ref. 170. Neurological disease worldwide impact. Mortality (a) and disability-adjusted life years lost (b) globally. a Global distribution of deaths from neurological conditions in 2012. Data grouped by deciles (WHO Global Health Estimates 2012). Map adapted from an image available on Wikimedia Commons (file: neurological conditions world map-Deaths per million persons-WHO2012.svg). Licensed under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0). b Age-standardised disability-adjusted life year rates (ASR) attributed to neurological disorders for all ages and both sexes across 204 countries and territories in 2019. Figure adapted from ref. 170. Eric L. Schwartz introduced the term “computational neuroscience” during a conference held in 1985 in California. Up to that moment, various names were used for the field, such as neural modelling, neural networks, and brain theory. The outcome of the conference was documented in the book Computational Neuroscience, published in 1990171. The convergence of computational neuroscience and neuronal communication with the advancement of new technology has unlocked a range of potential applications across diverse domains. These encompass healthcare, disease modelling, nanomachinery, pattern recognition, neuron-oriented hardware design, biological behaviour emulation, and data mining172. Computational neuroscience also deals with virtual reality173 and neurorobotics174. Moreover, within the spectrum of emerging applications, medical inquiries pertaining to the investigation of pathologies and potential therapeutic interventions have come to the fore. Therefore, today, we can speak of in silico neurology, or computational neurology, to describe this computational approach within the realm of medical research175,176. Recent advances have coupled virtual-reality-based neurorehabilitation with closed-loop brain-computer interfaces and physiologically realistic simulations, enabling patient-tailored therapy protocols and objective outcome prediction in stroke and motor disorders. These approaches often integrate multimodal data streams (EEG, kinematics, imaging) to adapt interventions in real time177,178. Neuroscience has highly complex, large-scale datasets (see ref. 179 for data-sharing issues). Neuroinformatics integrates these data (such as neuroimaging, electrophysiological recordings, and genetic data) and computational tools to collect, organise, analyse, and model neurological information171. Over the past few years, multiview learning methodologies—an ML technique–have gained significant popularity, and a considerable number of biomedical neurological applications that utilise multiview data have been documented in the literature, showing the potential of neuroinformatics applications for cutting-edge research problems180. In parallel, community standards and platforms (e.g., BIDS and OpenNeuro) plus model-description languages (e.g., NeuroML) have accelerated reproducibility and reuse by providing interoperable formats for neuroimaging, electrophysiology and multi-scale models; this FAIRification of neuroscience data underpins many successful large-scale and ML studies181–184. In silico studies revealed multiple neurological targets, e.g., microRNAs targets185 and for the antidepressant molecule ursolic acid186, identifying novel biomarkers187. Singh et al.187 conducted a comprehensive review of contemporary developments in in silico methodologies for biomarker discovery and validation. By integrating multi-omics strategies, these in silico techniques have demonstrated the capacity to swiftly and accurately identify new biomarkers. Moreover, recent initiatives on constructing extensive genomic databases and predictive computational platforms have streamlined the detection of potentially disease-causing genetic variants linked to rare human complex disorders188. The investigations related to the organisation of the brain network, both structural and functional, are one of the main addressed research in this field171, both at a microscopic level of brain dynamics (i.e., single neurons interconnected by synapses) and at a macroscopic level (i.e., brain regions and elementary processing units interconnected). In this context, neurodevelopment arises as a key player to consider, since understanding the process whereby the nervous system forms in humans is of capital relevance to identify possible defects that may arise during it and lead to disorders. Although neurodevelopment has been studied through the lenses of embryology and developmental biology, mathematics also has a role to play. Due to being a paradigmatic example of collective cell migration, neural crest formation has attracted the interest of mathematical modellers189,190. Neural tube formation has also been studied through mathematical modelling191. For a general review on mathematical models of neuronal growth, we refer the reader to ref. 192. Neuronal modelling falls under the umbrella of computational neuroscience. Its goal is to mathematically capture neuron behaviour by using mathematical equations to approximate important biological parameters. The integrate-and-fire (IF) model, the Hodgkin–Huxley (HH) model, the compartment model, and the 2D reduction model are widely used to model neurons172. The IF mechanism is not very complex, allowing the development of relatively simple models193. Neurons are modelled by exploiting an RC (Resistor-Capacitor) circuit. The neuron state is characterised by its membrane potential, which evolves according to the neuron’s synaptic inputs. Excitatory and inhibitory inputs increase and decrease, respectively, the membrane potential, while a time constant rules the potential decay toward a resting value in the absence of inputs (leaky neurons). When the potential reaches a threshold, an output spike is generated—the firing of the neuron—which causes the potential to reset at a basal value. A refractory period (the time period of complete inexcitability between depolarisation and repolarisation) can be explicitly considered. A few variants of the IF models exist, see ref. 194 for a review of IF neuron models. A highly comprehensive framework for describing neurons is the HH model195. Unlike the classical IF model, the HH model considers the complex biophysical properties of ion channels present in the neuronal cell membrane. Its adaptability extends to incorporating any number of ion channels within the membrane. Recognising the significance of their model, Hodgkin and Huxley were honoured with the Nobel Prize in 1963. A system of coupled differential equations characterising the dynamics of ion channel conductances, membrane capacitance, and the movement of ions across the membrane rules the model dynamics172. Neurons exhibit dendrites with diverse shapes and spatial arrangements, which are pivotal in shaping the neuron’s input and output characteristics. The compartment model approach serves as a valuable tool to explore how dendrite shape and structure impact the neuron’s behaviour172. The HH model, renowned for its complexity, presents challenges in both analysis and visualisation. In contrast, the simpler IF model assumes an action potential is triggered when the neuron’s potential reaches a constant threshold. However, certain observations are in opposition to this concept. For instance, using longer pulse currents for stimulation reduces the firing threshold, while shorter pulses increase it. Moreover, data-driven evidence suggests that the firing threshold potential depends on changes in neuron current during a step jump. To explore and understand neuron behaviours, it becomes necessary to condense the four nonlinear differential equations of the HH model into two: the 2D reduction model. The HH model offers greater empirical detail compared to any IF model, providing predictions about ionic currents and spike shape. Nevertheless, these predictions do not hold up against empirical observations. This outcome is not entirely unexpected, considering that spikes initiate in the axon, a factor not addressed by the single-compartment model. The surprising aspect is that the IF model, even with channel properties from patch-clamp measurements, turns out to be more realistic than the HH model when predicting neuronal spiking responses, particularly at high input frequencies196. The inherent complexity and high-dimensional parameter space of classical biophysical models, such as the IF and HH frameworks, have traditionally hindered their calibration from experimental data. However, models based on data science, statistics, and ML are seeing a surge in popularity due to the increasing availability and quality of neural data197. Recently, coupling these frameworks with modern ML has emerged as a powerful strategy to overcome parameter estimation challenges while maintaining biophysical interpretability198. Within this landscape, supervised methods provide a spectrum of tools to capture neuronal dynamics: from simple Linear Regression198 and complex support vector machines with radial basis function kernels199,200, to random forest and gradient boosting machines, which offer a crucial balance between predictive power and interpretability198. Complementing these are advanced signal analysis techniques such as the Frequency Modulated Möbius (FMM) approach and Fourier methods, which utilise amplitude-frequency decompositions201 to accurately fit diverse oscillatory patterns202. Specifically, the multicomponent FMM model has demonstrated significant potential in neuroscience for deciphering complex signals203, with proven applications in automatic signal analysis204. This synergy is further advanced by physics-informed ML and differentiable frameworks like Jaxley, which embed mechanistic knowledge into data-driven architectures205,206 to estimate latent physiological variables from noisy recordings207. By bridging the gap between mechanistic modelling and data-driven discovery, these hybrid approaches ensure that identified parameters remain biologically meaningful and clinically trustworthy208. In20, we use IF models to develop the first model of periodic leg movement, a sensory disorder present in many neurological diseases (for example, in restless legs syndrome). Neural networks are studied from the perspective of communication engineering in neural communication. Indeed, information theory is exploited to characterise the neuron network features. This includes understanding the error rate in information transmission within the network, determining the theoretical maximum information capacity of neuron networks, and identifying constraints imposed by the network on the flow of information172. Moreover, time-aware approaches to Information Theory uncover compelling evidence for the significance of processes (in contrast to states) in neural functioning, opening the way to generalised Information Theory209. All functions within the body, whether voluntary or involuntary, the origins of mental disorders, and any other aspects of brain activity, depend on the intricate operations of the neural network. Understanding this network requires knowledge of the communication process between neurons, commonly referred to as neuro-spike communication210. For identifying connectivity patterns, see ref. 211 for a review of the applications of Graph Theory to human brain networks. More recently, GNNs and dynamic graph models have extended classical graph-theoretic approaches by enabling end-to-end learning directly on brain graphs (structural or functional), improving diagnostic classification and allowing discovery of distributed patterns that are not accessible via pairwise statistics212,213. Advanced fusion mechanisms enable us to combine these GNNs with entropy-guided space state models, to help us handle complex, multiscale biological data in neuroscience121. The brain, a key argument in neurological research, is an incredibly complex neuronal network and has also received much attention from the in silico community, for example, see refs. 214–218. Indeed, we have witnessed a remarkable upsurge in large-scale initiatives and projects dedicated to the mapping, modelling, simulation, and creation of atlases for the human brain. Notable among these are endeavours such as the Human Brain Project, the BRAIN Initiative, the Human Connectome Project, the Blue Brain Project, the Big Brain, the Allen Brain Atlas, and the Brainnetome, to name a few219. Nowadays, it is possible to exploit cloud services to perform brain simulation, such as EBRAINS220. The simulations of brain function enable researchers to study brain dynamics under normal and disordered conditions. These simulations provide a deeper understanding of the underlying mechanisms causing neurological disorders and help develop targeted therapies. Large-scale simulators (e.g., The Virtual Brain) and cloud infrastructures permit the construction of subject-specific whole-brain models that combine individual connectivity, regional dynamics, and biophysical constraints; such models support hypothesis testing, virtual lesioning, and in silico trials of stimulation protocols, forming a bridge toward personalised neuromodulation strategies214,221,222. Computational neuroanatomy, the in silico modelling of brain anatomy, characterises brain shape and neuroanatomical configuration, helping to understand the structural organisation of the brain and how it relates to neurological disorders223. This includes the construction of brain atlases and the analysis of brain connectivity. G. Hickcok combined psycholinguistic and motor control techniques for speech production, developing a hierarchical state feedback control model based on neuroanatomy224. The interested reader on baby brain neuroanatomy could refer to Li et al.’s review225. Neuroimaging analysis applies in silico techniques to analyse and process neuroimaging data, such as magnetic resonance imaging, functional magnetic resonance imaging, and positron emission tomography scans. Müller and colleagues provided some practical rules to follow in meta-analysis226. Instead, in ref. 227, reproducibility is addressed. These methods aid in identifying patterns of brain activity and structural changes associated with different neurological conditions, such as eating in healthy228 and disorder-affected229 subjects, neurorehabilitation177, emotion230, and neuropsychiatric disorders231. Additionally, advances in deep learning have improved feature extraction and classification from neuroimaging, but concerns about interpretability and generalisability motivate hybrid pipelines that combine mechanistic models with data-driven components184,232. Many complex diseases are investigated in computational neuroscience, exploiting multiscale models. Lytton et al. perform an excellent clinical-oriented review covering epilepsy, traumatic brain injury, ischaemic disease, neurorehabilitation, drug addiction, schizophrenia and neurostimulation233. Zang and colleagues discuss probabilistic methods adopted in Computational Neuroscience234. Lustoza Rodrigues et al.235 highlights the importance of utilising computational methods, including quantitative structure-activity relationship prediction models, molecular docking, pharmacokinetic models, and molecular dynamics. Additionally, it showcases potential multi-target drug candidates. Multiscale and hybrid frameworks are now being used to build DTs of the brain–personalised, mechanistic computational avatars that integrate patient data across scales (molecular to systems) and can be used to test interventions in silico before clinical application236,237. Finally, important neurological disorders are highly investigated computationally. For example, Alzheimer’s disease238 and Parkinson’s disease19, two of the main neurodegenerative illnesses, have been investigated by exploiting in silico approaches. In general, the main computational contributions are achieved thanks to the analyses of complex neurological data, model simulations and analyses, and identifying potential targets for therapeutic interventions. In silico epidemiology When a higher-than-expected occurrence of an illness or health-related event within a community or region is observed, the experts say that an epidemic has begun. When an epidemic disseminates globally, it is designated as a pandemic. Epidemics have been extremely influential events, probably changing the history and course of humanity, e.g., the ancient plagues, the decline of the Han empire in the third century China, the almost 50% European population demise (estimated in 50 million people) caused by the Black Death in the fourteenth century, and the Aztec defeat by the Spanish empire in the sixteenth century7. In 2019, more than thirteen and a half million people worldwide died from infectious diseases, almost one-third of which co-occurred with non-communicable diseases (i.e., not spread through infection or through other people, but are typically caused by unhealthy behaviours)239. Figure 5 shows the death rate from infectious and parasitic diseases per million people in 2023 worldwide.Fig. 5Global distribution of deaths from infectious and parasitic diseases per million persons in 2023.The rates are age-standardised to enable comparisons between countries and over time. Data source: WHO Mortality Database (2025). It should be noted that not all deaths in each country may have been registered with a specified cause of death. Figure adapted from ref. 324. Global distribution of deaths from infectious and parasitic diseases per million persons in 2023. The rates are age-standardised to enable comparisons between countries and over time. Data source: WHO Mortality Database (2025). It should be noted that not all deaths in each country may have been registered with a specified cause of death. Figure adapted from ref. 324. Nowadays, the effects of infectious diseases in real life, such as social distancing, business and educational closures, increased levels of anxiety, depression, poverty, crime, and domestic violence, in addition to health consequences240, have been highly understood worldwide because of the Covid-19 pandemic. The Covid-19 experience also highlighted the crucial role of near real-time digital data streams (mobility, digital contact traces, health-care occupancy, and participatory surveillance) in enabling faster model updates and more targeted interventions241–243. At the same time, the pandemic exposed limitations in model communication and uncertainty quantification, fuelling a movement towards ensemble forecasting and clearer uncertainty reporting in public health decision support21,244. Moreover, the Covid-19 pandemic has caused more than 7.1 million confirmed deaths worldwide (as of November 2025), with the majority occurring in the first two years–over 5.5 million deaths by January 2022245. Epidemiological models246,247 are fundamental for understanding epidemic evolution and optimal control strategies to deploy limited resources when epidemics occur in various interconnected areas248–250. Researchers from diverse fields have employed various mathematical techniques and computational tools to study epidemic diffusion in human societies, utilising methodologies that span differential equations, agent-based models, stochastic processes, statistical analysis, graph theory, ML, AI, computer simulations, geographic information systems, and high-performance computing8. This interdisciplinary approach has led to the development of mathematical and computational frameworks for epidemic modelling, integrating epidemiology with mathematics, sociology, management science, control theory, network science, and computer science. Note that these approaches go beyond infectious diseases because they can be applied to describe a general reaction-diffusion process, for example, the spread of innovations, fake news, or computer viruses. In particular, modern computational epidemiology increasingly couples mechanistic transmission models with heterogeneous data (genomes, mobility, clinical records, wastewater, and social-media signals) to improve inference on transmission chains, variant emergence, and spatial spread241,251. These hybrid pipelines allow real-time phylogeographic analyses and phylodynamic-informed forecasts that were not commonplace prior to the Covid-19 pandemic251. The motivations for improving in silico approaches to epidemiology are multiple252. Future pandemics pose challenges due to emerging trends, such as an ever-more-connected (e.g., urbanisation, trade, and travel)253 and digital254 world, ageing population, and immune vulnerabilities. Public health epidemiology must comprehend the complex systems derived from the coevolution of epidemics, social interactions, behaviours255, and policies. Essential are mathematical models, big data23, and real-time capabilities to support decisions during epidemics7. Additionally, the community has prioritised improving transparency, reproducibility, and model comparison (benchmarking), as well as building robust ensemble forecast systems that combine disparate model types to increase reliability for policymakers21. Epidemiological models can be categorised into three broad classes: compartmental models, (complex) network models, and agent-based models8. Especially the latter two approaches are developed exploiting the in silico framework7. Table 2 compares the three approaches, highlighting their strengths and limitations.Table 2Comparative summary of mathematical and computational modelling paradigms in epidemiologyModel classApproachStrengthLimitationCompartmentalMarkov chain; ODEsHigh analytical tractability; calculation of thresholds (R0).Assumes a homogeneous and “well-mixed” population.Complex networksGraph theoryCaptures contact structures and spatial dissemination patterns; describes metapopulations.Requires high-resolution data on human mobility and connectivity.Agent-basedBottom-up simulationsAccounts for individual heterogeneity and emergent behaviours.High computational cost; challenges in parameter calibration. Table 2Comparative summary of mathematical and computational modelling paradigms in epidemiologyModel classApproachStrengthLimitationCompartmentalMarkov chain; ODEsHigh analytical tractability; calculation of thresholds (R0).Assumes a homogeneous and “well-mixed” population.Complex networksGraph theoryCaptures contact structures and spatial dissemination patterns; describes metapopulations.Requires high-resolution data on human mobility and connectivity.Agent-basedBottom-up simulationsAccounts for individual heterogeneity and emergent behaviours.High computational cost; challenges in parameter calibration. Comparative summary of mathematical and computational modelling paradigms in epidemiology Arguably, Bernoulli developed the first mathematical epidemiological models7,8. Compartmental approaches primarily focus on capturing macroscopic trends256. They employ assumptions such as population homogeneity and a well-mixed environment to portray epidemic spread at a general level. These models simplify the intricate process of epidemic propagation by considering only essential factors, which are represented through parameters with average values in equations. While this allows mathematical models to analyse overarching trends and characteristics (e.g., basic reproduction number and epidemic threshold), they may lack the nuanced details required to capture the intricate nature of epidemic spread. Deterministic and stochastic epidemic models are two broad subclasses of mathematical models. Deterministic compartmental models divide the population into compartments, or classes, grouping individuals based on their health states. The compartments are assumed to be well-mixed and homogeneous. ODEs rule the transitions between different compartments, corresponding to changes in the individual health states. The most famous and simple formulations are the SIR (susceptible, infectious, and recovered) model and the SIS (susceptible, infectious, and susceptible) model. Instead, stochastic epidemic models are developed by exploiting stochastic techniques, such as Markov Chains and the Monte Carlo method257. The Reed-Frost model is extensively utilised, such as in ref. 258. It describes the infection process between individuals as a binomial stochastic process, making specific assumptions about a closed population, all having equal contact opportunities. In each (discrete) time step, susceptible individuals can be infected by infectious individuals. The latter then becomes immune. Formulas allow the computation of flows into infectious and immune compartments. The process is reiterated until the entire population achieves immunity or there are no longer individuals capable of spreading the infection. The model can be iterated multiple times with variations in the starting conditions to investigate their impact on the epidemic course. Computational methods utilise a diverse array of simulation techniques to offer a more intricate representation of real-world scenarios8. These approaches often construct epidemic models based on complex network models259 and agent-based models240. However, the higher level of detail in computational models demands significant data availability and computational resources. Recent advancements in data accessibility and affordable high-performance computing have facilitated the increased use of computational methods to study the propagation of epidemics23,252,254. Concurrently, improvements in software infrastructure (open-source pipelines, cloud computing, and reproducible workflows) and in standardised data-sharing platforms have reduced barriers to deploying large-scale in silico experiments and rapid scenario analysis241,251. The past decade has been marked by a significant improvement in (complex) network-structured epidemiological models. These models exploit networks to describe individuals or populations (network nodes) and their interactions (network edges). Generally, these complex network models can be categorised into two main groups: mean-field models (focusing on the spreading dynamics within complex networks) and networked models (focusing on numerical simulations of epidemics in such networks)8. In the spreading dynamics models, the mean-field theory is employed to analyse complex networks, and then differential equations are formulated to describe the propagation of epidemics8. Like compartmental models, individuals in these complex network models are classified into different groups based on their health status. In homogeneous networks, the infectious node infectivity is derived from the average node degree, while in heterogeneous networks, nodes are further separated into subgroups based on their degrees8. Instead, numerical simulations in complex networks employ ODEs or AB models to represent contact patterns and infection spreads along the links. On the one hand, simulations with ODE models typically focus on homogeneous groups (compartments) of people, calculating the aggregated numbers of individuals that are in each state (e.g., susceptible, infected, and resistant) as a function of time. The population is divided into geographical areas, where local infections are governed by the ODEs that rule the people flows among the different model compartments (belonging to the same geographical area) based on the current model state. Thus, the network structure divides and interconnects the populations into regions, forming a network of sub-populations, with the connections between them representing the movement of individuals through transportation and mobility infrastructures and allowing the infection spreads256,257,260. On the other hand, simulations with AB models typically focus on heterogeneous individuals (with agent-specific characteristics), the network nodes, interacting through the network links. Among subsequent (discrete) time steps, probabilistic procedural rules update the health state of each agent (e.g., susceptible, infected, and resistant) based on the current agent state and the state of its neighbour (linked) agents. Thus, AB models are stochastic, discrete in time, and spatially explicit240,261. A significant focus of most epidemiological network-based models lies in understanding the impact of network topologies on the spread of the disease, particularly on networks such as small-world and scale-free networks. Researchers have also explored epidemic spread within weighted networks and adaptive networks in recent years, leading to intriguing insights into the dynamics of epidemics in various network structures8,256. Recent work has also integrated fine-grained mobility networks into metapopulation SEIR (with susceptible, exposed, infectious and recovered compartments) frameworks to explain inequities in exposure risk and to identify high-impact locations for interventions (e.g., specific points of interest that drive disproportionate transmission)242. In recent times, there has been a growing interest in delving deeper into the intricate evolutionary process of how epidemics spread, driven by an enhanced comprehension of the broader patterns in epidemic diffusion. As a result, there is an increasing demand for epidemic models that can provide a more nuanced and precise representation of the complex realities involved. Thus, with the goal of developing more realistic models, the AB framework is exploited8. Indeed, AB models have the great (potential) advantage of capturing the heterogeneity in individual attributes and behaviours, and also accounting for the stochastic nature inherent in epidemic spreads. By adopting a bottom-up approach, AB modelling places focus on a faithfully detailed representation of the behaviour and interactions among individuals, the micro-level, which ultimately leads to the emergence of the macro-level phenomena associated with the spread of the epidemic240,261,262. It is essential to note that, as usual, the AB framework leads to certain challenges: here, data availability and computational complexity. However, the ongoing progress in computational capabilities, the increasing availability of data, and the development of sophisticated computational algorithms have significantly boosted the popularity of agent-based models in gaining a more comprehensive understanding of the dynamics behind the spread of epidemics240,261,262. Nevertheless, these advantages must be balanced with rigorous calibration, sensitivity analysis, and methods to quantify structural uncertainty–areas that have seen substantial methodological development during and after the Covid-19 pandemic21,244. It is essential to underline that there is no perfect, correct, or superior modelling approach to transmission dynamics for infectious diseases in general. The model choice depends on factors such as the available data, the specific research question, the time and space scales of the study, and other relevant disease-specific considerations260. Commonly exploited methodologies include highly detailed agent-based models and large-scale spatial metapopulation models, i.e., ODEs network-structured models259. Indeed, the importance of spatial modelling has been recognised, especially in an integrated approach for early detection and intervention during epidemics263. Many different models have been developed to study specific infectious diseases, e.g., influenza262, West Nile virus261, tuberculosis264, bacterial265, HIV266, and pertussis267. Also, the toolbox has expanded to include genomic phylodynamics for pathogen surveillance (tracking variants and transmission chains) and wastewater-based epidemiology as complementary indicators for community-level infection trends251. During a pandemic, decision-making by local, national, and global public health authorities involves critical resource allocation. These resources go beyond routine disease monitoring, focusing on tracking and controlling potential pandemic outbreaks. This allocation hinges on two key factors: estimating the severity risk posed by the infectious pathogen and evaluating the likelihood of the outbreak reaching a specific region and causing widespread, severe illness7. Unfortunately, precise risk assessment, especially during the early pandemic stages, is highly challenging. Thus, the significance of modelling and decision support systems in analysing alternative scenarios becomes increasingly paramount. Various works have been proposed in the literature addressing optimal resource control during epidemics, for example, see refs. 22,268–272 for recent studies. Complementary to optimal control, fast scenario-generation and robust decision-making under deep uncertainty (including model ensembles and stress tests) have been emphasised as practical approaches to guide tactical and strategic choices21. Focusing on the well-known Covid-19 pandemic, many compartmental models extending the well-known SIR model have been proposed for the Covid-19 pandemic, e.g.,273–277. This modelling approach is reasonably predictive278 for infectious diseases that are transmitted among humans. A few works specifically investigated the effect of different COVID-19 vaccine allocation strategies via a model-informed approach279,280. In particular,279 analyses the effect of an equitable distribution of the vaccine doses administered during 2021, while280 studies the advantage of using an infection-dependent allocation strategy of the vaccines with respect to a constant one. In281, we propose an extended version of the model proposed by Gatto et al.274 at a global scale, including vaccinations and contact-tracing with new certified cases. An extended work also includes an autoregressive-exogenous (ARX) model to deduce the containment measures of national governments based on hospital occupancy25. We investigate optimal vaccine allocation strategies by varying both vaccine availability and their distribution. We conclude that to stop the COVID-19 (and similar) pandemic, we obviously need to increase vaccine access, but interestingly, this increase is less severe if global equitable access is granted (i.e., proportionally to population size). To bring this section to a close, in Table 3 we refer to the most relevant works discussed in this section. For each work listed, we specify the type of model employed, the diseases considered, whether the authors evaluated optimal strategies, and a brief note highlighting the main focus of the work.Table 3Summary of discussed references categorised by model type, disease, and focus on optimal strategiesRef.Model typeDiseaseOptimal strategyNotes267Comp.PertussisNoStochastic Petri Nets for age-structured vaccination strategies.275Comp.COVID-19NoSIRD model for forecasting case fatality and recovery rates.258Comp.GeneralNoExtension of Reed-Frost models to random graphs.269Comp.TuberculosisYesDiscrete age-structured model for child/adult populations.268Comp.GeneralYesSIS model numerical solution for optimal control (treatment/vaccination).248Comp.HIVYesOptimal resource allocation across independent populations.271Comp.FungiYesManagement of fungicide mixtures to control resistance.277Comp.COVID-19YesFeedback control of epidemic stability via daily testing reports.243Comp.COVID-19YesTransmission timing and impact of digital contact tracing.273Comp.COVID-19YesSEIR model with behavioural changes and governmental actions.22Comp. / CNCOVID-19YesTiming of one-shot interventions in metapopulations.280Comp. / CNCOVID-19YesPro rata vs optimal vaccine allocation via economic modelling.25Comp. / CNCOVID-19YesGlobal metapopulation with ARX feedback for vaccine access.242CNCOVID-19NoMetapopulation SEIR with dynamic mobility networks in US cities.249CNGeneralYesMetapopulation control directs treatment to low-prevalence regions.250CNGeneralYesAllocation across regions balancing efficiency and equity.276CNCOVID-19YesNetwork model for intermittent decentralised regional lockdowns.270CNCOVID-19YesSpatiotemporal resource allocation via network discovery.272CNCOVID-19YesSemidefinite programming for curing rates in arbitrary networks.279CNCOVID-19YesGlobal metapopulation model on equitable vaccine sharing.274CNCOVID-19YesSpatially explicit model assessing Italian containment measures.262ABInfluenza ANoABM addressing ecology and evolution of polymorphic pathogens.264ABTuberculosisNoABM focuses on age-structure and progression stages.Comp. compartmental, CN complex network, AB agent-based. Table 3Summary of discussed references categorised by model type, disease, and focus on optimal strategiesRef.Model typeDiseaseOptimal strategyNotes267Comp.PertussisNoStochastic Petri Nets for age-structured vaccination strategies.275Comp.COVID-19NoSIRD model for forecasting case fatality and recovery rates.258Comp.GeneralNoExtension of Reed-Frost models to random graphs.269Comp.TuberculosisYesDiscrete age-structured model for child/adult populations.268Comp.GeneralYesSIS model numerical solution for optimal control (treatment/vaccination).248Comp.HIVYesOptimal resource allocation across independent populations.271Comp.FungiYesManagement of fungicide mixtures to control resistance.277Comp.COVID-19YesFeedback control of epidemic stability via daily testing reports.243Comp.COVID-19YesTransmission timing and impact of digital contact tracing.273Comp.COVID-19YesSEIR model with behavioural changes and governmental actions.22Comp. / CNCOVID-19YesTiming of one-shot interventions in metapopulations.280Comp. / CNCOVID-19YesPro rata vs optimal vaccine allocation via economic modelling.25Comp. / CNCOVID-19YesGlobal metapopulation with ARX feedback for vaccine access.242CNCOVID-19NoMetapopulation SEIR with dynamic mobility networks in US cities.249CNGeneralYesMetapopulation control directs treatment to low-prevalence regions.250CNGeneralYesAllocation across regions balancing efficiency and equity.276CNCOVID-19YesNetwork model for intermittent decentralised regional lockdowns.270CNCOVID-19YesSpatiotemporal resource allocation via network discovery.272CNCOVID-19YesSemidefinite programming for curing rates in arbitrary networks.279CNCOVID-19YesGlobal metapopulation model on equitable vaccine sharing.274CNCOVID-19YesSpatially explicit model assessing Italian containment measures.262ABInfluenza ANoABM addressing ecology and evolution of polymorphic pathogens.264ABTuberculosisNoABM focuses on age-structure and progression stages.Comp. compartmental, CN complex network, AB agent-based. Summary of discussed references categorised by model type, disease, and focus on optimal strategies Comp. compartmental, CN complex network, AB agent-based. Shared paradigms and bottlenecks: cross-disciplinary synthesis The comparative analysis of oncology, neurology, and epidemiology reveals a profound methodological convergence that transcends the specific biology and scales of each field. Although the biological problems and clinical questions differ, the underlying mathematical structures are remarkably similar: all three domains deal with heterogeneous data, nonlinear dynamics, multiscale interactions, partial observability, and uncertainty-driven decision-making. In this sense, in silico modelling is not merely a domain-specific tool, but a unifying framework for extracting mechanistic insight, integrating multi-source information, and guiding intervention design across complex biomedical systems. Specifically, we first introduce the theoretical paradigm (the systems and control foundation), then its practical implementations (hybrid modelling and DTs) and shared pillars (the systems and control foundation). We conclude by highlighting universal bottlenecks in Section Universal bottlenecks. A graphical representation of the shared paradigms is given in Fig. 6, from patient data acquisition to complex-systems modelling, predictive simulation, and clinical translation, while also highlighting the main methodological bottlenecks that limit current implementations.Fig. 6Shared paradigms and bottlenecks.The top panel shows a cross-disciplinary Systems and Control architecture, linking multi-source patient data via hybrid modelling and DTs to clinical applications. Central to this is the feedback control loop, where predictive simulations drive adaptive treatments, which subsequently generate new clinical data used to refine the intervention. The bottom panels use bottleneck indicators to categorise structural and operational obstacles–ranging from data heterogeneity and numerical stiffness to practical identifiability and clinician scepticism–that must be addressed to ensure robust clinical translation. Icons used in the figure are provided by Flaticon and are used under their standard attribution license. Shared paradigms and bottlenecks. The top panel shows a cross-disciplinary Systems and Control architecture, linking multi-source patient data via hybrid modelling and DTs to clinical applications. Central to this is the feedback control loop, where predictive simulations drive adaptive treatments, which subsequently generate new clinical data used to refine the intervention. The bottom panels use bottleneck indicators to categorise structural and operational obstacles–ranging from data heterogeneity and numerical stiffness to practical identifiability and clinician scepticism–that must be addressed to ensure robust clinical translation. Icons used in the figure are provided by Flaticon and are used under their standard attribution license. The systems and control foundation A unifying analytical framework for these three domains is provided by Systems and Control theory. In each field, the modelling objective has shifted from descriptive observation to active system steering: that is, from passive characterisation of dynamics to the design of feedback-driven interventions acting on partially observed, nonlinear, and often stochastic systems155,282,283. In oncology, the use of evolutionary double binds51,149,151 and adaptive therapies146,152,153,155,157,284 aims to steer the tumour-host system toward a stable state, ideally a cancer-free equilibrium, typically preventing or delaying the emergence of resistant clones82,141,148,285,286. These strategies explicitly exploit eco-evolutionary dynamics, where treatment acts as a selective pressure shaping tumour composition over time, thus requiring control approaches that incorporate evolutionary game theory and population dynamics150,156. In neurology, closed-loop neurostimulation and brain-computer interfaces act as real-time feedback controllers to restore homeostatic neural activity287,288 in disorders such as Parkinson’s disease289 and psychiatric disorders290. In epidemiology, adaptive vaccine allocation and non-pharmaceutical interventions are designed as optimal control problems to stabilise population-level transmission dynamics during outbreaks22,25. Beyond these parallels, an emerging unifying perspective is provided by evolutionary control theory, which extends classical control frameworks to systems in which the state dynamics are shaped by selection, adaptation, and competition291,292. In oncology, this perspective is already central, but it is increasingly relevant in epidemiology through the control of pathogen evolution (e.g., vaccine escape or resistance), and in neurology through adaptive processes such as synaptic plasticity and network reorganisation. This common logic suggests that advances in one field—such as evolution-aware control algorithms in cancer–can provide a blueprint for managing pathogen evolution in the other areas292. More generally, this cross-fertilisation highlights that control strategies designed under a systems perspective are often transferable across scales and domains, provided that the underlying dynamical structure is preserved. In addition, successful modelling approaches in one field are derived in other fields; for instance, see ref. 293 for the application of an epidemic spreading model to neurology, and ref. 294 for the use of a SIR-like model in oncology. This methodological transfer reinforces the view of a shared mathematical backbone underpinning seemingly disparate biomedical problems. Hybrid modelling and DTs Biological systems are inherently multiscale, as they encompass processes ranging from the molecular to the population level. No single methodology can represent them accurately and effectively on its own. In this context, the hybrid paradigm has emerged to integrate complementary approaches, linking scales through effective interactions, reduced-order representations, or coarse-grained variables295,296. This integration often combines mechanistic, physics-based formulations with data-driven techniques, such as ML, to leverage high-dimensional data while enforcing consistency with biological principles. By embedding learning components within structured dynamical systems–often referred to as physics-informed learning–hybrid models improve generalisation and interpretability, which is particularly critical in biomedical applications where data are frequently sparse, noisy, or heterogeneous205,297. This hybrid foundation is central to the DT paradigm. A DT is a patient-specific computational model that integrates mechanistic knowledge with multiscale data to simulate biological evolution under both physiological and pathological conditions. Unlike static models, DTs can be adaptive, closed-loop systems continuously updated through data assimilation to refine predictions over time298,299. Within this framework, the twin acts as a predictive surrogate, enabling the design of personalised interventions through control-theoretic approaches300. Furthermore, hybrid DTs facilitate structured uncertainty quantification, combining probabilistic inference with mechanistic constraints to provide the reliable predictions necessary for clinical decision support301,302. The application of hybrid DTs across diverse domains addresses the shared need to model complex, nonlinear interactions and feedback loops. In oncology, DTs integrate intracellular signalling, tumour dynamics, and PK/PD to optimise dosing and predict resistance303. Recent developments also incorporate evolutionary and eco-dynamic principles to guide adaptive therapy strategies155. In neurology, personalised whole-brain models link structural and functional connectivity with dynamical systems to predict disease progression and neuromodulation responses236,237,304,305. Finally, in epidemiology, DTs couple host and population dynamics306. By combining mobility data and immunological states, these multilevel models serve as data-driven simulators for real-time surveillance and scenario analysis during outbreaks307,308. Despite their potential, hybrid DTs face challenges regarding model identifiability, computational complexity, and the trade-off between biological realism and transparency, which will be further discussed in the context of cross-disciplinary bottlenecks (Universal bottlenecks section). Addressing these hurdles is critical for translating these integrative strategies into robust, clinically validated tools across biomedical domains298,299. Shared methodological pillars The convergence of oncology, neurology, and epidemiology within the in silico framework rests upon four foundational methodological pillars that translate biological complexity into actionable mathematical models. First, the network-based paradigm provides a universal language to describe interactions across scales. Whether modelling intracellular signalling pathways and gene regulatory circuits in oncology11,91, the structural and functional connectome in neurology233, or the contact structures of metapopulations in epidemiology8,274, network science identifies the topological determinants of system behaviour. Despite the vast differences in physical scale–from biomolecules to global networks–properties such as small-world, scale-free distributions, and dynamical processes like synchronisation or diffusion remain the core drivers of system resilience and pathogen spread. Second, the data-informed paradigm bridges these structural representations with real-world observations. Rather than replacing physics-based approaches, this paradigm leverages high-dimensional data (imaging, omics, spatial data) to refine parameter estimation and calibrate model structures. Central to this paradigm is mechanistic learning, where data-driven components are constrained by governing equations or biological priors205,297. This integration ensures that models remain adaptive and personalised while maintaining biological plausibility and interpretability, a strategy now foundational across biomedical domains102,166,309,310. Third, the complexity paradigm addresses the emergent behaviours that arise from these networked and data-informed systems. AB models serve as the primary tool for exploring this bottom-up complexity, simulating how simple individual-level rules–be it single-cell interactions in oncology138, neuronal firing in neurology233, or human mobility in epidemiology240–scale up to shape macro-level phenomena such as clonal selection, neural oscillations, or epidemic trajectories (see Tables 1 and 3). Understanding these emergent dynamics is a prerequisite for effective system manipulation. Finally, the feedback control paradigm transforms these insights into a rigorous framework for designing targeted and adaptive interventions. In oncology, this is exemplified by adaptive therapies that use tumour response as a feedback signal to adjust dosing and prevent resistance146,148,152. This logic extends to neurology through closed-loop neurostimulation systems that monitor brain activity in real-time to suppress pathological states5, and to epidemiology, where control-theoretic approaches optimise vaccine allocation and non-pharmaceutical interventions based on dynamic transmission rates and hospital occupancy22,25. By closing the loop between observation and action, this pillar provides the operational link to the formal systems and control foundations of in silico personalised medicine. Universal bottlenecks The advancement of in silico medicine is hindered by a set of shared structural and operational obstacles that span across oncology, neurology, and epidemiology. The first and most prominent is the multiscale integration bottleneck. Bridging the gap between molecular-level events—such as DNA mutations, synaptic signalling, or pathogen evolution–and organism- or population-level outcomes remains computationally daunting. High-fidelity frameworks, including hybrid PDE-AB models in oncology (H in Table 1), HH frameworks in neurology, and AB models in epidemiology (Table 2), often suffer from numerical stiffness due to the vast differences in temporal and spatial scales. This disparity forces a constant trade-off between biological realism and computational tractability, necessitating more sophisticated coarse-graining techniques and surrogate modelling. The second barrier lies in data integration and standardisation. Despite the abundance of omics and imaging data, these often provide only snapshots that lack the longitudinal resolution required to calibrate truly dynamic models1. Furthermore, significant heterogeneity arising worldwide from different hospitals, devices, or acquisition protocols complicates the creation of cohesive datasets311. This fragmentation leads to critical issues in parameter inference and model identifiability, manifesting at two distinct levels. While a model might be structurally identifiable in theory–meaning its equations allow for unique parameter solutions under ideal conditions–it often fails the test of practical identifiability due to the limitations of real-world clinical data312. When datasets are sparse, noisy, or not properly integrated, the resulting lack of information prevents the precise estimation of parameters, even in structurally sound models. Optimal experimental design (how much data to collect and when to collect it) that ensures parameter identifiability (permitting confidence in model predictions), while minimising experimental time and costs, should be followed313. Otherwise, multiple, divergent parameter sets may yield the same output, hindering the model’s ability to uniquely capture the underlying biological mechanisms. While the FAIR (findable, accessible, interoperable, and reusable) data initiatives provide a vital theoretical roadmap to mitigate these issues, their full-scale clinical implementation still lags behind the requirements of high-performance modelling314. Directly linked to these challenges is the computational bottleneck. As models incorporate increasing biological detail–such as 3D spatial structures, stochasticity, or large-scale networks–their cost increases exponentially. Simulating a DT of the brain220,236, an H model in oncology (Table 1), or a city-scale epidemic7 requires high-performance computing resources seldom available in clinical settings. This also limits the application of optimal control theory in high-dimensional spaces82. Consequently, the development of accelerated solvers and reduced-order representations, such as mesoscopic simulators315–317, is a critical priority for enabling real-time clinical decision support. The final hurdle is the translational bottleneck, which governs the transition from simulation to practice. Regulatory frameworks for model-informed medical devices are still in their infancy10. Beyond legal hurdles, black-box models often face scepticism from clinicians. Overcoming this requires ensuring model interpretability, robustly quantifying uncertainty21, assessing the probability of toxicity318, and demonstrating clear cost-benefit advantages over current standards of care13,25. Addressing these multidimensional challenges is essential for the broad adoption and reliability of next-generation in silico frameworks. Conclusions In silico medicine has evolved from a supportive computational tool into a central pillar of modern biomedical research, enabling the systematic investigation of complex diseases across scales, modalities, and disciplines. This work has revisited advances in oncology, neurology, and epidemiology, demonstrating how shared principles, such as systems and control theory, network theory, nonlinear dynamics, and multiscale modelling, provide a unifying framework for understanding, predicting, and intervening in complex biological systems. This endeavour is inherently interdisciplinary, integrating fields from medicine and biology to mathematics, computer science, physics, and engineering. Such synergy is essential, as it transforms theoretical understanding into clinically actionable insights. The mathematical foundations of these frameworks serve as a critical bridge between abstract theory and practical implementation, enabling quantitative reasoning that guides medical decision-making. Central to this progress is sustained computational innovation. The development of advanced tools and high-performance computing has revolutionised our capacity for large-scale simulation and analysis, making detailed, multi-scale models and near real-time analysis a reality. These capabilities are amplified by data-driven and model-informed approaches. The intelligent integration of mechanistic models, which offer interpretability and causal insight, with data-driven and ML techniques enhances predictive power, scalability, and supports robust, transparent decision-making. The impact of this paradigm is vividly illustrated across key domains:In oncology, computational models yield profound insights into tumour growth, treatment response, and the evolution of drug resistance. They enable in silico testing of new drugs and regimens, accelerating the path toward personalised therapy.In neurology, models such as IF and HH allow detailed simulations of neuronal behaviour, advancing the study of neurodegenerative diseases and opening avenues for personalised neurotherapeutics when combined with clinical data.In epidemiology, computational models are indispensable for predicting disease spread, evaluating interventions, and informing public health policy, as recently underscored by global health crises. In oncology, computational models yield profound insights into tumour growth, treatment response, and the evolution of drug resistance. They enable in silico testing of new drugs and regimens, accelerating the path toward personalised therapy. In oncology, computational models yield profound insights into tumour growth, treatment response, and the evolution of drug resistance. They enable in silico testing of new drugs and regimens, accelerating the path toward personalised therapy. In neurology, models such as IF and HH allow detailed simulations of neuronal behaviour, advancing the study of neurodegenerative diseases and opening avenues for personalised neurotherapeutics when combined with clinical data. In neurology, models such as IF and HH allow detailed simulations of neuronal behaviour, advancing the study of neurodegenerative diseases and opening avenues for personalised neurotherapeutics when combined with clinical data. In epidemiology, computational models are indispensable for predicting disease spread, evaluating interventions, and informing public health policy, as recently underscored by global health crises. In epidemiology, computational models are indispensable for predicting disease spread, evaluating interventions, and informing public health policy, as recently underscored by global health crises. Looking ahead, the convergence of high-performance computing, big data analytics, and AI promises to further transform biomedical research. Future efforts will likely focus on enhancing model personalisation through the development of DTs, integrating multi-scale data, and improving real-time decision support using closed-loop control strategies. Key to this evolution will be the implementation of adaptive and evolution-aware control strategies to manage disease dynamics, alongside the refinement of explainable and hybrid mechanistic-ML workflows. Nevertheless, significant challenges remain, including data heterogeneity and availability, model validation, uncertainty quantification, computational complexity, and ethical considerations regarding data privacy and algorithmic transparency. Addressing these will require continued innovation, investment in infrastructure, and the establishment of collaborative, standardised frameworks. In summary, in silico medicine represents far more than a technological evolution. By enabling a systems-level, quantitative, and integrative approach to health and disease, it constitutes a true paradigm shift, following the vision proposed by the famous physicist and philosopher Thomas Kuhn in his noteworthy book The Structure of Scientific Revolutions319, where computation becomes an inseparable component of scientific discovery, clinical translation, and decision-making, poised to tackle the most pressing health challenges of our time.