Immune digital twins for complex human pathologies: applications, limitations, and challenges
Niarakis et al. (2024): Immune digital twins for complex human pathologies: applications, limitations, and challenges
The authors provide the following definition of medical digital twins:
Medical digital twins consist of computational models that represent the health state of individual patients over time, enabling optimal therapeutics and forecasting patient prognosis. [...] A digital twin (DT) in biomedicine is a virtual representation of a patient, or a patient's state, that allows communication and data feedback from the actual patient to the virtual patient and vice versa. This capability holds the potential to improve personalised care and patient-tailored treatments. However, implementing such a technology may only be feasible for some pathologies. [...] An IDT is a digital twin for a particular medical application with a significant immune system component.
One of the most important aspects of IDTs is interoperability:
The IDT design requires a concerted effort by the systems biology community to adopt and implement suggested community standards, such as Systems Biology Markup Language (SBML) for mathematical model exchange, Systems Biology Graphical Notation (SBGN) for model visualisation, Biological Pathway Exchange (BioPaX) for pathway descriptions, and Simulation Experiment Description Markup Language (SED-ML) for simulation specifications. IDTs will likely use various modelling platforms, including tools that support ODE, agent-based, discrete, stochastic, or data-driven models. Not all of them are currently supported by community standards. Thus, there is a critical need to create standards for model specification for a much broader class of models through close collaboration and discussions with the COMBINE community. A helpful resource is also the EDITH standards collection for Virtual Human Twins in Health.
IDTs should also be modular:
Each IDT module can be derived from previously built models designed to represent specific aspects of the immune system; the modules can be designed from scratch to match unanswered questions, enabling flexibility to mix and match models as needed. This architecture supports easy updates, replacements, or additions. It allows the integration of new data and new data types and formats as they are discovered or developed, ensuring that the IDT can adapt to new research findings and evolving medical and biological knowledge, keeping it always up-to-date.
Hybrid systems with both AI-based and mechanistic models have a great potential:
Mechanistic models excel in inferring causal relationships based on known biological mechanisms, while AI models can help identify patterns and correlations within extensive datasets. Hybrid IDTs could combine the robustness and interpretability of mechanistic models with the capability of AI models for extracting information from large data sets. Furthermore, hybrid models can address data scarcity while enhancing the robustness of the model, as exemplified in physics with physics-informed neural networks (PINNs), which showed that constraining neural networks with prior knowledge improves accuracy and generalisability even in data-limited scenarios. However, biological systems are typically described in qualitative terms, and how to effectively integrate qualitative prior knowledge with quantitative and mathematical models requires further investigation.
An image shows the relationship of the immune system to various classes of diseases:
Autoimmune diseases in which failure in non-self-recognition or negative feedback control of proinflammation leads to persistent inflammation and long-term tissue damage; Infections in which the immune response is responding to various types of microbes (viruses, bacteria and fungi); Ageing, where changes in the function of the immune response can lead to a host of diseases; Acute Illness, where a host of perturbations rapidly activates the immune response. Immune pathophysiological processes range in time scale from hours for acute illness and sepsis to years and decades in autoimmune diseases and cancer. We propose that nearly every disease process and its potential resolution involves to some degree, inflammation and immunity.
Scalability also plays a role:
A defining feature of the immune system is that it operates across scales, bridging molecules to organs' dynamics and spanning timescales from seconds to weeks, months, or years. This implies that digital twins incorporating immune system functions must be multiscale by default.
The authors provide examples for four paradigms of potential DT implementations: infectious pneumonia, rheumatoid arthritis, sepsis, and cancer.
Regarding infectious pneumonia the authors write:
A suitable implementation of an IP-IDT could consist of Ordinary Differential Equations (ODE) or discrete models at the intracellular scale, agent-based models (ABM) at the tissue scale and ODE models at the whole-organ scale.
Regarding rheumatoid arthritis:
The main aim of the RA-IDT would be to digitally represent the interplay between resident cells of the joint and immune cells in RA, which leads to bone erosion, cartilage breakdown, and inflammation, to the level where the RA-IDTs can propose personalised therapy interventions. [...] For the RA-IDT, a combination of large-scale Boolean models that govern cellular behaviour, with agent based models (ABM) for accounting for the interactions of multiple cellular types, could be a suitable implementation. In ABM, agents may receive signals and input from the environment and their neighbouring agents, provide output to the environment and their neighbours, and make 'decisions' based on the input from around them and their internal, sub-cellular decision making rules. An agent may grow, proliferate, enter a quiescent state, express cytokines/chemokines or undergo apoptosis or necrosis in response to surrounding environmental conditions. A first attempt to link Boolean models with ABMs was done during the COVID-19 Disease Map initiative.
Regarding sepsis:
The primary goal of the Sepsis-IDT is to provide the capability to treat sepsis by multimodal adaptive modulation of a patient's underlying cytokine milieu ("fit-for-purpose" as per the NASEM report). It will account for sepsis's heterogeneity and dynamic complexity by having an ongoing data link between the virtual and real twin and informing control/guiding therapy. [...] An early example of this approach simulates the gut-lung axis in sepsis. It consists of modular agent-based models of tissue/organ-specific epithelial cells interacting with and connected to an agent-based model of endothelial cells and circulating immune cells. These existing examples of dynamic multiscale molecule-to-organ integration form the basis of an initial Critical Illness Digital Twin (CIDT). This CIDT would be used to train (off-line) an artificial intelligence controller (AIC) offline. [...] This digital twin-trained AIC would be the "brain" of an integrated cyber-physical system that monitors real-time plasma cytokine/mediator levels [...] and uses the AIC to guide the administration of different amounts of pro- and anti-inflammatory mediators/monoclonal antibodies to steer an individual patient back to a state of health. [...] [T]he Sequential Organ Failure Score (SOFA) and its variants can be generated by readily obtainable clinical measurements and sequentially measured to update the IDT.
Regarding cancer:
An Onco-IDT should include elements like the TME, neo-angiogenesis, pre-metastatic and metastatic conditions, and system-level information like blood and lymphatic transport.
The paper also contains a chapter about IDTs in drug discovery.
[V]irtual populations of patient DTs can be used to run in silico clinical trials that can accompany or be used to design real-life trials. One recent example is the Universal Immune System Simulator (UISS). The European Medicines Agency (EMA) provided a letter of support for the use of the UISS as a simulation platform to predict how the circulating interferon-gamma changes over time as a function of the treatment dose in a cohort of virtual patients to select the doses to be tested in escalating dose phase IIa trials of new therapeutic whole cell / fragmented based vaccines against several diseases.