In Silico Studies Roadmap
The SPARC program offers experimental data and powerful computational modeling tools, which enable in silico studies. This is SPARC's modeling tools roadmap.
The SPARC Program aims to understand the role of the autonomic peripheral nervous system (ANS) in regulating physiological (organ) function, developing methodologies for interacting with the ANS to influence and control physiological function, and applying these methodologies for therapeutic purposes. In support of these goals, SPARC offers high-quality experimental data in conjunction with powerful computational modeling tools, which enable in silico studies. These mutually complement each other, as experimental data provides simulation input, forms the basis of data-driven modeling, and is needed for model validation, while computational modeling provides mechanistic insights, helps with designing experiments, and generates rich data under highly controlled conditions. The SPARC In Silico Studies Roadmap is summarized in the following.
In Silico Studies
Computational modeling is a powerful approach to support the SPARC goals. For example, it can be used to gain mechanistic understanding, to formulate testable hypotheses, to help interpret experimental data, to study the impact of selected factors under highly controlled conditions, to develop neural sensing and modulation technologies (e.g., neural interfaces), to develop control strategies (including controllers that can be integrated into devices), to evaluate safety and efficacy, to conduct in silico clinical trials, and to personalize therapies.
SPARC aims to develop models that can predictively simulate all aspects from physical interaction with ANS activity, e.g., through neural interfaces, to the resulting intended and unintended consequences on physiological function.
To this end, knowledge, activities, and/or functionalities are required in the following areas:
- Physical interaction modeling
- Neural response modeling
- Neural code and connectivity
- Physiological response modeling
- Anatomical models
- Image-based modeling
- Data-based modeling
- Model analysis and visualization
- Model composition
- Verification & validation
- Sensitivity & uncertainty quantification
- Quality assurance & regulatory concerns
- Optimization & control
- Safety and efficacy assessment, in silico trials
- ANS and CNS
- FAIRness, collaboration, and curation
- Mapping & knowledge management
- Health and disease, sex, population variability, and species
- Long-term response, plasticity, therapy
Much of the related development relies on the o²S²PARC computational modeling framework developed as part of the SPARC DRC for open, online, cloud-based, collaborative, sustainable, and FAIR (findable, accessible, interoperable, and reusable) simulations and data analysis. However, data and knowledge management and anatomical mapping are also closely linked. For example, data serves as model input, while computational modeling generates derived data, and such dependencies are captured as part of the knowledge management. Knowledge management helps find compatible data and models, and ensures that model quality and reliability is documented as metadata. Anatomical mapping provides the necessary PNS-organ connectivity information, helps identify neglected pathways in models, and provides proper anatomical environments, e.g., for finite element modeling.
Physical exposure modeling
Vision: It is possible to simulate physical exposure by neuromodulating devices across the relevant scales (e.g., body, organ, tissue, cellular, neuron, subcellular). Complex anatomical structure, heterogeneity, and material properties are considered. The relevant exposure physics is primarily low frequency electromagnetism -- though thermal, light (e.g., in combination with optogenetics) and acoustic exposure are also important.
This requires powerful, high-performance computing (HPC)-enabled physics solvers, robust discretization methods, and physical tissue properties, as well as anatomical models across the relevant scales or the ability to construct such models from image-data.
Example: An electrode is designed and the field distribution within a complex multi-fascicular nerve is calculated in the quasi-electrostatic regime, considering tissue anisotropy, thin layers, and interface effects (including, e.g., electrode polarization, fibrotic encapsulation).
Neural response modeling
Vision: The response of neurons (both single neurons and neural networks) to physical exposure can be simulated. This includes stimulation, inhibition (blocking), as well as modulation of natural activity.
Example: The modulation of spinal cord neural activity as a result of electric spinal cord stimulation can be simulated (blocking of signaling in nociceptive fibers, stimulation of spinal roots, modulation of reflex circuits).
Neural code and connectivity
Vision: There are models of how the information content of ANS signals is encoded, such that signaling can be designed to express an intended signal. Another example is how activity at different locations throughout the ANS relates to the signal (and signal information) received by target organs. Such models require knowledge of connectivity, ganglionic function, and organ innervation.
Example: Assuming a certain vagus nerve stimulation, what is the resulting modification of signals reaching the innervation of the heart, gut, etc.? How does the modulation need to be shaped to encode the desired physiological response?
Physiological response modeling
Vision: There are models of physiological function (organs and other) that are capable of accounting for the impact of ANS inputs. These models can take a wide variety of forms. For example, they can be mechanistic or phenomenological.
(see also Health and disease, sex, population variability, and species)
Example: A model can simulate cardiac activity (neural control at multiple levels, tissue electrophysiology, biomechanics, fluid-structure interaction) and how neuromodulation of the cardiac branch of the vagus nerve affects that activity.
Vision: Anatomical models provide a physical environment for exposure simulation (i.e., physical exposure modeling) and are a source of information about the distribution of material properties and boundary conditions. Anatomical models cover a wide range of scales (e.g., nerve structure, organ and whole-body scaffolds). In some cases, they consider the dynamic nature of anatomical geometries (e.g., breathing, motion, heartbeat).
(see also Health and disease, sex, population variability, and species)
Examples: A 3D model of a neurovascular bundle is provided, which is suitable for modeling a neural interface. A 4D scaffold of the stomach, suitable for biomechanical simulation of its physiological function, exists.
Vision: There is functionality that facilitates the conversion of structural and functional imaging data into components of computational models. This includes both fully automated and interactive functionality that can be used for general purposes (e.g., building a deep neural network model based on segmented imaging data) or for specialized purposes (e.g., segmenting specific anatomical structures from structural imaging data). Image-based modelling can be useful in generating personalized models.
(see also Health and disease, sex, population variability, and species)
Examples: Functionality exists that supports image-based extraction/generation of anatomical models, the assignment of image-based tissue property maps and/or boundary conditions in simulations, etc. Image data can also be used to train a data-based model.
Vision: There is functionality that facilitates the creation, training, validation, adaptation, and application of data-based models. This shall provide a powerful means to bridge the gap between experimental data acquisition and computational model generation. It can also be a means to generate efficient models from resource-intensive mechanistic models, e.g., as an internal model in a controller.
(see also Optimization & control strategies and Image-based modeling)
Example: A computational model of the intestinal response to neuromodulation is constructed from electrophysiological, myographic, and mechanical sensing data using deep learning techniques, with dedicated services for network creation, training, validation, adaptation, and application.
Model analysis and visualization
Vision: There are powerful tools that can be used to analyze and visualize computational models and their output. Analyses and visualizations can be published and shared, e.g., as supplementary material to a scientific paper.
Examples: Functionality is provided that supports result processing (e.g., spike detection and sorting), derived data extraction (e.g., through scripting), visualization, investigation of correlations, inverse problem solving, and more.
Vision: Functionality is available to assist users in assembling models and model components (e.g., services) into overarching models. This includes guidance on compatibility, search functionality, and means to handle multi-scale modeling.
Example: A map can be used to select model components and assemble them into a combined model.
Verification and validation
Vision: There is functionality available that facilitates verification and validation (V&V), e.g., by providing comparators and statistical tests. It shall be possible to assess the reliability of a model based on the V&V information provided. Verification is about determining whether an implementation correctly reproduces the intended model, while validation is about testing whether the model reproduces the real-world behavior of a relevant quantity-of-interest in a given context-of-use within the expected confidence interval. Validation is typically context-of-use specific and requires quantification of uncertainty.
(see also Sensitivity & uncertainty quantification)
Example: The minimal information standard ensures that the available information on V&V (including the extent and limits of the validated context-of-use) is readily apparent for published models.
Sensitivity and uncertainty quantification
Vision: Tools are available to facilitate the determination of uncertainty and the propagation of uncertainties through computational models. When available, information on the uncertainty associated with the models or model parameters is provided to the users. Knowledge of a model's sensitivity to underlying parameters and model uncertainty is fundamental in assessing the reliability of model predictions and the strength of its validation (benchmark sensitivity, agreement within expected uncertainty, etc.).
Example: The impact of uncertainty about tissue properties and model discretization on fiber recruitment by a neural interface can be determined.
Quality assurance (QA), regulatory compliance
Vision: Means are available to assess and ascertain the quality of models, and whether certain models are of regulatory grade. This includes, for example, the availability of test suites (from unit-test-level to regression testing) and means to ensure that services are used correctly within their validated context-of-use and that the limitations of the model are apparent. Of particular relevance are the ‘Ten Simple Rules’.
(see also Verification & validation)
Examples: Models can have associated tests that are run periodically, or whenever platform or model changes are published. Services determine if their inputs are within the validated range. QA information is clearly visible in the metadata.
Optimization and control strategies
Vision: There is extensive meta-modeling functionality that allow operations to be performed on computational models, such as uncertainty propagation, inverse problem solving, optimization, and reduced order model generation. Controllers can be constructed (using control algorithms from a library, as well as model-based control strategies the employ efficient internal models approximating behavior of the complex model to be controlled) and applied to computational models
Example: A reduced-order model is created based on a computationally demanding computational model and combined with control algorithms to build a smart controller, which is then applied to the full model before being experimentally tested as part of a device controller.
Safety and efficacy assessment, in silico trials
Vision: In silico studies can be performed in which the safety and efficacy of (therapeutic) interventions can be evaluated. Modeling shall provide information on the selectivity of stimulation, unwanted stimulation and/or side effects, efficacy in modulating the physiological response, and exposure-associated safety concerns (e.g., current-induced damage, tissue heating, friction). Population variability can be considered.
(see also Sensitivity and uncertainty quantification and Health and disease, sex, population variability, and species)
Example: The safety and efficacy of an implanted stimulator in terms of stimulation selectivity and current-induced tissue damage can be assessed, also considering population variability.
ANS and Central Nervous System (CNS)
Vision: The ANS does not function in isolation from the CNS. The interplay is understood and accounted for in models. For example, hybrid ANS-CNS models are provided.
Example: The role of central control in regulating micturition is considered in a bladder function model.
FAIRness, collaboration, and curation
Vision: FAIR stands for findable, accessible, interoperable, and reusable. SPARC data and models shall be FAIR. There is curation in place to ensure that published models adhere to the minimal information standard and are FAIR. The DRC platform ensures that models are sustainable, reproducible, shareable, and integratable, although the process is dynamic as there is always more that can be done to achieve a full FAIR specification. Functionality is in place that facilitates the collaborative elaboration of computational models and studies.
Example: After a computational model is submitted for publication, there is a curation process in place to ensure that the necessary meta-information is present.
Mapping & Knowledge Management (KM)
Vision: SPARC maps provide visual insight into anatomical and functional aspects of computational models (localization, dependencies, influences, related data). The maps can provide visual understanding for disease or treatment related changes and facilitate comparison between species. The knowledge management infrastructure enables flexible searches and queries on these models.
Example: A map allows the user to see which functional dependencies a model captures and which connections it neglects.
Health and disease, sex, population variability, and species
Vision: Data and models shall contain information about variability and ideally be parameterized. Important sources of variability that must be understood include interspecies, intersex, and intersubject variability, population variability, and differences between healthy and diseased conditions. This includes differences in anatomy, model parameters, tissue properties, and physiology.
Example: An organ scaffold is parameterized such that it can be used to evaluate the impact of population variability in organ shape in an in silico trial.
Long-term response, plasticity, therapy
Vision: Long term neuromodulation leads to (therapeutically) intended and unintended responses and adaptation, e.g., as a result of plasticity or chronic responses, such as fibrotic tissue formation. Models exist to help understand and predict such changes and their therapeutic implications.
Example: Modeling of spinal cord stimulation after spinal-cord injury, and how it repurposes the remaining circuits and restores control over bladder function.
In Silico Roadmap
Important past achievements
The following examples show what has already been achieved:
- o2S2PARC framework for open, online-accessible, cloud-based, collaborative, sustainable, and FAIR computational modeling and data analysis
- JupyterLab services for explorable and reproducible data analysis; distinct flavors for machine learning, neurosciences, finite-element modeling, and more
- Support for the Ten Simple Rules as a means to facilitate and encourage quality assurance in modeling
- Services to simulate physiological models from the Physiome Model Repository
- Onboarding and publication of models from the SPARC community (e.g., Control Core group); curation process
- Ability of using services and models as building blocks in larger studies; pipelining; tracking of model dependencies and detection when model parts are not up-to-date
- Complete workflow for image-based treatment planning of spinal cord stimulation (e.g., for bladder control)
- Guided Mode to convert modeling pipelines developed by modeling experts into step-by-step applications suitable for a broad user population, while hiding the underlying modeling complexity
- Detailed whole body anatomical models (NEUROCOUPLE (male and female), NEUROFAUNA (rat), multi-scale organ scaffolds
In Silico Roadmap: Short-Term (2022)
Functionality planned for roll out until the end of 2022
- Rollout of the o²S²PARC platform to the larger research community (i.e., beyond SPARC researchers)
- Graphical modeling/simulation/post-processing service to set up biophysical simulations (geometry, properties, discretization, etc.) and analyze/visualize results; initially for SIM-Core-supplied, high performance computing-enabled modeling; but also as a basis for future third-party solver integration
- Neural interface modeling (coupled EM-neuro simulation) in detailed nerve models (using provided nerve models or tools to construct models); stimulation selectivity assessment and optimization
- Model parameterization and support for sensitivity analysis, uncertainty propagation, as well as optimization; application to device design
- Reproducible, shareable, and explorable data analyses and visualizations; versioning and derived data
- Functionalities for data-based modeling (machine learning), data aggregation, and data-based model validation (comparators)
- Curation and information standard for computational models; quality assurance infrastructure (testing and meta-data)
- Knowledge management (intelligent queries)
- Explorable and queryable innervation maps
- Integration of organ scaffolds in whole-body scaffold
- First models ranging from neural interfaces to organ physiological responses (cardiac)
- Framework for the development of (model-based) closed-loop control strategies
- Infrastructure to facilitate model onboarding (more automation, less dependence on SIM-Core involvement)
- Scalable computational resources
- Support for units in model/service parameters
Continuously ongoing activities
- Onboarding and curation of PNS and organ physiology models from SPARC teams
- Extension and update of the scaffolds, maps, innervation and connectivity database (e.g., to support identification of potential interaction sites), and (neuro-functionalized) nerve models
- Generation of tutorials, demonstrators
In Silico Roadmap: Longer Term (beyond 2022)
- Stimulation optimization with regard to physiological responses
- Expert tools (treatment planning, standardized safety assessment...)
- Neural sensing modeling & optimization (e.g., for closed-loop control)
- More organ systems; models of ganglia and CNS
- Use of nerve cross-section libraries for high-throughput screening
- Information on sex/species differences, population variability, health & disease for in silico trials
- Community platform (forums, project associated discussions, review system, shared sessions), and support for collaborative modeling
- Support for generic multi-scale modeling, model order reduction, surrogate modeling
- Integration with other infrastructures and initiatives
- Interaction mechanisms and interface technologies beyond electromagnetic neuromodulation
- Identification of neglected pathways, based on maps; model compatibility assessment based on maps & knowledge management. These features can also provide a basis for the prediction of unexpected side-effects
- Validation tools; support for regulatory-grade modeling; identification of instances where services are used outside their validated context-of-use
- Presentation of data and model predictions with associated uncertainties; rare event (failure) detection
- Modeling of long term effects (e.g., plasticity)
- Understanding and applying the ‘neural code’
- Teaching material
- Image-based modeling, including, e.g., quantitative image analysis, AI-based segmentation
- Ability for users to integrate their own custom simulators in the modeling-simulation-postprocessing service
- Extensive infrastructure for creating services without SIM-Core assistance (functionality that facilitates and automates onboarding, code review, after-publication issue-tracking and version updating, and support by service authors)
Updated 2 months ago