About this Research Topic
Network-based representations are universally suited for the description of the structure and dynamics of real-world complex systems. For example, a personalized genome-scale metabolic network offers an efficient framework for omics data integration, analysis, and modeling. Similarly, network physiology outlines the necessary integrative framework to capture multi-scale physiological processes and the observed states that result from many component interactions and feedback loops. Such network-based representations could facilitate deeper understanding of the structure and multi-scale dynamics of complex biological systems. This understanding helps to resolve and solve innumerable important problems, such as medical treatment personalization, early abnormality detection and preventive-strategy optimization, diagnostics and treatment of unprecedented and/or complex abnormalities, and in-silico drug discovery. Efficiency of genome-only personalization is limited due to the multi-gene nature of common abnormalities and an ignorance of epigenetic factors, including lifestyles and environments. In contrast, physiological signals, now collectable by wearable devices and portable clinical devices, quantify dynamical-state aggregate impact from all factors, not just genetics, suggesting that physiological representations could play a more central role in mitigating biomedical challenges.
Many details of network physiology framework remain unknown and data required for full calibration are scarce. Even with reasonable assumptions for hard-to-calibrate parameters, a direct approach for simulation of multi-scale physiological processes is fundamentally unrealistic. Successfully simulating multi-scale spatiotemporal dynamics in plasma and space physics critically depend on proper formulation and coupling of physical models that describe processes on micro- and macro scales. Computational limitations, stability problems, and lack of detailed initial/boundary conditions are responsible for the lack of feasibility in modeling a wide range of scales from first principles. Even more coarse-grain formulations are often required in practical space weather simulations. Similar approaches of different granularities should be developed in network physiology or adopted from other domains.
Pure data-driven approaches that employ machine learning (ML) may look attractive because no explicit knowledge about underlying micro-processes is required. However, data limitations and absence of relevant problems for transfer or meta learning can significantly reduce capabilities of the best ML frameworks, including deep learning (DL) and boosting algorithms. Similarly, deep generative models (DGM) cannot be reliably trained on limited data to generate realistic synthetic data. However, optimal incorporation of the prior domain knowledge could drastically reduce requirements for training data. Since domain knowledge may come in very different forms, hybrid ML frameworks, the main scope of current proposal, should be capable of deep multi-level integration of various representations of the domain knowledge to discover acceptable solutions when analytical, simulated, or pure data-driven ML approaches fail if used separately.
We call for papers that introduce hybrid modeling frameworks, capable to solve the most challenging biomedical problems under significant data limitation and insufficient fundamental knowledge, including but not limited to:
- Realistic frameworks of different granularity for simulation of physiological processes suitable for applied studies or facilitating fundamental research
- Physics-based models for generation of synthetic data featuring real-data statistical properties and applicable for ML models pre-training with further fine-tuning on limited actual data
- ML algorithms intrinsically incorporating generic group-based symmetry constraints and various forms of application-specific constraints
- Deep generative models capable to discover disentangled representations from very limited data via incorporation of domain-specific constraints and/or pre-training on synthetic data
- Boosting-based frameworks for direct incorporation of self-contained structural domain-specific models
- One-shot learning frameworks incorporating domain-specific state representations
- Hybrid ML algorithms to discover models with minimal sensitivity to data resolution
- Implementing existing AI/ML methods to temporally extrapolate problematic physiological behavior from untrainably short record-length data sets to well beyond
- Developing new algorithms to build intuition and insight, and/or engaging AI/ML to explore or emulate physically based models.
- AI/ML methodologies posed in conjunction with other biomedical analysis methods to address a fundamental physiological question.
- Comparison or validation of the outputs of AI/ML techniques against outputs from other, traditional analytical methods or theoretical and experimental approaches.
Topic Editor Dr. Valeriy Gavrishchaka is co-founder and CEO of Aceekret LLC, United States. All other Topic Editors declare no competing interests with regards to the Research Topic subject
Many details of network physiology framework remain unknown and data required for full calibration are scarce. Even with reasonable assumptions for hard-to-calibrate parameters, a direct approach for simulation of multi-scale physiological processes is fundamentally unrealistic. Successfully simulating multi-scale spatiotemporal dynamics in plasma and space physics critically depend on proper formulation and coupling of physical models that describe processes on micro- and macro scales. Computational limitations, stability problems, and lack of detailed initial/boundary conditions are responsible for the lack of feasibility in modeling a wide range of scales from first principles. Even more coarse-grain formulations are often required in practical space weather simulations. Similar approaches of different granularities should be developed in network physiology or adopted from other domains.
Pure data-driven approaches that employ machine learning (ML) may look attractive because no explicit knowledge about underlying micro-processes is required. However, data limitations and absence of relevant problems for transfer or meta learning can significantly reduce capabilities of the best ML frameworks, including deep learning (DL) and boosting algorithms. Similarly, deep generative models (DGM) cannot be reliably trained on limited data to generate realistic synthetic data. However, optimal incorporation of the prior domain knowledge could drastically reduce requirements for training data. Since domain knowledge may come in very different forms, hybrid ML frameworks, the main scope of current proposal, should be capable of deep multi-level integration of various representations of the domain knowledge to discover acceptable solutions when analytical, simulated, or pure data-driven ML approaches fail if used separately.
We call for papers that introduce hybrid modeling frameworks, capable to solve the most challenging biomedical problems under significant data limitation and insufficient fundamental knowledge, including but not limited to:
- Realistic frameworks of different granularity for simulation of physiological processes suitable for applied studies or facilitating fundamental research
- Physics-based models for generation of synthetic data featuring real-data statistical properties and applicable for ML models pre-training with further fine-tuning on limited actual data
- ML algorithms intrinsically incorporating generic group-based symmetry constraints and various forms of application-specific constraints
- Deep generative models capable to discover disentangled representations from very limited data via incorporation of domain-specific constraints and/or pre-training on synthetic data
- Boosting-based frameworks for direct incorporation of self-contained structural domain-specific models
- One-shot learning frameworks incorporating domain-specific state representations
- Hybrid ML algorithms to discover models with minimal sensitivity to data resolution
- Implementing existing AI/ML methods to temporally extrapolate problematic physiological behavior from untrainably short record-length data sets to well beyond
- Developing new algorithms to build intuition and insight, and/or engaging AI/ML to explore or emulate physically based models.
- AI/ML methodologies posed in conjunction with other biomedical analysis methods to address a fundamental physiological question.
- Comparison or validation of the outputs of AI/ML techniques against outputs from other, traditional analytical methods or theoretical and experimental approaches.
Topic Editor Dr. Valeriy Gavrishchaka is co-founder and CEO of Aceekret LLC, United States. All other Topic Editors declare no competing interests with regards to the Research Topic subject
Keywords: Network physiology, machine learning, modeling frameworks, deep neural nets (DNN), hybrid algorithms, deep generative models (DGM), AI, biomedical analysis, hybrid ensemble learning, multi-expert systems, multi-objective optimization, multi-scale simulation of complex systems
Important Note: All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.
