AI, particularly ML and DL, is revolutionizing the biological and biomedical sciences by leveraging its nonpareil ability to identify patterns within massive datasets 4142. Traditional analytical methods, which generally rely on reducing or simplifying complexity, often fail to uncover subtle yet significant interactions concealed within large experimental datasets. AI excels at this task, offering diverse applications across numerous biological and biomedical disciplines. For example, it is indispensable for curating and analyzing datasets from high-throughput omics technologies 43. Also DL is used to analyze vast amounts of genomic data to identify genetic variations and elucidate complex gene expression patterns 44. Moreover, DL-based proteomics can accurately classify single-cell samples, providing insights into cellular states 4546.

Importantly, AI can help researchers investigate the architecture and behavior of biomolecular networks, leading to systems-level understanding of molecular mechanisms and processes within these networks 4748. Beyond molecular biology, AI plays an increasingly fecund role in ecology and biodiversity conservation. Remote sensing, computer vision, and ML-based habitat modeling enable real-time identification of species and prediction of species distributions 49. Also, biologists use AI to monitor climate impacts and ecosystem health, as well as to optimize wildlife conservation strategies by analyzing large datasets 50. AI has improved estimations of species diversity compared to traditional methods and can be used to model the impact of human activities on ecosystems to guide conservation planning 51. 

AI is also transforming the healthcare landscape. Convolutional neural networks are effective at image segmentation, classification, and detecting abnormalities in various medical images 52. For example, DL can be used to identify early-stage cancers in mammograms with performance comparable to expert clinicians 53. AI is also improving diagnosis and prognosis and is used in chronic disease research for surveillance, diagnosis, and prognosis. In cardiology, for example, ML was used to predict mortality rates in patients with suspected coronary artery disease 54. The use of AI to analyze medical datasets can support clinicians in making more accurate and timely diagnoses 55. Moreover, AI is accelerating drug discovery and development 56. ML can be used to screen large datasets to find biomolecules likely to bind to specific targets and predict their properties, which reduces the time and cost of experimental screening 57. AI is also used for de novo drug design and can be used to analyze patient data to identify biomarkers and predict treatment responses for personalized therapies 58. 

Integration of ST and AI is proposed to provide an effective approach for investigating living organisms as complex adaptive systems. Many biological and biomedical phenomena exhibit emergent properties, arising from nonlinear interactions and dynamic feedback loops, which severely challenge traditional reductionist, liner thinking 5960. Since ST emphasizes the interconnectedness of components and the importance of context, it is ideal for integrating with AI to identify and explicate complex patterns within large-scale, multidimensional datasets 61. Importantly, the synergy between ST and AI can advance biology and biomedicine from a descriptive to a predictive science.

The integrated ST-AI approach can guide framing and conceptualizing problems by clearly defining their components, boundaries, and levels of organization. ST emphasizes understanding how parts relate to a whole, which is critical when dealing with complex adaptive systems characterized by interdependencies and emergent behavior, rather than isolating elements in abstraction. When combined with AI’s capacity to process high-dimensional datasets and to recognize nonlinear patterns, ST helps clarify which components and interactions are essential to model and why, improving both problem framing and model testing and validity. For example, modeling the gut microbiome requires identifying microbial taxa, host epithelial cells, immune factors, and dietary influences as interacting components within a dynamic system, which AI can facilitate by identifying and modeling 6263. Moreover, defining boundaries helps to avoid model drift, where inclusion or exclusion of critical variables influence system modeling 64.

As for the second step, identifying key variables and relationships, an integrated ST-AI approach can direct attention toward causality and interdependence—how one variable influences another across scales. By identifying key variables and relationships, it can help to focus analysis on causality and interdependence by showing not just which components exist but how changes in one component influence others. ST emphasizes that components are interconnected, and outcomes arise from causal feedback loops rather than isolated factors, making it essential to map both direct and indirect relationships to understand system behavior holistically. And AI complements this by enabling computational discovery of complex, multivariate dependencies from large datasets, supporting causal inference and pattern recognition beyond simple correlations. Advanced methods, such as causal discovery algorithms, aim to distinguish cause-and-effect linkages rather than mere associations, which is crucial in high-dimensional systems where interdependent variables interact nonlinearly 65. For example, in cellular signaling networks, protein-protein interactions can be modeled as motifs such as feed-forward loops and feedback circuits that govern system behavior 66. Mapping these relationships through this integrated approach creates a structural framework for subsequent data-driven modeling, ensuring that the analyses are biologically causal rather than simply correlative.

Importantly, the integrated ST-AI approach can facilitate identifying feedback loops and emergent properties, which are crucial components of complex adaptive systems. Feedback loops—whether positive/negative or reinforcing/balancing—can generate nonlinear dynamics observed in gene regulation, metabolic control, and population ecology. For example, recurrent neural networks or dynamic Bayesian networks can be used to model such feedback loops 67. This integrated approach can also be helpful in recognizing emergent properties—novel system behaviors arising from complex interactions. ST can guide the use of graph-based models to investigate emergent properties within biological and biomedical phenomena 68. Moreover, an integrated ST-AI approach assures the broader, systematic context is considered, thereby safeguarding that the biological system is not analyzed in isolation. For environmental conditions and evolutionary pressures, contextual or social determinants often influence biological behavior 69. ST can assist researchers in using AI to analyze contextual datasets, such as environmental exposures, microbiota composition, or behavioral factors, and to model the ecology of health and disease.

Once a problem is conceptually framed and key variables and relationships identified, ST can then guide the third step of datasets acquisition using AI to curate and manage these datasets through multimodal integration 70. Biological systems operate across multiple organizational scales—from genes and proteins to populations and ecosystems—and across diverse dataset modalities 71. An integrated ST-AI approach allows researchers to model the multifaceted nature of biological regulation and adaptation. For example, integrating transcriptomic and proteomic datasets has revealed regulatory hubs not detectable from analyzing any single dataset 72.

An integrated ST-AI approach can also guide feature engineering by emphasizing mechanistic relevance 73. Instead of relying solely on automated feature extraction, researchers can incorporate domain knowledge about network topology, metabolic fluxes, or pathway dependencies 74. This integrated approach can help to combine data-driven discovery with mechanistic insight and thereby enhance both model interpretability and generalizability. For example, network-based features derived from protein interaction graphs can assist in using ML to identify biologically coherent gene clusters associated with disease 75.

Real-world biological datasets are frequently noisy, heterogeneous, and incomplete, posing challenges for robust system modeling and inference 76. The proposed integrated ST-AI approach addresses these challenges by situating data limitations within the broader system context and by applying AI methods such as data integration, uncertainty quantification, and missing-data imputation 77. Together, these strategies improve model reliability and interpretability in complex biological systems. For example, imputation methods guided by network topology or causal inference can better preserve system coherence than purely statistical techniques 78. Ensuring data quality through this integrated approach allows researchers to use AI to recognize patterns that remain faithful to biological facts rather than artifacts from preprocessing.

For the fourth step, ST can be used to direct AI in computing models for complex adaptive systems based on large datasets. Specifically, ST can efficiently guide model selection and training through configuring the experimental problem. Two examples include graph neural networks, which are well-suited for modeling biological interaction networks 79, and recurrent neural networks, which can model temporal gene expression dynamics 80. Ensemble methods, such as random forests or gradient boosting, can also be used to model nonlinear relationships while providing measures of variable importance 81.

And the integrated ST-AI approach can facilitate testing models by evaluating predictions generated through computational analyses and simulations. At this stage, hypotheses derived from the models are compared against empirical observations, experimental data, or independent validation datasets. In biological and biomedical contexts, this may involve testing predicted gene–gene interactions, pathway perturbations, or clinical outcomes under simulated interventions. Sensitivity analysis and uncertainty quantification are often applied to assess the robustness of model predictions and to identify parameters or interactions that exert disproportionate influence on system behavior 8283. Importantly, model testing is not limited to accuracy alone but also includes biological plausibility, internal consistency, and the model’s ability to reproduce known system dynamics across multiple conditions 84. Through this process, discrepancies between predictions and observed outcomes become valuable sources of insight, revealing gaps in system understanding or data representation.

The final step involves evaluating overall model performance and adapting the model to guide subsequent investigations of the complex adaptive system. Evaluation typically integrates quantitative metrics, such as predictive accuracy, stability, and generalizability, with qualitative assessments informed by domain expertise 85. For systems biology, this step emphasizes iterative refinement, where model structure, parameters, or assumptions are revised in response to validation results 86. Adaptive model revision reflects the inherently dynamic and evolving nature of living systems, acknowledging that no single model can fully depict biological complexity. Instead, models serve as learning tools that evolve alongside new data and theoretical advances. This adaptive cycle supports continuous improvement in explanatory power and predictive utility, ultimately enabling more reliable system-level insights and informing future experimental design and intervention strategies. Finally, models should not only achieve statistical accuracy but also reproduce emergent system behaviors, such as homeostasis or robustness, when simulated under perturbations 87.

Moreover, AI depends on explainability, interpretability, and transparency 88. For example, explainable AI—including attention mechanisms, feature attribution methods, and symbolic regression—allows researchers to trace the logic of model predictions back to biological hypotheses 89. The proposed integrated approach helps to uncover causal mechanisms rather than simply to predict correlations. For example, application of neural networks and DL to gene expression datasets resulted in the detection of structures involved in regulatory genomics and cellular imaging 9091. An integrated ST-AI approach ensures that computational outputs result in coherent and meaningful insights into biological and biomedical systems.

The explanatory power and clinical relevance of the proposed integrated ST-AI approach can be illustrated by returning to the blood glucose homeostasis example. In the first step, the integrated approach can assist in problem framing by analyzing large clinical and population datasets to identify relevant temporal scales (e.g., minutes vs. days), subpopulations (e.g., insulin-resistant phenotypes), and contextual factors (e.g., diet, sleep, or medication use). ML can reveal latent structures that help refine system boundaries, ensuring that the model includes biologically and clinically meaningful components without becoming intractably complex 92. For the second step, in terms of identifying key variables and relationships, feature selection algorithms and causal discovery methods, e.g., Bayesian networks, Granger causality–inspired approaches, can be used to identify influential variables and causal feedback loops among glucose, insulin, glucagon, cortisol, physical activity, and tissue-specific insulin sensitivity 9394. Moreover, ML can uncover nonlinear relationships and cross-scale interactions that are difficult to detect using traditional statistical approaches alone 95. With respect to the third step, collecting, curating, and integrating data, AI is especially powerful in handling heterogeneous data sources central to systems biology, while ML-based data harmonization techniques can integrate continuous glucose monitoring data, metabolomics, electronic health records, and lifestyle data. Automated anomaly detection and imputation methods improve data quality, while representation learning enables alignment of molecular, physiological, and behavioral data into a shared analytical space 96.

For the fourth step, hybrid modeling approaches combine mechanistic models, e.g., differential equations describing insulin–glucose dynamics, with ML components that identify unknown or poorly characterized functions, such as tissue-specific insulin responsiveness. Neural networks, Gaussian processes, or reinforcement learning agents can augment classical physiological models and thereby improve predictive accuracy while preserving biological interpretability 97. And for testing predictions and simulating interventions, AI-enabled models can simulate “what-if” scenarios at scale, such as varying meal composition, exercise timing, or pharmacologic interventions 98. Reinforcement learning can be used to explore optimal control strategies for maintaining euglycemia, providing insights relevant to personalized insulin dosing or lifestyle recommendations. These predictions can be tested against real-world or prospective clinical data. For the final step, evaluating, learning, and adapting the model, continuous learning frameworks allow models to be updated as new data become available, such as longitudinal glucose monitoring streams. And explainable AI techniques help to assess whether model behavior aligns with known physiology or pathology, supporting trust and scientific validity 99. Model performance metrics then inform refinement of assumptions, variable selection, or system boundaries, not only closing the ST-AI loop but also opening another for further investigations.

In summary, for the integrated ST–AI approach, AI operationalizes ST by enabling the integration, modeling, and iterative refinement of complex, multiscale datasets, while ST provides the conceptual architecture that guides the use of AI throughout each step of the approach. The approach’s overall value is ensuring that biological and biomedical phenomena are understood as dynamic, interconnected, multilevel systems. To that end, the value of step 1, (Re)Framing the Problem, consists of ST establishing an appropriate system boundary, context, purpose, and level of analysis before AI modeling begins, whereas the value of the next step, Identifying Key Variables, involves ST providing the structural and relational architecture needed to determine which variables and interactions AI should analyze. The value of the third step, Collecting and Curating Datasets, entails ST ensuring that AI receives contextually meaningful, multiscale, and appropriately integrated data rather than disconnected datasets, while the value of the following step, Computing and Testing the Model, comprises ST transforming AI from a primarily pattern-recognition tool into a system-level modeling and hypothesis-generation instrument. The value of the final step, Evaluating and Adapting Model, focuses on making the integrated ST–AI approach iterative, adaptive, and self-correcting, allowing the model and the investigator’s understanding of the system to evolve together. In the context of glucose homeostasis, the integrated ST-AI approach can transform the circular systems model into a learning system that evolves with data, deepens biological understanding, and supports precision approaches to metabolic health.