Using machine learning as a surrogate model for agent-based simulations

Artificial-intelligence-driven innovations in mechanistic computational modeling and digital twins for biomedical applications

Understanding of complex biological systems remains a significant challenge due to their high dimensionality, nonlinearity, and context-specific behavior. Artificial intelligence (AI) and mechanistic modeling are becoming essential tools for studying such complex systems. Mechanistic modeling can facilitate the construction of simulatable models that are interpretable but often struggle with scalability and parameters estimation. AI can integrate multi-omics data to create predictive models, but it lacks interpretability. The gap between these two modeling methods limits our ability to develop comprehensive and predictive models for biomedical applications. This article reviews the most recent advancements in the integration of AI and mechanistic modeling to fill this gap. Recently, with omics availability, AI has led to new discoveries in mechanistic computational modeling. The mechanistic models can also help in getting insight into the mechanism for prediction made by AI models. This integration is helpful in modeling complex systems, estimating the parameters that are hard to capture in experiments, and creating surrogate models to reduce computational costs because of expensive mechanistic model simulations. This article focuses on advancements in mechanistic computational models and AI models and their integration for scientific discoveries in biology, pharmacology, drug discovery and diseases. The mechanistic models with AI integration can facilitate biological discoveries to advance our understanding of disease mechanisms, drug development, and personalized medicine. The article also highlights the role of AI and mechanistic model integration in the development of more advanced models in the biomedical domain, such as medical digital twins and virtual patients for pharmacological discoveries.

Improving clinical decision making by creating surrogate models from health technology assessment models: A case study on Type 1 Diabetes Melitus

BACKGROUND AND OBJECTIVE

Computerized clinical decision support systems (CDSS) that incorporate the latest scientific evidence are essential for enhancing patient care quality. Such systems typically rely on some kind of model to accurately represent the knowledge required to assess the clinicians. Although the use of complex and computationally demanding simulation models is common in this field, such models limit the potential applications of CDSSs, both in real-time applications and in simulation-in-the-loop optimization tools. This paper presents a case study on Type 1 Diabetes Mellitus (T1DM) to demonstrate the development of surrogate models from health technology assessment models, with the aim of enhancing the potential of CDSSs.

METHODS

The paper details the process of developing machine learning (ML) based surrogate models, including the generation of a dataset for training and testing, and the comparison of different ML techniques. A number of distinct groupings of comorbidities were utilized in the creation of models, which were trained to predict confidence intervals for the time to develop each complication.

RESULTS

The results of the intersection over union (IoU) analysis between the simulation model output and the surrogate models output for the comorbidities under study were greater than 0.9.

CONCLUSION

The study concludes that ML-based surrogate models are a viable solution for real-time clinical decision-making, offering a substantial speedup in execution time compared to traditional simulation models.

Expanding interpretability through complexity reduction in machine learning-based modelling of cardiovascular disease: A myocardial perfusion imaging PET/CT prognostic study

BACKGROUND

Machine learning-based analysis can be used in myocardial perfusion imaging data to improve risk stratification and the prediction of major adverse cardiovascular events for patients with suspected or established coronary artery disease. We present a new machine learning approach for the identification of patients who develop major adverse cardiovascular events. The new method is robust against the deleterious effect of outliers in the training set stratification and training process.

METHODS

The proposed sum-of-sigmoids model is obtained by averaging the contributions of various input variables in an ensemble of XGBoost models. To illustrate its performance, we have applied it to predict major adverse cardiovascular events from advanced imaging data extracted from rest and adenosine stress 13N-ammonia positron emission tomography myocardial perfusion imaging polar maps. There were 1185 individual studies performed, and the event occurrence was tracked over a follow-up period of 2 years.

RESULTS

The sum-of-sigmoids model achieved a prediction accuracy of .83 on the test set, matching the performance of significantly more complex and less interpretable models (whose accuracies were .83-.84).

CONCLUSION

The sum-of-sigmoids model is interpretable and simple, while achieving similar prediction accuracy to significantly more complex machine learning models in the considered prediction task. It should be suitable for applications such as automated clinical risk stratification, where clear and explicit justification of the classification procedure is highly pertinent.

Bayesian Calibration of Stochastic Agent Based Model via Random Forest

Agent-based models (ABM) provide an excellent framework for modeling outbreaks and interventions in epidemiology by explicitly accounting for diverse individual interactions and environments. However, these models are usually stochastic and highly parametrized, requiring precise calibration for predictive performance. When considering realistic numbers of agents and properly accounting for stochasticity, this high-dimensional calibration can be computationally prohibitive. This paper presents a random forest-based surrogate modeling technique to accelerate the evaluation of ABMs and demonstrates its use to calibrate an epidemiological ABM named CityCOVID via Markov chain Monte Carlo (MCMC). The technique is first outlined in the context of CityCOVID's quantities of interest, namely hospitalizations and deaths, by exploring dimensionality reduction via temporal decomposition with principal component analysis (PCA) and via sensitivity analysis. The calibration problem is then presented, and samples are generated to best match COVID-19 hospitalization and death numbers in Chicago from March to June in 2020. These results are compared with previous approximate Bayesian calibration (IMABC) results, and their predictive performance is analyzed, showing improved performance with a reduction in computation.

Towards an interpretable deep learning model of cancer

Cancer is a manifestation of dysfunctional cell states. It emerges from an interplay of intrinsic and extrinsic factors that disrupt cellular dynamics, including genetic and epigenetic alterations, as well as the tumor microenvironment. This complexity can make it challenging to infer molecular causes for treating the disease. This may be addressed by system-wide computer models of cells, as they allow rapid generation and testing of hypotheses that would be too slow or impossible to perform in the laboratory and clinic. However, so far, such models have been impeded by both experimental and computational limitations. In this perspective, we argue that they can now be achieved using deep learning algorithms to integrate omics data and prior knowledge of molecular networks. Such models would have many applications in precision oncology, e.g., for identifying drug targets and biomarkers, predicting resistance mechanisms and toxicity effects of drugs, or simulating cell-cell interactions in the microenvironment.

Comprehensive Sepsis Risk Prediction in Leukemia Using a Random Forest Model and Restricted Cubic Spline Analysis

Background

Sepsis is a severe complication in leukemia patients, contributing to high mortality rates. Identifying early predictors of sepsis is crucial for timely intervention. This study aimed to develop and validate a predictive model for sepsis risk in leukemia patients using machine learning techniques.

Methods

This retrospective study included 4310 leukemia patients admitted to the Affiliated Hospital of Guangdong Medical University from 2005 to 2024, using 70% for training and 30% for validation. Feature selection was performed using univariate logistic regression, LASSO, and the Boruta algorithm, followed by multivariate logistic regression analysis. Seven machine learning models were constructed and evaluated using receiver operating characteristic (ROC) curves and decision curve analysis (DCA). Shapley additive explanations (SHAP) were applied to interpret the results, and restricted cubic spline (RCS) regression explored the nonlinear relationships between variables and sepsis risk. Furthermore, we examined the interactions among predictors to better understand their potential interrelationships.

Results

The random forest (RF) model outperformed all others, achieving an AUC of 0.765 in the training cohort and 0.700 in the validation cohort. Key predictors of sepsis identified by SHAP analysis included C-reactive protein (CRP), procalcitonin (PCT), neutrophil count (Neut), lymphocyte count (Lymph), thrombin time (TT), red blood cell count (RBC), total bile acid (TBA), and systolic blood pressure (SBP). RCS analysis revealed significant non-linear associations between CPR, PCT, Neut, Lymph, TT, RBC and SBP with sepsis risk. Pairwise correlation analysis further revealed interactions among these variables.

Conclusion

The RF model exhibited robust predictive power for sepsis in leukemia patients, providing clinicians with a valuable tool for early risk assessment and the optimization of treatment strategies.

Agent-based modeling in cancer biomedicine: applications and tools for calibration and validation

Computational models are not just appealing because they can simulate and predict the development of biological phenomena across multiple spatial and temporal scales, but also because they can integrate information from well-established in vitro and in vivo models and test new hypotheses in cancer biomedicine. Agent-based models and simulations are especially interesting candidates among computational modeling procedures in cancer research due to the capability to, for instance, recapitulate the dynamics of neoplasia and tumor - host interactions. Yet, the absence of methods to validate the consistency of the results across scales can hinder adoption by turning fine-tuned models into black boxes. This review compiles relevant literature that explores strategies to leverage high-fidelity simulations of multi-scale, or multi-level, cancer models with a focus on verification approached as simulation calibration. We consolidate our review with an outline of modern approaches for agent-based models' validation and provide an ambitious outlook toward rigorous and reliable calibration.

Forecasting and Predicting Stochastic Agent-Based Model Data with Biologically-Informed Neural Networks

Collective migration is an important component of many biological processes, including wound healing, tumorigenesis, and embryo development. Spatial agent-based models (ABMs) are often used to model collective migration, but it is challenging to thoroughly predict these models' behavior throughout parameter space due to their random and computationally intensive nature. Modelers often coarse-grain ABM rules into mean-field differential equation (DE) models. While these DE models are fast to simulate, they suffer from poor (or even ill-posed) ABM predictions in some regions of parameter space. In this work, we describe how biologically-informed neural networks (BINNs) can be trained to learn interpretable BINN-guided DE models capable of accurately predicting ABM behavior. In particular, we show that BINN-guided partial DE (PDE) simulations can (1) forecast future spatial ABM data not seen during model training, and (2) predict ABM data at previously-unexplored parameter values. This latter task is achieved by combining BINN-guided PDE simulations with multivariate interpolation. We demonstrate our approach using three case study ABMs of collective migration that imitate cell biology experiments and find that BINN-guided PDEs accurately forecast and predict ABM data with a one-compartment PDE when the mean-field PDE is ill-posed or requires two compartments. This work suggests that BINN-guided PDEs allow modelers to efficiently explore parameter space, which may enable data-driven tasks for ABMs, such as estimating parameters from experimental data. All code and data from our study is available at https://github.com/johnnardini/Forecasting_predicting_ABMs .

Machine learning-based surrogate modelling of a robust, sustainable development goal (SDG)-compliant land-use future for Australia at high spatial resolution

We developed a high-resolution machine learning based surrogate model to identify a robust land-use future for Australia which meets multiple UN Sustainable Development Goals. We compared machine learning models with different architectures to pick the best performing model considering the data type, accuracy metrics, ability to handle uncertainty and computational overhead requirement. The surrogate model, called ML-LUTO Spatial, was trained on the Land-Use Trade-Offs (version 1.0) model of Australian agricultural land system sustainability. Using the surrogate model, we generated projections of land-use futures at 1.1 km resolution with 95% classification accuracy, and which far surpassed the computational benchmarks of the original model. This efficiency enabled the generation of numerous SDG-compliant (SDGs 2, 6, 7, 13, 15) future land-use maps on a standard laptop, a task previously dependent upon high-performance computing clusters. Combining these projections, we derived a single, robust land-use future and quantified the uncertainty. Our findings indicate that while agricultural land-use remains dominant in all Australian regions, extensive carbon plantings were identified in Queensland and environmental plantings played a role across the study area, reflecting a growing urgency for offsetting greenhouse gas emissions and the restoration of ecosystems to support biodiversity across Australia to meet the 2050 Sustainable Development Goals.

Perspectives on computational modeling of biological systems and the significance of the SysMod community

Motivation

In recent years, applying computational modeling to systems biology has caused a substantial surge in both discovery and practical applications and a significant shift in our understanding of the complexity inherent in biological systems.

Results

In this perspective article, we briefly overview computational modeling in biology, highlighting recent advancements such as multi-scale modeling due to the omics revolution, single-cell technology, and integration of artificial intelligence and machine learning approaches. We also discuss the primary challenges faced: integration, standardization, model complexity, scalability, and interdisciplinary collaboration. Lastly, we highlight the contribution made by the Computational Modeling of Biological Systems (SysMod) Community of Special Interest (COSI) associated with the International Society of Computational Biology (ISCB) in driving progress within this rapidly evolving field through community engagement (via both in person and virtual meetings, social media interactions), webinars, and conferences.

Availability and implementation

Additional information about SysMod is available at https://sysmod.info.

Evolutionary Approach to Optimal Oil Skimmer Assignment for Oil Spill Response: A Case Study

We propose a genetic algorithm for optimizing oil skimmer assignments, introducing a tailored repair operation for constrained assignments. Methods essentially involve simulation-based evaluation to ensure adherence to South Korea's regulations. Results show that the optimized assignments, compared to current ones, reduced work time on average and led to a significant reduction in total skimmer capacity. Additionally, we present a deep neural network-based surrogate model, greatly enhancing efficiency compared to simulation-based optimization. Addressing inefficiencies in mobilizing locations that store oil skimmers, further optimization aimed to minimize mobilized locations and was validated through scenario-based simulations resembling actual situations. Based on major oil spills in South Korea, this strategy significantly reduced work time and required locations. These findings demonstrate the effectiveness of the proposed genetic algorithm and mobilized location minimization strategy in enhancing oil spill response operations.

Incorporating temporal information during feature engineering bolsters emulation of spatio-temporal emergence

MOTIVATION

Emergent biological dynamics derive from the evolution of lower-level spatial and temporal processes. A long-standing challenge for scientists and engineers is identifying simple low-level rules that give rise to complex higher-level dynamics. High-resolution biological data acquisition enables this identification and has evolved at a rapid pace for both experimental and computational approaches. Simultaneously harnessing the resolution and managing the expense of emerging technologies-e.g. live cell imaging, scRNAseq, agent-based models-requires a deeper understanding of how spatial and temporal axes impact biological systems. Effective emulation is a promising solution to manage the expense of increasingly complex high-resolution computational models. In this research, we focus on the emulation of a tumor microenvironment agent-based model to examine the relationship between spatial and temporal environment features, and emergent tumor properties.

RESULTS

Despite significant feature engineering, we find limited predictive capacity of tumor properties from initial system representations. However, incorporating temporal information derived from intermediate simulation states dramatically improves the predictive performance of machine learning models. We train a deep-learning emulator on intermediate simulation states and observe promising enhancements over emulators trained solely on initial conditions. Our results underscore the importance of incorporating temporal information in the evaluation of spatio-temporal emergent behavior. Nevertheless, the emulators exhibit inconsistent performance, suggesting that the underlying model characterizes unique cell populations dynamics that are not easily replaced.

AVAILABILITY AND IMPLEMENTATION

All source codes for the agent-based model, emulation, and analyses are publicly available at the corresponding DOIs: 10.5281/zenodo.10622155, 10.5281/zenodo.10611675, 10.5281/zenodo.10621244, respectively.

A novel adaptive-weight ensemble surrogate model base on distance and mixture error

Surrogate models are commonly used as a substitute for the computation-intensive simulations in design optimization. However, building a high-accuracy surrogate model with limited samples remains a challenging task. In this paper, a novel adaptive-weight ensemble surrogate modeling method is proposed to address this challenge. Instead of using a single error metric, the proposed method takes into account the position of the prediction sample, the mixture error metric and the learning characteristics of the component surrogate models. The effectiveness of proposed ensemble models are tested on five highly nonlinear benchmark functions and a finite element model for the analysis of the frequency response of an automotive exhaust pipe. Comparative results demonstrate the effectiveness and promising potential of proposed method in achieving higher accuracy.

Probabilistic Neural Computing with Stochastic Devices

The brain has effectively proven a powerful inspiration for the development of computing architectures in which processing is tightly integrated with memory, communication is event-driven, and analog computation can be performed at scale. These neuromorphic systems increasingly show an ability to improve the efficiency and speed of scientific computing and artificial intelligence applications. Herein, it is proposed that the brain's ubiquitous stochasticity represents an additional source of inspiration for expanding the reach of neuromorphic computing to probabilistic applications. To date, many efforts exploring probabilistic computing have focused primarily on one scale of the microelectronics stack, such as implementing probabilistic algorithms on deterministic hardware or developing probabilistic devices and circuits with the expectation that they will be leveraged by eventual probabilistic architectures. A co-design vision is described by which large numbers of devices, such as magnetic tunnel junctions and tunnel diodes, can be operated in a stochastic regime and incorporated into a scalable neuromorphic architecture that can impact a number of probabilistic computing applications, such as Monte Carlo simulations and Bayesian neural networks. Finally, a framework is presented to categorize increasingly advanced hardware-based probabilistic computing technologies.

HPOSS: A hierarchical portfolio optimization stacking strategy to reduce the generalization error of ensembles of models

Surrogate models are frequently used to replace costly engineering simulations. A single surrogate is frequently chosen based on previous experience or by fitting multiple surrogates and selecting one based on mean cross-validation errors. A novel stacking strategy will be presented in this paper. This new strategy results from reinterpreting the model selection process based on the generalization error. For the first time, this problem is proposed to be translated into a well-studied financial problem: portfolio management and optimization. In short, it is demonstrated that the individual residues calculated by leave-one-out procedures are samples from a given random variable ϵi, whose second non-central moment is the i-th model's generalization error. Thus, a stacking methodology based solely on evaluating the behavior of the linear combination of the random variables ϵi is proposed. At first, several surrogate models are calibrated. The Directed Bubble Hierarchical Tree (DBHT) clustering algorithm is then used to determine which models are worth stacking. The stacking weights can be calculated using any financial approach to the portfolio optimization problem. This alternative understanding of the problem enables practitioners to use established financial methodologies to calculate the models' weights, significantly improving the ensemble of models' out-of-sample performance. A study case is carried out to demonstrate the applicability of the new methodology. Overall, a total of 124 models were trained using a specific dataset: 40 Machine Learning models and 84 Polynomial Chaos Expansion models (which considered 3 types of base random variables, 7 least square algorithms for fitting the up to fourth order expansion's coefficients). Among those, 99 models could be fitted without convergence and other numerical issues. The DBHT algorithm with Pearson correlation distance and generalization error similarity was able to select a subgroup of 23 models from the 99 fitted ones, implying a reduction of about 77% in the total number of models, representing a good filtering scheme which still preserves diversity. Finally, it has been demonstrated that the weights obtained by building a Hierarchical Risk Parity (HPR) portfolio perform better for various input random variables, indicating better out-of-sample performance. In this way, an economic stacking strategy has demonstrated its worth in improving the out-of-sample capabilities of stacked models, which illustrates how the new understanding of model stacking methodologies may be useful.

Efficient Bayesian inference for stochastic agent-based models

The modelling of many real-world problems relies on computationally heavy simulations of randomly interacting individuals or agents. However, the values of the parameters that underlie the interactions between agents are typically poorly known, and hence they need to be inferred from macroscopic observations of the system. Since statistical inference rests on repeated simulations to sample the parameter space, the high computational expense of these simulations can become a stumbling block. In this paper, we compare two ways to mitigate this issue in a Bayesian setting through the use of machine learning methods: One approach is to construct lightweight surrogate models to substitute the simulations used in inference. Alternatively, one might altogether circumvent the need for Bayesian sampling schemes and directly estimate the posterior distribution. We focus on stochastic simulations that track autonomous agents and present two case studies: tumour growths and the spread of infectious diseases. We demonstrate that good accuracy in inference can be achieved with a relatively small number of simulations, making our machine learning approaches orders of magnitude faster than classical simulation-based methods that rely on sampling the parameter space. However, we find that while some methods generally produce more robust results than others, no algorithm offers a one-size-fits-all solution when attempting to infer model parameters from observations. Instead, one must choose the inference technique with the specific real-world application in mind. The stochastic nature of the considered real-world phenomena poses an additional challenge that can become insurmountable for some approaches. Overall, we find machine learning approaches that create direct inference machines to be promising for real-world applications. We present our findings as general guidelines for modelling practitioners.