Model Complex Research Articles

Bayesian blockwise inference for joint models of longitudinal and multistate data with application to longitudinal multimorbidity analysis.

Multistate models provide a useful framework for modelling complex event history data in clinical settings and have recently been extended to the joint modelling framework to appropriately handle endogenous longitudinal covariates, such as repeatedly measured biomarkers, which are informative about health status and disease progression. However, the practical application of such joint models faces considerable computational challenges. Motivated by a longitudinal multimorbidity analysis of large-scale UK health records, we introduce novel Bayesian inference approaches for these models that are capable of handling complex multistate processes and large datasets with straightforward implementation. These approaches decompose the original estimation task into smaller inference blocks, leveraging parallel computing and facilitating flexible model specification and comparison. Using extensive simulation studies, we show that the proposed approaches achieve satisfactory estimation accuracy, with notable gains in computational efficiency compared to the standard Bayesian estimation strategy. We illustrate our approaches by analysing the coevolution of routinely measured systolic blood pressure and the progression of three important chronic conditions, using a large dataset from the Clinical Practice Research Datalink Aurum database. Our analysis reveals distinct and previously lesser-known association structures between systolic blood pressure and different disease transitions.

Leveraging the variational Bayes autoencoder for survival analysis

Survival analysis in medical research has witnessed a growing interest in applying deep learning techniques to model complex, high-dimensional, heterogeneous, incomplete, and censored data. Current methods make assumptions about the relations between data that may not be valid in practice. Therefore, we introduce SAVAE (Survival Analysis Variational Autoencoder). SAVAE, based on Variational Autoencoders, contributes significantly to the field by introducing a tailored Evidence Lower BOund formulation, supporting various parametric distributions for covariates and survival time (if the log-likelihood is differentiable). It offers a general method that demonstrates robustness and stability through different experiments. Our proposal effectively estimates time-to-event, accounting for censoring, covariate interactions, and time-varying risk associations. We validate our model in diverse datasets, including genomic, clinical, and demographic tabular data, with varying levels of censoring. This approach demonstrates competitive performance compared to state-of-the-art techniques, as assessed by the Concordance Index and the Integrated Brier Score. SAVAE also offers an interpretable model that parametrically models covariates and time. Moreover, its generative architecture facilitates further applications such as clustering, data imputation, and synthetic patient data generation through latent space inference from survival data. This approach fosters data sharing and collaboration, improving medical research and personalized patient care.

Harnessing Quantum Materials for Enhanced Energy Efficiency and Storage: A Comprehensive Study of Electronic Properties and Performance Optimization

This research explores the exciting potential of various quantum materials to enhance energy efficiency and storage capabilities, filling in a crucial gap in our understanding of their electronic properties and interactions. A major hurdle in this field is the lack of empirical data on key factors such as conductivity, charge mobility, and thermal stability—elements that are vital for optimizing the performance of these materials in energy applications. Our study aims to blend experimental data with computational simulations to create a robust framework for modelling how quantum materials behave under different conditions. This comprehensive approach helps lay a solid foundation for their practical use in energy technologies. We focus on an array of fascinating quantum materials, including topological insulators, transition metal dichalcogenides, and superconductors. To investigate these materials, we employ advanced techniques like scanning tunnelling microscopy and spectroscopy, which allow us to delve deep into their unique properties. While pursuing this research, we do encounter challenges. Variations in sample quality, environmental factors during experimentation, and the complexities of theoretical modelling can affect the consistency of our results and the interpretation of our data. Despite these obstacles, our findings suggest that certain quantum materials hold remarkable potential for enhancing energy storage and conversion systems through improved charge mobility and thermal stability. In fact, some materials have demonstrated conductivity levels that surpass those of traditional options, hinting at transformative applications in energy technologies. These insights highlight the importance of continuing our exploration of quantum materials, positioning them as promising candidates for tackling today’s energy challenges and paving the way for advancements in sustainable energy solutions.

Multi-faceted sustainability improvement in low voltage power distribution network employing DG and capacitor bank

Efficient distribution of active and reactive power in distribution networks is crucial for ensuring that consumer demand is met and that electrical power quality is maintained. This requires careful planning, particularly in terms of optimally placing and sizing distributed generator (DG) and capacitor bank (CB) units. These units play a key role in minimizing power losses and voltage fluctuations within the network. However, implementing DG and CB units in unbalanced distribution networks presents unique challenges due to inherent system unbalances. To address these challenges, this article proposes a method for allocating and sizing DG and CB units in unbalanced distribution networks. The approach accounts for the network's unbalance and targets multiple objectives, such as reducing power losses, improving multi-phase voltage stability, and minimizing phase-to-phase voltage unbalance. The fast and flexible radial power flow (FFRPF) technique is employed to model complex interactions and constraints within the network, leading to a multi-objective optimization problem. This optimization problem is solved using the weight aggregated particle swarm optimization (WA-PSO) method, a variant of particle swarm optimization tailored for multi-objective functions. WA-PSO simplifies the process by combining multiple objectives into a single function using weighted aggregation. The efficacy of this approach is validated on 19, 34, and 123-node unbalanced radial distribution networks (URDNs). The results show significant improvements in power delivery efficiency across all tested networks, especially when DG and CB units are operated simultaneously. Specifically, DG and CB integration led to a reduction in system losses by 89.90 %, 88.42 %, and 86.87 %, and a decrease in three-phase voltage unbalance indices by 88.75 %, 81.68 %, and 39.05 %, for the 19, 34, and 123-node systems, respectively, while maintaining voltage stability within acceptable limits. Additionally, CO2 emissions were reduced by 52 %, 62 %, and 53 % when microturbines were utilized instead of coal-based thermal power plants, further highlighting the environmental benefits of the proposed approach.

Machine learning-enhanced stochastic uncertainty and sensitivity analysis of the ICRP human respiratory tract model for an inhaled radionuclide

The International Commission on Radiological Protection (ICRP) has developed the reference Human Respiratory Tract Model (HRTM), detailed in ICRP Publications 66 and 130, to estimate the deposition and clearance of inhaled radionuclides. These models utilize reference anatomical and physiological parameters for particle deposition (PD). Biokinetic models further estimate retention and excretion of internalized particulates, aiding the derivation of inhalation dose coefficients (DC). This study aimed to assess variability in deterministic131I biokinetic and dosimetry models through stochastic analysis using the updated HRTM from ICRP Publication 130. The complexities of the ICRP PD model were reconstructed into a new, independent computational model. Comparison with reference data for total PD fractions for reference worker, solely a nose breather, covering activity median aerodynamic diameters from 0.3μm to 20μm, showed a 1.04% relative and 0.7% absolute difference, demonstrating good agreement with ICRP deposition fractions. The deterministic DC module was reconstructed in Python and expanded for stochastic analysis, systematically expanding deposition components from HRTM and assigning probability distribution functions to uncertain parameters. These were integrated into an in-house stochastic radiological exposure dose calculator, utilizing latin hypercube sampling. A case of an occupational radionuclide intake was explored, in which biodistribution and committed effective DC (CEDC) were computed for131I type F, considering a lognormal particle size distribution with a median of 5μm. Results showed the published ICRP reference CEDC marginally exceeds the 75th percentile of observed samples, with log-gamma distribution as the best-fit probability distribution. A Random Forest regression model with SHapley Additive exPlanations was employed for sensitivity analysis to predict feature importance. The analysis identified the HRTM particle transport rates scaling factor, followed by the aerodynamic deposition efficiency in the alveolar interstitial region as the most impactful parameters. This study offers a unique stochastic approach on inhaled particulate metabolism, enhancing radiation consequence management, medical countermeasures, and dose reconstruction for epidemiological studies.

A Comprehensive Survey on Deep Clustering: Taxonomy, Challenges, and Future Directions

Clustering is a fundamental machine learning task which aims at assigning instances into groups so that similar samples belong to the same cluster while dissimilar samples belong to different clusters. Shallow clustering methods usually assume that data are collected and expressed as feature vectors within which clustering is performed. However, clustering high-dimensional data, such as images, texts, videos, and graphs, poses significant challenges for clustering tasks, such as indiscriminate representation and intricate relationships among instances. Over the past decades, deep learning has achieved remarkable success in effective representation learning and modeling complex relationships. Motivated by these advancements, Deep Clustering seeks to improve clustering outcomes through deep learning techniques, garnering considerable interest from both academia and industry. Despite many contributions to this vibrant area of research, the lack of systematic analysis and a comprehensive taxonomy has hindered progress in this field. In this survey, we first explore how deep learning can be integrated into deep clustering and identify two fundamental components: the representation learning module and the clustering module. Then we summarize and analyze the representative design of these two modules. Furthermore, we introduce a novel taxonomy of deep clustering based on how these two modules interact, specifically through multistage, generative, iterative, and simultaneous approaches. In addition, we present well-known benchmark datasets, evaluation metrics, and open-source tools to clearly demonstrate different experimental approaches. Finally, we examine the practical applications of deep clustering and propose challenging areas for future research.

Towards complexity in de Sitter space from the doubled-scaled Sachdev-Ye-Kitaev model

How can we define complexity in dS space from microscopic principles? Based on recent developments pointing towards a correspondence between a pair of double-scaled Sachdev-Ye-Kitaev (DSSYK) models/ 2D Liouville-de Sitter (LdS2) field theory/ 3D Schwarzschild de Sitter (SdS3) space in [1–3], we study concrete complexity proposals in the microscopic models and their dual descriptions. First, we examine the spread complexity of the maximal entropy state of the doubled DSSYK model. We show that it counts the number of entangled chord states in its doubled Hilbert space. We interpret spread complexity in terms of a time difference between antipodal observers in SdS3 space, and a boundary time difference of the dual LdS2 CFTs. This provides a new connection between entanglement and geometry in dS space. Second, Krylov complexity, which describes operator growth, is computed for physical operators on all sides of the correspondence. Their late time evolution behaves as expected for chaotic systems. Later, we define the query complexity in the LdS2 model as the number of steps in an algorithm computing n-point correlation functions of boundary operators of the corresponding antipodal points in SdS3 space. We interpret query complexity as the number of matter operator chord insertions in a cylinder amplitude in the DSSYK, and the number of junctions of Wilson lines between antipodal static patch observers in SdS3 space. Finally, we evaluate a specific proposal of Nielsen complexity for the DSSYK model and comment on its possible dual manifestations.

Advancing sustainable mobility: Dynamic predictive modeling of charging cycles in electric vehicles using machine learning techniques and predictive application development

The main goal in this research is to train various machine learning models to predict charging cycles in EV Electric Vehicles) battery systems. The considered models are gradient boosting, random forests, decision trees, and linear regression. Each of these was assessed based on its R-squared score, which is an important statistical measure in indicating the variance proportion yielded by the model. In contrast, the Random Forest model significantly improved, with an R-squared value of 0.83, thereby doing an excellent job in capturing nuances of the data. Only surpassed by the Gradient Boosting model at an astonishing R-squared score of 0.87, it is this excellent score that underlines its capability to predict the outcome quite accurately by modeling complex interrelations. In other words, gradient boosting outran the rest and provided the most robust results concerning drivers of students' performance. It also underlines how important choosing a good model is in educational analytics in order to increase the accuracy of the predictions. The use of these models in the proposed EV Battery Charging Cycle Predictor App results in accurate predictions to aid schedule maintenance and energy-related decisions. This research brings light to the future of advanced machine learning methods in enhancing the battery efficiencies of EVs and the development of electric mobility technologies. It is possible that the future work will imply the additional inclusion of real data and the integration of the application to general energy systems.

Considerations for using tree-based machine learning to assess causation between demographic and environmental risk factors and health outcomes.

Evaluation of the heterogeneous treatment effect (HTE) allows for the assessment of the causal effect of a therapy or intervention while considering heterogeneity in individual factors within a population. Machine learning (ML) methods have previously been employed for HTE evaluation, addressing the limitations associated with modelling complex systems. In this work, three tree-based ML algorithms, causal random forest (CRF), causal Bayesian additive regression trees (CBART), and causal rule ensemble (CRE), are used to analyze the potential causation of benzene exposure to cause childhood acute myeloid leukemia (AML). Data for this analysis is generated by drawing samples from a previously developed model that estimates AML probability given as input demographic information and benzene exposure. Comparison is drawn between the three tree-based algorithms in terms of the predicted average treatment effect (ATE), the regression coefficient of determination, and the computational time of each algorithm. Minimal difference is reported between the three tree-based algorithms in terms of the ATE, as well as the regression coefficient of determination. However, CRF outperforms CBART in terms of algorithm computational time. Moreover, CRF allows for both continuous and binary treatment variables, as opposed to CBART and CRE, making it better suited to environmental health studies, where exposure levels of pollutants shall be considered continuous. Following the comparison of all three algorithms, the influence of adding Gaussian noise to the treatment and outcome variables, as well as outliers, is investigated using CRF. A set of considerations is drawn to guide researchers in using these algorithms. These considerations detail the simulation settings, applications, and results interpretation and aim to provide prompt information in decision-making surrounding the establishment of pollutant exposure thresholds in environmental risk assessments.

Heterobimetallic NiFe Complex for Photocatalytic CO2 Reduction: United Efforts of NiFe Dual Sites.

Catalytic CO2 reduction poses a significant challenge for the conversion of CO2 into chemicals and fuels. Ni-Fe carbon monoxide dehydrogenase ([NiFe]-CODH) effectively mediates the reversible conversion of CO2 and CO at a nearly thermodynamic equilibrium potential, highlighting the heterobimetallic cooperation for the design of CO2 reduction catalysts. However, numerous NiFe biomimetic model complexes have realized little success in CO2 reduction catalysis, which underscores the crucial role of precise bimetallic configuration and functionality. Herein, we presented a heterobimetallic NiFe complex for the photocatalytic reduction of CO2 to CO, demonstrating significantly enhanced catalytic performance compared to the homonuclear NiNi catalyst. Photocatalytic and mechanistic investigations revealed that with the assistance of a redox-active phenanthroline ligand, NiFe achieves dual-site activation of CO2 through a pivotal intermediate, NiII(μ-CO22--κC:κO)FeII, where the Lewis acidity of the FeII site plays an important role, as corroborated in the homonuclear FeFe system. This study introduces the first heteronuclear NiFe molecular catalyst capable of efficiently catalyzing the reduction of CO2 to CO, deepening insights into heterobimetallic cooperation and offering a novel strategy for designing highly active and selective CO2 reduction catalysts.

Perspectives of Experts in the Selection of Contents for a Course of Differential Equations in Engineering

This study investigates the teaching of Differential Equations in engineering education, focusing on expert opinions from diverse institutional backgrounds. The primary objectives were to understand expert perspectives on essential course content, justify their content choices, measure institutional gaps, and highlight the range of professional profiles. Through qualitative interviews with nine experts from academia and industry, we explored their views on the integration of Differential Equations in engineering curricula. The findings reveal a consensus on the importance of basic concepts to support physical models. Notably, experts highlighted variational calculus and energy representations as critical yet underrepresented areas in Differential Equations courses, essential for modeling complex, nonlinear problems. The study also states on the growing importance of computational methods and software in engineering education, advocating for an integrated approach aligning theoretical concepts with real-world applications. The expert opinions revealed a significant institutional influence on content perception, demonstrating a gap between traditional theoretical focus and modern, application-oriented approaches. The study suggests a need for curricular innovation in differential equations education, emphasizing practical applications and computational methods to bridge this gap between the two institutions considered in which the experts are adhered to.

A PNP ion channel deep learning solver with local neural network and finite element input data

Abstract This paper presents a deep learning method for solving an improved one-dimensional Poisson–Nernst–Planck ion channel (PNPic) model, called the PNPic deep learning solver. The solver combines a novel local neural network, adapted from the neural network with local converging inputs, with an efficient PNPic finite element solver, developed in this work. In particular, the local neural network is extended to handle the complexities of the PNPic model—a system of nonlinear convection–diffusion and elliptic equations with multiple subdomains connected by interface conditions. The PNPic finite element solver efficiently generates input and reference datasets for fast training the local neural network, as well as input datasets for quickly predicting PNPic solutions with high accuracy for a family of PNPic models. Initial numerical tests, involving perturbations of model parameters and interface locations, demonstrate that the PNPic deep learning solver can generate highly accurate numerical solutions.

Structural Snapshots, Trajectory and Thermodynamics of the Reversible µ-1,2-Peroxo/µ-1,1-Hydroperoxo Dicopper(II) Interconversion.

Hydrogen bonds involving the oxygen atoms of intermediates that result from copper-mediated O2 activation play a key role for controlling the reactivity of Cux/O2 active sites in metalloenzymes and synthetic model complexes. However, structural insight into H-bonding in such transient species as well as thermodynamic information about proton transfer to or from the O2-derived ligands is scarce. Here we present a detailed study of the reversible interconversion of a μ1,2-peroxodicopper(II) complex ([1]+) and its μ1,1-hydroperoxo congener ([2]+) via (de)protonation, including the isolation and structural characterization of several H-bond donor (HBD) adducts of [1]+ and the determination of binding constants. For one of these adducts a temperature-dependent μ1,2-peroxo/μ1,1-hydroperoxo equilibrium associated with reversible H+-translocation is observed, its thermodynamics investigated experimentally and computationally, and effects of H-bonding on spectroscopic parameters of the CuII2(μ1,2-O2) species are revealed. DFT calculations allowed to fully map and correlate the trajectories of H+-transfer and μ1,2-peroxo→μ1,1-peroxo rearrangement. These findings enhance our understanding of two key intermediates in bioinspired Cu2/O2 chemistry.

An analytical dynamic stiffness method for efficient broadband vibro-acoustic analysis of complex built-up structures

An analytical modelling technique is presented, combining the dynamic stiffness method (DSM) with the spectral DSM (SDSM) for the broadband vibro-acoustic analysis of complex built-up structures. In particular, the structures are modelled exactly by the DSM based on particular solutions tailored for general acoustic or distributed force excitations; whereas the acoustic cavities are modelled by the SDSM where the boundary conditions are described by modified Fourier series with a rapid convergence rate. The vibro-acoustic coupling along the interfaces between the structures and the cavities is achieved by applying the normal velocity continuity condition, leading to a coupling matrix in an analytical manner. Both structural and acoustic cavity elements are assembled directly to model complex vibro-acoustic systems and no extra element discretization is required for both structural and acoustic domains. Consequently, the proposed method uses very few degrees of freedom (DoFs) but describes the broadband vibro-acoustic behaviour highly accurately. It is demonstrated that the proposed method exhibits a predominant advantage over the finite element package COMSOL in terms of accuracy and computational efficiency. This approach also has the potential to serve as the benchmark for a wide range of vibro-acoustic problems.

BEMM-GEN: A Toolkit for Generating a Biomolecular Environment-Mimicking Model for Molecular Dynamics Simulation.

Understanding the influence of the cellular environment on protein conformations is crucial for elucidating protein functions within living cells. In studies using molecular dynamics (MD) simulation, carbon nanotubes and hydrophobic cages have been widely used to emulate the cellular environment inside specific large biomolecules such as ribosome tunnels and chaperones. However, recent studies suggest that these uniform hydrophobic models may not adequately capture the environmental effects inside each biomolecule. Based on these facts, it is necessary to generate spherical and cylindrical models with varied chemical properties corresponding to the components within target biomolecules. We developed a biomolecular environment-mimicking model generator (BEMM-GEN) that generates spherical and cylindrical models with user-specified chemical properties and allows the integration of arbitrary protein conformations into the generated models. BEMM-GEN provides model and protein complex structures, along with the corresponding parameter files for MD simulation (AMBER and GROMACS), and users immediately run their MD simulation based on the generated input files. BEMM-GEN can be freely downloaded and installed via a Python package manager (pip install BEMM-gen). The source code files and a user manual for operation are provided on GitHub (https://github.com/y4suda/BEMM-GEN).

An artificial neural network based approach for harmonic component prediction in a distribution line

With the increasing use of nonlinear devices in both generation and consumption of power, it is essential that we develop accurate and quick control for active filters to suppress harmonics. Time delays between input and output are catastrophic for such filters which rely on real-time operation. Artificial Neural Networks (ANNs) are capable of modeling complex nonlinear systems through adjustments in their learned parameters. Once properly trained, they can produce highly accurate predictions at an instantaneous time frame. Leveraging these qualities, various complex control systems may be replaced or aided by neural networks to provide quick and precise responses. This paper proposes an ANN-based approach for the prediction of individual harmonic components using minimal inputs. By extracting and analyzing the nature of harmonic component magnitudes obtained from the survey of a particular area through real-time measurements, a sequential pattern in their occurrence is observed. Various neural network architectures are trained using the collected data and their performances are evaluated. The best-performing model, whose losses are minimal, is then used to observe the harmonic cancellation for multiple unseen cases through a simplified simulation in hardware-in-the-loop. These neural network structures, which produce instantaneous and accurate outputs, are effective in harmonic filtering.

Fitness-guided particle swarm optimization with adaptive Newton-Raphson for photovoltaic model parameter estimation

This study introduces a new approach for parameter optimization in the four-diode photovoltaic (PV) model, employing a Dynamic Fitness-Guided Particle Swarm Optimization (DFGPSO) algorithm and Enhanced Newton-Raphson (ENR) method. The new DFGPSO algorithm is specifically designed to address the intrinsic challenges in PV modelling, such as local optima entrapment and slow convergence rates that typically hinder traditional optimization methods. By integrating a dynamically evolving fitness function derived from advanced swarm intelligence, the proposed approach significantly enhances global search capabilities. This new fitness function adapts continuously to the search landscape, facilitating rapid convergence towards optimal solutions and effectively navigating the complex, non-linear, and multi-modal parameter space of the PV model. Moreover, the robustness of the DFGPSO algorithm is substantially improved through the strategic incorporation of the ENR method. This integration not only provides accurate initial guesses for the particle positions, thus expediting the convergence process, but also minimizes computational burden, making the method more efficient. Comprehensive simulation studies across various case scenarios demonstrate that the proposed method markedly outperforms existing state-of-the-art optimization algorithms. It delivers faster convergence, enhanced accuracy, and robust performance under diverse environmental conditions, establishing a reliable and precise tool for optimizing PV system performance. This advancement promises significant improvements in energy yield and system reliability for the PV industry.

Optimizing parameters for additive manufacturing: a study on the vibrational performance of 3D printed cantilever beams using material extrusion

Purpose This study aims to investigate the impact of different printing parameters on the free vibration characteristics of 3D printed cantilever beams. Through a comprehensive analysis of material extrusion (ME) variables such as extrusion rate, printing pattern and layer thickness, the study seeks to enhance the understanding of how these parameters influence the vibrational properties, particularly the natural frequency, of printed components. Design/methodology/approach The experimental design involves conducting a series of experiments using a central composite design approach to gather data on the vibrational response of ABS cantilever beams under diverse ME parameters. These parameters are systematically varied across different levels, facilitating a thorough exploration of their effects on the vibrational behavior of the printed specimens. The collected data are then used to develop a predictive model leveraging a hybrid artificial neural network (ANN)/ particle swarm optimization (PSO) approach, which combines the strengths of ANN in modeling complex relationships and PSO in optimizing model parameters. Findings The developed ANN/PSO hybrid model demonstrates high accuracy in predicting the natural frequency of 3D printed cantilever beams, with a correlation ratio (R) of 0.9846 when tested against experimental data. Through iterative fine-tuning with PSO, the model achieves a low mean square error (MSE) of 1.1353e-5, underscoring its precision in estimating the vibrational characteristics of printed specimens. Furthermore, the model’s transformation into a regression model enables the derivation of surface response characteristics governing the vibration properties of 3D printed objects in response to input parameters, facilitating the identification of optimal parameter configurations for maximizing vibration characteristics in 3D printed products. Originality/value This study introduces a novel predictive model that combines ANNs with PSO to analyze the vibrational behavior of 3D printed ABS cantilever beams produced under various ME parameters. By integrating these advanced methodologies, the research offers a pioneering approach to precisely estimating the natural frequency of 3D printed objects, contributing to the advancement of predictive modeling in additive manufacturing.

Learning accurate neighborhood- and self-information for higher-order relation prediction in Heterogeneous Information Networks

Heterogeneous Information Networks (HINs) are commonly employed to model complex real-world scenarios with diverse node and edge types. However, due to constraints in data collection and processing, constructed networks often lack certain relations. Consequently, various methods have emerged, particularly recently, leveraging heterogeneous graph neural networks (HGNNs) to predict missing relations. Nevertheless, these methods primarily focus on pairwise relations between two nodes. Real-world interactions, however, often involve multiple nodes and diverse types, extending beyond simple pairwise relations. For instance, academic collaboration networks may entail interactions among authors, papers, and conferences simultaneously. Despite their prevalence, higher-order relations are often overlooked. While HGNNs are effective at learning network structures, they may suffer from over-smoothing, resulting in similar representations for nodes and their neighbors. The learned inaccurate proximity among nodes impedes the discernment of higher-order relations. Furthermore, observed edges among a target group of nodes can provide valuable evidence for predicting higher-order relations. To address these challenges, we propose a novel model called Accurate Neighborhood- and Self-information Enhanced Heterogeneous Graph Neural Network (ANSE-HGN). Building upon HGNNs to encode network structure and attributes, we introduce a relation-based neighborhood encoder to capture information within multi-hop neighborhoods in heterogeneous higher-order relations. This enables the calculation of accurate proximity among target groups of nodes, thereby enhancing prediction accuracy. Additionally, we leverage self-information from observed higher-order relations as an auxiliary loss to reinforce the learning process. Extensive experiments on four real-world datasets demonstrate the superiority of our proposed method in higher-order relation prediction tasks.

Development of polarizable and hydration-focused water models for the Martini 3 force field

Coarse-grained simulations have been increasingly employed in many fields of molecular chemistry. These models enable the evaluation of complex molecular structures by simplifying their description compared to all-atoms models. Specifically, the Martini 3 force field has been developed to describe biomolecular structures such as proteins, lipids, and carbohydrates. Their “building block” approach has been successfully used in modeling complex systems ranging from carbohydrate solutions, biological membranes, and even deep eutectic solvents. However, one main issue arises when considering interfacial systems: water is represented as a single particle with no effect of electrostatic interactions. In this work, we developed a polarizable water model for the Martini 3 force field, introducing charged particles to model the electrostatic interactions present in water systems accurately. Gaussian Processes were trained with simulation data for both bulk water and cross-interaction parameterization to consistently explore the parameter space and fine-tune our model to the target properties. The developed model showed good performance in reproducing the density and dielectric constant of bulk water. Regarding cross-interaction with other beads, two different sets of parameters were developed. The first, called the polarizable model, is parameterized to reproduce the hydration free energies of the original Martini 3 organic beads. This allows for a good description of free energies of transfer between water and organic phases, the main target property for the Martini 3 force field. The second, called the polarizable-hydration model, is parameterized to match the experimental hydration free energy of selected molecules. This allows for more accurate evaluation of hydration properties, which can be useful for water solutions. In particular, we show that the partitioning and interfacial behavior of nonionic surfactants is improved by the polarizable-hydration model when compared to the standard Martini 3 water model.