Parkinson's disease(PD)exhibits significant phenotypic heterogeneity,which complicates clinical management and underscores the need for precise subtyping.Existing subtyping approaches often rely on a single modali...Parkinson's disease(PD)exhibits significant phenotypic heterogeneity,which complicates clinical management and underscores the need for precise subtyping.Existing subtyping approaches often rely on a single modality,such as clinical assessments,failing to capture the complex,multi-faceted nature of the disease.This paper proposes a novel computational framework that integrates multi-modal data,specifically preprocessed functional MRI,DNA methylation,and clinical behavioral assessments,for PD subtyping.The methodology involves constructing individual hypergraphs for each modality using K-nearest neighbors(KNN),followed by the integration of these hypergraphs into a unified,multi-modal hypergraph using similarity network fusion(SNF).This consolidated hypergraph is then processed via a hypergraph neural network(HGNN)utilizing hyperedge convolution to cluster patients into distinct subtypes.Our experimental results demonstrate that this approach effectively identifies PD subtypes with significant clinical and biological relevance.We provide a comprehensive analysis of the model's performance and further validate the reliability of the identified subtypes through post-hoc statistical tests.This study highlights the potential of graph-based machine learning in disentangling disease heterogeneity,paving the way for personalized therapeutic strategies and improved patient outcomes.展开更多
Solvation free energies play a fundamental role in various fields of chemistry and biology.Accurately determining the solvation Gibbs free energy(ΔGsolv)of a molecule in a given solvent requires a deep understandi...Solvation free energies play a fundamental role in various fields of chemistry and biology.Accurately determining the solvation Gibbs free energy(ΔGsolv)of a molecule in a given solvent requires a deep understanding of the intrinsic relationships between solute and solvent molecules.While deep learning methods have been developed forΔGsolv prediction,few explicitly model intermolecular interactions between solute and solvent molecules.The molecular modeling graph neural network more closely aligns with real-world chemical processes by explicitly capturing atomic-level interactions,such as hydrogen bonding.It achieves this by initially establishing indiscriminate connections between intermolecular atoms,which are subsequently refined using an attention-based aggregation mechanism tailored to specific solute–solvent pairs.However,its sharply increasing computational complexity limits its scalability and broader applicability.Here,we introduce an improved framework,molecular merged hypergraph neural network(MMHNN),which leverages a predefined subgraph set and replaces subgraphs with supernodes to construct a hypergraph representation.This design effectively mitigates model complexity while preserving key molecular interactions.Furthermore,to handle noninteractive or repulsive atomic interactions,MMHNN incorporates an interpretation mechanism for nodes and edges within the merged graph,leveraging the graph information bottleneck theory to enhance model explainability.Extensive experimental validation demonstrates the efficiency of MMHNN and its improved interpretability in capturing solute–solvent interactions.展开更多
semantics information while maintaining spatial detail con-texts.Long-range context information plays a crucial role in this scenario.How-ever,the traditional convolution kernel only provides the local and small size ...semantics information while maintaining spatial detail con-texts.Long-range context information plays a crucial role in this scenario.How-ever,the traditional convolution kernel only provides the local and small size of the receptivefield.To address the problem,we propose a plug-and-play module aggregating both local and global information(aka LGIA module)to capture the high-order relationship between nodes that are far apart.We incorporate both local and global correlations into hypergraph which is able to capture high-order rela-tionships between nodes via the concept of a hyperedge connecting a subset of nodes.The local correlation considers neighborhood nodes that are spatially adja-cent and similar in the same CNN feature maps of magnetic resonance(MR)image;and the global correlation is searched from a batch of CNN feature maps of MR images in feature space.The influence of these two correlations on seman-tic segmentation is complementary.We validated our LGIA module on various CNN segmentation models with the cardiac MR images dataset.Experimental results demonstrate that our approach outperformed several baseline models.展开更多
现有的电子健康记录(electronic health records,EHR)的图表示学习方法多依赖单个患者的局部信息,忽视了群体患者在疾病演化和诊疗路径上的潜在关联,从而限制了模型的泛化性与鲁棒性.针对这一问题,本文提出一种混合多层级图神经网络(hyb...现有的电子健康记录(electronic health records,EHR)的图表示学习方法多依赖单个患者的局部信息,忽视了群体患者在疾病演化和诊疗路径上的潜在关联,从而限制了模型的泛化性与鲁棒性.针对这一问题,本文提出一种混合多层级图神经网络(hybrid multi-level graph neural network,H-MGNN)模型,并将其应用于重症监护室(intensive care unit,ICU)患者的死亡预测.该模型通过构建宏观层面的患者关系图(patient-patient graph,P-P)、微观层面的分类-笔记-词汇超图(taxonomy-note-word hypergraph,T-N-W),结合超图的时序依赖关系,实现多尺度上的患者特征融合.同时,本文设计了融合算法(hybrid embedding,Hybrid-E),用于提取和整合患者嵌入的潜在特征,以提升预测准确性.实验结果表明,H-MGNN在MIMIC-Ⅲ(medical information mart for intensive care Ⅲ)数据集上的住院死亡率预测等任务中显著优于现有方法,验证了其在复杂EHR数据挖掘中的有效性和先进性.展开更多
点云分割是场景理解、目标识别、文化遗产保护等领域的一项基础且关键的技术。然而,由于点云数据的复杂性,如何从点云中提取深层特征对目前的研究提出了很大的挑战。为了解决这一问题,提出了一个结合超图卷积和生成对抗网络的新框架,用...点云分割是场景理解、目标识别、文化遗产保护等领域的一项基础且关键的技术。然而,由于点云数据的复杂性,如何从点云中提取深层特征对目前的研究提出了很大的挑战。为了解决这一问题,提出了一个结合超图卷积和生成对抗网络的新框架,用于点云分割。首先,使用超图建模点云之间的几何拓扑关系,捕捉点云中的高阶相关性。其次,以多层超图卷积作为鉴别器,结合生成对抗模型构建超图生成对抗网络,提取点云稀疏区域的细节特征。最后,利用逐点损失和对抗损失相结合对网络进行训练,提升网络训练的稳定性和标签预测的准确性。在ShapeNet Part数据集上进行了点云分割实验,结果表明所提方法在16个类别上的平均交并比(Intersection over Union,IoU)为84.5%,较大地提升了分割精度。展开更多
摘要Parkinson's disease(PD)exhibits significant phenotypic heterogeneity,which complicates clinical management and underscores the need for precise subtyping.Existing subtyping approaches often rely on a single modality,such as clinical assessments,failing to capture the complex,multi-faceted nature of the disease.This paper proposes a novel computational framework that integrates multi-modal data,specifically preprocessed functional MRI,DNA methylation,and clinical behavioral assessments,for PD subtyping.The methodology involves constructing individual hypergraphs for each modality using K-nearest neighbors(KNN),followed by the integration of these hypergraphs into a unified,multi-modal hypergraph using similarity network fusion(SNF).This consolidated hypergraph is then processed via a hypergraph neural network(HGNN)utilizing hyperedge convolution to cluster patients into distinct subtypes.Our experimental results demonstrate that this approach effectively identifies PD subtypes with significant clinical and biological relevance.We provide a comprehensive analysis of the model's performance and further validate the reliability of the identified subtypes through post-hoc statistical tests.This study highlights the potential of graph-based machine learning in disentangling disease heterogeneity,paving the way for personalized therapeutic strategies and improved patient outcomes.
基金partially supported by the Project of Stable Support for Youth Team in Basic Research Field,CAS(YSBR-005)the Anhui Science Foundation for Distinguished Young Scholars(No.1908085J24)+1 种基金the Natural Science Foundation of China(No.62072427)the Jiangsu Natural Science Foundation(No.BK2019119).
摘要Solvation free energies play a fundamental role in various fields of chemistry and biology.Accurately determining the solvation Gibbs free energy(ΔGsolv)of a molecule in a given solvent requires a deep understanding of the intrinsic relationships between solute and solvent molecules.While deep learning methods have been developed forΔGsolv prediction,few explicitly model intermolecular interactions between solute and solvent molecules.The molecular modeling graph neural network more closely aligns with real-world chemical processes by explicitly capturing atomic-level interactions,such as hydrogen bonding.It achieves this by initially establishing indiscriminate connections between intermolecular atoms,which are subsequently refined using an attention-based aggregation mechanism tailored to specific solute–solvent pairs.However,its sharply increasing computational complexity limits its scalability and broader applicability.Here,we introduce an improved framework,molecular merged hypergraph neural network(MMHNN),which leverages a predefined subgraph set and replaces subgraphs with supernodes to construct a hypergraph representation.This design effectively mitigates model complexity while preserving key molecular interactions.Furthermore,to handle noninteractive or repulsive atomic interactions,MMHNN incorporates an interpretation mechanism for nodes and edges within the merged graph,leveraging the graph information bottleneck theory to enhance model explainability.Extensive experimental validation demonstrates the efficiency of MMHNN and its improved interpretability in capturing solute–solvent interactions.
基金supported by the Sichuan Science and Technology Program(Grant No.2019ZDZX0005,2019YFG0496,2020YFG0143,2019JDJQ0002 and 2020YFG0009).
摘要semantics information while maintaining spatial detail con-texts.Long-range context information plays a crucial role in this scenario.How-ever,the traditional convolution kernel only provides the local and small size of the receptivefield.To address the problem,we propose a plug-and-play module aggregating both local and global information(aka LGIA module)to capture the high-order relationship between nodes that are far apart.We incorporate both local and global correlations into hypergraph which is able to capture high-order rela-tionships between nodes via the concept of a hyperedge connecting a subset of nodes.The local correlation considers neighborhood nodes that are spatially adja-cent and similar in the same CNN feature maps of magnetic resonance(MR)image;and the global correlation is searched from a batch of CNN feature maps of MR images in feature space.The influence of these two correlations on seman-tic segmentation is complementary.We validated our LGIA module on various CNN segmentation models with the cardiac MR images dataset.Experimental results demonstrate that our approach outperformed several baseline models.
摘要现有的电子健康记录(electronic health records,EHR)的图表示学习方法多依赖单个患者的局部信息,忽视了群体患者在疾病演化和诊疗路径上的潜在关联,从而限制了模型的泛化性与鲁棒性.针对这一问题,本文提出一种混合多层级图神经网络(hybrid multi-level graph neural network,H-MGNN)模型,并将其应用于重症监护室(intensive care unit,ICU)患者的死亡预测.该模型通过构建宏观层面的患者关系图(patient-patient graph,P-P)、微观层面的分类-笔记-词汇超图(taxonomy-note-word hypergraph,T-N-W),结合超图的时序依赖关系,实现多尺度上的患者特征融合.同时,本文设计了融合算法(hybrid embedding,Hybrid-E),用于提取和整合患者嵌入的潜在特征,以提升预测准确性.实验结果表明,H-MGNN在MIMIC-Ⅲ(medical information mart for intensive care Ⅲ)数据集上的住院死亡率预测等任务中显著优于现有方法,验证了其在复杂EHR数据挖掘中的有效性和先进性.
摘要点云分割是场景理解、目标识别、文化遗产保护等领域的一项基础且关键的技术。然而,由于点云数据的复杂性,如何从点云中提取深层特征对目前的研究提出了很大的挑战。为了解决这一问题,提出了一个结合超图卷积和生成对抗网络的新框架,用于点云分割。首先,使用超图建模点云之间的几何拓扑关系,捕捉点云中的高阶相关性。其次,以多层超图卷积作为鉴别器,结合生成对抗模型构建超图生成对抗网络,提取点云稀疏区域的细节特征。最后,利用逐点损失和对抗损失相结合对网络进行训练,提升网络训练的稳定性和标签预测的准确性。在ShapeNet Part数据集上进行了点云分割实验,结果表明所提方法在16个类别上的平均交并比(Intersection over Union,IoU)为84.5%,较大地提升了分割精度。