Robust deep learning from incomplete annotation for accurate lung nodule detection

Deep learning plays a significant role in the detection of pulmonary nodules in low-dose computed tomography (LDCT) scans, contributing to the diagnosis and treatment of lung cancer. Nevertheless, its effectiveness often relies on the availability of extensive, meticulously annotated dataset. In this paper, we explore the utilization of an incompletely annotated dataset for pulmonary nodules detection and introduce the FULFIL (Forecasting Uncompleted Labels For Inexpensive Lung nodule detection) algorithm as an innovative approach. By instructing annotators to label only the nodules they are most confident about, without requiring complete coverage, we can substantially reduce annotation costs. Nevertheless, this approach results in an incompletely annotated dataset, which presents challenges when training deep learning models. Within the FULFIL algorithm, we employ Graph Convolution Network (GCN) to discover the relationships between annotated and unannotated nodules for self-adaptively completing the annotation. Meanwhile, a teacher-student framework is employed for self-adaptive learning using the completed annotation dataset. Furthermore, we have designed a Dual-Views loss to leverage different data perspectives, aiding the model in acquiring robust features and enhancing generalization. We carried out experiments using the LUng Nodule Analysis (LUNA) dataset, achieving a sensitivity of 0.574 at a False positives per scan (FPs/scan) of 0.125 with only 10% instance-level annotations for nodules. This performance outperformed comparative methods by 7.00%. Experimental comparisons were conducted to evaluate the performance of our model and human experts on test dataset. The results demonstrate that our model can achieve a comparable level of performance to that of human experts. The comprehensive experimental results demonstrate that FULFIL can effectively leverage an incomplete pulmonary nodule dataset to develop a robust deep learning model, making it a promising tool for assisting in lung nodule detection.

Predicting disease-gene associations through self-supervised mutual infomax graph convolution network

Illuminating associations between diseases and genes can help reveal the pathogenesis of syndromes and contribute to treatments, but a large number of associations remained unexplored. To identify novel disease-gene associations, many computational methods have been developed using disease and gene-related prior knowledge. However, these methods remain of relatively inferior performance due to the limited external data sources and the inevitable noise among the prior knowledge. In this study, we have developed a new method, Self-Supervised Mutual Infomax Graph Convolution Network (MiGCN), to predict disease-gene associations under the guidance of external disease-disease and gene-gene collaborative graphs. The noises within the collaborative graphs were eliminated by maximizing the mutual information between nodes and neighbors through a graphical mutual infomax layer. In parallel, the node interactions were strengthened by a novel informative message passing layer to improve the learning ability of graph neural network. The extensive experiments showed that our model achieved performance improvement over the state-of-art method by more than 8 % on AUC. The datasets, source codes and trained models of MiGCN are available at https://github.com/biomed-AI/MiGCN.

Predicting potential microbe-disease associations based on auto-encoder and graph convolution network

The increasing body of research has consistently demonstrated the intricate correlation between the human microbiome and human well-being. Microbes can impact the efficacy and toxicity of drugs through various pathways, as well as influence the occurrence and metastasis of tumors. In clinical practice, it is crucial to elucidate the association between microbes and diseases. Although traditional biological experiments accurately identify this association, they are time-consuming, expensive, and susceptible to experimental conditions. Consequently, conducting extensive biological experiments to screen potential microbe-disease associations becomes challenging. The computational methods can solve the above problems well, but the previous computational methods still have the problems of low utilization of node features and the prediction accuracy needs to be improved. To address this issue, we propose the DAEGCNDF model predicting potential associations between microbes and diseases. Our model calculates four similar features for each microbe and disease. These features are fused to obtain a comprehensive feature matrix representing microbes and diseases. Our model first uses the graph convolutional network module to extract low-rank features with graph information of microbes and diseases, and then uses a deep sparse Auto-Encoder to extract high-rank features of microbe-disease pairs, after which the low-rank and high-rank features are spliced to improve the utilization of node features. Finally, Deep Forest was used for microbe-disease potential relationship prediction. The experimental results show that combining low-rank and high-rank features helps to improve the model performance and Deep Forest has better classification performance than the baseline model.

TGDAUNet: Transformer and GCNN based dual-branch attention UNet for medical image segmentation

Accurate and automatic segmentation of medical images is a key step in clinical diagnosis and analysis. Currently, the successful application of Transformers' model in the field of computer vision, researchers have begun to gradually explore the application of Transformers in medical segmentation of images, especially in combination with convolutional neural networks with coding-decoding structure, which have achieved remarkable results in the field of medical segmentation. However, most studies have combined Transformers with CNNs at a single scale or processed only the highest-level semantic feature information, ignoring the rich location information in the lower-level semantic feature information. At the same time, for problems such as blurred structural boundaries and heterogeneous textures in images, most existing methods usually simply connect contour information to capture the boundaries of the target. However, these methods cannot capture the precise outline of the target and ignore the potential relationship between the boundary and the region. In this paper, we propose the TGDAUNet, which consists of a dual-branch backbone network of CNNs and Transformers and a parallel attention mechanism, to achieve accurate segmentation of lesions in medical images. Firstly, high-level semantic feature information of the CNN backbone branches is fused at multiple scales, and the high-level and low-level feature information complement each other's location and spatial information. We further use the polarised self-attentive (PSA) module to reduce the impact of redundant information caused by multiple scales, to better couple with the feature information extracted from the Transformers backbone branch, and to establish global contextual long-range dependencies at multiple scales. In addition, we have designed the Reverse Graph-reasoned Fusion (RGF) module and the Feature Aggregation (FA) module to jointly guide the global context. The FA module aggregates high-level semantic feature information to generate an original global predictive segmentation map. The RGF module captures non-significant features of the boundaries in the original or secondary global prediction segmentation graph through a reverse attention mechanism, establishing a graph reasoning module to explore the potential semantic relationships between boundaries and regions, further refining the target boundaries. Finally, to validate the effectiveness of our proposed method, we compare our proposed method with the current popular methods in the CVC-ClinicDB, Kvasir-SEG, ETIS, CVC-ColonDB, CVC-300,datasets as well as the skin cancer segmentation datasets ISIC-2016 and ISIC-2017. The large number of experimental results show that our method outperforms the currently popular methods. Source code is released at https://github.com/sd-spf/TGDAUNet.

Attention based spatio-temporal graph convolutional network with focal loss for crash risk evaluation on urban road traffic network based on multi-source risks

The urban road transportation has presented a high probability of crash occurrence, and the aim of the present study is to evaluate the crash risk for urban road networks. However, the irregular structure of urban road networks, the high-dimensional spatio-temporal correlations among multi-source risks (i.e., the contributing risks from traffic flow, meteorological conditions, road design, and so forth), and the issue of data imbalance have brought challenges to this topic. To solve these issues, an Attention based Spatio-Temporal Graph Convolutional Network (ASTGCN) model with focal loss function is used for the first time to evaluate crash risk on an urban road network. This work can be summarized as (1) adopting the spatio-temporal graph convolution structure to capture the spatio-temporal properties and characterize the multi-source risks; (2) utilizing an attention mechanism network to address the critical contributing risks during crash risk evaluation; (3) introducing the focal loss function to improve the model performance impacted by the imbalanced data; and (4) investigating the different contributions of multi-source risks to model performance. The evaluation performance is tested in a real-world urban road traffic network. The raw data consists of 1239 crash records with corresponding datasets of traffic flow characteristics, meteorological conditions, road attributes and the topological structure of the road network. At the same time, three baseline models Artificial Neural Network (ANN), Random Forest (RF), and Deep Spatio-Temporal Graph Convolutional Network (DSTGCN) are compared to the proposed ASTGCN on the same datasets. Overall, the results show that ASTGCN outperforms the baseline models in several evaluation metrics. ASTGCN with focal loss function further improves performance by tackling the issues of dataset imbalance. Additionally, it is also found that the traffic flow risk is most crucial to model performance. The findings of the present study indicate that the proposed model can efficiently evaluate dynamic crash risk for urban road networks, which will benefit the safety management of urban road transportation.

Joint learning of feature and topology for multi-view graph convolutional network

Graph convolutional network has been extensively employed in semi-supervised classification tasks. Although some studies have attempted to leverage graph convolutional networks to explore multi-view data, they mostly consider the fusion of feature and topology individually, leading to the underutilization of the consistency and complementarity of multi-view data. In this paper, we propose an end-to-end joint fusion framework that aims to simultaneously conduct a consistent feature integration and an adaptive topology adjustment. Specifically, to capture the feature consistency, we construct a deep matrix decomposition module, which maps data from different views onto a feature space obtaining a consistent feature representation. Moreover, we design a more flexible graph convolution that allows to adaptively learn a more robust topology. A dynamic topology can greatly reduce the influence of unreliable information, which acquires a more adaptive representation. As a result, our method jointly designs an effective feature fusion module and a topology adjustment module, and lets these two modules mutually enhance each other. It takes full advantage of the consistency and complementarity to better capture the more intrinsic information. The experimental results indicate that our method surpasses state-of-the-art semi-supervised classification methods.

CEGAT: A CNN and enhanced-GAT based on key sample selection strategy for hyperspectral image classification

In recent years, the application of convolutional neural networks (CNNs) and graph convolutional networks (GCNs) in hyperspectral image classification (HSIC) has achieved remarkable results. However, the limited label samples are still a major challenge when using CNN and GCN to classify hyperspectral images. In order to alleviate this problem, a double branch fusion network of CNN and enhanced graph attention network (CEGAT) based on key sample selection strategy is proposed. First, a linear discrimination of spectral inter-class slices (LD_SICS) module is designed to eliminate spectral redundancy of HSIs. Then, a spatial spectral correlation attention (SSCA) module is proposed, which can extract and assign attention weight to the spatial and spectral correlation features. On the graph attention (GAT) branch, the HSI is segmented into some super pixels as input to reduce the amount of network parameters. In addition, an enhanced graph attention (EGAT) module is constructed to enhance the relationship between nodes. Finally, a key sample selection (KSS) strategy is proposed to enable the network to achieve better classification performance with few labeled samples. Compared with other state-of-the-art methods, CEGAT has better classification performance under limited label samples.

Glimpse and focus: Global and local-scale graph convolution network for skeleton-based action recognition

In the 3D skeleton-based action recognition task, learning rich spatial and temporal motion patterns from body joints are two foundational yet under-explored problems. In this paper, we propose two methods for improving these problems: (I) a novel glimpse-focus action recognition strategy that captures multi-range pose features from the whole body and key body parts jointly; (II) a powerful temporal feature extractor JD-TC that enriches trajectory features by inferring different inter-frame correlations for different joints. By coupling these two proposals, we develop a powerful skeleton-based action recognition system that extracts rich pose and trajectory features from a skeleton sequence and outperforms previous state-of-the-art methods on three large-scale datasets.

Monitor water quality through retrieving water quality parameters from hyperspectral images using graph convolution network with superposition of multi-point effect: A case study in Maozhou River

Quantitative prediction by unmanned aerial vehicle (UAV) remote sensing on water quality parameters (WQPs) including phosphorus, nitrogen, chemical oxygen demand (COD), biochemical oxygen demand (BOD), and chlorophyll a (Chl-a), total suspended solids (TSS), and turbidity provides a flexible and effective approach to monitor the variation in water quality. In this study, a deep learning-based method integrating graph convolution network (GCN), gravity model variant, and dual feedback machine involving parametric probability analysis and spatial distribution pattern analysis, named Graph Convolution Network with Superposition of Multi-point Effect (SMPE-GCN) has been developed to calculate concentrations of WQPs through UAV hyperspectral reflectance data on large scale efficiently. With an end-to-end structure, our proposed method has been applied to assisting environmental protection department to trace potential pollution sources in real time. The proposed method is trained on a real-world dataset and its effectiveness is validated on an equal amount of testing dataset with respect to three evaluation metrics including root of mean squared error (RMSE), mean absolute percent error (MAPE), and coefficient of determination (R²). The experimental results demonstrate that our proposed model achieves better performance in comparison with state-of-the-art baseline models in terms of RMSE, MAPE, and R². The proposed method is applicable for quantifying seven various WQPs and has achieved good performance for each WQP. The resulting MAPE ranges from 7.16% to 10.96% and R² ranges from 0.80 to 0.94 for all WQPs. This approach brings a novel and systematic insight into real-time quantitative water quality monitoring of urban rivers, and provides a unified framework for in-situ data acquisition, feature engineering, data conversion, and data modeling for further research. It provides fundamental support to assist environmental managers to efficiently monitor water quality of urban rivers.

GraphLSurv: A scalable survival prediction network with adaptive and sparse structure learning for histopathological whole-slide images

BACKGROUND AND OBJECTIVE

Predicting patients' survival from gigapixel Whole-Slide Images (WSIs) has always been a challenging task. To learn effective WSI representations for survival prediction, existing deep learning methods have explored utilizing graphs to describe the complex structure inner WSIs, where graph node is respective to WSI patch. However, these graphs are often densely-connected or static, leading to some redundant or missing patch correlations. Moreover, these methods cannot be directly scaled to the very-large WSI with more than 10,000 patches. To address these, this paper proposes a scalable graph convolution network, GraphLSurv, which can efficiently learn adaptive and sparse structures to better characterize WSIs for survival prediction.

METHODS

GraphLSurv has three highlights in methodology: (1) it generates adaptive and sparse structures for patches so that latent patch correlations could be captured and adjusted dynamically according to prediction tasks; (2) based on the generated structure and a given graph, GraphLSurv further aggregates local microenvironmental cues into a non-local embedding using the proposed hybrid message passing network; (3) to make this network suitable for very large-scale graphs, it adopts an anchor-based technique to reduce theorical computation complexity.

RESULTS

The experiments on 2268 WSIs show that GraphLSurv achieves a concordance-index of 0.66132 and 0.68348, with an improvement of 3.79% and 3.41% compared to existing methods, on NLST and TCGA-BRCA, respectively.

CONCLUSIONS

GraphLSurv could often perform better than previous methods, which suggests that GraphLSurv could provide an important and effective means for WSI survival prediction. Moreover, this work empirically shows that adaptive and sparse structures could be more suitable than static or dense ones for modeling WSIs.

Interpretable classification of pathology whole-slide images using attention based context-aware graph convolutional neural network

BACKGROUND AND OBJECTIVE

Whole slide image (WSI) classification and lesion localization within giga-pixel slide are challenging tasks in computational pathology that requires context-aware representations of histological features to adequately infer nidus. The existing weakly supervised learning methods mainly treat different locations in the slide as independent regions and cannot learn potential nonlinear interactions between instances based on i.i.d assumption, resulting in the model unable to effectively utilize context-ware information to predict the labels of WSIs and locate the region of interest (ROI).

METHODS

Here, we propose an interpretable classification model named bidirectional Attention-based Multiple Instance Learning Graph Convolutional Network (ABMIL-GCN), which hierarchically aggregates context-aware features of instances into a global representation in a topology fashion to predict the slide labels and localize the region of lymph node metastasis in WSIs.

RESULTS

We verified the superiority of this method on the Camelyon16 dataset, and the results show that the average predicted ACC and AUC of the proposed model after flooding optimization can reach 90.89% and 0.9149, respectively. The average accuracy and ACC score are improved by more than 7% and 4% compared with the existing state-of-the-art algorithms.

CONCLUSIONS

The results demonstrate that context-aware GCN outperforms existing weakly supervised learning methods by introducing spatial correlations between the neighbor image patches, which also addresses the 'accuracy-interpretability trade-off' problem. The framework provides a novel paradigm for the clinical application of computer-aided diagnosis and intelligent systems.

Automatic diagnosis and grading of Prostate Cancer with weakly supervised learning on whole slide images

BACKGROUND

The workflow of prostate cancer diagnosis and grading is cumbersome and the results suffer from substantial inter-observer variability. Recent trials have shown potential in using machine learning to develop automated systems to address this challenge. Most automated deep learning systems for prostate cancer Gleason grading focused on supervised learning requiring demanding fine-grained pixel-level annotations.

METHODS

A weakly-supervised deep learning model with slide-level labels is presented in this study for the diagnosis and grading of prostate cancer with whole slide image (WSI). WSIs are first cropped into small patches and then processed with a deep learning model to extract patch-level features. A graph convolution network (GCN) is used to aggregate the features for classifications. Throughout the training process, the noisy labels are progressively filtered out to reduce inter-observer variations in clinical reports. Finally, multi-center independent test cohorts with 6,174 slides are collected to evaluate the prostate cancer diagnosis and grading performance of our model.

RESULTS

The cancer diagnosis (2-level classification) results on two external test sets (n= 4,675, n= 844) show an area under the receiver operating characteristic curve (AUC) of 0.985 and 0.986. The Gleason grading (6-level classification) results reach 0.931 quadratic weighted kappa on the internal test set (n= 531). It generalizes well on the external test dataset (n= 844) with 0.801 quadratic weighted kappa with the reference standard set independently. The model enables pathological meaningful interpretability by visualizing the most attended lesions which are highly consistent with expert annotations.

CONCLUSION

The proposed model incorporates a graph network in weakly supervised learning with only slide-level reports. A robust learning strategy is also employed to correct the label noise. It is highly accurate (>0.985 AUC for diagnosis) and also interpretable with intuitive heatmap visualization. It can be unified with a digital pathology pipeline to deliver prostate cancer metrics for a pathology report.

GCNCPR-ACPs: a novel graph convolution network method for ACPs prediction

BACKGROUND

Anticancer peptide (ACP) inhibits and kills tumor cells. Research on ACP is of great significance for the development of new drugs, and the prediction of ACPs and non-ACPs is the new hotspot.

RESULTS

We propose a new machine learning-based method named GCNCPR-ACPs (a Graph Convolutional Neural Network Method based on collapse pooling and residual network to predict the ACPs), which automatically and accurately predicts ACPs using residual graph convolution networks, differentiable graph pooling, and features extracted using peptide sequence information extraction. The GCNCPR-ACPs method can effectively capture different levels of node attributes for amino acid node representation learning, GCNCPR-ACPs uses node2vec and one-hot embedding methods to extract initial amino acid features for ACP prediction.

CONCLUSIONS

Experimental results of ten-fold cross-validation and independent validation based on different metrics showed that GCNCPR-ACPs significantly outperformed state-of-the-art methods. Specifically, the evaluation indicators of Matthews Correlation Coefficient (MCC) and AUC of our predicator were 69.5% and 90%, respectively, which were 4.3% and 2% higher than those of the other predictors, respectively, in ten-fold cross-validation. And in the independent test, the scores of MCC and SP were 69.6% and 93.9%, respectively, which were 37.6% and 5.5% higher than those of the other predictors, respectively. The overall results showed that the GCNCPR-ACPs method proposed in the current paper can effectively predict ACPs.

GCN-GENE: A novel method for prediction of coronary heart disease-related genes

Coronary heart disease is the most common heart disease, it can induce myocardial infarction, and the cause of the disease has a lot to do with life and eating habits. The results of a large number of epidemiological studies at home and abroad show that the incidence of coronary heart disease has an obvious familial tendency. However, little is known about the genetic factors of coronary heart disease. Although genome-wide association analysis and gene knockout experiments have found some genes related to coronary heart disease, there are still a large number of genes potentially related to coronary heart disease that have not been discovered. If it is confirmed by biological experimental means, the time and money cost is too high. Therefore, it is urgent to identify genes related to coronary heart disease on a large scale by computational means, so as to conduct targeted biological experimental verification. This paper proposes a deep learning method based on biological networks for the identification of coronary heart disease-related genes. We constructed gene interaction networks and extracted gene expression levels in different tissues as features. Through the association information and expression characteristics between genes, we constructed a model of coronary heart disease-related genes. Through cross-validation, we found that our proposed GCN-GENE that has AUC as 0.75 and AUPR as 0.78, which is more accurate than other methods and is a reliable method for predicting coronary heart disease-related genes.

Structure enhanced deep clustering network via a weighted neighbourhood auto-encoder

Structural deep clustering involves the use of neural networks for fusing semantic and structural representations for clustering tasks, and it has been receiving increasing attention. In some pioneering works, auto-encoder (AE)-specific representations were integrated with a graph convolutional network (GCN)-specific representation by delivering semantic information to the GCN module layer-by-layer. Although promising performance has been achieved in various applications, we observed that a vital aspect was overlooked in these works: the structural information may vanish in the learning process because of the over-smoothing problem of the GCN module, leading to non-representative features and, thus, deteriorating clustering performance. In this study, we address this issue by proposing a structure enhanced deep clustering network. The GCN-specific structural data representation is enhanced and supervised by its structural information. Specifically, the GCN-specific structural data representation is strengthened during the learning process by combining it with a structure enhanced semantic (SES) representation. A novel structure enhanced AE, named the weighted neighbourhood AE (wNAE), is employed to learn the SES representation for each data sample. Finally, we design a joint supervision strategy to uniformly guide the simultaneous learning of the wNAE and GCN modules and the clustering assignment. Experimental results for different datasets empirically validate the importance of semantic and neighbour-wise structure learning.

MTAGCN: predicting miRNA-target associations in Camellia sinensis var. assamica through graph convolution neural network

Background

MircoRNAs (miRNAs) play a central role in diverse biological processes of Camellia sinensis var.assamica (CSA) through their associations with target mRNAs, including CSA growth, development and stress response. However, although the experiment methods of CSA miRNA-target identifications are costly and time-consuming, few computational methods have been developed to tackle the CSA miRNA-target association prediction problem.

Results

In this paper, we constructed a heterogeneous network for CSA miRNA and targets by integrating rich biological information, including a miRNA similarity network, a target similarity network, and a miRNA-target association network. We then proposed a deep learning framework of graph convolution networks with layer attention mechanism, named MTAGCN. In particular, MTAGCN uses the attention mechanism to combine embeddings of multiple graph convolution layers, employing the integrated embedding to score the unobserved CSA miRNA-target associations.

Discussion

Comprehensive experiment results on two tasks (balanced task and unbalanced task) demonstrated that our proposed model achieved better performance than the classic machine learning and existing graph convolution network-based methods. The analysis of these results could offer valuable information for understanding complex CSA miRNA-target association mechanisms and would make a contribution to precision plant breeding.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-022-04819-3.

DeepPN: a deep parallel neural network based on convolutional neural network and graph convolutional network for predicting RNA-protein binding sites

Background

Addressing the laborious nature of traditional biological experiments by using an efficient computational approach to analyze RNA-binding proteins (RBPs) binding sites has always been a challenging task. RBPs play a vital role in post-transcriptional control. Identification of RBPs binding sites is a key step for the anatomy of the essential mechanism of gene regulation by controlling splicing, stability, localization and translation. Traditional methods for detecting RBPs binding sites are time-consuming and computationally-intensive. Recently, the computational method has been incorporated in researches of RBPs. Nevertheless, lots of them not only rely on the sequence data of RNA but also need additional data, for example the secondary structural data of RNA, to improve the performance of prediction, which needs the pre-work to prepare the learnable representation of structural data.

Results

To reduce the dependency of those pre-work, in this paper, we introduce DeepPN, a deep parallel neural network that is constructed with a convolutional neural network (CNN) and graph convolutional network (GCN) for detecting RBPs binding sites. It includes a two-layer CNN and GCN in parallel to extract the hidden features, followed by a fully connected layer to make the prediction. DeepPN discriminates the RBP binding sites on learnable representation of RNA sequences, which only uses the sequence data without using other data, for example the secondary or tertiary structure data of RNA. DeepPN is evaluated on 24 datasets of RBPs binding sites with other state-of-the-art methods. The results show that the performance of DeepPN is comparable to the published methods.

Conclusion

The experimental results show that DeepPN can effectively capture potential hidden features in RBPs and use these features for effective prediction of binding sites.

Forecasting PM2.5 using hybrid graph convolution-based model considering dynamic wind-field to offer the benefit of spatial interpretability

Air pollution is a complex process and is affected by meteorological conditions and other chemical components. Numerous studies have demonstrated that data-driven spatio-temporal prediction models of PM2.5 concentration are comparable with the model-driven model. However, data-driven models are usually depending on the statistical correlation between PM2.5 and other factors and have challenges in dealing with causality in complex systems. In this paper, we argue that domain knowledge should be incorporated into data-driven models to enhance prediction accuracy and make the model more physically realistic. We focus on the influence of dynamic wind-field on PM2.5 concentration distribution and fuse the pollution diffusion distance with the deep learning model based on a wind-field surface. In order to model spatial dependence between monitoring stations, which is dynamic and anisotropic because of the wind-field, we proposed a hybrid deep learning framework, dynamic directed spatio-temporal graph convolution networks (DD-STGCN). It expanded the ability to deal with space-time prediction in the continuous and dynamic wind-field. We used a directed graph time-series to describe the vertex state and topological relationship between vertices and replaced traditional Euclidean distance with wind-field diffusion distance to describe the proximity relationship between vertices. Our experiment results demonstrated that the DD-STGCN model achieved a better prediction ability than LSTM, GC-LSTM, and STGCN models. Compared to the best comparison model, MAPE, MAE, and RMSE were improved by 10.2%, 9.7%, and 9.6% in 12 h on an average, respectively. The performance of our model was further tested during a haze period. In the case that two models both considered the effect of wind, compared with the pure data-driven model, our model performed better in prediction distribution and showed the benefit of spatial interpretability provided by domain knowledge.

Graph convolution network with similarity awareness and adaptive calibration for disease-induced deterioration prediction

Graph convolution networks (GCN) have been successfully applied in disease prediction tasks as they capture interactions (i.e., edges and edge weights on the graph) between individual elements. The interactions in existing works are constructed by fusing similarity between imaging information and distance between non-imaging information, whereas disregarding the disease status of those individuals in the training set. Besides, the similarity is being evaluated by computing the correlation distance between feature vectors, which limits prediction performance, especially for predicting significant memory concern (SMC) and mild cognitive impairment (MCI). In this paper, we propose three mechanisms to improve GCN, namely similarity-aware adaptive calibrated GCN (SAC-GCN), for predicting SMC and MCI. First, we design a similarity-aware graph using different receptive fields to consider disease status. The labelled subjects on the graph are only connected with those labelled subjects with the same status. Second, we propose an adaptive mechanism to evaluate similarity. Specifically, we construct initial GCN with evaluating similarity by using traditional correlation distance, then pre-train the initial GCN by using training samples and use it to score all subjects. Then, the difference between these scores replaces correlation distance to update similarity. Last, we devise a calibration mechanism to fuse functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI) information into edges. The proposed method is tested on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Experimental results demonstrate that our proposed method is useful to predict disease-induced deterioration and superior to other related algorithms, with a mean classification accuracy of 86.83% in our prediction tasks.

ScGSLC: An unsupervised graph similarity learning framework for single-cell RNA-seq data clustering

Accurate clustering of cells from single-cell RNA sequencing (scRNA-seq) data is an essential step for biological analysis such as putative cell type identification. However, scRNA-seq data has high dimension and high sparsity, which makes traditional clustering methods less effective to reflect the similarity between cells. Since genetic network fundamentally defines the functions of cell and deep learning shows strong advantages in network representation learning, we propose a novel scRNA-seq clustering framework ScGSLC based on graph similarity learning. ScGSLC effectively integrates scRNA-seq data and protein-protein interaction network to a graph. Then graph convolution network is employed by ScGSLC to embedding graph and clustering the cells by the calculated similarity between graphs. Unsupervised clustering results of nine public data sets demonstrate that ScGSLC shows better performance than the state-of-the-art methods.

Adversarial brain multiplex prediction from a single brain network with application to gender fingerprinting

Brain connectivity networks, derived from magnetic resonance imaging (MRI), non-invasively quantify the relationship in function, structure, and morphology between two brain regions of interest (ROIs) and give insights into gender-related connectional differences. However, to the best of our knowledge, studies on gender differences in brain connectivity were limited to investigating pairwise (i.e., low-order) relationships across ROIs, overlooking the complex high-order interconnectedness of the brain as a network. A few recent works on neurological disorders addressed this limitation by introducing the brain multiplex which is composed of a source network intra-layer, a target intra-layer, and a convolutional interlayer capturing the high-level relationship between both intra-layers. However, brain multiplexes are built from at least two different brain networks hindering their application to connectomic datasets with single brain networks (e.g., functional networks). To fill this gap, we propose Adversarial Brain Multiplex Translator (ABMT), the first work for predicting brain multiplexes from a source network using geometric adversarial learning to investigate gender differences in the human brain. Our framework comprises: (i) a geometric source to target network translator mimicking a U-Net architecture with skip connections, (ii) a conditional discriminator which distinguishes between predicted and ground truth target intra-layers, and finally (iii) a multi-layer perceptron (MLP) classifier which supervises the prediction of the target multiplex using the subject class label (e.g., gender). Our experiments on a large dataset demonstrated that predicted multiplexes significantly boost gender classification accuracy compared with source networks and unprecedentedly identify both low and high-order gender-specific brain multiplex connections. Our ABMT source code is available on GitHub at https://github.com/basiralab/ABMT.