Final Representation Research Articles

Dual-Attention Multiple Instance Learning Framework for Pathology Whole-Slide Image Classification

Conventional methods for tumor diagnosis suffer from two inherent limitations: they are time-consuming and subjective. Computer-aided diagnosis (CAD) is an important approach for addressing these limitations. Pathology whole-slide images (WSIs) are high-resolution tissue images that have made significant contributions to cancer diagnosis and prognosis assessment. Due to the complexity of WSIs and the availability of only slide-level labels, multiple instance learning (MIL) has become the primary framework for WSI classification. However, most MIL methods fail to capture the interdependence among image patches within a WSI, which is crucial for accurate classification prediction. Moreover, due to the weak supervision of slide-level labels, overfitting may occur during the training process. To address these issues, this paper proposes a dual-attention-based multiple instance learning framework (DAMIL). DAMIL leverages the spatial relationships and channel information between WSI patches for classification prediction, without detailed pixel-level tumor annotations. The output of the model preserves the semantic variations in the latent space, enhances semantic disturbance invariance, and provides reliable class identification for the final slide-level representation. We validate the effectiveness of DAMIL on the most commonly used public dataset, Camelyon16. The results demonstrate that DAMIL outperforms the state-of-the-art methods in terms of classification accuracy (ACC), area under the curve (AUC), and F1-Score. Our model also allows for the examination of its interpretability by visualizing the dual-attention weights. To the best of our knowledge, this is the first attempt to use a dual-attention mechanism, considering both spatial and channel information, for whole-slide image classification.

VLFATRollout: Fully transformer-based classifier for retinal OCT volumes

Background and Objective:Despite the promising capabilities of 3D transformer architectures in video analysis, their application to high-resolution 3D medical volumes encounters several challenges. One major limitation is the high number of 3D patches, which reduces the efficiency of the global self-attention mechanisms of transformers. Additionally, background information can distract vision transformers from focusing on crucial areas of the input image, thereby introducing noise into the final representation. Moreover, the variability in the number of slices per volume complicates the development of models capable of processing input volumes of any resolution while simple solutions like subsampling may risk losing essential diagnostic details. Methods:To address these challenges, we introduce an end-to-end transformer-based framework, variable length feature aggregator transformer rollout (VLFATRollout), to classify volumetric data. The proposed VLFATRollout enjoys several merits. First, the proposed VLFATRollout can effectively mine slice-level fore-background information with the help of transformer’s attention matrices. Second, randomization of volume-wise resolution (i.e. the number of slices) during training enhances the learning capacity of the learnable positional embedding (PE) assigned to each volume slice. This technique allows the PEs to generalize across neighboring slices, facilitating the handling of high-resolution volumes at the test time. Results:VLFATRollout was thoroughly tested on the retinal optical coherence tomography (OCT) volume classification task, demonstrating a notable average improvement of 5.47% in balanced accuracy over the leading convolutional models for a 5-class diagnostic task. These results emphasize the effectiveness of our framework in enhancing slice-level representation and its adaptability across different volume resolutions, paving the way for advanced transformer applications in medical image analysis. The code is available at https://github.com/marziehoghbaie/VLFATRollout/.

LDAGM: prediction lncRNA-disease asociations by graph convolutional auto-encoder and multilayer perceptron based on multi-view heterogeneous networks

BackgroundLong non-coding RNAs (lncRNAs) can prevent, diagnose, and treat a variety of complex human diseases, and it is crucial to establish a method to efficiently predict lncRNA-disease associations.ResultsIn this paper, we propose a prediction method for the lncRNA-disease association relationship, named LDAGM, which is based on the Graph Convolutional Autoencoder and Multilayer Perceptron model. The method first extracts the functional similarity and Gaussian interaction profile kernel similarity of lncRNAs and miRNAs, as well as the semantic similarity and Gaussian interaction profile kernel similarity of diseases. It then constructs six homogeneous networks and deeply fuses them using a deep topology feature extraction method. The fused networks facilitate feature complementation and deep mining of the original association relationships, capturing the deep connections between nodes. Next, by combining the obtained deep topological features with the similarity network of lncRNA, disease, and miRNA interactions, we construct a multi-view heterogeneous network model. The Graph Convolutional Autoencoder is employed for nonlinear feature extraction. Finally, the extracted nonlinear features are combined with the deep topological features of the multi-view heterogeneous network to obtain the final feature representation of the lncRNA-disease pair. Prediction of the lncRNA-disease association relationship is performed using the Multilayer Perceptron model. To enhance the performance and stability of the Multilayer Perceptron model, we introduce a hidden layer called the aggregation layer in the Multilayer Perceptron model. Through a gate mechanism, it controls the flow of information between each hidden layer in the Multilayer Perceptron model, aiming to achieve optimal feature extraction from each hidden layer.ConclusionsParameter analysis, ablation studies, and comparison experiments verified the effectiveness of this method, and case studies verified the accuracy of this method in predicting lncRNA-disease association relationships.

Improving Text Classification in Agricultural Expert Systems with a Bidirectional Encoder Recurrent Convolutional Neural Network

With the rapid development of internet and AI technologies, Agricultural Expert Systems (AESs) have become crucial for delivering technical support and decision-making in agricultural management. However, traditional natural language processing methods often struggle with specialized terminology and context, and they lack the adaptability to handle complex text classifications. The diversity and evolving nature of agricultural texts make deep semantic understanding and integration of contextual knowledge especially challenging. To tackle these challenges, this paper introduces a Bidirectional Encoder Recurrent Convolutional Neural Network (AES-BERCNN) tailored for short-text classification in agricultural expert systems. We designed an Agricultural Text Encoder (ATE) with a six-layer transformer architecture to capture both preceding and following word information. A recursive convolutional neural network based on Gated Recurrent Units (GRUs) was also developed to merge contextual information and learn complex semantic features, which are then combined with the ATE output and refined through max-pooling to form the final feature representation. The AES-BERCNN model was tested on a self-constructed agricultural dataset, achieving an accuracy of 99.63% in text classification. Its generalization ability was further verified on the Tsinghua News dataset. Compared to other models such as TextCNN, DPCNN, BiLSTM, and BERT-based models, the AES-BERCNN shows clear advantages in agricultural text classification. This work provides precise and timely technical support for intelligent agricultural expert systems.

Diversifying Collaborative Filtering via Graph Spreading Network and Selective Sampling.

Graph neural network (GNN) is a robust model for processing non-Euclidean data, such as graphs, by extracting structural information and learning high-level representations. GNN has achieved state-of-the-art recommendation performance on collaborative filtering (CF) for accuracy. Nevertheless, the diversity of the recommendations has not received good attention. Existing work using GNN for recommendation suffers from the accuracy-diversity dilemma, where slightly increases diversity while accuracy drops significantly. Furthermore, GNN-based recommendation models lack the flexibility to adapt to different scenarios' demands concerning the accuracy-diversity ratio of their recommendation lists. In this work, we endeavor to address the above problems from the perspective of aggregate diversity, which modifies the propagation rule and develops a new sampling strategy. We propose graph spreading network (GSN), a novel model that leverages only neighborhood aggregation for CF. Specifically, GSN learns user and item embeddings by propagating them over the graph structure, utilizing both diversity-oriented and accuracy-oriented aggregations. The final representations are obtained by taking the weighted sum of the embeddings learned at all layers. We also present a new sampling strategy that selects potentially accurate and diverse items as negative samples to assist model training. GSN effectively addresses the accuracy-diversity dilemma and achieves improved diversity while maintaining accuracy with the help of a selective sampler. Moreover, a hyper-parameter in GSN allows for adjustment of the accuracy-diversity ratio of recommendation lists to satisfy the diverse demands. Compared to the state-of-the-art model, GSN improved R @20 by 1.62%, N @20 by 0.67%, G @20 by 3.59%, and E @20 by 4.15% on average over three real-world datasets, verifying the effectiveness of our proposed model in diversifying overall collaborative recommendations.

Regulation-aware graph learning for drug repositioning over heterogeneous biological network

Drug repositioning (DR) is crucial for identifying new disease indications for existing drugs and enhancing their clinical utility. Despite the effectiveness of various artificial intelligence techniques in discovering novel drug-disease associations (DDAs), many algorithms primarily focus on incorporating biological knowledge of drugs and diseases into DDA networks, often overlooking the rich connectivity patterns inherent in heterogeneous biological networks. In this study, we leveraged diverse connectivity patterns to gain new insights into the regulatory mechanisms of drugs acting on target proteins in diseases. We defined a set of meta-paths to reveal different regulatory mechanisms, each corresponding to distinct connectivity patterns. For each meta-path, we constructed a regulation graph through random-walk sampling of its instances in the network and obtained drug and disease embeddings through regulation-aware graph representation learning. Subsequently, we proposed a novel multi-view attention mechanism to enhance drug and disease representations. The task of predicting DDAs was accomplished using the XGBoost classifier based on the final representations of drugs and diseases. The experimental results demonstrated the superior performance of our method, RGLDR, on three benchmark datasets under ten-fold cross-validation, outperforming state-of-the-art DR algorithms across several evaluation metrics. Furthermore, case studies on two diseases indicated that RGLDR is a promising DR tool that leverages meaningful connectivity patterns for improved efficacy.

An end-to-end image-text matching approach considering semantic uncertainty

We propose a novel end-to-end image-text matching approach considering semantic uncertainty (SU-ITM), aiming to deal with the one-to-many semantic diversity involved in image-text matching in order to capture the associations between them more comprehensively and improve the robustness of the model. Traditional methods map images and texts as definite points in an embedding space to measure cross-modal similarity. However, the point-based embedding cannot capture the semantic uncertainty, leading to a large bias in the matching results. To address this problem, we model the one-to-many associations between image and text in a way that establishes a probability distribution, incorporating the uncertainty information into the final semantic representation of the text. In addition, we optimize the image-text matching loss so that the different text features approximate the image features in a distributed manner while maintaining the discriminative nature of the semantic representation, effectively reducing the matching uncertainty. Notably, our method achieves end-to-end training by not using pre-trained target detection branches throughout the training process. We fully demonstrate the excellent performance of our method in the image-text matching task through experimental validation on Flickr30k and MSCOCO. Excellent performance levels of 546.1 and 545.0 are achieved on the R@SUM metric for Flickr30k and MSCOCO 1k, respectively.

A multi-task framework based on decomposition for multimodal named entity recognition

Given a text-image pair, Multimodal Named Entity Recognition (MNER) is the task of identifying and categorizing entities in the text. Most existing work performs named entity labeling directly using final token representations derived by fusing image and text representations. Although they achieve promising results, these work may fail to effectively exploit text and image modalities. This is because they neglect the difference in the role of the two modalities: text modality can detect the boundary of an entity, while image modality is introduced to disambiguate the category of the entity. Based on these findings, in this paper, we construct two auxiliary tasks based on the decomposition strategy and propose a multi-task framework for MNER. Specifically, we first decompose MNER into two auxiliary tasks: entity boundary detection task and entity category classification task. Here, the former treats only the text modality as input and outputs the boundary labels, since it can achieve satisfactory boundary results by itself. The latter uses two modalities to yield category labels where image modality is dedicated to disambiguating categories. These two auxiliary tasks allow the effective exploitation of text and image modalities and put them back into their respective roles. Then, we vectorize their results to improve entity recognition using label clues from auxiliary tasks. Finally, we fuse features from text and image modalities and label embeddings from auxiliary tasks to fulfill MNER. Experimental results on two widely used MNER datasets show that our framework can yield new SOTA performance.

STDNet: Rethinking Disentanglement Learning With Information Theory.

Disentangled representation learning is typically achieved by a generative model, variational encoder (VAE). Existing VAE-based methods try to disentangle all the attributes simultaneously in a single hidden space, while the separation of the attribute from irrelevant information varies in complexity. Thus, it should be conducted in different hidden spaces. Therefore, we propose to disentangle the disentanglement itself by assigning the disentanglement of each attribute to different layers. To achieve this, we present a stair disentanglement net (STDNet), a stair-like structure network with each step corresponding to the disentanglement of an attribute. An information separation principle is employed to peel off the irrelevant information to form a compact representation of the targeted attribute within each step. Compact representations, thus, obtained together form the final disentangled representation. To ensure the final disentangled representation is compressed as well as complete with respect to the input data, we propose a variant of the information bottleneck (IB) principle, the stair IB (SIB) principle, to optimize a tradeoff between compression and expressiveness. In particular, for the assignment to the network steps, we define an attribute complexity metric to assign the attributes by the complexity ascending rule (CAR) that dictates a sequencing of the attribute disentanglement in ascending order of complexity. Experimentally, STDNet achieves state-of-the-art results in representation learning and image generation on multiple benchmarks, including Mixed National Institute of Standards and Technology database (MNIST), dSprites, and CelebA. Furthermore, we conduct thorough ablation experiments to show how the strategies employed here contribute to the performance, including neurons block, CAR, hierarchical structure, and variational form of SIB.

Online trajectory anomaly detection model based on graph neural networks and variational autoencoder

Abstract To efficiently determine whether an entire trajectory exhibits abnormal behavior, we introduce an online trajectory anomaly detection model known as GeoGNFTOD, which employs graph neural networks for road segment representation, creating a directed graph by mapping trajectories onto the road network. The graph representation is constructed based on the road segments in this directed graph. By utilizing Transformer sequence encoding, the trajectory representation is derived and hierarchical geographic encoding captures the GPS mapping of the original trajectories. Merging these two representations produces the final trajectory representation, serving as input for an LSTM-based variational autoencoder to reconstruct the trajectory representation, facilitating rapid online anomaly detection. Experimental findings on a large-scale taxi trajectory dataset illustrate the superior performance of GeoGNFTOD compared to baseline algorithms.

HGBER: Heterogeneous Graph Neural Network With Bidirectional Encoding Representation.

Heterogeneous graphs with multiple types of nodes and link relationships are ubiquitous in many real-world applications. Heterogeneous graph neural networks (HGNNs) as an efficient technique have shown superior capacity of dealing with heterogeneous graphs. Existing HGNNs usually define multiple meta-paths in a heterogeneous graph to capture the composite relations and guide neighbor selection. However, these models only consider the simple relationships (i.e., concatenation or linear superposition) between different meta-paths, ignoring more general or complex relationships. In this article, we propose a novel unsupervised framework termed Heterogeneous Graph neural network with bidirectional encoding representation (HGBER) to learn comprehensive node representations. Specifically, the contrastive forward encoding is firstly performed to extract node representations on a set of meta-specific graphs corresponding to meta-paths. We then introduce the reversed encoding for the degradation process from the final node representations to each single meta-specific node representations. Moreover, to learn structure-preserving node representations, we further utilize a self-training module to discover the optimal node distribution through iterative optimization. Extensive experiments on five open public datasets show that the proposed HGBER model outperforms the state-of-the-art HGNNs baselines by 0.8%-8.4% in terms of accuracy on most datasets in various downstream tasks.

GraphADT: empowering interpretable predictions of acute dermal toxicity with multi-view graph pooling and structure remapping.

Accurate prediction of acute dermal toxicity (ADT) is essential for the safe and effective development of contact drugs. Currently, graph neural networks, a form of deep learning technology, accurately model the structure of compound molecules, enhancing predictions of their ADT. However, many existing methods emphasize atom-level information transfer and overlook crucial data conveyed by molecular bonds and their interrelationships. Additionally, these methods often generate "equal" node representations across the entire graph, failing to accentuate "important" substructures like functional groups, pharmacophores, and toxicophores, thereby reducing interpretability. We introduce a novel model, GraphADT, utilizing structure remapping and multi-view graph pooling (MVPool) technologies to accurately predict compound ADT. Initially, our model applies structure remapping to better delineate bonds, transforming "bonds" into new nodes and "bond-atom-bond" interactions into new edges, thereby reconstructing the compound molecular graph. Subsequently, we use MVPool to amalgamate data from various perspectives, minimizing biases inherent to single-view analyses. Following this, the model generates a robust node ranking collaboratively, emphasizing critical nodes or substructures to enhance model interpretability. Lastly, we apply a graph comparison learning strategy to train both the original and structure remapped molecular graphs, deriving the final molecular representation. Experimental results on public datasets indicate that the GraphADT model outperforms existing state-of-the-art models. The GraphADT model has been demonstrated to effectively predict compound ADT, offering potential guidance for the development of contact drugs and related treatments. Our code and data are accessible at: https://github.com/mxqmxqmxq/GraphADT.git.

Multi-view Heterogeneous Graph Neural Networks for Node Classification

Recently, with graph neural networks (GNNs) becoming a powerful technique for graph representation, many excellent GNN-based models have been proposed for processing heterogeneous graphs, which are termed Heterogeneous graph neural networks (HGNNs). However, existing HGNNs tend to aggregate information from either direct neighbors or those connected by short metapaths, thereby neglecting the higher-order information and global feature similarity information in heterogeneous graphs. In this paper, we propose a Multi-View Heterogeneous graph neural network (MV-HGNN) to aggregate these information. Firstly, two auxiliary views, specifically a global feature similarity view and a graph diffusion view, are generated from the original heterogeneous graph. Secondly, MV-HGNN performs two message-passing strategies to get the representation of different views. Subsequently, a transformer-based aggregator is used to get the semantic information. Subsequently, the representations of the three views are fused into a final composite representation. We evaluate our method on the node classification task over three commonly used heterogeneous graph datasets, and the results demonstrate that our proposed MV-HGNN significantly outperforms state-of-the-art baselines.

Prior semantic-embedding representation learning for on-the-fly FG-SBIR

In fine-grained sketch-based image retrieval (FG-SBIR) framework, sketch drawing is a time-consuming and skill-intensive process that greatly restricts its practical application. Our study aims to enhance the early-stage retrieval efficiency by utilizing partial sketches with a minimum number of strokes in the on-the-fly FG-SBIR framework. Nevertheless, partial sketches with few strokes usually lack global information and consist only of sparse local details. This lack of detail results in ambiguous semantic content and indistinct differences among images, which consequently impairs early-stage retrieval. To acquire an efficient representation for these sketches, we propose a multi-granularity feature representation model that utilizes prior knowledge embedding to deal with the sketch drawing process. In the first stage,we design a triplet network to learn the joint a coarse representation embedding space that is common between the complete sketch and the image. In the second stage, We enhanced the semantic representation of sketches by utilizing prior knowledge, which aids in supplementing missing fine-grained features. This was achieved by formalizing the prior knowledge and integrating it into the feature vector obtained in the initial stage, thereby acquiring the final representation of a subset of sketches. Experiments conducted on two public datasets indicate that our method outperforms the state-of-the-art baseline during the early retrieval phase. In practical applications, our approach has also achieved good results.

Biological direct-shortcut deep residual learning for sparse visual processing

We show, based on the following three grounds, that the primary visual cortex (V1) is a biological direct-shortcut deep residual learning neural network (ResNet) for sparse visual processing: (1) We first highlight that Gabor-like sets of basis functions, which are similar to the receptive fields of simple cells in the primary visual cortex (V1), are excellent candidates for sparse representation of natural images; i.e., images from the natural world, affirming the brain to be optimized for this. (2) We then prove that the intra-layer synaptic weight matrices of this region can be reasonably first-order approximated by identity mappings, and are thus sparse themselves. (3) Finally, we point out that intra-layer weight matrices having identity mapping as their initial approximation, irrespective of this approximation being also a reasonable first-order one or not, resemble the building blocks of direct-shortcut digital ResNets, which completes the grounds. This biological ResNet interconnects the sparsity of the final representation of the image to that of its intra-layer weights. Further exploration of this ResNet, and understanding the joint effects of its architecture and learning rules, e.g. on its inductive bias, could lead to major advancements in the area of bio-inspired digital ResNets. One immediate line of research in this context, for instance, is to study the impact of forcing the direct-shortcuts to be good first-order approximations of each building block. For this, along with the ℓ1\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${{\\ell}}_{1}$$\\end{document}-minimization posed on the basis function coefficients the ResNet finally provides at its output, another parallel one could e.g. also be posed on the weights of its residual layers.

Fusing differentiable rendering and language–image contrastive learning for superior zero-shot point cloud classification

Zero-shot point cloud classification involves recognizing categories not encountered during training. Current models often exhibit reduced accuracy on unseen categories without 3D pre-training, emphasizing the need for improved precision and interoperability. We propose a novel approach integrating differentiable rendering with contrastive language–image pre-training. Initially, differentiable rendering autonomously learns representative viewpoints from the data, enabling the transformation of point clouds into multi-view images while preserving key visual information. This transformation facilitates optimized viewpoint selection during training, refining the final feature representation. Features are extracted from the multi-view images and integrated into a global multi-view feature using a cross-attention mechanism. On the textual side, a large language model (LLM) is provided with 3D heuristic prompts to generate 3D-specific text reflecting category-specific traits, from which textual features are derived. The LLM’s extensive pre-trained knowledge enables it to capture abstract notions and categorical features relevant to distinct point cloud categories. Visual and textual features are aligned in a unified embedding space, enabling zero-shot classification. Throughout training, the Structural Similarity Index (SSIM) is integrated into the loss function to encourage the model to discern more distinctive viewpoints, reduce redundancy in multi-view imagery, and enhance computational efficiency. Experimental results on the ModelNet10, ModelNet40, and ScanObjectNN datasets demonstrate classification accuracies of 75.68%, 66.42%, and 52.03%, respectively, surpassing prevailing methods in zero-shot point cloud classification accuracy.

Complex Relation Embedding for Scene Graph Generation.

Given an input image, scene graph generation (SGG) aims to generate comprehensive visual relationships between objects in the form of graphs. Recently, more attention to the design of complex networks and complicated strategies has been paid to the long tail issue caused by the imbalanced class distribution. However, most existing methods adopt the concatenated features of two objects in real space as the final relation representation for a given triplet. We mainly argue that such a simple concatenation may neglect the importance of complex interactions between objects, which results in the diversity of visual relations. In addition, the representation learning in real space is also inadequate to express this property. To alleviate these issues, we seamlessly incorporate Hermitian inner product into existing models to facilitate the generation of scene graphs by learning Relation Embedding in Complex space (CoRE). More specifically, we first introduce the concept of complex-valued representations for entities and then formulate the relation triplets with Hermitian inner product in complex space. Finally, we investigate the effect of utilizing only real component or both of Hermitian inner product on inferring more reasonable interaction between objects for scene graphs. Comprehensive experiments on two widely used benchmark datasets, Visual Genome (VG) and Open Image, demonstrate our effectiveness, superiority, and generalization on various metrics for biased or unbiased inference.

DCNeT: A disease comorbidity network-based temporal deep learning framework to predict cardiovascular risk in patients with mental disorders

Patients with mental disorders (MDs) are at higher subsequent risk of developing cardiovascular diseases (CVDs) than the general population. Early identification of cardiovascular risk in patients with MDs is beneficial for timely intervention and reducing disease burden. Recently, deep learning approaches have been increasingly applied in CVDs risk prediction. However, these methods have three major issues: 1) mostly relying on multiple types of clinical data, 2) not sufficiently mining and utilizing comorbidity patterns hidden in complex correlations among various diseases, and 3) not fully leveraging the time information, including the irregular intervals. To address these issues, we propose a disease comorbidity network-based temporal deep learning framework (DCNeT) to predict the subsequent CVDs risk for patients with MDs based on routinely collected administrative health data. Firstly, to identify the comorbidity patterns of MDs, we construct a disease comorbidity network (DCN) for MDs and apply graph embedding method to generate disease embeddings for each disease in the DCN. Then, a code attention mechanism is proposed to obtain the weight of each disease which is embedded into dense vectors based on the structure of DCN. We present a view attention mechanism to compute the attention weights of different types of features including disease embeddings, basic features, and disease indicators for generating the final representations of patients’ hospitalizations. Furthermore, to fully utilize the information on the irregular time intervals between hospitalizations, a time encoding module is designed, and the time-aware LSTM is adopted to model the irregular time intervals and capture the temporal patterns of patients’ hospitalizations. The experimental results show that DCNeT outperforms the state-of-the-art methods, with the area under the receiver operating characteristic curve of 0.7658, 0.8143, 0.8110, and 0.7839 on four datasets, respectively. The ablation experiments further demonstrate that each module of DCNeT, including the code attention, view attention, and time encoding module, contributes to its superior performance, with average improvements of 1.20 %, 1.65 %, and 1.13 % in accuracy, respectively. Our DCNeT could be utilized as an efficient framework for identifying high-risk groups of CVDs among patients with MDs that may benefit from screening and preventive strategies.

MASMDDI: multi-layer adaptive soft-mask graph neural network for drug-drug interaction prediction.

Accurately predicting Drug-Drug Interaction (DDI) is a critical and challenging aspect of the drug discovery process, particularly in preventing adverse reactions in patients undergoing combination therapy. However, current DDI prediction methods often overlook the interaction information between chemical substructures of drugs, focusing solely on the interaction information between drugs and failing to capture sufficient chemical substructure details. To address this limitation, we introduce a novel DDI prediction method: Multi-layer Adaptive Soft Mask Graph Neural Network (MASMDDI). Specifically, we first design a multi-layer adaptive soft mask graph neural network to extract substructures from molecular graphs. Second, we employ an attention mechanism to mine substructure feature information and update latent features. In this process, to optimize the final feature representation, we decompose drug-drug interactions into pairwise interaction correlations between the core substructures of each drug. Third, we use these features to predict the interaction probabilities of DDI tuples and evaluate the model using real-world datasets. Experimental results demonstrate that the proposed model outperforms state-of-the-art methods in DDI prediction. Furthermore, MASMDDI exhibits excellent performance in predicting DDIs of unknown drugs in two tasks that are more aligned with real-world scenarios. In particular, in the transductive scenario using the DrugBank dataset, the ACC and AUROC and AUPRC scores of MASMDDI are 0.9596, 0.9903, and 0.9894, which are 2% higher than the best performing baseline.

Adaptive propagation deep graph neural networks

Graph neural networks (GNNs) with adaptive propagation combinations represent a specialized deep learning paradigm, engineered to capture complex nodal interconnections within graph data. The primary challenge of this model lies in distilling and representing features extracted over varying nodal distances. This paper delves into an array of adaptive propagation strategies, with a focus on the influence of nodal distances and information aggregation on model efficacy. Our investigation identifies a critical performance drop in scenarios featuring overly brief propagation paths or an insufficient number of layers. Addressing this, we propose an innovative adaptive propagation technique in deep graph neural networks, named AP-DGNN, aimed at reconstructing high-order graph convolutional neural networks (GCNs). The AP-DGNN model assigns unique aggregation combination weights to each node and category, culminating in a final model representation through a process of weighted aggregation. Notably, these weights are capable of assimilating both subjective and objective information characteristics within the network. To substantiate our model’s effectiveness and scalability, we employed often-used benchmark datasets for experimental validation. A notable aspect of our AP-DGNN model is its minimal training parameter requirement and reduced computational demand. Furthermore, we demonstrate the model’s enhanced performance, which remains consistent across various hyperparameter configurations. This aspect was rigorously tested under diverse hyperparameter settings. Our findings contribute significantly to the evolution of graph neural networks, potentially revolutionizing their application across multiple domains. The research presented herein not only advances the understanding of GNNs but also paves the way for their robust application in varied scenarios. Codes are available at https://github.com/CW112/AP_DGNN.