SPANN: annotating single-cell resolution spatial transcriptome data with scRNA-seq data.

MOTIVATION: The rapid development of spatial transcriptome technologies has enabled researchers to acquire single-cell-level spatial data at an affordable price. However, computational analysis tools, such as annotation tools, tailored for these data are still lacking. Recently, many computational frameworks have emerged to integrate single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics datasets. While some frameworks can utilize well-annotated scRNA-seq data to annotate spatial expression patterns, they overlook critical aspects. First, existing tools do not explicitly consider cell type mapping when aligning the two modalities. Second, current frameworks lack the capability to detect novel cells, which remains a key interest for biologists. RESULTS: To address these problems, we propose an annotation method for spatial transcriptome data called SPANN. The main tasks of SPANN are to transfer cell-type labels from well-annotated scRNA-seq data to newly generated single-cell resolution spatial transcriptome data and discover novel cells from spatial data. The major innovations of SPANN come from two aspects: SPANN automatically detects novel cells from unseen cell types while maintaining high annotation accuracy over known cell types. SPANN finds a mapping between spatial transcriptome samples and RNA data prototypes and thus conducts cell-type-level alignment. Comprehensive experiments using datasets from various spatial platforms demonstrate SPANN's capabilities in annotating known cell types and discovering novel cell states within complex tissue contexts. AVAILABILITY: The source code of SPANN can be accessed at https://github.com/ddb-qiwang/SPANN-torch. CONTACT: dengmh@math.pku.edu.cn.

Likelihood-based feature representation learning combined with neighborhood information for predicting circRNA-miRNA associations.

Connections between circular RNAs (circRNAs) and microRNAs (miRNAs) assume a pivotal position in the onset, evolution, diagnosis and treatment of diseases and tumors. Selecting the most potential circRNA-related miRNAs and taking advantage of them as the biological markers or drug targets could be conducive to dealing with complex human diseases through preventive strategies, diagnostic procedures and therapeutic approaches. Compared to traditional biological experiments, leveraging computational models to integrate diverse biological data in order to infer potential associations proves to be a more efficient and cost-effective approach. This paper developed a model of Convolutional Autoencoder for CircRNA-MiRNA Associations (CA-CMA) prediction. Initially, this model merged the natural language characteristics of the circRNA and miRNA sequence with the features of circRNA-miRNA interactions. Subsequently, it utilized all circRNA-miRNA pairs to construct a molecular association network, which was then fine-tuned by labeled samples to optimize the network parameters. Finally, the prediction outcome is obtained by utilizing the deep neural networks classifier. This model innovatively combines the likelihood objective that preserves the neighborhood through optimization, to learn the continuous feature representation of words and preserve the spatial information of two-dimensional signals. During the process of 5-fold cross-validation, CA-CMA exhibited exceptional performance compared to numerous prior computational approaches, as evidenced by its mean area under the receiver operating characteristic curve of 0.9138 and a minimal SD of 0.0024. Furthermore, recent literature has confirmed the accuracy of 25 out of the top 30 circRNA-miRNA pairs identified with the highest CA-CMA scores during case studies. The results of these experiments highlight the robustness and versatility of our model.

Spatially contrastive variational autoencoder for deciphering tissue heterogeneity from spatially resolved transcriptomics.

Recent advances in spatially resolved transcriptomics (SRT) have brought ever-increasing opportunities to characterize expression landscape in the context of tissue spatiality. Nevertheless, there still exist multiple challenges to accurately detect spatial functional regions in tissue. Here, we present a novel contrastive learning framework, SPAtially Contrastive variational AutoEncoder (SpaCAE), which contrasts transcriptomic signals of each spot and its spatial neighbors to achieve fine-grained tissue structures detection. By employing a graph embedding variational autoencoder and incorporating a deep contrastive strategy, SpaCAE achieves a balance between spatial local information and global information of expression, enabling effective learning of representations with spatial constraints. Particularly, SpaCAE provides a graph deconvolutional decoder to address the smoothing effect of local spatial structure on expression's self-supervised learning, an aspect often overlooked by current graph neural networks. We demonstrated that SpaCAE could achieve effective performance on SRT data generated from multiple technologies for spatial domains identification and data denoising, making it a remarkable tool to obtain novel insights from SRT studies.

scAMAC: self-supervised clustering of scRNA-seq data based on adaptive multi-scale autoencoder.

Cluster assignment is vital to analyzing single-cell RNA sequencing (scRNA-seq) data to understand high-level biological processes. Deep learning-based clustering methods have recently been widely used in scRNA-seq data analysis. However, existing deep models often overlook the interconnections and interactions among network layers, leading to the loss of structural information within the network layers. Herein, we develop a new self-supervised clustering method based on an adaptive multi-scale autoencoder, called scAMAC. The self-supervised clustering network utilizes the Multi-Scale Attention mechanism to fuse the feature information from the encoder, hidden and decoder layers of the multi-scale autoencoder, which enables the exploration of cellular correlations within the same scale and captures deep features across different scales. The self-supervised clustering network calculates the membership matrix using the fused latent features and optimizes the clustering network based on the membership matrix. scAMAC employs an adaptive feedback mechanism to supervise the parameter updates of the multi-scale autoencoder, obtaining a more effective representation of cell features. scAMAC not only enables cell clustering but also performs data reconstruction through the decoding layer. Through extensive experiments, we demonstrate that scAMAC is superior to several advanced clustering and imputation methods in both data clustering and reconstruction. In addition, scAMAC is beneficial for downstream analysis, such as cell trajectory inference. Our scAMAC model codes are freely available at https://github.com/yancy2024/scAMAC.

Deeply integrating latent consistent representations in high-noise multi-omics data for cancer subtyping.

Cancer is a complex and high-mortality disease regulated by multiple factors. Accurate cancer subtyping is crucial for formulating personalized treatment plans and improving patient survival rates. The underlying mechanisms that drive cancer progression can be comprehensively understood by analyzing multi-omics data. However, the high noise levels in omics data often pose challenges in capturing consistent representations and adequately integrating their information. This paper proposed a novel variational autoencoder-based deep learning model, named Deeply Integrating Latent Consistent Representations (DILCR). Firstly, multiple independent variational autoencoders and contrastive loss functions were designed to separate noise from omics data and capture latent consistent representations. Subsequently, an Attention Deep Integration Network was proposed to integrate consistent representations across different omics levels effectively. Additionally, we introduced the Improved Deep Embedded Clustering algorithm to make integrated variable clustering friendly. The effectiveness of DILCR was evaluated using 10 typical cancer datasets from The Cancer Genome Atlas and compared with 14 state-of-the-art integration methods. The results demonstrated that DILCR effectively captures the consistent representations in omics data and outperforms other integration methods in cancer subtyping. In the Kidney Renal Clear Cell Carcinoma case study, cancer subtypes were identified by DILCR with significant biological significance and interpretability.

Effective multi-modal clustering method via skip aggregation network for parallel scRNA-seq and scATAC-seq data.

In recent years, there has been a growing trend in the realm of parallel clustering analysis for single-cell RNA-seq (scRNA) and single-cell Assay of Transposase Accessible Chromatin (scATAC) data. However, prevailing methods often treat these two data modalities as equals, neglecting the fact that the scRNA mode holds significantly richer information compared to the scATAC. This disregard hinders the model benefits from the insights derived from multiple modalities, compromising the overall clustering performance. To this end, we propose an effective multi-modal clustering model scEMC for parallel scRNA and Assay of Transposase Accessible Chromatin data. Concretely, we have devised a skip aggregation network to simultaneously learn global structural information among cells and integrate data from diverse modalities. To safeguard the quality of integrated cell representation against the influence stemming from sparse scATAC data, we connect the scRNA data with the aggregated representation via skip connection. Moreover, to effectively fit the real distribution of cells, we introduced a Zero Inflated Negative Binomial-based denoising autoencoder that accommodates corrupted data containing synthetic noise, concurrently integrating a joint optimization module that employs multiple losses. Extensive experiments serve to underscore the effectiveness of our model. This work contributes significantly to the ongoing exploration of cell subpopulations and tumor microenvironments, and the code of our work will be public at https://github.com/DayuHuu/scEMC.

FusionNW, a potential clinical impact assessment of kinases in pan-cancer fusion gene network.

Kinase fusion genes are the most active fusion gene group in human cancer fusion genes. To help choose the clinically significant kinase so that the cancer patients that have fusion genes can be better diagnosed, we need a metric to infer the assessment of kinases in pan-cancer fusion genes rather than relying on the sample frequency expressed fusion genes. Most of all, multiple studies assessed human kinases as the drug targets using multiple types of genomic and clinical information, but none used the kinase fusion genes in their study. The assessment studies of kinase without kinase fusion gene events can miss the effect of one of the mechanisms that enhance the kinase function in cancer. To fill this gap, in this study, we suggest a novel way of assessing genes using a network propagation approach to infer how likely individual kinases influence the kinase fusion gene network composed of ~5K kinase fusion gene pairs. To select a better seed of propagation, we chose the top genes via dimensionality reduction like a principal component or latent layer information of six features of individual genes in pan-cancer fusion genes. Our approach may provide a novel way to assess of human kinases in cancer.

Attention-guided variational graph autoencoders reveal heterogeneity in spatial transcriptomics.

The latest breakthroughs in spatially resolved transcriptomics technology offer comprehensive opportunities to delve into gene expression patterns within the tissue microenvironment. However, the precise identification of spatial domains within tissues remains challenging. In this study, we introduce AttentionVGAE (AVGN), which integrates slice images, spatial information and raw gene expression while calibrating low-quality gene expression. By combining the variational graph autoencoder with multi-head attention blocks (MHA blocks), AVGN captures spatial relationships in tissue gene expression, adaptively focusing on key features and alleviating the need for prior knowledge of cluster numbers, thereby achieving superior clustering performance. Particularly, AVGN attempts to balance the model's attention focus on local and global structures by utilizing MHA blocks, an aspect that current graph neural networks have not extensively addressed. Benchmark testing demonstrates its significant efficacy in elucidating tissue anatomy and interpreting tumor heterogeneity, indicating its potential in advancing spatial transcriptomics research and understanding complex biological phenomena.

PMiSLocMF: predicting miRNA subcellular localizations by incorporating multi-source features of miRNAs.

The microRNAs (miRNAs) play crucial roles in several biological processes. It is essential for a deeper insight into their functions and mechanisms by detecting their subcellular localizations. The traditional methods for determining miRNAs subcellular localizations are expensive. The computational methods are alternative ways to quickly predict miRNAs subcellular localizations. Although several computational methods have been proposed in this regard, the incomplete representations of miRNAs in these methods left the room for improvement. In this study, a novel computational method for predicting miRNA subcellular localizations, named PMiSLocMF, was developed. As lots of miRNAs have multiple subcellular localizations, this method was a multi-label classifier. Several properties of miRNA, such as miRNA sequences, miRNA functional similarity, miRNA-disease, miRNA-drug, and miRNA-mRNA associations were adopted for generating informative miRNA features. To this end, powerful algorithms [node2vec and graph attention auto-encoder (GATE)] and one newly designed scheme were adopted to process above properties, producing five feature types. All features were poured into self-attention and fully connected layers to make predictions. The cross-validation results indicated the high performance of PMiSLocMF with accuracy higher than 0.83, average area under the receiver operating characteristic curve (AUC) and area under the precision-recall curve (AUPR) exceeding 0.90 and 0.77, respectively. Such performance was better than all previous methods based on the same dataset. Further tests proved that using all feature types can improve the performance of PMiSLocMF, and GATE and self-attention layer can help enhance the performance. Finally, we deeply analyzed the influence of miRNA associations with diseases, drugs, and mRNAs on PMiSLocMF. The dataset and codes are available at https://github.com/Gu20201017/PMiSLocMF.

Predicting potential interactions between lncRNAs and proteins via combined graph auto-encoder methods.

Long noncoding RNA (lncRNA) is a kind of noncoding RNA with a length of more than 200 nucleotide units. Numerous research studies have proven that although lncRNAs cannot be directly translated into proteins, lncRNAs still play an important role in human growth processes by interacting with proteins. Since traditional biological experiments often require a lot of time and material costs to explore potential lncRNA-protein interactions (LPI), several computational models have been proposed for this task. In this study, we introduce a novel deep learning method known as combined graph auto-encoders (LPICGAE) to predict potential human LPIs. First, we apply a variational graph auto-encoder to learn the low dimensional representations from the high-dimensional features of lncRNAs and proteins. Then the graph auto-encoder is used to reconstruct the adjacency matrix for inferring potential interactions between lncRNAs and proteins. Finally, we minimize the loss of the two processes alternately to gain the final predicted interaction matrix. The result in 5-fold cross-validation experiments illustrates that our method achieves an average area under receiver operating characteristic curve of 0.974 and an average accuracy of 0.985, which is better than those of existing six state-of-the-art computational methods. We believe that LPICGAE can help researchers to gain more potential relationships between lncRNAs and proteins effectively.

Benchmark of embedding-based methods for accurate and transferable prediction of drug response.

Prediction of therapy response has been a major challenge in cancer precision medicine due to the extensive tumor heterogeneity. Recently, several deep learning methods have been developed to predict drug response by utilizing various omics data. Most of them train models by using the drug-response screening data generated from cell lines and then use these models to predict response in cancer patient data. In this study, we focus on and evaluate deep learning methods using transcriptome data for the long-standing question of personalized drug-response prediction. We developed an embedding-based approach for drug-response prediction and benchmarked similar methods for their performance. For all methods, we used pretreatment transcriptome data to train models and then conducted a comprehensive evaluation and comparison of the models using cross-panels, cross-datasets and target genes. We further validated the methods using three independent datasets assessing multiple compounds for their predictive capability of drug response, survival outcome and cell line status. As a result, the methods building on gene embeddings had an overall competitive performance with reduced overfitting when we applied evaluation parameters for model fitting as well as the correlation with clinical outcomes in the validation data. We further developed an ensemble model to combine the results from the three most competitive methods for an overall prediction. Finally, we developed DrVAEN (https://bioinfo.uth.edu/drvaen), a user-friendly and easy-accessible web-server that hosts all these methods for drug-response prediction and model comparison for broad use in cancer research, method evaluation and drug development.

Bubble: a fast single-cell RNA-seq imputation using an autoencoder constrained by bulk RNA-seq data.

Single-cell RNA-sequencing technology (scRNA-seq) brings research to single-cell resolution. However, a major drawback of scRNA-seq is large sparsity, i.e. expressed genes with no reads due to technical noise or limited sequence depth during the scRNA-seq protocol. This phenomenon is also called 'dropout' events, which likely affect downstream analyses such as differential expression analysis, the clustering and visualization of cell subpopulations, cellular trajectory inference, etc. Therefore, there is a need to develop a method to identify and impute these dropout events. We propose Bubble, which first identifies dropout events from all zeros based on expression rate and coefficient of variation of genes within cell subpopulation, and then leverages an autoencoder constrained by bulk RNA-seq data to only impute those values. Unlike other deep learning-based imputation methods, Bubble fuses the matched bulk RNA-seq data as a constraint to reduce the introduction of false positive signals. Using simulated and several real scRNA-seq datasets, we demonstrate that Bubble enhances the recovery of missing values, gene-to-gene and cell-to-cell correlations, and reduces the introduction of false positive signals. Regarding some crucial downstream analyses of scRNA-seq data, Bubble facilitates the identification of differentially expressed genes, improves the performance of clustering and visualization, and aids the construction of cellular trajectory. More importantly, Bubble provides fast and scalable imputation with minimal memory usage.

LncRNA-disease association identification using graph auto-encoder and learning to rank.

Discovering the relationships between long non-coding RNAs (lncRNAs) and diseases is significant in the treatment, diagnosis and prevention of diseases. However, current identified lncRNA-disease associations are not enough because of the expensive and heavy workload of wet laboratory experiments. Therefore, it is greatly important to develop an efficient computational method for predicting potential lncRNA-disease associations. Previous methods showed that combining the prediction results of the lncRNA-disease associations predicted by different classification methods via Learning to Rank (LTR) algorithm can be effective for predicting potential lncRNA-disease associations. However, when the classification results are incorrect, the ranking results will inevitably be affected. We propose the GraLTR-LDA predictor based on biological knowledge graphs and ranking framework for predicting potential lncRNA-disease associations. Firstly, homogeneous graph and heterogeneous graph are constructed by integrating multi-source biological information. Then, GraLTR-LDA integrates graph auto-encoder and attention mechanism to extract embedded features from the constructed graphs. Finally, GraLTR-LDA incorporates the embedded features into the LTR via feature crossing statistical strategies to predict priority order of diseases associated with query lncRNAs. Experimental results demonstrate that GraLTR-LDA outperforms the other state-of-the-art predictors and can effectively detect potential lncRNA-disease associations. Availability and implementation: Datasets and source codes are available at http://bliulab.net/GraLTR-LDA.

CapsNet-LDA: predicting lncRNA-disease associations using attention mechanism and capsule network based on multi-view data.

Cumulative studies have shown that many long non-coding RNAs (lncRNAs) are crucial in a number of diseases. Predicting potential lncRNA-disease associations (LDAs) can facilitate disease prevention, diagnosis and treatment. Therefore, it is vital to develop practical computational methods for LDA prediction. In this study, we propose a novel predictor named capsule network (CapsNet)-LDA for LDA prediction. CapsNet-LDA first uses a stacked autoencoder for acquiring the informative low-dimensional representations of the lncRNA-disease pairs under multiple views, then the attention mechanism is leveraged to implement an adaptive allocation of importance weights to them, and they are subsequently processed using a CapsNet-based architecture for predicting LDAs. Different from the conventional convolutional neural networks (CNNs) that have some restrictions with the usage of scalar neurons and pooling operations. the CapsNets use vector neurons instead of scalar neurons that have better robustness for the complex combination of features and they use dynamic routing processes for updating parameters. CapsNet-LDA is superior to other five state-of-the-art models on four benchmark datasets, four perturbed datasets and an independent test set in the comparison experiments, demonstrating that CapsNet-LDA has excellent performance and robustness against perturbation, as well as good generalization ability. The ablation studies verify the effectiveness of some modules of CapsNet-LDA. Moreover, the ability of multi-view data to improve performance is proven. Case studies further indicate that CapsNet-LDA can accurately predict novel LDAs for specific diseases.

Dimensionality reduction and visualization of single-cell RNA-seq data with an improved deep variational autoencoder.

Single-cell RNA sequencing (scRNA-seq) is a revolutionary breakthrough that determines the precise gene expressions on individual cells and deciphers cell heterogeneity and subpopulations. However, scRNA-seq data are much noisier than traditional high-throughput RNA-seq data because of technical limitations, leading to many scRNA-seq data studies about dimensionality reduction and visualization remaining at the basic data-stacking stage. In this study, we propose an improved variational autoencoder model (termed DREAM) for dimensionality reduction and a visual analysis of scRNA-seq data. Here, DREAM combines the variational autoencoder and Gaussian mixture model for cell type identification, meanwhile explicitly solving 'dropout' events by introducing the zero-inflated layer to obtain the low-dimensional representation that describes the changes in the original scRNA-seq dataset. Benchmarking comparisons across nine scRNA-seq datasets show that DREAM outperforms four state-of-the-art methods on average. Moreover, we prove that DREAM can accurately capture the expression dynamics of human preimplantation embryonic development. DREAM is implemented in Python, freely available via the GitHub website, https://github.com/Crystal-JJ/DREAM.

rcCAE: a convolutional autoencoder method for detecting intra-tumor heterogeneity and single-cell copy number alterations.

Intra-tumor heterogeneity (ITH) is one of the major confounding factors that result in cancer relapse, and deciphering ITH is essential for personalized therapy. Single-cell DNA sequencing (scDNA-seq) now enables profiling of single-cell copy number alterations (CNAs) and thus aids in high-resolution inference of ITH. Here, we introduce an integrated framework called rcCAE to accurately infer cell subpopulations and single-cell CNAs from scDNA-seq data. A convolutional autoencoder (CAE) is employed in rcCAE to learn latent representation of the cells as well as distill copy number information from noisy read counts data. This unsupervised representation learning via the CAE model makes it convenient to accurately cluster cells over the low-dimensional latent space, and detect single-cell CNAs from enhanced read counts data. Extensive performance evaluations on simulated datasets show that rcCAE outperforms the existing CNA calling methods, and is highly effective in inferring clonal architecture. Furthermore, evaluations of rcCAE on two real datasets demonstrate that it is able to provide a more refined clonal structure, of which some details are lost in clonal inference based on integer copy numbers.

Predicting potential microbe-disease associations based on multi-source features and deep learning.

Studies have confirmed that the occurrence of many complex diseases in the human body is closely related to the microbial community, and microbes can affect tumorigenesis and metastasis by regulating the tumor microenvironment. However, there are still large gaps in the clinical observation of the microbiota in disease. Although biological experiments are accurate in identifying disease-associated microbes, they are also time-consuming and expensive. The computational models for effective identification of diseases related microbes can shorten this process, and reduce capital and time costs. Based on this, in the paper, a model named DSAE_RF is presented to predict latent microbe-disease associations by combining multi-source features and deep learning. DSAE_RF calculates four similarities between microbes and diseases, which are then used as feature vectors for the disease-microbe pairs. Later, reliable negative samples are screened by k-means clustering, and a deep sparse autoencoder neural network is further used to extract effective features of the disease-microbe pairs. In this foundation, a random forest classifier is presented to predict the associations between microbes and diseases. To assess the performance of the model in this paper, 10-fold cross-validation is implemented on the same dataset. As a result, the AUC and AUPR of the model are 0.9448 and 0.9431, respectively. Furthermore, we also conduct a variety of experiments, including comparison of negative sample selection methods, comparison with different models and classifiers, Kolmogorov-Smirnov test and t-test, ablation experiments, robustness analysis, and case studies on Covid-19 and colorectal cancer. The results fully demonstrate the reliability and availability of our model.

scDCCA: deep contrastive clustering for single-cell RNA-seq data based on auto-encoder network.

The advances in single-cell ribonucleic acid sequencing (scRNA-seq) allow researchers to explore cellular heterogeneity and human diseases at cell resolution. Cell clustering is a prerequisite in scRNA-seq analysis since it can recognize cell identities. However, the high dimensionality, noises and significant sparsity of scRNA-seq data have made it a big challenge. Although many methods have emerged, they still fail to fully explore the intrinsic properties of cells and the relationship among cells, which seriously affects the downstream clustering performance. Here, we propose a new deep contrastive clustering algorithm called scDCCA. It integrates a denoising auto-encoder and a dual contrastive learning module into a deep clustering framework to extract valuable features and realize cell clustering. Specifically, to better characterize and learn data representations robustly, scDCCA utilizes a denoising Zero-Inflated Negative Binomial model-based auto-encoder to extract low-dimensional features. Meanwhile, scDCCA incorporates a dual contrastive learning module to capture the pairwise proximity of cells. By increasing the similarities between positive pairs and the differences between negative ones, the contrasts at both the instance and the cluster level help the model learn more discriminative features and achieve better cell segregation. Furthermore, scDCCA joins feature learning with clustering, which realizes representation learning and cell clustering in an end-to-end manner. Experimental results of 14 real datasets validate that scDCCA outperforms eight state-of-the-art methods in terms of accuracy, generalizability, scalability and efficiency. Cell visualization and biological analysis demonstrate that scDCCA significantly improves clustering and facilitates downstream analysis for scRNA-seq data. The code is available at https://github.com/WJ319/scDCCA.

A feature extraction method based on noise reduction for circRNA-miRNA interaction prediction combining multi-structure features in the association networks.

MOTIVATION: A large number of studies have shown that circular RNA (circRNA) affects biological processes by competitively binding miRNA, providing a new perspective for the diagnosis, and treatment of human diseases. Therefore, exploring the potential circRNA-miRNA interactions (CMIs) is an important and urgent task at present. Although some computational methods have been tried, their performance is limited by the incompleteness of feature extraction in sparse networks and the low computational efficiency of lengthy data. RESULTS: In this paper, we proposed JSNDCMI, which combines the multi-structure feature extraction framework and Denoising Autoencoder (DAE) to meet the challenge of CMI prediction in sparse networks. In detail, JSNDCMI integrates functional similarity and local topological structure similarity in the CMI network through the multi-structure feature extraction framework, then forces the neural network to learn the robust representation of features through DAE and finally uses the Gradient Boosting Decision Tree classifier to predict the potential CMIs. JSNDCMI produces the best performance in the 5-fold cross-validation of all data sets. In the case study, seven of the top 10 CMIs with the highest score were verified in PubMed. AVAILABILITY: The data and source code can be found at https://github.com/1axin/JSNDCMI.

Predicting metabolite-disease associations based on auto-encoder and non-negative matrix factorization.

Metabolism refers to a series of orderly chemical reactions used to maintain life activities in organisms. In healthy individuals, metabolism remains within a normal range. However, specific diseases can lead to abnormalities in the levels of certain metabolites, causing them to either increase or decrease. Detecting these deviations in metabolite levels can aid in diagnosing a disease. Traditional biological experiments often rely on a lot of manpower to do repeated experiments, which is time consuming and labor intensive. To address this issue, we develop a deep learning model based on the auto-encoder and non-negative matrix factorization named as MDA-AENMF to predict the potential associations between metabolites and diseases. We integrate a variety of similarity networks and then acquire the characteristics of both metabolites and diseases through three specific modules. First, we get the disease characteristics from the five-layer auto-encoder module. Later, in the non-negative matrix factorization module, we extract both the metabolite and disease characteristics. Furthermore, the graph attention auto-encoder module helps us obtain metabolite characteristics. After obtaining the features from three modules, these characteristics are merged into a single, comprehensive feature vector for each metabolite-disease pair. Finally, we send the corresponding feature vector and label to the multi-layer perceptron for training. The experiment demonstrates our area under the receiver operating characteristic curve of 0.975 and area under the precision-recall curve of 0.973 in 5-fold cross-validation, which are superior to those of existing state-of-the-art predictive methods. Through case studies, most of the new associations obtained by MDA-AENMF have been verified, further highlighting the reliability of MDA-AENMF in predicting the potential relationships between metabolites and diseases.