Front. Microbiol.Frontiers in MicrobiologyFront. Microbiol.1664-302XFrontiers Media S.A.10.3389/fmicb.2023.1244527MicrobiologyOriginal ResearchPredicting potential microbe-disease associations with graph attention autoencoder, positive-unlabeled learning, and deep neural networkPengLihong12†HuangLiangliang1†TianGeng3WuYan3LiGuang4567CaoJianying4567WangPeng8LiZejun8*DuanLian4567*1School of Computer Science, Hunan University of Technology, Zhuzhou, China2College of Life Sciences and Chemistry, Hunan University of Technology, Zhuzhou, China3Geneis (Beijing) Co. Ltd., Beijing, China4Faculty of Pediatrics, The Chinese PLA General Hospital, Beijing, China5Department of Pediatric Surgery, The Seventh Medical Center of PLA General Hospital, Beijing, China6National Engineering Laboratory for Birth Defects Prevention and Control of Key Technology, Beijing, China7Beijing Key Laboratory of Pediatric Organ Failure, Beijing, China8School of Computer Science, Hunan Institute of Technology, Hengyang, China
Edited by: Qi Zhao, University of Science and Technology Liaoning, China
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Background
Microbes have dense linkages with human diseases. Balanced microorganisms protect human body against physiological disorders while unbalanced ones may cause diseases. Thus, identification of potential associations between microbes and diseases can contribute to the diagnosis and therapy of various complex diseases. Biological experiments for microbe–disease association (MDA) prediction are expensive, time-consuming, and labor-intensive.
Methods
We developed a computational MDA prediction method called GPUDMDA by combining graph attention autoencoder, positive-unlabeled learning, and deep neural network. First, GPUDMDA computes disease similarity and microbe similarity matrices by integrating their functional similarity and Gaussian association profile kernel similarity, respectively. Next, it learns the feature representation of each microbe–disease pair using graph attention autoencoder based on the obtained disease similarity and microbe similarity matrices. Third, it selects a few reliable negative MDAs based on positive-unlabeled learning. Finally, it takes the learned MDA features and the selected negative MDAs as inputs and designed a deep neural network to predict potential MDAs.
Results
GPUDMDA was compared with four state-of-the-art MDA identification models (i.e., MNNMDA, GATMDA, LRLSHMDA, and NTSHMDA) on the HMDAD and Disbiome databases under five-fold cross validations on microbes, diseases, and microbe-disease pairs. Under the three five-fold cross validations, GPUDMDA computed the best AUCs of 0.7121, 0.9454, and 0.9501 on the HMDAD database and 0.8372, 0.8908, and 0.8948 on the Disbiome database, respectively, outperforming the other four MDA prediction methods. Asthma is the most common chronic respiratory condition and affects ~339 million people worldwide. Inflammatory bowel disease is a class of globally chronic intestinal disease widely existed in the gut and gastrointestinal tract and extraintestinal organs of patients. Particularly, inflammatory bowel disease severely affects the growth and development of children. We used the proposed GPUDMDA method and found that Enterobacter hormaechei had potential associations with both asthma and inflammatory bowel disease and need further biological experimental validation.
Conclusion
The proposed GPUDMDA demonstrated the powerful MDA prediction ability. We anticipate that GPUDMDA helps screen the therapeutic clues for microbe-related diseases.
microbe-disease associationsgraph attention autoencoderpositive-unlabeled learningK-meansXGBoostdeep neural network6180315162172158815003912023JJ50201National Natural Science Foundation of China10.13039/501100001809Natural Science Foundation of Hunan Province10.13039/501100004735section-at-acceptanceSystems Microbiology1. Introduction
Microorganisms or microbes exist in the form of single cell or a group of cells. Microbes mainly contain bacteria, archaea, fungi, viruses, and protozoa (Wen et al., 2021). They widely distribute on the human skin, oral cavity, respiratory tract, and gastrointestinal tract (Holmes et al., 2015). Most of human microbes are beneficial to human health. They can promote nutrient absorption, protect human body against pathogens, and strengthen metabolic capability. In addition, they have the similar metabolic ability to the liver and are even taken as “forgotten organ” of human body (Gill et al., 2006). However, their imbalance or dysbiosis could cause human diseases (Peng et al., 2022c; Tian et al., 2022), such as inflammatory bowel disease (IBD) (El Mouzan et al., 2018), diabetes (Wen et al., 2008), asthma (Demirci et al., 2019), liver diseases (Henao-Mejia et al., 2013), and cancer (Schwabe and Jobin, 2013). Although many evidence demonstrated that microbes have close relationships with human diseases, a comprehensive understanding about how microbes influence human healths and produce diseases remains unknown.
Microbe–disease association (MDA) identification not only help us to capture the mechanisms of complex diseases but also provide multiple possible biomarkers for their diagnosis and therapy. However, traditional wet lab remains costly, laborious, and time-consuming (Chen et al., 2019, 2020; Shen et al., 2022; Chen and Huang, 2023). With the advance of single cell sequencing (Peng et al., 2022d, 2023a,b; Wu et al., 2022; Hu et al., 2023a,b; Xu et al., 2023) and wide application of artificial intelligence (Chen et al., 2021; Lihong et al., 2022; Peng et al., 2022a; Wang et al., 2022, 2023; Zhang et al., 2022a,b; Zhang and Wu, 2023), many computational methods have been developed to discover potential MDAs. These methods mainly contain network-based algorithms and machine learning-based algorithms.
Network-based algorithms take MDA prediction as a random walk or label propagation problem. For example, to decode underlying MDAs, BRWMDA fused similarity networks and bi-random walk (Yan et al., 2019), NBLPIHMDA developed a bidirectional label propagation algorithm (Wang et al., 2019), MHEN constructed a multiplex heterogeneous network (Ma and Jiang, 2020), WMGHMDA implemented iteratively weighted meta-graph search model (Long and Luo, 2019), RWHMDA was a hypergraph-based random walk method (Niu et al., 2019), BDHNS formulated a bi-directional heterogeneous MDA network (Guan et al., 2022), and MNNMDA used low-rank matrix completion (Liu et al., 2023).
Machine learning-based algorithms take MDA prediction as a classification problem. For example, to discover potential MDAs, BPNNHMDA (Li et al., 2020) adopted a neural network structure, GATMDA (Long et al., 2021) exploited a graph attention network with inductive matrix completion, DMFMDA (Liu et al., 2020) utilized a deep neural network-based deep matrix factorization model, NinimHMDA (Ma and Jiang, 2020) explored an end-to-end graph convolutional neural network structure, KGNMDA (Jiang et al., 2022) used a graph neural network model, MGATMDA (Liu et al., 2021) comprised decomposer, combiner, and predictor where the decomposer captured the latent components using node-level attention mechanism, the combiner obtained unified embedding using component-level attention mechanism, and unknown microbe–disease pairs were classified by a fully connected network. HNGFL (Wang et al., 2022) designed an embedding algorithm for feature learning and used support vector machine for MDA classification.
Although computational methods significantly improved MDA prediction and uncovered many potential MDAs, there are some limitations presented in this study. For example, network-based MDA inference methods cannot find associated entities for a new microbe or disease. Machine learning-based inference methods need reliable negative MDAs for implementing the MDA classification task. Here, we developed an MDA prediction method called GPUDMDA by combining feature extraction based on graph attention autoencoder (GATE), reliable negative MDA selection based on positive-unlabeled (PU) learning, and MDA classification based on deep neural network (DNN).
2. Materials and methods2.1. Data preparation
We used two MDA databases to implement MDA prediction. One database is from the Human microbe–disease Association Database (HMDAD; http://www.cuilab.cn/hmdad) and contain 450 MDAs between 292 microbes and 39 diseases (Ma et al., 2017). The other comes from Disbiome (https://disbiome.ugent.be/home) (Janssens et al., 2018) and contains 4,351 MDAs between 218 diseases and 1,052 microbes. Moreover, an MDA network Y∈ℜn×m is constructed by Eq. (1) as follows:
yij={1, if microbe mi associates with disease dj0, otherwise 2.2. Pipeline for MDA prediction
In this manuscript, we developed an MDA prediction method called GPUDMDA by combining graph attention autoencoder, positive-unlabeled learning, and deep neural network. First, GPUDMDA computes disease similarity and microbe similarity matrices by integrating their functional similarity and Gaussian association profile kernel (GAPK) similarity, respectively. Next, it learns features of each microbe–disease pair using GATE. Third, it selects several reliable negative MDAs based on PU learning. Finally, it takes the extracted MDA features and the selected negative MDAs as inputs and proposes a DNN for discovering potential MDAs. Figure 1 shows the pipeline of GPUDMDA.
The pipeline of the MDA framework GPUDMDA.
2.3. Similarity computation2.3.1. Functional similarity of microbes
In the GATMDA, Long et al. (2021) computed microbe functional similarity according to their co-occurrences (Kamneva, 2017). Similarly, we use the microbe function similarity method in GATMDA and then compute a functional similarity matrix Smfun between m microbes, where Smfun(mi,mj) denotes the similarity between two microbes mi and mj.
2.3.2. Functional similarity of diseases
We use the disease functional similarity assessment method proposed by Long et al. (2021) and compute functional similarity matrix Sdfun between n diseases, where Sdfun(di,dj) denotes the similarity between two diseases di and dj.
2.3.3. Gaussian association profile kernel similarity
GAPK function is a symmetric function along the radial direction. It can better cluster similar examples with linearly separable form (Wang et al., 2020). Let Vdi (the ith row of Y) and Vdj (the jth row of Y) denote two diseases di and dj, respectively, their similarity can be computed by Eq. (2) as follows:
Gd(di,dj)=exp(−θd‖Vdi−Vdj‖2)
where
θd=1n∑i=1n‖Vdi‖2
Similarly, microbe GAPK similarity Gm is computed.
2.3.4. Similarity integration
Functional similarity is used to measure microbe/disease similarity from the aspect of biological properties. GAPK similarity is used to evaluate microbe/disease similarity from the topological structure of MDA network. As compared with two individual similarity measurements, the combination of functional similarity and GAPK similarity can more accurately assess microbe/disease similarity and further improve MDA identification performance. Thus, we use the two types of information for microbe/disease similarity evaluation. Moreover, the final disease similarity matrix Sd is computed by integrating their functional similarity and GAPK similarity by Eq. (4) as follows:
Sd(di,dj)={Sdfun(di,dj)+Gd(di,dj)2 if Sdfun(di,dj)≠0Gd(di,dj) otherwise.
Similarly, microbe similarity matrix Sm is computed by Eq. (5) as follows:
Sm(mi,mj)={Smfun(mi,mj)+Gm(mi,mj)2 if Smfun(mi,mj)≠0Gm(mi,mj) otherwise. 2.4. Feature extraction
GATE can efficiently learn features from structured graph data by stacking encoders and decoders (Deng et al., 2022). In this study, we use GATE to extract features for each microbe–disease pair. The GATE structure contain multiple encoders and decoders. In the encoders, each encoder uses a self-attention mechanism to generate new representations for nodes based on their neighborhood information (Veličković et al., 2017). In the kth layer of encoder, relationship between node i and its neighbor node j is computed by Eq. (6) as follows:
where W(k), Vs(k), and Vr(k) denote the trainable parameters in the kth layer of encoder with the sigmoid activation function. hi(k-1) and hj(k-1) denote the feature representations of nodes i and j in the (k−1)th layer, respectively. For the ith node, its associations with the other nodes are taken as its initial representation, that is, hi(0)=xi, and its representation in the kth layer is generated by Eq. (7) as follows:
hi(0)=xihik=∑j∈Niαij(k)σ(W(k)hj(k-1))
We use the softmax function to normalize coefficients of node i's neighbors and solve the comparability problem by Eq. (8) as follows:
αij(k)=exp(cij(k))∑l∈Niexp(cil(k))
where Ni represents node i and its all neighbors. Moreover, the output in the final layer of encoder is considered the node representations.
In the decoder, the initial attributes of each node are reconstructed. Its input comes from the output in the final layer of encoder. Each neighbor of the current node is assigned to different weights by the attention mechanism. The normalized relevance between node i and its neighbor j in the kth layer of decoder is computed by Eqs (9) and (10) as follows:
where Ŵ(k), V^s(k)T, and V^r(k)T denote the trainable parameters in the kth layer of decoder. The kth layer in decoder reconstructs the node representations in the (k−1)th layer by Eq. (11) as follows:
h^ik−1=∑j∈Niα^ij(k)σ(W^(k)h^j(k))
The loss function is defined by Eq. (12) as follows:
Loss=∑i=1N‖xi−x^i‖2−λ∑j∈Nilog(11+exp(−hiThj))
where the first and second terms denote the reconstruction loss of node features and one of graph structure, respectively. λ is a hyperparameter used to balance the contribution of two reconstruction loss terms. xi and x^i represent the initial features and the reconstructed features of node i, respectively. hj is the representation of a neighboring node j of node i.
Finally, we compute microbe feature vectors and disease feature vectors using GATE, and then, a microbe-disease pair is characterized as a a-dimensional vector by concatenating features of both the microbe and the disease.
2.5. Reliable negative MDA selection
In the area of machine learning, negative samples are equally important to final classification performance. However, there are lack of reliable negative MDAs on existing MDA databases due to the limitations of biological experiments. Thus, we design a reliable negative MDA selection method based on PU learning.
PU learning can efficiently identify high-quality negative samples from unlabeled samples and has been widely used in various practical situations (Li et al., 2022). The K-means clustering approach is one of the most popular unsupervised learning algorithms (Peng et al., 2022b). In the HMDAD and Disbiome databases, there are a few positive MDAs and multiple unknown microbe–disease pairs; that is, the two MDA databases are imbalanced. XGBoost has extremely fast parallel computation speed and demonstrates better performance in both balanced and imbalanced databases (Abdu-Aljabar and Awad, 2021).
In this manuscript, we propose a PU learning algorithm to select reliable negative MDAs by combining K-means clustering and XGBoost. Let that positive sample set P and unlabeled example set U denote known MDAs and unknown microbe–disease pairs, respectively. To select reliable negative MDAs from U, as shown in Algorithm 1, we design a PU learning algorithm.
A PU learning algorithm for selecting reliable negative MDAs.
1: Clustering each MDA sample with the K-means clustering algorithm based on the extracted MDA features using GATE.
2: Selecting the first t samples in P which have the smallest distance with cluster centroid as S and adding S into U.
2: Taking P−S as positive samples, and U+S as negative samples.
3: Calculating association score matrix A for all microbe–disease pairs based on XGBoost.
4: Ranking microbe–disease pairs in S based on association scores in A and obtaining the minimum score Amin in S.
5: For every sample x in U
6: If Ax satisfies Ax<Amin
7: then RN = RN∪x
8: Endfor
9: Obtaining reliable negative MDA samples RN.
Particularly, during PU learning, if spy samples are randomly selected from positive sample set P and placed into U, the obtained spy samples could be located at the boundary of the class cluster composed of samples in the entire P and belong to outliers. These spy samples have low spatial similarity with unknown positive examples in U. If a large number of noise or outliers are selected as spy samples, it will greatly affect the evaluation of the classifier on unlabeled samples, which could directly cause decreasing classification performance. Thus, we use K-means clustering algorithm for spy sample selection.
2.6. MDA prediction
We build a DNN to classify unknown microbe–disease pairs based on the extracted MDA features, the selected reliable negative MDAs, and known MDAs. The DNN contains an input layer, multiple hidden layers, and an output layer. In the input layer with a neurons, each MDA sample x with a-dimensional features is fed into the model by Eq. (13) as follows:
x=[x1,x2,…,xa]
where xi denotes the ith feature in x.
The jth hidden layer outputs the results by Eq. (14) as follows:
hj=∑i=1awixi+bjf(hj)=ReLU(hj)
where f denotes the ReLU activation function. Finally, the output layer with the sigmoid activation function outputs MDA classification results by Eq. (15) as follows:
σ(h)=11+e-h′
where h′ denotes the output in the final hidden layer.
3. Result3.1. Experimental settings
To evaluate the MDA prediction performance of our proposed GPUDMDA method, we compared it with other MDA identification methods (LRLSHMDA, NTSHMDA, GATMDA, and MNMDA) under five-fold cross validation (CV) on diseases, microbes, and microbe–disease pairs for 20 times. LRLSHMDA (Wang et al., 2017) is Laplacian regularized least square-based MDA identification algorithm, NTSHMDA (Luo and Long, 2018) is integrated random walk and network topology similarity, GATMDA (Long et al., 2021) combined inductive matrix completion and graph attention networks to complete missing MDAs, and MNNMDA (Liu et al., 2023) used a low-rank matrix completion model for identifying possible MDAs. During MDA prediction, it is not enough to reflect the MDA identification performance of a computational model only through cross-validation on microbe–disease pairs. Thus, in the study, we implemented cross-validations on microbes, diseases, and microbe–disease pairs to comprehensively assess the model's performance. The detailed definitions about the above three cross-validations have been proposed by Peng et al. (2020). AUC and AUPR were applied to measure the performance of MDA prediction methods.
In this study, we used GATE to extract features of microbes and diseases from their similarity networks, both of which are 64 dimensional vectors. We selected t samples from positive sample set P to place unlabeled example set U. When t was set to 15% of P on the HMDAD database and 20% of P on the Disbiome database, GPUDMDA obtained the best performance. Thus, we set t to 15 and 20% of P on the two databases, respectively. For DNN with four layers, the input layer, the following three hidden layer, and the output layer have 128, 100, 100, 50, and one nodes, respectively. Learning rate and “dropout” were set to 0.001 and 0.2. The parameter “epoch_num,” denoting the number of training, was set to 300 and 1,500 on the two databases, respectively. Disbiome is a larger dataset, and the proposed computational model needs to be trained for enough times to obtain better classification performance; thus, the “epoch_num” value was much larger on the Disbiome database.
Additionally, the number of positive samples is the same as one of the known MDAs.The number of selected credible negative MDAs is related to the computed smallest association probability score Amin. Since the credible negative MDAs were selected from unknown microbe–disease pairs, unknown microbe–disease pairs were decreased but accounted for most of all microbe–disease pairs.
3.2. Performance comparison under CV on diseases
Under CV on diseases, 80% diseases were taken as the training set and the remaining was test set. Figure 2 elucidates the receiver operating characteristic (ROC) and precision-recall (PR) curves of the five MDA prediction methods on the HMDAD and Disbiome databases under CV on diseases. Under CV on diseases, GPUDMDA obtained the best AUCs of 0.7121 and 0.8372, and the best AUPRs of 0.2022 and better AUPR of 0.2030 on the HMDAD and Disbiome databases, respectively, significantly outperforming LRLSHMDA, NTSHMDA, GATMDA, and MNMDA.
Performance comparison of five MDA prediction methods under five-fold CV on diseases. (A, B) The ROC and PR curves of the five methods on HMDAD. (C, D) The ROC and PR curves of the five methods on Disbiome.
3.3. Performance comparison under CV on microbes
Under CV on microbes, 80% microbes were taken as the training set and the remaining was test set. Figure 3 shows the ROC and PR curves of the five methods under CV on microbes. Under CV on microbes, GPUDMDA obtained better AUCs of 0.9454 and 0.8908 and AUPRs of 0.8529 and 0.4367 than LRLSHMDA, NTSHMDA, GATMDA, and MNMDA.
Performance comparison of five MDA prediction methods under five-fold CV on microbes. (A, B) The ROC and PR curves of the five methods on HMDAD. (C, D) The ROC and PR curves of the five methods on Disbiome.
3.4. Performance comparison under CV on microbe–disease pairs
Under CV on microbe–disease pairs, 80% microbe–disease pairs were taken as the training set and the remaining was test set. Figure 4 illustrates the ROC and PR curves of the five MDA prediction methods under CV on microbe–disease pairs. Under the CV, GPUDMDA computed better AUCs of 0.9501 and 0.8948, and the best AUPRs of 0.8545 and 0.4464 among the five methods.
Performance comparison of the five MDA prediction methods under five-fold CV on microbe–disease pairs. (A, B) The ROC and PR curves of the five methods on HMDAD. (C, D) The ROC and PR curves of the five methods on Disbiome.
3.5. The affect of PU learning on performance
Reliable negative samples can improve the classification performance of a model. To evaluate the reliability of the identified negative MDAs by GPUDMDA, we compared its performance under negative sample selection. Figure 5 demonstrates the affect of negative samples selected by PU learning on performance. The results elucidated that GPUDMDA with PU learning outperformed one without PU learning. Particularly, the performance of GPUDMDA with PU learning obtained significant improvement on Disbiome. The results suggested that reliable negative MDAs selected by PU learning can boost the MDA prediction ability.
The impact of PU learning on performance in two databases.
3.6. Case study
In the above sections, we have confirmed the MDA identification accuracy of GPUDMDA. Next, we intend to find new microbes for asthma and IBD.
3.6.1. Identifying new microbes for asthma
Asthma is a heterogeneous disease with respect to respiratory symptoms including chest tightness, shortness of breath, wheeze, and cough. It is the most common chronic respiratory condition and affects ~339 million people worldwide. Approximately 5%–10% of these patients have severe asthma. More than 10% of adults and 2.5% of children suffered from asthma have severe asthma (Brusselle and Koppelman, 2022; Reddel et al., 2022; Rattu et al., 2023).
We used the proposed GPUDMDA method to find new microbes associated with asthma. Tables 1, 2 show the predicted top 30 microbes that may associate with asthma on the HMDAD and Disbiome databases. The predicted 30 asthma-associated microbes included microbes with known association information with asthma and microbes without association information with asthma on the two databases. As shown in Table 1, 23 and 29 microbes can be validated by each or both of two databases or existing literatures among the identified top 30 potential asthma-associated microbes on the two databases, respectively. Furthermore, we found that Enterobacter hormaechei could associate with asthma with the ranking of 15 on the HMDAD database. On the Disbiome database, GPUDMDA predicted that Enterobacter may be a sole and unknown asthma-associated microbe among the predicted top 30 microbes associated with asthma.
The predicted top 30 microbes associated with Asthma on HMDAD.
Rank
Microbe
Evidence
1
Proteobacteria
Confirmed by HMDAD and Disbiome
2
Prevotella
Confirmed by HMDAD and Disbiome
3
Staphylococcus
Confirmed by HMDAD and Disbiome
4
Bacteroidetes
Confirmed by HMDAD and Disbiome
5
Enterobacteriaceae
Confirmed by Disbiome
6
Clostridium coccoides
PMID: 1477358
7
Firmicutes
PMID: 23265859
8
Comamonadaceae
Confirmed by HMDAD and Disbiome
9
Oxalobacteraceae
Confirmed by HMDAD and Disbiome
10
Sphingomonadaceae
Confirmed by HMDAD and Disbiome
11
Haemophilus
Confirmed by HMDAD and Disbiome
12
Helicobacter pylori
Confirmed by HMDAD
13
Enterococcus
PMID: 29788027
14
Enterobacter aerogenes
PMID: 23842440
15
Enterobacter hormaechei
Unconfirmed
16
Klebsiella pneumoniae
PMID: 26953325
17
Shigella dysenteriae
Unconfirmed
18
Lactobacillus
Confirmed by Disbiome
19
Clostridia
PMID: 21477358
20
Veillonella
Confirmed by Disbiome
21
Klebsiella
Confirmed by Disbiome
22
Prevotella copri
Unconfirmed
23
Actinobacteria
PMID: 28947029
24
Shuttleworthia
Unconfirmed
25
Desulfovibrio
PMID: 29198875
26
Clostridium difficile
PMID: 21872915
27
Oxalobacter formigenes
Unconfirmed
28
Fusobacteria
Unconfirmed
29
Porphyromonadaceae
Confirmed by Disbiome
30
Verrucomicrobiaceae
Unconfirmed
The predicted top 30 microbes associated with Asthma on Disbiome.
Rank
Microbe
Evidence
1
Actinomyces
Confirmed by Disbiome
2
Bacteroides stercoris
Confirmed by Disbiome
3
Bifidobacterium
Confirmed by Disbiome
4
Blautia
Confirmed by Disbiome
5
Clostridiaceae
Confirmed by Disbiome
6
Clostridium neonatale
Confirmed by Disbiome
7
Comamonadaceae
Confirmed by HMDAD and Disbiome
8
Corynebacterium
Confirmed by Disbiome
9
Faecalibacterium
Confirmed by Disbiome
10
Gallibacterium
Confirmed by Disbiome
11
Gammaproteobacteria
Confirmed by Disbiome
12
Gemella
Confirmed by Disbiome
13
Klebsiella
Confirmed by Disbiome
14
Leclercia
Confirmed by Disbiome
15
Moraxella
Confirmed by Disbiome
16
Neisseria
Confirmed by Disbiome
17
Nitrosomonadaceae
Confirmed by Disbiome
18
Oxalobacteraceae
Confirmed by HMDAD and Disbiome
19
Planococcaceae
Confirmed by Disbiome
20
Prevotella
Confirmed by HMDAD and Disbiome
21
Pseudomonadaceae
Confirmed by Disbiome
22
Sphingomonadaceae
Confirmed by HMDAD and Disbiome
23
Staphylococcus
Confirmed by HMDAD and Disbiome
24
Stenotrophomonas
Confirmed by Disbiome
25
Streptococcus
Confirmed by Disbiome
26
Sutterella wadsworthensis
Confirmed by Disbiome
27
Veillonella
Confirmed by Disbiome
28
Weeksella
Confirmed by Disbiome
29
Propionibacterium
PMID: 13268970
30
Enterobacter
Unconfirmed
Enterobacter hormaechei (Yeh et al., 2022) is a member and the most common nosocomial pathogen of the Enterobacter cloacae complex. It plays a key role in infectious diseases including, urinary tract infections, pneumonia, biliary tract infections, bacteremia, colitis and cellulitis. It is commonly found to be a high-pathogenicity island on its chromosome and is more virulent compared with other E. cloacae complex. In this study, GPUDMDA identified that E. hormaechei could associate with asthma.
Figure 6 shows the association network between the predicted top 53 asthma-associated microbes and asthma, after removing the repeated associations on the two databases. In Figure 6, the gray solid lines and blue dashed lines denote known associations between microbes and asthma and the predicted associations between microbes and asthma, respectively.
The predicted top 53 microbes associated with asthma on the two databases.
3.6.2. Identifying new microbes for inflammatory bowel disease
IBD is a class of globally chronic intestinal disease (Chang, 2020; Kaplan and Windsor, 2021). It widely exists in the gut and gastrointestinal tract and extraintestinal organs in many patients (Rogler et al., 2021). Up to 2 million Europeans and 1.5 million North Americans suffer from this disease (Jairath and Feagan, 2020). It mainly comprises Crohn's disease, ulcerative colitis, and indeterminate colitis (Flynn and Eisenstein, 2019). Many studies thought that it is the result of interactions between microbial, environmental, and immune-mediated factors. In particular, microbiome has been reported to have potential roles in the development, progression, and treatment of IBD. The gut microbiome is different in the IBD patients from one in healthy bodies (Glassner et al., 2020).
In particular, IBD is very common in children. Many pediatricians and the other pediatric clinicians meet children suffered from IBD. The IBD pediatric populations demonstrate the classic features of abdominal pain, bloody diarrhea, and weight loss as well as non-classic features of anemia, isolated poor growth, or the other extraintestinal symptoms. Recently, the IBD children patients show a rising incidence (Rosen et al., 2015; Oliveira and Monteiro, 2017). In total, 25%–30% of patients with Crohn's disease and 20% of patients with ulcerative colitis have been diagnosed in < 20 years of age. Moreover, 4% of pediatric IBD patients have been detected before 5 years (Kelsen and Baldassano, 2008). IBD severely affects normal growth and development of children. When treating children with newly diagnosed IBD, we need to consider their affects on growth and development and bone health (Rosen et al., 2015).
In this manuscript, we used the proposed GPUDMDA method to find potential microbes associated with IBD. Tables 3, 4 show the predicted top 30 IBD-associated microbes on the two MDA databases. The predicted 30 IBD-associated microbes included microbes with known association information with IBD and microbes without association information with IBD. In total, 20 and 28 predicted IBD-associated microbes can be validated by databases or existing publications among all predicted top 30 microbes on the two databases, respectively. On HMDAD, GPUDMDA predicted that E. hormaechei could associate with IBD with the ranking of 7. On Disbiome, the former 28 microbes have been confirmed to associate with IBD, and GPUDMDA also identified that E. hormaechei could link with IBD with the ranking of 29.
The predicted top 30 microbes associated with IBD on HMDAD.
Rank
Microbe
Evidence
1
Clostridium difficile
PMID: 27499718
2
Helicobacter pylori
PMID: 22221289
3
Staphylococcus
PMID: 27239107
4
Clostridia
PMID: 31142855
5
Clostridium coccoides
PMID: 19235886
6
Enterobacter aerogenes
PMID: 4061480
7
Enterobacter hormaechei
Unconfirmed
8
Klebsiella pneumoniae
PMID: 9930068
9
Shigella dysenteriae
Unconfirmed
10
Prevotella copri
Unconfirmed
11
Enterococcus
PMID: 24629344
12
Klebsiella
PMID: 29573336
13
Actinobacteria
Confirmed by HMDAD
14
Bifidobacterium
PMID: 24478468
15
Dietzia maris
Unconfirmed
16
Staphylococcus epidermidis
Unconfirmed
17
Oxalobacter formigenes
Unconfirmed
18
Tropheryma whipplei
Unconfirmed
19
Staphylococcus aureus
PMID: 11424320
20
Bacteroides vulgatus
PMID: 29454108
21
Actinomyces
PMID: 30545401
22
Porphyromonas gingivalis
PMID: 31652577
23
Selenomonas
Unconfirmed
24
Treponema
PMID: 31851086
25
Fusobacterium nucleatum
PMID: 26718210
26
Bacteroides ovatus
PMID: 30666959
27
Verrucomicrobiaceae
PMID: 22572638
28
Desulfovibrio
Confirmed by Disbiome
29
Clostridiales
Unconfirmed
30
Escherichia coli
PMID: 29573336
The predicted top 30 microbes associated with IBD on Disbiome.
Rank
Microbe
Evidence
1
Anaerostipes
Confirmed by Disbiome
2
Bacillus licheniformis
Confirmed by Disbiome
3
Blautia
Confirmed by Disbiome
4
Bradyrhizobiaceae
Confirmed by Disbiome
5
Butyricimonas
Confirmed by Disbiome
6
Comamonadaceae
Confirmed by Disbiome
7
Christensenellaceae
Confirmed by Disbiome
8
Dehalobacter
Confirmed by Disbiome
9
Desulfovibrio
Confirmed by Disbiome
10
Dorea formicigenerans
Confirmed by Disbiome
11
Eubacterium biforme
Confirmed by Disbiome
12
Gemella
Confirmed by Disbiome
13
Gluconobacter oxydans
Confirmed by Disbiome
14
Lachnobacterium
Confirmed by Disbiome
15
Methanobrevibacter smithii
Confirmed by Disbiome
16
Mogibacterium
Confirmed by Disbiome
17
Moraxellaceae
Confirmed by Disbiome
18
Pseudomonas straminea
Confirmed by Disbiome
19
Ruminococcus bromii
Confirmed by Disbiome
20
Saccharomyces cerevisiae
Confirmed by Disbiome
21
Streptococcus anginosus
Confirmed by Disbiome
22
Bacteroides ovatus
PMID: 30666959
23
Enterobacter aerogenes
PMID: 4061480
24
Fusobacterium
PMID: 25307765
25
Klebsiella pneumoniae
PMID: 9930068
26
Paraprevotella
PMID: 25307765
27
Propionibacterium acnes
PMID: 28630242
28
Staphylococcus
PMID: 27239107
29
Oxalobacteraceae
Unconfirmed
30
Enterobacter hormaechei
Unconfirmed
Figure 7 shows the association network between the predicted top 54 IBD-associated microbes and IBD, after removing the repeated associations on the two databases. In Figure 7, the gray solid lines and blue dashed lines denote known associations between microbes and IBD and the predicted associations between microbes and IBD, respectively.
The predicted top 54 microbes associated with IBD on the two databases.
4. Discussion and conclusion
Microbes manifest dense relationships with various human complex diseases. Predicting underlying MDAs can contribute to analyzing complex disease-causing mechanisms and screening potential biomarkers for the diagnosis and therapy of these diseases. Traditional wet lab methods are expensive, time-consuming, and laborious. Consequently, in silico methods have been increasingly developed as an efficient complementary to experimental methods.
In this study, we developed a deep learning model called GPUDMDA to capture new linkages between microbes and various human complex diseases. GPUDMDA first computed disease similarity and microbe similarity matrices based on their functional similarity and GIPK similarity, respectively. Next, it extracted features for each microbe–disease pair with GATE. Third, it selected a few reliable negative MDAs based on PU learning with K-means clustering and XGBoost. Finally, it took the extracted MDA features and the selected negative MDAs as inputs and designed a DNN to predict potential MDAs.
GPUDMDA was compared with four state-of-the-art MDA identification models (i.e., MNNMDA, GATMDA, LRLSHMDA, and NTSHMDA) on the HMDAD and Disbiome databases under five-fold CVs on microbes, diseases, and microbe–disease pairs. Under the three CVs, GPUDMDA computed the best AUCs and AUPRs on the two databases, suggesting that GPUDMDA could improve MDA prediction performance. Finally, we implemented case studies for asthma and IBD. The results showed that E. hormaechei could densely associate with asthma and IBD and need further biological experimental validation.
In future, we will combine biological features of microbe, diseases, and MDA network to design more accurate negative MDA selection method. In addition, we will also develop novel deep learning model to improve MDA classification performance based on the selected reliable negative MDA samples. Interestingly, we have conducted several computational models including existing classical MDA prediction methods. But the results elucidated that many models failed to compute better AUPR on the Disbiome database. It may be caused by different data structures of Disbiome. In future, we will further design a better robust computational method to improve MDA prediction on the Disbiome database. We hope that the proposed GPUDMDA method helps to identify microbes associated with related diseases and further contributes to mining the clues of treatment.
Data availability statement
The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding authors. The HMDAD and Disbiome databases are available at: http://www.cuilab.cn/hmdad and https://disbiome.ugent.be/, respectively. Accession numbers can be downloaded at https://github.com/plhhnu/GPUDMDA.
Author contributions
LP and LH: conceptualization and methodology. LP, ZL, and LD: funding acquisition. LP, GT, YW, ZL, and LD: project administration. LH: writing—original draft and software. LP, LH, PW, and ZL: writing—reviewing and editing. LP, LH, GT, GL, JC, and LD: investigation. LH, GL, JC, and LD: validation. All authors contributed to the article and approved the submitted version.
Funding
LD was supported by the National Natural Science Foundation of China under Grant No. 81500391. LP was supported by the National Natural Science Foundation of China under Grant No. 61803151 and the Natural Science Foundation of Hunan province Grant No. 2023JJ50201. ZL was supported by the National Natural Science Foundation of China under Grant No. 62172158. PW was supported by the Excellent Youth Project of Hunan Provincial Education Department Grant No. 21B0802.
The authors would like to thank all the authors of the cited references.
Conflict of interest
GT and YW were employed by Geneis (Beijing) Co. Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
ReferencesAbdu-AljabarR. D.AwadO. A. (2021). “A comparative analysis study of lung cancer detection and relapse prediction using XGBoost classifier,” in IOP Conference Series: Materials Science and Engineering, volume 1076 (Bristol: IOP Publishing), 012048. 10.1088/1757-899X/1076/1/012048BrusselleG. G.KoppelmanG. H. (2022). Biologic therapies for severe asthma. N. Engl. J. Med. 386, 157–171. 10.1056/NEJMra203250635020986ChangJ. T. (2020). Pathophysiology of inflammatory bowel diseases. N. Engl. J. Med. 383, 2652–2664. 10.1056/NEJMra200269733382932ChenX.HuangL. (2023). Computational model for disease research. Brief. Bioinformatics 24, bbac615. 10.1093/bib/bbac61536642407ChenX.LiT.-H.ZhaoY.WangC.-C.ZhuC.-C. (2021). Deep-belief network for predicting potential mirna-disease associations. Brief. Bioinformatics 22, bbaa186. 10.1093/bib/bbaa18634020550ChenX.LiuH.ZhaoQ. (2020). Editorial: Bioinformatics in microbiota. Front. Microbiol. 11, 100. 10.3389/fmicb.2020.0010032117130ChenX.XieD.ZhaoQ.YouZ.-H. (2019). Micrornas and complex diseases: from experimental results to computational models. Brief. Bioinformatics20, 515–539. 10.1093/bib/bbx13029045685DemirciM.TokmanH.UysalH.DemiryasS.KarakullukcuA.SaribasS.. (2019). Reduced akkermansia muciniphila and faecalibacterium prausnitzii levels in the gut microbiota of children with allergic asthma. Allergol. Immunopathol. 47, 365–371. 10.1016/j.aller.2018.12.00930765132DengL.LiuZ.QianY.ZhangJ. (2022). Predicting circrna-drug sensitivity associations via graph attention auto-encoder. BMC Bioinformatics23, 1–15. 10.1186/s12859-022-04694-y35508967El MouzanM. I.WinterH. S.AssiriA. A.KorolevK. S.Al SarkhyA. A.DowdS. E.. (2018). Microbiota profile in new-onset pediatric crohn's disease: data from a non-western population. Gut Pathog. 10, 1–10. 10.1186/s13099-018-0276-330519287FlynnS.EisensteinS. (2019). Inflammatory bowel disease presentation and diagnosis. Surg. Clin. 99, 1051–1062. 10.1016/j.suc.2019.08.00131676047GillS. R.PopM.DeBoyR. T.EckburgP. B.TurnbaughP. J.SamuelB. S.. (2006). Metagenomic analysis of the human distal gut microbiome. Science312, 1355–1359. 10.1126/science.112423416741115GlassnerK. L.AbrahamB. P.QuigleyE. M. (2020). The microbiome and inflammatory bowel disease. J. Allergy Clin. Immunol. 145, 16–27. 10.1016/j.jaci.2019.11.00331910984GuanJ.ZhangZ. G.LiuY.WangM. (2022). A novel bi-directional heterogeneous network selection method for disease and microbial association prediction. BMC Bioinformatics23, 1–15. 10.1186/s12859-022-04961-y36376802Henao-MejiaJ.ElinavE.ThaissC. A.Licona-LimonP.FlavellR. A. (2013). Role of the intestinal microbiome in liver disease. J. Autoimmun. 46, 66–73. 10.1016/j.jaut.2013.07.00128249266HolmesE.WijeyesekeraA.Taylor-RobinsonS. D.NicholsonJ. K. (2015). The promise of metabolic phenotyping in gastroenterology and hepatology. Nat. Rev. Gastroenterol. Hepatol. 12, 458–471. 10.1038/nrgastro.2015.11426194948HuH.FengZ.LinH.ChengJ.LyuJ.ZhangY.. (2023a). Gene function and cell surface protein association analysis based on single-cell multiomics data. Comput. Biol. Med. 157, 106733. 10.1016/j.compbiomed.2023.10673336924730HuH.FengZ.LinH.ZhaoJ.ZhangY.XuF.. (2023b). Modeling and analyzing single-cell multimodal data with deep parametric inference. Brief. Bioinformatics 24, bbad005. 10.1093/bib/bbad00536642414JairathV.FeaganB. G. (2020). Global burden of inflammatory bowel disease. Lancet Gastroenterol. Hepatol. 5, 2–3. 10.1016/S2468-1253(19)30358-931648974JanssensY.NielandtJ.BronselaerA.DebunneN.VerbekeF.WynendaeleE.. (2018). Disbiome database: linking the microbiome to disease. BMC Microbiol. 18, 1–6. 10.1186/s12866-018-1197-529866037JiangC.TangM.JinS.HuangW.LiuX. (2022). KGNMDA: a knowledge graph neural network method for predicting microbe-disease associations. IEEE/ACM Trans. Comput. Biol. Bioinform. 20, 1147–1115. 10.1109/TCBB.2022.318436235724280KamnevaO. K. (2017). Genome composition and phylogeny of microbes predict their co-occurrence in the environment. PLoS Comput. Biol. 13, e1005366. 10.1371/journal.pcbi.100536628152007KaplanG. G.WindsorJ. W. (2021). The four epidemiological stages in the global evolution of inflammatory bowel disease. Nat. Rev. Gastroenterol. Hepatol. 18, 56–66. 10.1038/s41575-020-00360-x33033392KelsenJ.BaldassanoR. N. (2008). Inflammatory bowel disease: the difference between children and adults. Inflamm. Bowel Dis. 14(suppl_2), S9–S11. 10.1002/ibd.2056030881893LiF.DongS.LeierA.HanM.GuoX.XuJ.. (2022). Positive-unlabeled learning in bioinformatics and computational biology: a brief review. Brief. Bioinformatics 23, bbab461. 10.1093/bib/bbab46134729589LiH.WangY.ZhangZ.TanY.ChenZ.WangX.. (2020). “Identifying microbe-disease association based on a novel back-propagation neural network model,” in IEEE/ACM Trans. Comput. Biol. Bioinformatics18, 2502–2513. 10.1109/TCBB.2020.298645932305935LihongP.WangC.TianX.ZhouL.LiK. (2022). Finding lncrna-protein interactions based on deep learning with dual-net neural architecture. IEEE/ACM Trans. Comput. Biol. Bioinformatics19, 3456–3468. 10.1109/TCBB.2021.311623234587091LiuD.LiuJ.LuoY.HeQ.DengL. (2021). Mgatmda: predicting microbe-disease associations via multi-component graph attention network. IEEE/ACM Trans. Comput. Biol. Bioinformatics19, 3578–3585. 10.1109/TCBB.2021.311631834587092LiuH.BingP.ZhangM.TianG.MaJ.LiH.. (2023). MNNMDA: predicting human microbe-disease association via a method to minimize matrix nuclear norm. Comput. Struct. Biotechnol. J. 21, 1414–1423. 10.1016/j.csbj.2022.12.05336824227LiuY.WangS.-L.ZhangJ.-F.ZhangW.ZhouS.LiW.. (2020). DMFMDA: prediction of microbe-disease associations based on deep matrix factorization using bayesian personalized ranking. IEEE/ACM Trans. Comput. Biol. Bioinformatics18, 1763–1772. 10.1109/TCBB.2020.301813832816678LongY.LuoJ. (2019). WMGHMDA: a novel weighted meta-graph-based model for predicting human microbe-disease association on heterogeneous information network. BMC Bioinformatics20, 1–18. 10.1186/s12859-019-3066-031675979LongY.LuoJ.ZhangY.XiaY. (2021). Predicting human microbe-disease associations via graph attention networks with inductive matrix completion. Brief. Bioinformatics 22, bbaa146. 10.1093/bib/bbaa14632725163LuoJ.LongY. (2018). Ntshmda: prediction of human microbe-disease association based on random walk by integrating network topological similarity. IEEE/ACM Trans. Comput. Biol. Bioinformatics17, 1341–1351. 10.1109/TCBB.2018.288304130489271MaW.ZhangL.ZengP.HuangC.LiJ.GengB.. (2017). An analysis of human microbe-disease associations. Brief. Bioinformatics18, 85–97. 10.1093/bib/bbw00526883326MaY.JiangH. (2020). Ninimhmda: neural integration of neighborhood information on a multiplex heterogeneous network for multiple types of human microbe-disease association. Bioinformatics36, 5665–5671. 10.1093/bioinformatics/btaa108033416850NiuY.-W.QuC.-Q.WangG.-H.YanG.-Y. (2019). Rwhmda: random walk on hypergraph for microbe-disease association prediction. Front. Microbiol. 10, 1578. 10.3389/fmicb.2019.0157831354672OliveiraS. B.MonteiroI. M. (2017). Diagnosis and management of inflammatory bowel disease in children. BMJ357, j2083. 10.1136/bmj.j208328566467PengC.-X.ZhouX.-G.XiaY.-H.LiuJ.HouM.-H.ZhangG.-J.. (2022a). Structural analogue-based protein structure domain assembly assisted by deep learning. Bioinformatics38, 4513–4521. 10.1093/bioinformatics/btac55335962986PengL.ShenL.LiaoL.LiuG.ZhouL. (2020). RNMFMDA: a microbe-disease association identification method based on reliable negative sample selection and logistic matrix factorization with neighborhood regularization. Front. Microbiol. 11, 592430. 10.3389/fmicb.2020.59243033193260PengL.TanJ.XiongW.ZhangL.WangZ.YuanR.. (2023a). Deciphering ligand-receptor-mediated intercellular communication based on ensemble deep learning and the joint scoring strategy from single-cell transcriptomic data. Comput. Biol. Med. 163, 107137. 10.1016/j.compbiomed.2023.10713737364528PengL.WangC.TianG.LiuG.LiG.LuY.. (2022c). Analysis of ct scan images for COVID-19 pneumonia based on a deep ensemble framework with densenet, swin transformer, and regnet. Front. Microbiol. 13, 995323. 10.3389/fmicb.2022.99532336212877PengL.WangF.WangZ.TanJ.HuangL.TianX.. (2022d). Cell-cell communication inference and analysis in the tumour microenvironments from single-cell transcriptomics: data resources and computational strategies. Brief. Bioinformatics 23, bbac234. 10.1093/bib/bbac23435753695PengL.YuanR.HanC.HanG.TanJ.WangZ.. (2023b). Cellenboost: a boosting-based ligand-receptor interaction identification model for cell-to-cell communication inference. IEEE Trans. Nanobioscience. 10.1109/TNB.2023.327868537216267PengL.HeX.ZhangL.PengX.LuY.LiZ.. (2022b). “A deep learning-based unsupervised learning method for spatially resolved transcriptomic data analysist,” in 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) (Las Vegas, NV: IEEE), 281–286. 10.1109/BIBM55620.2022.9995207RattuA.KhalevaE.BrightlingC.DahlénS.-E.BossiosA.FlemingL.. (2023). Identifying and appraising outcome measures for severe asthma: a systematic review. Eur. Respir. J. 61, 2201231. 10.1183/13993003.01231-202236549712ReddelH. K.BacharierL. B.BatemanE. D.BrightlingC. E.BrusselleG. G.BuhlR.. (2022). Global initiative for asthma strategy 2021: executive summary and rationale for key changes. Am. J. Respir. Crit. Care Med. 205, 17–35. 10.1164/rccm.202109-2205PP34658302RoglerG.SinghA.KavanaughA.RubinD. T. (2021). Extraintestinal manifestations of inflammatory bowel disease: current concepts, treatment, and implications for disease management. Gastroenterology161, 1118–1132. 10.1053/j.gastro.2021.07.04234358489RosenM. J.DhawanA.SaeedS. A. (2015). Inflammatory bowel disease in children and adolescents. JAMA Pediatr. 169, 1053–1060. 10.1001/jamapediatrics.2015.198226414706SchwabeR. F.JobinC. (2013). The microbiome and cancer. Nat. Rev. Cancer13, 800–812. 10.1038/nrc361024132111ShenL.LiuF.HuangL.LiuG.ZhouL.PengL.. (2022). VDA-RWLRLS: an anti-SARS-CoV-2 drug prioritizing framework combining an unbalanced bi-random walk and laplacian regularized least squares. Comput. Biol. Med. 140, 105119. 10.1016/j.compbiomed.2021.10511934902608TianX.ShenL.GaoP.HuangL.LiuG.ZhouL.. (2022). Discovery of potential therapeutic drugs for COVID-19 through logistic matrix factorization with Kernel diffusion. Front. Microbiol. 13, 740382. 10.3389/fmicb.2022.74038235295301VeličkovićP.CucurullG.CasanovaA.RomeroA.LioP.BengioY. (2017). Graph attention networks. arXiv. [preprint]. 10.48550/arXiv.1710.10903WangF.HuangZ.-A.ChenX.ZhuZ.WenZ.ZhaoJ.. (2017). LRLSHMDA: Laplacian regularized least squares for human microbe-disease association prediction. Sci. Rep. 7, 7601. 10.1038/s41598-017-08127-228790448WangL.WangY.LiH.FengX.YuanD.YangJ.. (2019). A bidirectional label propagation based computational model for potential microbe-disease association prediction. Front. Microbiol. 10, 684. 10.3389/fmicb.2019.0068431024481WangL.YouZ.-H.HuangY.-A.HuangD.-S.ChanK. C. (2020). An efficient approach based on multi-sources information to predict circrna-disease associations using deep convolutional neural network. Bioinformatics36, 4038–4046. 10.1093/bioinformatics/btz82531793982WangT.SunJ.ZhaoQ. (2023). Investigating cardiotoxicity related with herg channel blockers using molecular fingerprints and graph attention mechanism. Comput. Biol. Med. 153, 106464. 10.1016/j.compbiomed.2022.10646436584603WangW.ZhangL.SunJ.ZhaoQ.ShuaiJ. (2022). Predicting the potential human lncrna-mirna interactions based on graph convolution network with conditional random field. Brief. Bioinformatics 23, bbac463. 10.1093/bib/bbac46336305458WenL.LeyR. E.VolchkovP. Y.StrangesP. B.AvanesyanL.StonebrakerA. C.. (2008). Innate immunity and intestinal microbiota in the development of type 1 diabetes. Nature455, 1109–1113. 10.1038/nature0733618806780WenZ.YanC.DuanG.LiS.WuF.-X.WangJ.. (2021). A survey on predicting microbe-disease associations: biological data and computational methods. Brief. Bioinformatics 22, bbaa157. 10.1093/bib/bbaa15734020541WuH.WuY.JiangY.ZhouB.ZhouH.ChenZ.. (2022). SCHICSTACKL: a stacking ensemble learning-based method for single-cell Hi-C classification using cell embedding. Brief. Bioinformatics 23, bbab396. 10.1093/bib/bbab39634553746XuJ.XuJ.MengY.LuC.CaiL.ZengX.. (2023). Graph embedding and gaussian mixture variational autoencoder network for end-to-end analysis of single-cell RNA sequencing data. Cell Rep. Methods3, 100382. 10.1016/j.crmeth.2022.10038236814845YanC.DuanG.WuF.-X.PanY.WangJ. (2019). BRWMDA: predicting microbe-disease associations based on similarities and bi-random walk on disease and microbe networks. IEEE/ACM Trans. Comput. Biol. Bioinformatics17, 1595–1604. 10.1109/TCBB.2019.290762630932846YehT.-K.LinH.-J.LiuP.-Y.WangJ.-H.HsuehP.-R. (2022). Antibiotic resistance in Enterobacter hormaechei. Int. J. Antimicrob. Agents60, 106650. 10.1016/j.ijantimicag.2022.10665035934231ZhangP.WuY.ZhouH.ZhouB.ZhangH.WuH.. (2022a). CLNN-loop: a deep learning model to predict CTCF-mediated chromatin loops in the different cell lines and CTCF-binding sites (CBS) pair types. Bioinformatics38, 4497–4504. 10.1093/bioinformatics/btac57535997565ZhangP.ZhangH.WuH. (2022b). iPro-WAEL: a comprehensive and robust framework for identifying promoters in multiple species. Nucleic Acids Res. 50, 10278–10289. 10.1093/nar/gkac82436161334ZhangP.WuH. (2023). ICHROM-deep: an attention-based deep learning model for identifying chromatin interactions. IEEE J. Biomed. Health Inf . 10.1109/JBHI.2023.329229937402191