The dataset in this study was sourced from the HMDAD database (http://www.cuilab.cn/hmdad), which collects known associations between microbes and diseases by searching past research literature (Ma et al., 2017). HMDAD adapted a systematic approach by only including associations that have been experimentally validated and published in reputable journals. This ensures a high level of reliability in the dataset. Past researchers commonly employ metagenomic sequencing techniques to analyze fluctuations in microbial community abundance within specific diseases, contrasting them with the microbial compositions of healthy individuals, thus exploring the associations between microbes and diseases. In the HMDAD dataset, a microbe-disease association pair may contain multiple entries from different research literature sources. Therefore, here, we regard the same microbe-disease association from different evidences as a pair, further removing the redundant information present in the HMDAD dataset. Finally, for this study, we employed a dataset consisting of 450 microbe-disease associations, encompassing 39 human diseases and 292 microbes.

For the convenience of clarity in describing the subsequent research methods, we provide a simple problem definition for the task of predicting associations between microbes and diseases here. We denote M = {m₁, m₂, …, m_nm} and D = {d₁, d₂, …, d_nd} as the sets representing n_m microbes and n_d diseases, respectively. The matrix A∈ℝnm×nd represents the known associations between microbes and diseases, where A_ij = 1 if microbe m_i is associated with disease d_j, otherwise A_ij = 0. However, A_ij = 0 does not mean that microbe m_i has no relation with disease d_j. It may be the reason that their association has not yet been discovered. Therefore, the task of predicting associations between microbes and diseases aims to find microbe m_i for each disease d_j where A_ij = 0 in the known association matrix, but microbe m_i is actually related to disease d_j.

Figure 1 illustrates the workflow of GCATCMDA, a model based on graph neural networks and contrastive learning for predicting effective candidate sets of microbe-disease associations. First microbe-microbe and disease-disease Gaussian kernel similarity networks are constructed using known associations. The model then integrates graph neural networks and contrastive learning principles to extract meaningful feature representations of microbes and diseases from the association networks. Last the obtained microbe and disease features are fed into an inner product decoder to compute their corresponding association scores. A detailed description of the key components of this model is elucidated below.

The GCATCMDA model proposed in this study is a microbe-disease association prediction model based on graph neural networks and contrastive learning. It aims to predict potential associations between microbes and diseases from known microbial-disease association dataset. The hyperparameter settings required for this model are described as follows. Firstly, the Gaussian kernel similarity threshold t needs to be set for constructing microbe and disease similarity networks. Secondly, parameters need to be set for the GCAT feature encoder module, including the dimensionality F of node features, the number of network layers L for graph convolution, and the number of attention heads heads for the graph attention mechanism. Then, in the contrastive learning loss module, the temperature hyperparameter τ and the weight parameter λ relative to the total loss are adjusted. Finally, the GCATCMDA model is trained using the Adam (Kingma and Ba, 2014) optimizer, with parameters including the learning rate lr, weight decay wd, and the number of training iterations epochs.

This study determines the optimal parameter settings of the GCATCMDA model on the dataset by enumerating different parameter combinations. Subsequently, there is an analysis of key parameters t, F, L, and heads. After comparing experimental results, the optimal hyperparameter settings for the GCATCMDA model on the HMDAD dataset are determined as follows: t = 0.4, F = 128, L = 3, heads = 2, τ = 1, λ = 0.2, lr = 0.00001, wd = 0.001, and epochs = 100.

In order to verify the effectiveness of the proposed GCATCMDA model, we compares it with KATZHMDA (Chen et al., 2017), LRLSHMDA (Wang et al., 2017), NTSHMDA (Luo and Long, 2018), and KGNMDA (Jiang et al., 2022). These five methods are recognized for their outstanding performance in this task in past studies and provide research methods with open-source code. For negative samples required in model training, this study randomly selects an equal number of negative samples from all unknown microbe-disease association pairs. The number of negative samples matched the number of positive samples as to maintain a balanced dataset. In each cross-validation experiment, the Gaussian kernel similarity scores for microbes and diseases are recalculated based on the training set to ensure the effectiveness of evaluating model performance through the test set. In this experiment, we employ the same dataset and follow the hyperparameter settings used in the original papers or provide open source codes for other compared models. We adopted the same evaluation metrics as the previous study (Jiang et al., 2022), including the area under the ROC curve (AUC) and the area under the precision-recall curve (AUPR) to assess the performance of the models. To evaluate the performance of these models in predicting potential associations between microbes and diseases, this study conducted 10 repetitions of five-fold cross-validation experiments and 10 repetitions of ten-fold cross-validation experiments by setting different random seeds, and then computed the average to ensure the accuracy of our results.

The comparative results of the two cross-validation experiments conducted on the HMDAD dataset for the five aforementioned models are presented in Table 1. The optimal performance is highlighted in bold, with standard deviations indicated in parentheses. To provide readers with a clearer visualization of the performance of the models, this study further plotted the ROC curve and PR curve, as shown in Figures 3, 4, respectively.

From the experimental data presented above, it can be observed that the GCATCMDA model proposed in this study has achieved excellent predictive performance in the task of predicting associations between microbes and diseases, surpassing methods proposed in previous studies. For instance, in the five-fold cross-validation experiment, the model obtained an approximate 1.3% improvement in AUC compared to the best previous predictive performance. Similarly, in the ten-fold cross-validation experiment, the model obtained an approximate 1.0% enhancement in AUC compared to the best previous predictive performance. The improvement in predictive performance was slightly more pronounced in the fold-fold cross-validation compared to the 10-fold cross-validation. This can be attributed to the larger training sets used in the 10-fold validation, which reduce variability across folds and provide more comprehensive data for model training. However, the reduced variability can lead to subtler improvements in performance metrics, as the model benefits from a more stable but less varied dataset. In contrast, the five-fold validation, with its larger test sets, introduces more variability, making performance improvements more apparent.

Graph transformer models offer strong capabilities in capturing global node features through their self-attention mechanisms (Li et al., 2024a,b). This allows them to handle complex and non-local structures, which can be beneficial for highly heterogeneous datasets. However, these models come with significant computational complexity, scaling quadratically with the number of nodes, making them less practical for large datasets like microbe-disease networks.

While the GCATCMDA model combines GCN and GAT to effectively capture both local features and selective attention on relevant neighbors, Graph Transformer models are designed to capture these relationships on a broader scale. The full attention mechanism of Graph Transformers allows them to dynamically weigh the importance of distant nodes, offering more flexibility in feature extraction across large and complex networks. In contrast, our GCATCMDA model, which combines GCNs and GATs, is more computationally efficient and particularly suited to smaller, sparser datasets like the HMDAD database. While graph transformers excel in capturing global relationships, our approach balances local feature aggregation and attention, offering a more efficient solution. Future work could explore integrating graph transformers to leverage their global feature-capturing capabilities alongside our model's efficiency in handling localized data.

The GCATCMDA model proposed in this study possesses several crucial parameters, such as the Gaussian kernel similarity threshold t for constructing microbe and disease similarity networks, the dimensionality F of node features, the number of network layers L for graph convolution, and the number of attention heads heads for the graph attention mechanism. Therefore, this study conducted training with different parameter combinations on the HMDAD dataset and utilized the experimental results from 10 repetitions of five-fold cross-validation to analyze the impact of these parameters on the model's performance.

As shown in Figure 5, the model fails to achieve the best predictive performance when the Gaussian kernel similarity threshold t is either set too high or too low, the optimal predictive performance of the model is attained when t = 0.4. Moreover, as the dimensionality of node features increases, the predictive performance of the model gradually improves, with the best performance observed when F = 128. Additionally, the model exhibits its best predictive performance when the number of network layers for graph convolution L = 3. Furthermore, it is observed that the evaluation metrics AUC and AUPR attain their maximum values when the number of attention heads for the graph attention mechanism heads = 2.

To further validate the impact of each module in the GCATCMDA model on the prediction performance of microbe-disease associations, this study conducted ablation experiments on the HMDAD dataset. The evaluation metrics included AUC, AUPR, Precision, Recall, and F1 score. These metrics aimed to comprehensively analyze the influence of different modules on the performance of the GCATCMDA model. The experimental results represent the average scores of 10 repetitions of five-fold cross-validation experiments. Initially, given that the GCATCMDA model simultaneously utilizes microbe similarity networks, disease similarity networks, and microbe-disease association networks to learn the feature representations of microbes and diseases, this study assessed the impact of node features from different association networks on the model's prediction performance. The experimental results are illustrated in Figure 6, where GCATCMDA_sim denotes learning the feature representations of microbes and diseases only from microbe and disease similarity networks, while GCATCMDA_asso denotes learning the feature representations only from the microbe-disease association network. It can be observed from Figure 6 that integrating feature representations of microbes and diseases from different association networks effectively enhances the model's predictive performance.

Next, given that the GCATCMDA model mainly consists of GCN, GAT, feature dual fusion module, and contrastive learning module, this study attempted to remove each module individually to investigate the impact of different modules on the model's prediction performance. The experimental results are presented in Table 2, where “GCATCMDA_GCN” denotes the removal of the graph convolutional network from the original model, “GCATCMDA_GAT” denotes the removal of the graph attention mechanism, “GCATCMDA_SUM” denotes replacing the feature dual fusion module of the original model with a simple summation operation, and “GCATCMDA_CL” denotes the removal of the contrastive learning module from the original model. From the results in Table 2, it can be observed that both “GCATCMDA_GCN” and “GCATCMDA_GAT” exhibit lower predictive performance compared to the original model, indicating that the integration of graph convolutional networks and graph attention mechanisms for node feature learning is effective in obtaining more informative node feature representations from the network. The predictive performance of “GCATCMDA_SUM” is also lower than that of the original model, suggesting that the designed feature dual fusion module effectively fuses node feature information outputted by the graph attention layers. Similarly, the predictive performance of “GCATCMDA_CL” is slightly lower than that of the original model, indicating that the addition of the contrastive learning module can improve the model's predictive performance to some extent.

Finally, to investigate the impact of different operations for fusing node features outputted by the graph attention layers on the GCATCMDA model prediction performance, this study sophisticatedly combined three common feature vector operations: concatenation, sum, and element-wise product. The combined fusion feature formulas are similar to the feature dual fusion formula described earlier. The experimental results are illustrated in Figure 7. GCATCMDA_C represents the use of concatenation only, GCATCMDA_S represents the use of sum only, and GCATCMDA_H represents the use of element-wise product only. GCATCMDA_CS represents the combination of concatenation and sum, GCATCMDA_CH represents the combination of concatenation and element-wise product, GCATCMDA_SH represents the combination of sum and element-wise product, and GCATCMDA_CSH represents the combination of concatenation, sum, and element-wise product. From the experimental results in Figure 7, it can be observed that selecting the combination operations of concatenation and element-wise product in the feature dual fusion module can most effectively fuse node features outputted by the graph attention layers.

To further validate whether the GCATCMDA model can predict associations between microbes and diseases, this study initially trained the model using all known microbial-disease associations in the HMDAD dataset. Subsequently, obesity and inflammatory bowel disease (IBD), two common diseases, were selected as subjects for case analysis. The model predicted microbial associations with obesity and IBD by sorting the predicted association scores from high to low and retaining the top 20 unknown microbial associations with high scores for these two diseases. Finally, employing a literature search approach, this study validated whether these microbial associations with diseases existed by examining relevant publications in the biomedical literature database PubMed. This validation process aimed to assess the accuracy of the microbial-disease associations predicted by the GCATCMDA model.

From Table 3, it can be observed that among the top 20 associated microbes identified by the GCATCMDA model for obesity, 16 of them have been previously documented in the literature to be associated with obesity. For instance, Xu et al. (2022), by reviewing literature on gut microbiota and obesity, identified an association between Prevotella and obesity. Baradaran et al. (2021) experimentally demonstrated that individuals positive for Helicobacter pylori infection are more likely to suffer from obesity, with an increased risk of Helicobacter pylori infection among obese individuals. From Table 4, it can be observed that in IBD, among the top 20 associated microbes identified by the GCATCMDA model, 15 have been previously demonstrated to be associated with IBD in the literature. For example, Quaglio et al. (2022) demonstrated that the abundance of Bacteroidetes and Firmicutes in patients with IBD undergoes significant changes. Cardoneanu et al. (2021) experimental research showed a significant decrease in the abundance of Clostridium coccoides in patients with IBD compared to healthy individuals.

In summary, it can be observed from Tables 3, 4 that the GCATCMDA model achieves an accuracy of over 75% in predicting potential associated microbes for both obesity and inflammatory bowel disease. Therefore, this study concludes that the GCATCMDA model can provide effective and accurate candidate sets of microbes associated with diseases, thereby reducing the research costs and duration of traditional biological experiments.