A lightweight single-view contrastive learning hypergraph neural network for food–microbe–disease association prediction

Abstract

Background

Identifying potential associations among food, gut microbiota and disease is fundamental for elucidating interaction mechanisms and advancing personalized healthy dietary strategies. While computational methods have been extensively applied to predict microbiota–disease associations, methods on predicting food–microbiota relationships remain limited, particularly regarding higher-order food–microbiota–disease interactions.

Results

In this work, we construct a food–microbe–disease (FMD) database encompassing 190 food items, 219 gut microbiota species, and 163 disease entities, resulting in 17,065 FMD associations. We then propose a lightweight single-view contrastive learning hypergraph neural network (LSCHNN) for FMD association prediction on the sparse FMD dataset. LSCHNN formulates ternary FMD interactions as a hypergraph, in which foods, microbes, and diseases are represented by nodes and FMD triplets are represented by hyperedges, and leverages the biological features of foods, microbes, and diseases as node attributes. Subsequently, a hypergraph neural network is designed to learn the embeddings of foods, microbes, and diseases from the hypergraph and predict potential ternary FMD associations. Additionally, we incorporate a single-view contrastive learning mechanism that enhances the model’s ability to extract discriminative features and improves generalization on sparse data. Comprehensive comparison experiments demonstrate that LSCHNN outperforms other state-of-the-art methods in terms of the precision of predicting ternary FMD associations and discovering more potential FMD associations. Case studies on two microbes further confirm the effectiveness of LSCHNN in identifying potential FMD associations.

Conclusions

A novel computational model, LSCHNN, is proposed, marking the first integration of hypergraph neural networks with lightweight single-view contrastive learning for ternary FMD association prediction, providing a groundbreaking framework for precision nutrition and personalized dietary interventions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-025-06283-1.

Background

The human gut harbors a complex microbial ecosystem, composed of approximately 100 trillion microorganisms, collectively known as the “gut microbiota” [1]. These microorganisms are essential for maintaining intestinal homeostasis and promoting host health through the production of various metabolites that influence intestinal barrier function and immunity [2]. Recent advances in gut microbiome research have identified diet as a primary determinant of gut microbiota composition. For instance, high-fat food consumption increases the abundance of Bacteroides and Alistipes while reducing Faecalibacterium [3] populations. Moreover, disruptions in gut microbial communities are strongly associated with numerous pathological conditions, including cardiovascular and neurodegenerative diseases [4–6], underscoring the importance of maintaining a balanced microbial ecosystem. Therefore, understanding the intricate interactions among food, gut microbiota, and disease is essential for developing personalized nutritional strategies to enhance health outcomes.

Existing studies have extensively investigated associations between gut microbiota and disease entities, predominantly employing traditional graph neural networks (GNNs) such as graph attention networks (GATs) and graph convolutional networks (GCNs) to predict pairwise associations involving gut microbiota [7–15]. For example, Long et al. developed a GCN-based model for predicting human microbe–drug associations [16], and Lu proposed an innovative graph contrastive learning model incorporating sparse relation augmentation and a cascaded multi-kernel fusion network for microbe–disease association prediction [17]. Additionally, Ha constructed a graph convolutional neural network with neural collaborative filtering for effectively predicting disease-related miRNAs [18] and a deepwalk-based graph embedding method for predicting miRNA-disease associations [19]. These graph neural networks have demonstrated considerable accuracy in predicting potential associations between gut microbiota-related bipartite entities [20]. However, traditional graph networks are inherently constrained by their pairwise relationships, which restrict their ability to capture the higher-order complexity of microbiological systems, such as food–microbe–disease associations, where interactions frequently involve three or more entities simultaneously.

Basically, understanding food–microbe–disease interactions has become critical for advancing personalized dietary interventions. However, computational models capable of predicting ternary FMD associations remain absent from current research. To bridge this gap, hypergraphs offer a powerful framework by extending traditional graphs through hyperedges, which connect multiple entities simultaneously, thereby capturing higher-order relationships more effectively. In hypergraphs, vertices represent entities, and hyperedges represent higher-order associations among multiple entities [21]. Due to their flexibility and expressive power, hypergraph-based approaches have been widely adopted in association prediction and recommendation systems, where they excel at modeling multi-layered data structures. Therefore, this study proposes hypergraph-based modeling to decode intricate FMD interactions, enabling the prediction of potential ternary associations and establishing a computational foundation for personalized dietary strategies.

Specifically, we first compile an FMD dataset from available literature sources, comprising 190 food types, 219 gut microbiota species, 163 diseases, and 17,065 unique associations that integrate food-microbiota and microbiota–disease relationships. This dataset provides a foundation for investigating complex FMD associations. However, the dataset remains sparse, limiting model generalizability and predictive accuracy. To address this challenge, we propose a lightweight single-view contrastive learning hypergraph neural network (LSCHNN) for FMD association prediction. Our framework incorporates two key innovations to overcome the difficulties associated with predicting potential ternary associations. First, we utilize a hypergraph-based representation to model high-order interactions among food, gut microbiota, and diseases. Unlike traditional graph models restricted to pairwise relationships, LSCHNN employs hyperedges to simultaneously connect multiple entities, thus capturing the complex, multi-dimensional associations within the FMD network. This enables LSCHNN to better characterize the intricate interactions between food, microbiota, and diseases. Second, we integrate a single-view contrastive learning mechanism to enhance feature discriminability. Compared to traditional multi-view contrastive learning, which is susceptible to noise and overfitting in sparse data, our single-view approach focuses on learning discriminative features from one perspective. This significantly enhances LSCHNN’s ability to generalize, reducing the impact of noisy data and improving predictive performance on sparse FMD data. Together, these innovations allow LSCHNN to effectively address the limitations posed by data sparsity and high-dimensional complexity, making it a powerful method for predicting ternary FMD associations.

In summary, this work presents a comprehensive solution to the challenge of predicting potential ternary FMD associations. Our work delivers four key advancements beyond the state-of-the-art:

• We firstly constructed a comprehensive, large-scale FMD database, providing 17,065 unique associations among food, gut microbiota, and disease. This extensive resource facilitates the exploration of complex, high-dimensional interactions in future dietary and microbiome research.

• Pioneering a novel ensemble framework utilizing hypergraph neural networks to predict potential ternary FMD associations, we overcome the pairwise limitation of GCNs. To our knowledge, this is the first investigation of hypergraph methodology to address this significant challenge.

• Incorporating a lightweight single-view contrastive learning with microbiota-level negative sampling, our framework significantly reduces noise and improves predictive performance by 8.91% in terms of AUPR compared to multi-view methods.

• Extensive experimental validation confirms LSCHNN’s effectiveness, demonstrating consistent superiority over traditional methods. Our approach not only addresses sparse data challenges but also delivers more accurate and interpretable predictions of food-gut microbiota–disease relationships, establishing a foundation for personalized dietary interventions.

Materials and methods

Datasets

Data sources

While numerous gut microbiota databases, such as MicroPhenoDB and HMDAD [22, 23], exist to catalog microbes and their interactions with diseases or interventions (such as pharmaceuticals and surgical procedures), publicly available resources documenting food-gut microbiota or food-gut microbiota–disease relationships remain scarce. To address this gap, we collected literature data from PubMed through May 2024 using keywords including “intestinal flora,” “fecal microbiota,” “gut microbiota,” “intestinal microbiota,” “disease,” “dietary pattern,” “dietary components,” “dietary patterns,” “nutrients,” “foods,” and related terms. Our search yielded 584 studies investigating gut microbiota–disease relationships and 314 studies exploring gut microbiota-food connections. These literature types include observational studies, randomized controlled trials, pathological reports, and so on. Following the initial screening, we removed 15 duplicate papers from the dataset.

Selection criteria

We implemented a rigorous two-stage screening process, applying strict quality controls to ensure the inclusion of only the most reliable and relevant studies. The initial phase involved a systematic review of titles and abstracts to identify potentially eligible articles for database inclusion. The subsequent evaluation entailed manual extraction of conclusive statements, with particular attention to methodological rigor and statistical validity of reported results.

Inclusion criteria were established through interdisciplinary consensus, ensuring a focus on studies that best represent generalizable patterns in food-gut microbiota–disease interactions. For microbiota-host relationship studies, we retained only large-scale cohort investigations to minimize confounding from individual variability, recognizing the sensitivity of microbial profiling to sample size effects. For dietary intervention trials, the literature was limited to oral administration protocols (dietary patterns, specific foods/nutrients) without concurrent pharmacological or surgical interventions. We prioritized studies reporting statistically significant microbial alterations (p < 0.05) explicitly in their concluding sections and excluded those that performed responder/non-responder subgroup analyses, as these could introduce biases.

Our conflict resolution protocols prioritized evidence from peer-reviewed journals demonstrating population generalizability, ensuring that the database captured prevalent host-microbiota interaction patterns rather than individual-specific or extreme-case phenomena. As a result, the final curation yielded 71 articles documenting gut microbiota-food relationships and 115 articles describing gut microbiota–disease associations, forming the foundation of our database for subsequent computational modeling.

Preprocessing steps

We initially established separate microbe–Disease and microbe–Food databases using a unified data format to ensure interoperability. Despite integrating all related databases, associations between food, microbes, and diseases remained sparse due to the limited overlap of gut microbiota species across databases. This necessitated manual extraction of disease–microbe associations from published sources. To enhance coverage, we incorporated the Gold-Standard Corpus (GSC) [24], a manually annotated dataset of high-confidence microbe–disease interactions curated from biomedical literature, into our microbe–disease database. GSC contains 1028 relationships involving 165 diseases and 355 gut microbiota species. The microbe–Food database comprises two components: the microbe–Dietary database (4062 relationships involving 214 gut microbiota species and 49 foods) and the microbe–Nutrients database (7570 relationships involving 177 gut microbiota species and 44 nutrients). Table 1 summarizes the details of these database components.

Table 1.

The components of the FMD databases

Database		Relationship	Disease/food/nutrient	Microbe
Microbe–disease		1028	165	355
Microbe–food	Microbe–dietary	4062	49	214
	Microbe–nutrients	7570	44	177

To address variations in data storage formats and inconsistencies in microbiota and disease nomenclature across databases, we employed Medical Subject Headings (MeSH) from the National Library of Medicine and the Taxonomy database from the National Center for Biotechnology Information for standardization. For example, we used “TAXID:1579” to standardize Lactobacillus acidophilus and “D065766” to standardize Atypical Hemolytic Uremic Syndrome from different literature sources. To facilitate the representation of high-order interactions among entities, we structured the data in a standardized format (< Food, Microbe, Disease>), yielding 17,065 unique FMD triples after deduplication. This comprehensive dataset encompasses 190 foods, 219 gut microbiota species, and 163 diseases. Notably, the collected FMD associations represent only 0.2516% of all possible FMD combinations (190 × 219 × 163), highlighting the significant sparsity of the overall dataset.

During the construction of the FMD database, we conducted a detailed examination of its potential biases. We found that for all potential sources of bias, including literature sources, screening criteria, manual extraction, data cleaning, and data integration, the procedures employed in building the FMD database complied with established standards. Consequently, the quality of the FMD database also meets the corresponding standards.

Methods

This section details the architecture of LSCHNN for FMD association prediction. As illustrated in Fig. 1, LSCHNN comprises four sequential components. First, a hypergraph is constructed to capture the complex, high-order associations among food, gut microbiota, and disease entities, establishing an informative foundation for subsequent representation learning. Next, a single-view contrastive learning approach is implemented to enhance the discriminative power of learned representations. This approach incorporates gut microbiota-level perturbations, which are particularly beneficial for sparse, high-dimensional data. By focusing on a single view, the model avoids the challenges of noise and overfitting that can arise in multi-view contrastive learning, making it more effective in extracting relevant features from the data. Subsequently, the third component involves hypergraph convolutional layers that encode the high-order interactions among entities. These layers leverage the hypergraph structure to improve the accuracy of the learned representations, capturing the intricate relationships between food, gut microbiota, and diseases in a more sophisticated manner. Finally, a decoder constructed using a Multilayer Perceptron (MLP) maps the extracted high-order interaction features into FMD association prediction scores. These scores represent the likelihood of specific interactions between food, microbiota, and disease entities. Based on these scores, an end-to-end optimization process is employed to adjust the network parameters during training, ensuring high-precision modeling for FMD association prediction.

Fig. 1.

Workflow of LSCHNN. A FMD hypergraph construction, B Single-view contrastive learning via gut microbiota-level negative sampling, C Hypergraph representation learning, D FMD association prediction

FMD hypergraph construction

Given a food set , a microbiota set , and a disease set , their Cartesian product represents all possible ternary FMD associations. For each triplet , we assign a binary label , where indicates confirmed food–microbe and microbe–disease associations, and otherwise. Importantly, does not necessarily indicate the absence of association but rather represents an unknown relationship that is potentially a valid association that remains undiscovered.

To model these complex high-order FMD associations, we construct a hypergraph , where nodes represent food, gut microbiota, or disease entities, and hyperedges correspond to known FMD associations. These hyperedges specifically represent triplets labeled as 1, where . Formally, is structured as an attributed hypergraph comprising an adjacency matrix and node attributes , with serving as the feature matrix.

The adjacency matrix stores hyperedges with entries defined as:

1

Node attributes encompass food features , microbe features , and disease features . We process the similarity matrices of nodes through the Random Walk with Restart (RWR) [16] method and transform them into the feature matrix. Finally, we convert it into node features through a fully connected layer.

For microbe nodes, we compute the microbe Gaussian Interaction Profile (GIP) kernel similarities as the microbe similarity matrix . The underlying principle of GIP similarity is that microbes associated with functionally similar diseases (or foods) tend to exhibit semantic similarity. Following the approach in reference [25], we construct two binary vectors and documenting associations between microbe and all foods and diseases in the FMD dataset, respectively. The corresponding value equals 1 when a supported association exists between the two entities and 0 otherwise. Consequently, the GIP kernel similarity between microbes and is defined as:

2

3

where and represent the normalized kernel bandwidth, calculated as:

4

5

Here, indicates the total number of microbes, and is set to 1.

Finally, we formulate the integrated microbe similarity matrix by combining derived from microbe–food associations and derived from microbe–disease associations to generate the final disease GIP kernel similarity calculated as:

6

For food nodes, we compute the food GIP kernel similarities as the food similarity matrix . The GIP kernel similarity between foods and is defined as:

7

8

where is a binary vector documenting associations between food and all microbes in the FMD dataset, represents the normalized kernel bandwidth, indicates the total number of foods, and is set to 1.

For disease nodes, we integrate disease GIP kernel similarity with disease semantic similarity to create a comprehensive disease similarity matrix . Likewise, the GIP kernel similarity between diseases and is defined as:

9

10

where is a binary vector documenting associations between disease and all microbes in the FMD dataset, represents the normalized kernel bandwidth, indicates the total number of diseases, and is set to 1.

Each disease is represented by a unique directed acyclic graph from the Medical Subject Headings (MeSH) ontology. Specifically, we characterize a disease using , where comprises the disease itself along with all its ancestor diseases, and consists of direct edges from parent (general) terms to child (specific) terms. The semantic contribution of disease to is calculated as:

11

where represents a factor affecting the semantic contribution between the parent node and its child node . In our experiment, all ancestor nodes in the directed acyclic graph of disease are assigned a semantic contribution factor of 0.5 to [26], while contributes a semantic contribution factor of 1 to itself. The semantic value of disease is calculated based on contributions from both ancestral diseases and the disease itself:

12

For any given disease , nodes further from the disease contribute less semantically, while nodes at identical hierarchical levels contribute equally. These semantic values are compiled into a disease semantic similarity matrix , where the semantic similarity between diseases and is calculated as:

13

Finally, we formulate the integrated disease similarity matrix by combining the disease GIP kernel similarity matrix and the disease semantic similarity matrix . Thus, is defined as:

14

As previously noted, , , and represent food, microbe, and disease similarity matrices, respectively. However, directly using these similarity matrices as node feature matrices is suboptimal due to potential noise from false positives and methodological limitations. Therefore, we employ an RWR-based [16] approach to extract refined features from these similarity matrices, effectively capturing both local and global topological network characteristics while reducing noise. RWR is a widely adopted network-based method for noise reduction in image processing and preservation of neighborhood information in feature learning [27]. Formally, RWR is defined as:

15

where denotes the transfer probability matrix, represents the restart probability, and is the initial probability vector for the -th node. If , then ; otherwise, . The vector indicates the probability of reaching other nodes from the -th node at time , and we adopt as the feature vector for the -th node at steady state. We apply RWR to process the similarity matrices of diseases, microbes, and foods to generate their respective feature matrices.

Single-view contrastive learning via gut microbiota-level negative sampling

The sparse associations in the FMD database constrain the ability of the hypergraph convolutional network to learn comprehensive node representations, particularly for low-degree nodes, thereby affecting our model’s overall performance in predicting potential FMD associations. To address this limitation, we incorporate the graph contrastive learning technique within the Deep Graph Infomax (DGI) framework, which maximizes mutual information between node-level and graph-level representations. This enhancement enables the model to identify node representations that better align with underlying graph structures.

For implementing graph contrastive learning on FMD hypergraphs within the DGI framework, we first generate negative samples by corrupting the hypergraph, which can be achieved through either multi-view or single-view contrastive learning. For instance, in utilizing the hypergraph neural network to predict drug-gut microbiota–disease associations, Liu et al. employed a multi-view contrastive learning technique to perturb the hypergraph for generating negative samples [28]. Specifically, they designed four perturbation strategies: drug-level perturbation, microbiota-level perturbation, disease-level perturbation, and random perturbation, to disrupt each triplet and obtain one or multiple “fake” triplets not present in the original hyperedge set . After that, they can maintain “fake” hyperedge sets and construct the corresponding negative hypergraph , where Z is the number of views they consider. Then, the augumented multi-view node embeddings are further computed from these perturbed hypergraphs through the same encoder as HGCN deployed on the original hypergraph. is the z-th counterpart of the node representation . The objective of the multi-view contrastive learning task is formulated as:

16

where serves as the positive example while ( functions as the negative example. is the graph-level representation obtained through a global mean pooling layer: .And denotes the contrastive discriminator constructed by a simple bilinear function that estimates similarities between node-level and graph-level representations.

However, Liu et al. overlooked a critical aspect: unlike standard DGI on bipartite graphs where edge perturbation is straightforward, perturbing hyperedges in the hypergraph requires careful consideration due to the complex, high-order interactions they represent. Traditional multi-view contrastive learning introduces noise and generalization errors when applied to large-scale data, while simultaneously increasing computational costs and memory requirements [29]. Compared to multi-view approaches, single-view contrastive learning reduces generalization errors by implementing negative view sampling for specific scenarios while maintaining comparable alignment quality. Additionally, single-view contrastive learning proves more cost-effective and computationally efficient in terms of both memory utilization and processing demands. We therefore employ single-view negative hypergraph contrastive learning in our model.

Specifically, for each hyperedge (or triple-wise association < F, M, D>) in the FMD hypergraph, we generate a “fake” triplet not present in the original hyperedge set . We then form the “false” hyperedge set , and construct the corresponding negative hypergraph . Our perturbation scheme generates a negative hypergraphic view through gut microbiota-level perturbation: we disconnect the gut microbiota m from the food-disease pair (f, d) and connect (f, d) to another gut microbiota to create the ‘fake’ hyperedge (f, m’, d). The augmented single-view node embeddings are then computed from these perturbed hypergraphs using the same hypergraph convolutional network (HGCN) encoder deployed on the original hypergraph, resulting in as the single-view node representation within the set of single-view node representations .

Following the DGI approach, serves as the positive example while ( functions as the negative example, where represents the graph-level representation obtained through a global mean pooling layer: . The objective of our single-view CL task is thus formulated as:

17

where denotes the contrastive discriminator constructed by a simple bilinear function that estimates similarities between node-level and graph-level representations.

Hypergraph representation learning

To encode high-order interactions within FMD associations and leverage biological domain knowledge embedded in node attributes, we implement an HGCN on our constructed FMD hypergraph . The HGCN updates node embeddings by aggregating information from connected nodes via hyperedges, thereby capturing complex interactions. The first step of hypergraph convolution is specifically formulated as [30]:

18

where represents a nonlinear activation function, is a learnable weighted matrix, denotes the node embedding at the l-th layer and is initialized with , represents the degree matrix of the vertices, indicates the degree matrix of the hyperedges, is the adjacency matrix of the hypergraph, and is the diagonal matrix storing hyperedge weights. We denote the final learned node representation as , and for a given node , its representation is extracted from .

FMD association prediction

For predicting FMD associations, our model utilizes the learned embeddings of food , microbe , and disease to generate the association probability through a decoder implemented as a multilayer perceptron unit:

19

where is a 4-layer perceptron with sigmoid activations, and is the concatenation operator.

The loss of the supervised prediction task is formulated as:

20

where represents the training sets and denotes the true label.

During training, we jointly optimize the model using both the contrastive learning objective and the supervised prediction loss:

21

where serves as a hyperparameter balancing the contrastive and supervised loss components.

Results

Experimental setup

In our experimental framework, we employed established FMD associations as positive samples and generated negative samples (labeled as 0) through gut microbiota-model perturbations. Specifically, during training, one negative triplet was generated for each observed positive triplet. During testing, each positive triplet was paired with 29 negative triplets. This differential sampling approach mitigates model overfitting while enabling a robust evaluation of discriminative performance. We partitioned the FMD dataset into training (90%) and test (10%) sets. For the 5-fold Cross-Validation (5-fold CV), the training set was randomly shuffled and split into 5 mutually exclusive groups. During each round, four groups of FMD associations with the same number of negative triplets were used to train the model, and the remaining group with a set of negative triplets that was 29 times larger was used for validation. For the independent test, we used the training set to train LSCHNN and the test set to test LSCHNN’s predictive performance, adopting the same negative sample generation scheme. Then, the performance is evaluated by eight metrics, including the Area Under the Receiver Operating Characteristic Curve (AUC), the Area Under the precision–recall Curve (AUPR), Accuracy, Precision, F1-score, Recall, Hit Ratio (HR), and Normalized Discounted Cumulative Gain (NDCG). These metrics are defined by Eq. (22) to (26).

22

23

24

25

26

where TP represents instances that are genuinely positive and correctly predicted as positive by the model. TN indicates instances that are genuinely negative and correctly predicted as negative by the model. FP represents instances that are genuinely negative but erroneously predicted as positive by the model. FN shows instances that are genuinely positive but erroneously predicted as negative by the model. The number of samples is denoted by , which can be interpreted as the number of user requirements. indicates whether the -th requirement item appears in the item list recommended by LSCHNN, with a value of 1 if included and 0 if excluded. denotes the position of the -th requirement item in the LSCHNN-recommended item list. For the -th requirement item not included in the recommended list, .

To ensure our method LSCHNN is not overfit to the FMD dataset, independent testing was conducted. Table 2 presents LSCHNN’s performance through 5-fold CV and independent testing on the FMD dataset.

Table 2.

The LSCHNN performance through 5-fold CV and independent testing on the FMD dataset

5-fold cross-validation
Fold	AUC	AUPR	Accuracy	Precision	F1-score	Recall	Hits@1	NDCG@1	Hits@3	NDCG@3	Hits@5	NDCG@5
1	0.9912	0.9122	0.9897	0.7844	0.8607	0.9535	0.9444	0.9444	0.9824	0.9670	0.9876	0.9692
2	0.9940	0.9336	0.9920	0.8415	0.8857	0.9349	0.9541	0.9541	0.9883	0.9747	0.9909	0.9757
3	0.9926	0.9400	0.9908	0.7968	0.8751	0.9704	0.9496	0.9496	0.9844	0.9709	0.9899	0.9731
4	0.9928	0.9228	0.9879	0.7448	0.8426	0.9701	0.9466	0.9466	0.9860	0.9706	0.9896	0.9721
5	0.9923	0.9151	0.9909	0.8683	0.8625	0.8567	0.9515	0.9515	0.9876	0.9734	0.9906	0.9746
Average.	0.9926	0.9247	0.9918	0.8072	0.8653	0.9371	0.9492	0.9492	0.9857	0.9713	0.9897	0.9730

Independent testing
Metrics	AUC	AUPR	Accuracy	Precision	F1-score	Recall	Hits@1	NDCG@1	Hits@3	NDCG@3	Hits@5	NDCG@5
Values	0.9961	0.9412	0.9918	0.8159	0.8880	0.9742	0.9566	0.9566	0.9906	0.9774	0.9941	0.9789

Comparison with other methods

As FMD association prediction is a new problem, few computational approaches have been presented for this important task. We compare our method LSCHNN with five state-of-the-art methods that were proposed for different link prediction problems in the field of computational biology.

• EGATMDA [31] is an ensembling graph attention network with a hierarchical attention mechanism for human microbe–drug association prediction.

• LAGCN [32] is a layer attention graph convolutional network for the drug-disease association prediction.

• GCNMDA [16] is a graph convolutional network with a conditional random field for human microbe–drug association prediction.

• NAGTLDA [33] is a node-adaptive graph transformer method with structural encoding for IncRNA-disease association prediction.

• MCHNN [28] is a multi-view contrastive learning hypergraph neural network for drug–microbe–disease association prediction.

For a scientifically rigorous evaluation, we ran five state-of-the-art methods with their optimized parameter settings and validated performance via 5-fold CV. For the binary association prediction methods (NAGTLDA, LAGCN, EGATMDA, and GCNMDA), we broke the FMD ternary associations into two binary associations(food–microbe and microbe–disease association), and then we modeled them using these methods separately. Subsequently, we integrated the two sets of pairwise predictions based on shared microbes to reconstruct ternary FMD association prediction and selected the minimum between the two predicted values as the final prediction value. In contrast, MCHNN, which inherently supports ternary interaction modeling, was directly applied to the FMD dataset without decomposition to perform end-to-end ternary FMD association prediction. As shown in Table 3, LSCHNN outperforms all comparison methods in terms of all six evaluation metrics. Compared with the GNN-based methods (EGATMDA, LAGCN, GCNMDA, NAGTLDA), the hypergraph neural network (HGNN)-based methods (LSCHNN and MCHNN) achieve 37.03% (p < 0.001) and 56.74% (p < 0.001) improvement on average over them in terms of AUC and AUPR. This performance advantage stems from HGNN’s inherent capability to effectively model ternary interactions through hyperedges, whereas the conventional GNN-based methods struggle to transcend binary association modeling due to the edge-centric paradigm when predicting ternary FMD associations. Specifically, the decomposition of ternary FMD associations into disjoint food–microbe and microbe–disease associations inevitably discards critical synergistic information embedded in the original ternary FMD structure and introduces error propagation during prediction assembly. In contrast, HGNN’s hypergraph architecture preserves the holistic ternary information through hyperedges, enabling direct learning of nonlinear interaction information among food, microbe, and disease entities. Compared with the multi-view contrastive learning HGNN-based method MCHNN, our method LSCHNN outperforms MCHNN in all evaluation metrics. Specifically, on the FMD dataset, LSCHNN surpasses MCHNN in AUC (0.71%, p < 0.05), AUPR (8.91%, p < 0.05), Accuracy (1.44%, p < 0.01), Precision (17.24%, p < 0.01), F1-score (12.46%, p < 0.01), and Recall (4.67%, p = 0.3361). This demonstrates that our microbiota-perturbed single-view contrastive learning approach is a biologically grounded strategy for generating negative samples by selectively disconnecting microbes. This focuses learning on discriminative features, reducing generalization error, and improving LSCHNN’s predictive performance on the sparse FMD dataset.

Table 3.

Performance of LSCHNN and other baselines on the FMD dataset by 5-fold CV

Method	AUC	AUPR	Accuracy	Precision	F1-score	Recall
LSCHNN	0.9926 ± 0.0010	0.9247 ± 0.0119	0.9903 ± 0.0015	0.8072 ± 0.0486	0.8653 ± 0.0162	0.9371 ± 0.0472
EGATMDA	0.7144 ± 0.0574	0.2300 ± 0.0652	0.5991 ± 0.0261	0.1933 ± 0.0531	0.3019 ± 0.0744	0.6992 ± 0.1122
LAGCN	0.5353 ± 0.0267	0.4284 ± 0.0404	0.4006 ± 0.0354	0.3023 ± 0.0290	0.4357 ± 0.0299	0.7860 ± 0.0475
GCNMDA	0.4619 ± 0.0417	0.3138 ± 0.0444	0.5175 ± 0.0264	0.2828 ± 0.0260	0.3282 ± 0.0206	0.3964 ± 0.0506
NAGTLDA	0.7554 ± 0.0277	0.2789 ± 0.0531	0.8750 ± 0.0075	0.4279 ± 0.1464	0.0876 ± 0.0399	0.0495 ± 0.0238
MCHNN	0.9855 ± 0.0051	0.8356 ± 0.0620	0.9759 ± 0.0042	0.6347 ± 0.0369	0.7407 ± 0.0434	0.8904 ± 0.0646

To comprehensively compare the generalization performance of LSCHNN and MCHNN in predicting ternary associations, we implemented both models on the FMD and Drug–microbe–Disease [28] (DMD) datasets, and evaluated their predictive capabilities through 5-fold CV tests. The DMD dataset comprises 2763 DMD triple associations involving 270 drugs, 58 microbes, and 167 diseases. As displayed in Fig. 2, LSCHNN surpasses MCHNN in the FMD dataset by AUC (0.71%, p < 0.05), AUPR (8.91%, p < 0.05), Accuracy (1.44%, p < 0.01), F1-score (12.46%, p < 0.01), Hits@1/3/5 (average 8.11%, p < 0.001), and NDCG@1/3/5 (average 10.95%, p < 0.001). Similarly, on the DMD dataset, LSCHNN outperforms MCHNN in six evaluation metrics: AUC (5.77%, p < 0.01), AUPR (16.72%, p < 0.001), Accuracy (0.73%, p < 0.05), F1-score (6.15%, p < 0.05), Hits@1/3/5 (average 3.93%, p < 0.01), and NDCG@1/3/5 (average 4.19%, p < 0.01). In general scenarios, multi-view contrastive learning is expected to enhance model performance. However, this conventional wisdom does not hold for triplet association prediction tasks. We hypothesize that in sparse triplet association datasets, the noise interference introduced through multi-view contrastive learning disproportionately impacts the model’s predictive capability for triplet associations. In contrast, single-view contrastive learning effectively mitigates such noise interference while improving the generalization performance of LSCHNN in predicting triplet associations.

Fig. 2.

Performance of LSCHNN and MCHNN for ternary association prediction on different datasets by 5-fold CV. A Average common parameter comparison of LSCHNN and MCHNN, B Average recommended parameter comparison of LSCHNN and MCHNN

Parameter analysis

Hyperparameters can affect model prediction performance. Therefore, we employed a 5-fold CV to analyze the influence of the following parameters on the prediction results:

Feature embedding dimensions: The embedding dimension of features impacts the dimension of their corresponding learnable parameter matrices. Figure 3A, B illustrate the evaluation index scores for different Bio_embedding dimensions [32, 64, 96] and HGNN_embedding dimensions [256, 512, 1024]. To optimize LSCHNN’s association prediction performance, we set the Bio_embedding dimension for all nodes to 64 and the HGNN_embedding dimension to 512.

Fig. 3.

The parameter analysis for LSCHNN. A Parameter analysis for different Bioencoder embedding dimensions, B Parameter analysis for different HGNN encoder embedding dimensions, C Parameter analysis for different HGNN layers, D Parameter analysis for different learning rates

The number of HGCN layers: Varying the number of HGCN layers affects both the depth and complexity of data extraction, ultimately influencing LSCHNN’s prediction performance. Figure 3C demonstrates the impact of different HGCN layer configurations [2, 3, 4] on LSCHNN performance. Based on these findings, we determined that 3 HGCN layers provide optimal association prediction performance.

The learning rates: The learning rate plays a crucial role in model convergence. An excessively high learning rate may prevent convergence during training, while an overly low rate leads to prohibitively slow convergence. Figure 3D demonstrates how different learning rates [0.001, 0.0001, 0.0005] affect LSCHNN performance. To maximize LSCHNN’s association prediction capabilities, we established 0.0005 as the optimal learning rate.

To validate the discriminative capability of LSCHNN on ternary associations, we performed t-SNE and PCA analyses on the learned embeddings of FMD triplets. Specifically, we visualized the embeddings of positive FMD triplets and negative FMD triplets from the test set after training the model with optimal parameter settings (Bio_embedding dimension = 64, HGNN_embedding dimension = 512, 3 HGCN layers, learning rate = 0.0005). The t-SNE and PCA visualizations (Fig. 4) demonstrate that the embeddings of positive triplets form distinct clusters compared to negative triplets, indicating that LSCHNN effectively captures the underlying patterns of FMD associations.

Fig. 4.

The t-SNE and PCA visualizations of LSCHNN. A The t-SNE visualization, B The PCA visualization

Ablation experiments

There are four key modules in our method LSCHNN: single-view contrastive learning (CL), hypergraph convolution network (HGCN), random walk with restart (RWR), and GIP kernel similarity of food, microbe, and disease (GIP). Here, we conducted an ablation study to investigate the contribution of each component to the LSCHNN model’s performance. We created variants by removing each component individually:

LSCHNN without CL (w/o CL): We remove the contrastive learning objective and optimize the model using the supervised prediction loss , while keeping the rest unchanged.

LSCHNN without HGCN (w/o HGCN): We preserve feature processing, bio_encoder, and decoder, and replace HGCN with fully connected networks while keeping the rest unchanged.

LSCHNN without RWR (w/o RWR): We remove the Random Restart Walk previously employed to process the similarity matrices of food, microbe, and disease, and directly utilize these similarity matrices as feature matrices while keeping the rest unchanged.

LSCHNN without GIP (w/o GIP): We remove the GIP kernel similarity of food, microbe, and disease used in this paper and replace it with a one-hot matrix while keeping the rest unchanged.

As shown in Fig. 5, the findings demonstrate that all components significantly contribute to the model’s predictive capabilities. On the FMD dataset, the complete LSCHNN model achieved scores of 0.9961, 0.9412, 0.9918, and 0.8881 for AUC, AUPR, Accuracy, and F1-score, respectively. Among the ablated variants, LSCHNN w/o RWR showed the second-best performance, while LSCHNN w/o CL exhibited the poorest performance, followed by LSCHNN w/o HGCN. We also plotted the confusion matrices of these models from the ablation study and found that these matrices highly align with their prediction metrics. For example, we found that the w/o CL model achieved an AUC of 0.9789, AUPR of 0.6543, Accuracy of 0.9317, and F1-score of 0.4802. And its confusion matrix indicates that after the removal of single-view contrastive learning, the w/o CL model, which was trained on the extremely imbalanced dataset dominated by “fake” hyperedges (label 0), can identify nearly all positive hyperedges. However, among all predicted positive hyperedges, only a small proportion are actually correct. This suggests that the w/o CL model serves as an effective screening model but poorly as a precise classifier. Similarly, the same reasoning applies to the w/o HGCN model. In contrast, our model that retains single-view contrastive learning and HGCN maintains strong screening performance while significantly improving its predictive accuracy, as reflected in substantially enhanced Precision and F1 scores (as shown in Supplementary Fig. 1–5). Overall, these results highlight that single-view contrastive learning substantially enhances LSCHNN’s ability to handle sparse datasets and improves its learning capacity. Additionally, hypergraph convolution plays a crucial role in the model’s predictive performance.

Fig. 5.

Results from ablation experiments conducted on five different models

Computational performance

To evaluate the computational efficiency of our method LSCHNN, we conducted comparative experiments with five models on the FMD dataset: LSCHNN, its multi-view variants (2/3/4-view CL), and MCHNN using an independent test set. As shown in Fig. 6, LSCHNN demonstrates optimal efficiency, requiring only 6.0477 s per epoch and 0.2820 GB of CPU memory. Both computational time and memory usage generally increase with the number of views. Interestingly, the 3-view variant consumes approximately 5% more memory than the 4-view variant (0.3689 GB vs. 0.3523 GB), which we attribute to inherent stochasticity in deep learning training processes. In contrast, MCHNN requires 58.8185 s per epoch (9.7 times slower than LSCHNN) and 0.4208 GB of CPU memory (1.5 times higher than LSCHNN). In addition, on the DMD dataset, LSCHNN requires 2.9822 s per epoch and 0.1636 GB of CPU memory. In comparison, MCHNN requires 13.8141 s per epoch and 0.1167 GB of CPU memory. These results confirm that our proposed single-view contrastive learning framework offers a lightweight and computationally efficient solution.

Fig. 6.

Computational performance of LSCHNN and MCHNN on the FMD dataset by the independent test set. A The running time (sec/epoch) of LSCHNN, LSCHNN’s multi-view variants (2/3/4-view CL), and MCHNN, B The memory (GB/CPU) of LSCHNN, LSCHNN’s multi-view variants (2/3/4-view CL), and MCHNN

Overall, our model demonstrates high computational efficiency on the DMD dataset and maintains low computational complexity and high computational efficiency on the larger-scale FMD dataset. Therefore, we conclude that LSCHNN is capable of maintaining low computational complexity and good scalability on larger datasets.

Case study

To further validate the reliability of LSCHNN for FMD association prediction in real situations, we conducted case studies focusing on two key gut microbiota: Escherichia coli and Bifidobacterium. For each target microbe, all the known food-target microbe–disease association information will first be excluded. Then, all the candidate foods and diseases will be ranked in a descending manner according to LSCHNN’s predicted scores. Finally, the top 20 ranked associations were verified by published literature, with results presented in Tables 4 and 5.

Table 4.

The top 20 predicted Escherichia coli-associated foods and diseases of LSCHNN

Microbe	Food	Disease	Food–microbe evidence	Microbe–disease evidence
Escherichia coli	Coffee	Hemolytic–uremicre Syndrome	PMID:37194345	PMID:37819955
	Mediterranean Diet	Intraabdominal Infections	PMID:37771255	PMID:29254478
	Coffee	Ileal Diseases	PMID:37194345	PMID:22508665
	Mediterranean Diet	Colitis, Ulcerative	PMID:37771255	PMID:26697577
	Breastfeeding	Ileal Diseases	PMID:28742106	PMID:22508665
	Mediterranean Diet	Irritable Bowel Syndrome	PMID:37771255	PMID:32754068
	Breastfeeding	Intraabdominal Infections	PMID:28742106	PMID:29254478
	Breastfeeding	Inflammatory Bowel Diseases	PMID:28742106	PMID:37771255
	Red meat	Liver Cirrhosis, Alcoholic	PMID:35760086	PMID:37361234
	Red meat	Ileal Diseases	PMID:35760086	PMID:22508665
	Fermented dairy drink	Irritable Bowel Syndrome	PMID:24140320	PMID:32754068
	Red meat	Hamartoma Syndrome, Multiple	PMID:35760086	Unconfirmed
	Tea polyphenols	Colitis, Ulcerative	PMID:33232763	PMID:26697577
	Red meat	Intraabdominal Infections	PMID:35760086	PMID:29254478
	Fermented dairy drink	Colitis, Ulcerative	PMID:24140320	PMID:26697577
	Specific carbohydrate diet	Inflammatory Bowel Diseases	PMID:33291229	PMID:37771255
	Oxidized konjac glucomannan	Colitis, Ulcerative	Unconfirmed	PMID:26697577
	Whole grain	Colitis, Ulcerative	PMID: 37786251	PMID:26697577
	Ganpu tea	Colitis, Ulcerative	Unconfirmed	PMID:26697577
	Low fermentable oligosaccharides, Disaccharides, monosaccharides, and polyols (FODMAP) diet	Inflammatory Bowel Diseases	PMID: 32235316	PMID:37771255

Table 5.

The top 20 predicted Bifidobacterium-associated foods and diseases of LSCHNN

Microbe	Food	Disease	Food–microbe evidence	Microbe–disease evidence
Bifidobacterium	Grains	Diabetes Mellitus	PMID:36737921	PMID:36794003
	Citrus fruits	Diabetes Mellitus	PMID:35685878	PMID:36794003
	Western diet	Diabetes Mellitus	PMID:35071544	PMID:36794003
	Meat	Diabetes Mellitus	PMID:37701742	PMID:36794003
	Mediterranean Diet	Meningioma	PMID:35264213	PMID:35291914
	Dehydrated red wine	Diabetes Mellitus	Unconfirmed	PMID:36794003
	Mediterranean Diet	Major Depressive Disorder 1	PMID:35264213	PMID:29731182
	Grains	Fanconi Anemia	PMID:36737921	Unconfirmed
	Fiber and prebiotics	Major Depressive Disorder 1	PMID:23609775	PMID:29731182
	Fiber and prebiotics	Antibodies, Antiphospholipid	PMID:23609775	PMID:36505427
	Western diet	Fatty Liver	PMID:35071544	PMID:37459922
	flaxseed/fish oil	Fatty Liver	PMID:24555449	PMID:37459922
	Whole grape products	HIV Infections	Unconfirmed	PMID:36604616
	Inulin-type fructans	Fatty Liver	PMID:34555168	PMID:37459922
	Inulin	Crohn Disease	PMID:35381290	PMID:28467461
	fish oil (unsaturated fats)	Fatty Liver	PMID:34291263	PMID:37459922
	Western diet	HIV Infections	PMID:35071544	PMID:36604616
	plant proteins	Fanconi Anemia	PMID:32751533	Unconfirmed
	plant proteins	Fatty Liver	PMID:32751533	PMID:37459922
	walnut	Meningioma	PMID:35262409	PMID:35291914

Escherichia coli (E. coli) is a Gram-negative bacillus that is part of the normal flora but can cause intestinal and extraintestinal illnesses in humans [34]. Previous studies have indicated that E. coli has a close relationship with many foods [35–38]. For example, McConnell et al. [35] indicated that coffee (the 1st candidate food) can significantly inhibit the growth of E. coli. Furthermore, E. coli also plays a crucial role in Hemolytic–uremicre Syndrome (HUS) [39–41], which is predicted by our method to be the 1st candidate disease. Page [40] et al. reported that Enterohemorrhagic E. coli triggers HUS via Shiga toxin binding to endothelial Gb3 receptors, inducing thrombotic microangiopathy. As shown in Table 4, we found that 18 of the top 20 predicted candidate foods and 19 of the top 20 predicted candidate diseases related to E. coli were validated by the literature.

Bifidobacterium is among the first microbes to colonize the human gastrointestinal tract and is believed to exert positive health benefits on its host. An increasing number of reports have demonstrated that dietary patterns significantly influence the abundance of Bifidobacterium [42–44]. For example, Ceballos [36] et al. demonstrated that the Western diet (the 3rd candidate food) significantly inhibits the growth of Bifidobacterium. Furthermore, Bifidobacterium has been closely associated with diabetes mellitus [45, 46] (the 1st candidate disease), where it improves type 2 diabetes mellitus by enhancing host metabolism and regulating immune-inflammatory responses through short-chain fatty acid production. As shown in Table 5, our analysis revealed that 18 of the top 20 predicted food candidates and 18 of the top 20 predicted disease candidates associated with Bifidobacterium are supported by existing literature.

To further validate the potential of LSCHNN in discovering novel FMD ternary associations, comparative analyses with 5 baseline models were conducted on E. coli and Bifidobacterium. As illustrated in Fig. 7, LSCHNN identified 131 novel FMD ternary associations in E. coli (ranking 2nd), among which 114 (87.02%) were exclusively found by our method and 17 were commonly predicted by other baseline models. Notably, MCHNN failed to discover novel FMD ternary associations in E. coli. For Bifidobacterium, LSCHNN discovered 202 novel FMD associations (ranking 2nd ), including 182 exclusive predictions (90.10%) and 20 common predictions. MCHNN and NAGTLDA failed to predict Bifidobacterium-related FMD ternary associations. Interestingly, although LAGCN exhibited suboptimal performance in evaluation metrics, it discovered the highest novel FMD associations in both E. coli and Bifidobacterium across six models. We infer this discrepancy originates from LAGCN’s elevated false positive rate, which has been corroborated by existing literature through manual verification of predicted associations. The encouraging results demonstrate the effectiveness of our method LSCHNN in identifying novel FMD associations in real situations.

Fig. 7.

Comparative analysis of novel FMD associations predicted by all six models. A Comparative analysis of Escherichia coli-related FMD association prediction by six models. B Comparative analysis of Bifidobacterium-related FMD association prediction by six models

Discussion and conclusion

Recent investigations have established the pivotal role of intestinal microbiota in human health. Predicting food-gut microbiota–disease ternary associations offers the potential for developing tailored dietary interventions that precisely address nutritional requirements, thereby enabling personalized health management and potentially delaying or preventing chronic disease progression. Compared to traditional large-scale cohort studies, computational approaches such as graph neural networks provide more efficient and cost-effective methods for predicting potential binary associations. However, the application of these computational techniques to address this specific challenge remains limited.

In this study, we introduce LSCHNN, a lightweight single-view contrastive learning model based on a hypergraph architecture. LSCHNN represents FMD associations through hypergraph construction and encodes them using HGCNs. To overcome the expressiveness limitations imposed by sparse hypergraphs on HGCNs, we implement single-view contrastive learning on FMD hypergraphs with an enhancement scheme based on gut microbiota perturbation modes. Our experimental results demonstrate that LSCHNN outperforms the state-of-the-art baseline methods in terms of the precision of predicting ternary FMD associations and discovering more potential FMD associations. LSCHNN establishes a new direction for ternary biomedical association prediction. Its single-view contrastive learning framework resolves a key dilemma: multi-view methods require abundant data for view augmentation, whereas sparse datasets benefit from targeted, biologically informed perturbations. Our method extends beyond FMD prediction to drug–microbe–disease networks, demonstrating good performance on generalizability.

For future research, we propose three key directions to enhance FMD association prediction: First, enriching hypergraph representation by incorporating multi-dimensional entity features, including microbial genome sequence characteristics and disease phenotypic attributes to strengthen biological representation capabilities; Second, integrating broader biological interaction networks such as gut microbiota-metabolite interactions and metabolite-disease relationships to construct more comprehensive association patterns; Third, implementing graph attention mechanisms to dynamically weight hyperedges and nodes, thereby improving the model’s capacity to capture complex multi-way interactions while enhancing its generalization performance across diverse biological contexts. These synergistic enhancements aim to establish a more robust computational framework for deciphering the intricate relationships between food, gut microbiota, and disease.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

LSCHNN

Lightweight single-view contrastive learning hypergraph neural network

AUC

Area under the receiver operating characteristic curve

AUPR

Area under the precision–recall curve

CV

Cross-validation

CL

Single-view contrastive learning

DGI

Deep graph infomax

DMD

Drug–microbe–disease

FMD

Food–microbe–disease

GNNs

Graph neural networks

GCNs

Graph convolutional networks

GATs

Graph attention networks

GSC

Gold-standard corpus

GIP

Gaussian interaction profile

HGNN

Hypergraph neural network

HGCN

Hypergraph convolutional network

HR

Hit ratio

HUS

Hemolytic–uremicre syndrome

MeSH

Medical subject headings

MLP

Multilayer perceptron

NDCG

Normalized discounted cumulative gain

RWR

Random walk with restart

Author contributions

J.Z., the corresponding author of this paper, supervised this project and advised on all parts of this paper. J.H. took the lead in designing, carrying out the computational experiments, and writing the manuscript with support from M.H., S.M., B.W., D.H., Y.G., Y.W., and S.G. provided help in analysis and discussion on results. All authors reviewed and approved the final version of the manuscript.

Funding

This work was supported by the Special Research Project of the Jiangsu Higher Education Association for 2024, "AI-Enabled Construction of a Characteristic Textbook System for the Strategic Emerging Major of Food Nutrition and Health" [No. 2024JCSZ33];the National Natural Science Foundation of China [32372345]; the Fundamental Research Funds for the Central Universities [JUSRP622034]; and the Wuxi Taihu Talent Project. The funders have no role in study design, data analysis, data interpretation, or writing of the manuscript.

Data availability

The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author. The source code of LSCHNN is available at https://github.com/Hujianqiang-scientificedition/LSCHNN.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.