Representation learning from similarity networks improves the prediction of disease-associated microRNAs

Abstract

Background

MicroRNAs (miRNAs) play critical roles in disease mechanisms and represent promising biomarkers and therapeutic targets, yet experimentally identifying disease–miRNA associations remains costly and incomplete. Existing computational methods often rely on known disease–miRNA associations during similarity construction or representation learning, potentially leading to information reuse between training and evaluation and limiting generalization to unseen diseases or miRNAs. To address these limitations, we propose SimNetRLMDA, an association-independent representation learning framework that predicts disease–miRNA associations by learning embeddings exclusively from association-independent similarity networks. Multiple miRNA functional similarity networks derived from miRNA–target interactions are integrated with a MeSH-based disease similarity network, while disease and miRNA representations are learned independently using graph attention networks. Known disease–miRNA associations are incorporated only at the final supervised prediction stage through a Multi-Layer Perceptron.

Results

Extensive five-fold cross-validation experiments demonstrate that SimNetRLMDA consistently outperforms state-of-the-art network-based and deep learning methods (RWRHMDA, MHMDA, MHXGMDA, and MAMFGAT), achieving AUROC and AUPRC values of up to 0.983. Ablation and sensitivity analyses confirm the robustness of the framework and the contribution of each model component. Independent validation using HMDD v3/v4, dbDEMC, and published literature further verifies 516 novel disease–miRNA associations, including predictions involving diseases and miRNAs without previously known associations.

Conclusions

SimNetRLMDA provides a robust and generalizable framework for disease–miRNA association prediction by learning representations independently of known disease–miRNA associations. Its strong predictive performance and ability to identify associations for previously unseen diseases and miRNAs highlight its potential to support biomarker discovery, disease mechanism studies, and precision medicine.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12859-026-06554-5.

Introduction

MicroRNAs (miRNAs) are short, non-coding RNAs that modulate gene expression post-transcriptionally [1, 2], playing essential roles in the molecular pathways of both common [3–7] and rare diseases [8]. Aberrant miRNA expression is associated with diverse conditions, such as cancer, metabolic disorders, and neurodegenerative diseases, positioning them as valuable biomarkers and potential therapeutic targets [9]. For instance, hsa-miR-21 is overexpressed in breast cancer, driving tumor progression, while hsa-miR-146a is involved in immune responses in Alzheimer’s disease [8, 10]. Identifying miRNAs linked to diseases is crucial for elucidating disease mechanisms and advancing precision medicine. However, experimental validation of these associations is labor-intensive and expensive, necessitating efficient computational approaches to predict novel miRNA-disease associations.

Computational methods for predicting disease-miRNA associations are divided into machine learning-based and network-based approaches [11, 12]. Machine learning methods, including traditional models like Support Vector Machines [13], k-nearest neighbor [14], and Decision Trees [15, 16], as well as deep learning models such as Deep Belief Networks [17], Convolutional Neural Networks [18], and Graph Convolutional Networks [19, 20], extract complex features from biological data. For example, KDFGMDA [21] uses a knowledge graph and deep learning for robust predictions. However, these methods often require large, high-quality datasets, which are limited in miRNA studies, and their complexity can lead to overfitting [12].

Network-based methods, leveraging the “disease module” principle, which posits that functionally related miRNAs are associated with phenotypically similar diseases [22, 23], use biological interaction networks to enhance interpretability and reduce data dependency. Local similarity methods [24–26] are limited to direct neighbors, while global methods based on Random Walk with Restart (RWR) algorithm, such as RWRMDA [27] and RWRHMDA [28] capture broader network interactions. RWRHMDA integrates disease and miRNA similarity networks, improving performance over RWRMDA. Our recent work, MHMDA [29], further enhances accuracy by integrating multiple miRNA similarity networks with a disease similarity network, thereby reducing reliance on a single similarity source.

Network-based methods effectively utilize the disease module principle by integrating disease and miRNA similarity networks (RWRHMDA [28] and MHMDA [29]), but they lack the ability to learn hidden representations, unlike deep learning-based approaches. Recent studies, such as AMHMDA [30], MHXGMDA [31], and MAMFGAT [32], combine similarities between miRNAs and between diseases with representation learning to predict disease-miRNA associations. More recently, advances in representation learning have further improved disease–miRNA association prediction. Graph-based deep learning models, such as GCNCF [33], combine Graph Convolutional Networks with neural collaborative filtering to jointly model network structure and latent interactions. Similarly, NCMD [34] and DWMDA [35] employ random-walk-based graph embedding methods, including Node2Vec and DeepWalk, followed by deep neural networks to infer disease-related miRNAs. In addition, matrix-factorization-based frameworks such as SMAP [36] exploit similarity information to learn low-dimensional latent representations for association prediction.

While these approaches demonstrate strong predictive performance, most rely on known disease–miRNA associations during either similarity construction or representation learning, which may lead to information reuse between representation learning and evaluation and restrict generalization to diseases or miRNAs without prior associations. Moreover, random-walk-based embeddings and convolution-based encoders typically assume uniform importance across neighboring nodes, limiting their ability to capture heterogeneous similarity contributions. Consequently, despite architectural advances, the fundamental learning protocol of many existing methods remains tightly coupled to known associations, raising concerns about limited applicability to truly novel prediction scenarios. These limitations motivate the development of methods that integrate similarity networks derived from independent biological evidence sources while learning expressive representations without directly using known disease–miRNA associations, as proposed in this study.

In this study, we propose SimNetRLMDA, an association-independent representation learning framework for disease–miRNA association prediction that emphasizes how similarity information is used and when known associations are introduced. SimNetRLMDA combines the strengths of network- and deep learning-based approaches while addressing their limitations, particularly association-dependent learning protocols and restricted generalizability. Specifically: (i) We construct miRNA functional similarity networks exclusively from miRNA–target interaction datasets [37, 38], where similarity is based on shared target genes, independent of known disease-miRNA associations. Unlike AMHMDA [30], MHXGMDA [31], and MAMFGAT [32], which rely on similarity matrices limited to miRNAs with known disease associations, this design enables prediction for previously unassociated miRNAs and prevents information leakage during representation learning. (ii) We integrate multiple miRNA similarity networks, as demonstrated in MHMDA [29], leveraging diverse miRNA-target datasets (e.g., experimentally validated datasets: miRTarBase [39], miRWalk [40]; and predicted datasets: TargetScan [41–44], miRDB [45]) within a unified representation learning framework, allowing complementary functional signals to be captured. (iii) We incorporate a disease similarity network based on MeSH vocabulary, independent of disease-miRNA associations, aligning with the disease module principle used in network-based methods, RWRHMDA [28] and MHMDA [29]. (iv) Disease and miRNA representations are learned independently on homogeneous similarity networks using graph attention networks (GATs), rather than on heterogeneous association graphs, thereby avoiding the propagation of known disease–miRNA links into embeddings, a limitation of methods such as KDFGMDA [21]. These representations are subsequently combined in a Multi-Layer Perceptron (MLP) classifier for supervised prediction.

By decoupling representation learning from known disease–miRNA associations, SimNetRLMDA provides a principled solution to potential information reuse across learning stages while enabling prediction for diseases and miRNAs without prior known associations. Comprehensive experiments using five-fold cross-validation demonstrate that SimNetRLMDA consistently outperforms representative network-based methods (RWRHMDA and MHMDA) and recent deep learning approaches (MHXGMDA and MAMFGAT) across single and integrated miRNA similarity networks in terms of AUROC and AUPRC. Extensive hyperparameter sensitivity and ablation studies further confirm the robustness of the proposed framework and the effectiveness of attention-based representation learning. Moreover, predictions were validated using independent datasets (HMDD v3 and v4, dbDEMC) and large-scale literature mining, revealing numerous biologically meaningful and previously unreported disease–miRNA associations. Together, these results demonstrate that SimNetRLMDA achieves both strong predictive performance and improved generalizability through association-independent learning from biologically grounded similarity networks.

Materials and methods

In this section, we first outline the construction of disease and miRNA similarity networks. Next, we explain the process of learning representations for diseases and miRNAs from these networks, which are then used to predict novel disease-miRNA associations (Fig. 1).

Fig. 1.

Overview of SimNetRLMDA. The process begins with constructing similarity networks: a DiSimNet, a disease similarity network derived from the MeSH disease vocabulary database; b miRNA similarity networks (miRNANet_S, miRNANet_B, and miRNANet_W) built from miRNA-target databases. Next, c representations of diseases and miRNAs are learned using a graph attention network (GAT) encoder, with d detailing the GAT encoder architecture. Finally, e a Multi-Layer Perceptron (MLP) predicts novel disease-miRNA associations, with top-scoring pairs validated against independent disease-miRNA association datasets and literature

Network construction

Construction of disease similarity network

We constructed a disease similarity network, DiSimNet, using the MeSH disease vocabulary database, which annotates diseases in OMIM records [46, 47]. Pairwise disease similarities were computed to form a similarity matrix, with each element representing phenotypic similarity normalized to [0,1]. To ensure a sparse and reliable network, we selected the five largest neighbors (kLN = 5) with the highest similarity scores for each disease, resulting in DiSimNet with 5080 phenotypes and 19,729 interactions (Fig. 1a). The kLN = 5 setting balances network connectivity and specificity, prioritizing strong associations.

Construction of miRNA similarity networks

miRNA similarity networks

To construct miRNA functional similarity networks, we utilized three complementary miRNA–target datasets that differ in coverage and evidence type: TargetScan [48] for predicted interactions, miRTarBase [49] for experimentally validated interactions, and miRWalk [50] for curated validated interactions. These datasets were used to build three miRNA similarity networks, denoted as miRNANet_S, miRNANet_B, and miRNANet_W, respectively (Fig. 1b).

It is important to note that each miRNA similarity network contains only the subset of miRNAs supported by its underlying dataset, rather than enforcing a fixed set of miRNAs across all networks. As a result, the number of miRNAs varies between networks, reflecting the natural coverage differences of predicted and experimentally validated data sources. The detailed statistics of each network, including the number of miRNAs and interactions, are summarized in Table 1.

Table 1.

Summary of miRNA functional similarity networks

miRNA similarity network	Dataset	Number of miRNAs	Number of interactions
miRNANetS	TargetScan	1428	46,118
miRNANetB	miRTarBase	2988	107,568
miRNANetW	miRWalk	730	29,089
miRNANetSB	TargetScan + miRTarBase	2988	144,001
miRNANetWS	miRWalk + TargetScan	1814	74,368
miRNANetWB	miRWalk + miRTarBase	3017	135,475
miRNANetWSB	miRWalk + TargetScan + miRTarBase	3019	171,283

Following previous studies [28, 29, 51], miRNA functional similarity was defined based on shared target genes, under the assumption that miRNAs regulating common targets are likely to participate in similar biological processes and disease mechanisms. Specifically, an edge was established between two miRNAs if they shared at least one target gene, and the functional similarity between miRNAs i and j was computed as:

where T_i and T_j denote the sets of target genes regulated by miRNAs i and j, respectively. In the resulting networks, nodes represent miRNAs and weighted edges represent functional similarity.

This procedure yielded miRNANet_S (1,428 miRNAs, 46,118 interactions), miRNANet_B (2988 miRNAs, 107,568 interactions), and miRNANet_W (730 miRNAs, 29,089 interactions), capturing both predicted and validated miRNA–target relationships from different data sources.

Integration of miRNA similarity networks

To assess the benefit of combining multiple similarity sources, we integrated individual miRNA functional similarity networks using a per-edge averaging strategy. For each pair of miRNAs i and j, the integrated edge weight was computed as the average of their similarity scores across all networks in which an interaction exists:

where M is the number of networks with an interaction between miRNAs i and j, and (w_ij)^k is the edge weight in network k. This integration strategy preserves the strength of consistent interactions while preventing any single network from dominating the fused representation.

Using this approach, we constructed integrated miRNA similarity networks miRNANet_SB (TargetScan + miRTarBase), miRNANet_WS (miRWalk + TargetScan), miRNANet_WB (miRWalk + miRTarBase), and miRNANet_WSB (miRWalk + TargetScan + miRTarBase). Their updated statistics are reported in Table 1. In addition, individual miRNA similarity networks were also organized as multiplex networks, with each network forming a separate layer, for fair comparison with multiplex-based methods such as MHMDA [29].

Database of known disease-miRNA associations

We retrieved 5,618 known disease-miRNA associations from HMDD v2 [52], involving 243 diseases and 574 miRNAs. These associations were used to construct a bipartite network connecting diseases and miRNAs for network-based methods and served as the ground truth for training and evaluation.

Representation learning from similarity networks

We learn low-dimensional representations of diseases and miRNAs from their respective similarity networks using a weighted graph attention network (GAT) framework (Fig. 1c–d). Each similarity network is modeled as a weighted undirected graph

 G=(V,E,W)

where V denotes the set of nodes (diseases or miRNAs), E the set of edges representing similarity relationships, and the corresponding edge weights derived from normalized similarity scores.

Each node v_i ∈ V is initialized with a feature vector x_i ∈ , where is the embedding dimension. The graph structure and edge weights are preserved during representation learning.

Weighted graph attention layer

To incorporate edge weights into representation learning, we extend the standard GAT by explicitly integrating similarity weights into the attention computation. For a given attention head ℎ ∈ {1, …, H}, node features are first linearly transformed:

where is a learnable weight matrix.

The unnormalized attention coefficient between node i and its neighbor j is then computed as:

where a^(ℎ) is a learnable attention vector, ∥ denotes concatenation, and is the similarity weight between nodes i and j. By multiplying with , the attention mechanism prioritizes stronger biological similarities while suppressing weaker or noisy connections.

The normalized attention coefficient is obtained using a softmax over the neighborhood ):

Node representations are updated by aggregating neighborhood information:

where σ(⋅) denotes the ELU activation function.

For multi-head attention, outputs from all heads are concatenated:

GAT encoder

Our GAT encoder consists of two stacked WeightedGATConv layers (Fig. 1d). The first layer employs H = 4 attention heads with hidden dimension d_hid, followed by ELU activation and dropout (p = 0.5). The second layer uses a single attention head (H = 1) without concatenation to produce the final node embeddings of dimension d_embed.

Given a weighted similarity graph  G=(V,E,W), with node feature matrix , the forward propagation of the encoder is defined as:

Z = GAT₂(Dropout(ELU(GAT₁(X, E, W)))).

where denotes the learned embedding matrix for all nodes.

Unsupervised training via graph reconstruction

Disease and miRNA similarity networks are trained independently in an unsupervised manner using a graph reconstruction objective. Observed edges are treated as positive samples, while an equal number of randomly sampled non-edge pairs serve as negatives. The model is optimized to distinguish true edges from non-edges using binary cross-entropy loss. Training is performed with the Adam optimizer at a learning rate of 0.01.

This procedure yields embedding matrices and , which are subsequently used for disease–miRNA association prediction.

Disease–MiRNA association prediction

For each disease-miRNA pair (d_i, r_j), we constructed a feature vector:

where and are the embedding vectors of disease d_i and miRNA r_j, respectively.

Negative examples were generated by randomly sampling disease–miRNA pairs that are not reported as known associations, without applying any additional negative sampling heuristics, ensuring an equal number of positive and negative samples.

We employed a Multi-Layer Perceptron (MLP) classifier to predict disease–miRNA associations (Fig. 1e). The MLP consisted of three hidden layers with 256, 128, and 64 neurons, respectively, each using the rectified linear unit (ReLU) activation function. Model training was performed using the Adam optimizer with an initial learning rate of 0.001. To reduce overfitting, L₂ regularization was applied with a penalty parameter of 1 × 10⁻⁴. Prior to model fitting, all input features were standardized using z-score normalization.

Performance evaluation

We evaluated the model using 5-fold stratified cross-validation, maintaining the proportion of positive and negative pairs in each fold. For each fold, we trained on 80% of data, and tested on 20%, then computed AUROC (Area under the Receiver Operating Characteristic curve) and AUPRC (Area under the Precision–Recall curve). Final performance was reported as mean ± standard deviation for each metric over the 5 folds. Because the disease and miRNA similarity networks are constructed exclusively from external biological knowledge sources and do not use disease–miRNA association labels, graph construction, normalization, k-largest-neighbor filtering, and network fusion were performed globally prior to cross-validation without introducing label leakage. Known disease–miRNA associations were used only during supervised MLP training within each fold.

Results

Representation learning from similarity networks improves the prediction performance

In this section, we assessed SimNetRLMDA’s prediction performance by varying the embedding size (d_embed) for disease and miRNA representations and the number of training epochs (T). We tested d_embed∈{128, 256, 512}and T∈{100, 200, 400} to generate embeddings learned exclusively from association-independent similarity networks. Performance was evaluated for each combination across miRNA similarity networks (Figs. 2 and 3) and compared against representative network-based methods (RWRHMDA and MHMDA), which rely on random-walk-based score propagation without representation learning, and deep learning-based methods (MHXGMDA and MAMFGAT), which perform representation learning on similarity matrices or graphs constructed using known disease–miRNA associations. This comparison highlights the effect of association-independent representation learning in SimNetRLMDA relative to both propagation-based and association-dependent learning paradigms.

Fig. 2.

Prediction performance of SimNetRLMDA on single miRNA similarity networks according to the change of epoch and embedding size: a and b are AUROC and AUPRC for miRNANet_B; and c and d are AUROC and AUPRC for miRNANet_S

Fig. 3.

Prediction performance of SimNetRLMDA on integrated miRNA similarity networks according to the change of epochs and embedding sizes: a and b are AUROC and AUPRC for miRNANet_SB; and c and d are AUROC and AUPRC for miRNANet_WSB

Comparison on single miRNA similarity networks

We compared the prediction performance of SimNetRLMDA with the network-based method RWRHMDA [28] and recent deep learning-based methods, MHXGMDA [31] and MAMFGAT [32]. RWRHMDA integrates disease and miRNA similarity networks into a heterogeneous network and predicts disease–miRNA associations using a Random Walk with Restart (RWR) variant, without learning explicit low-dimensional representations. MHXGMDA constructs multi-view similarity matrices of miRNAs and diseases and applies a multi-layer heterogeneous graph Transformer (HGT) to learn embeddings, which are then concatenated and fed into an XGBoost classifier for association prediction. However, the similarity matrices used in MHXGMDA are restricted to diseases and miRNAs with known associations, and the known disease–miRNA links implicitly influence representation learning. Similarly, MAMFGAT derives embeddings from similarity matrices using a two-path Graph Attention Network (GAT) and predicts association scores via a Multi-Layer Perceptron (MLP), where the similarity matrices are constructed using known disease–miRNA associations, leading to information leakage between training and prediction stages.

In contrast, SimNetRLMDA learns disease and miRNA representations independently from association-free similarity networks and uses known disease–miRNA associations only at the final supervised prediction stage. Importantly, the proposed miRNA similarity networks were incorporated into the baseline methods whenever their original implementations supported external similarity matrices. Therefore, the observed performance improvements of SimNetRLMDA are not solely attributable to the use of the proposed similarity networks themselves, but primarily arise from the association-independent representation learning and prediction strategy employed by the framework. For fair comparison, we accessed the original implementations of all baseline methods and followed their preprocessing guidelines, converting similarity network edge lists into matrices and including only entities with known associations, as required by these methods. All approaches were evaluated under the same 5-fold cross-validation protocol, with performance results summarized in Table 2.

Table 2.

Performance comparison of SimNetRLMDA, MAMFGAT, MHXGMDA, and RWRHMDA on single miRNA similarity networks

miRNA similarity network	Methods	AUROC	AUPRC
miRNANetB	SimNetRLMDA	0.980 (± 0.003)	0.980 (± 0.006)
	MAMFGAT	0.954 (± 0.005)	0.949 (± 0.006)
	MHXGMDA	0.951 (± 0.005)	0.936 (± 0.007)
	RWRHMDA	0.943 (± 0.145)	0.961 (± 0.078)
miRNANetS	SimNetRLMDA	0.979 (± 0.003)	0.981 (± 0.004)
	MAMFGAT	0.952 (± 0.007)	0.947 (± 0.005)
	MHXGMDA	0.948 (± 0.006)	0.930 (± 0.011)
	RWRHMDA	0.912 (± 0.195)	0.949 (± 0.093)

Best performance values are shown in bold

Figure 2 shows that SimNetRLMDA achieved consistently strong performance on the two single miRNA similarity networks, miRNANet_B and miRNANet_S, with optimal results obtained using an embedding size of 512 and 100–200 training epochs. Under these settings, SimNetRLMDA reached AUROC values of approximately 0.979–0.980 and AUPRC values of 0.980–0.981, while the lowest performance across all configurations remained above 0.969 in AUROC and 0.968 in AUPRC. Figure S1 (Supplementary Materials) shows the ROC and PR curves for the optimal results for miRNANet_S.

Across all embedding sizes and epochs, SimNetRLMDA consistently outperformed MAMFGAT, MHXGMDA, and the network-based method RWRHMDA on both miRNA similarity networks (Fig. 2, Table 2). In particular, deep learning–based baselines relying on known disease–miRNA associations for similarity construction, such as MAMFGAT and MHXGMDA, exhibited notably lower AUROC and AUPRC values, highlighting the benefit of SimNetRLMDA’s association-independent representation learning strategy.

Comparison on multiple miRNA similarity networks

Our prior work, MHMDA [29], demonstrated that integrating multiple miRNA similarity networks improves prediction performance. To evaluate if this holds for SimNetRLMDA, MAMFGAT and MHXGMDA, we compared their performance, along with MHMDA, on the integrated miRNA similarity network miRNANet_SB. For SimNetRLMDA and MAMFGAT, miRNANet_SB was constructed using the per-edge average method (see Materials and Methods), while MHMDA used a multiplex network approach, with each miRNA similarity network as a layer. To further explore the impact of integration, we incorporated an additional miRNA similarity network, miRNANet_W, to create miRNANet_WSB. Notably, because the same integrated miRNA similarity information was provided to multiple compared methods where applicable, the comparative results primarily reflect differences in representation learning and prediction strategies rather than differences in input similarity data. The consistently superior performance of SimNetRLMDA therefore supports the effectiveness of its association-independent representation learning framework rather than merely the use of richer similarity networks. Figure 3 illustrates SimNetRLMDA’s performance across all combinations of embedding sizes and epochs for miRNANet_SB and miRNANet_WSB. Table 3 presents the performance of MAMFGAT, MHXGMDA, MHMDA, and the best results for SimNetRLMDA on these networks.

Table 3.

Performance comparison of SimNetRLMDA, MAMFGAT, MHXGMDA, and MHMDA on integrated/multiplex miRNA similarity networks

miRNA similarity networks	Methods	AUROC	AUPRC
miRNANetSB	SimNetRLMDA	0.980 (± 0.005)	0.980 (± 0.005)
	MAMFGAT	0.955 (± 0.006)	0.949 (± 0.006)
	MHXGMDA	0.949 (± 0.003)	0.929 (± 0.006)
	MHMDA	0.953 (± 0.138)	0.969 (± 0.068)
miRNANetWSB	SimNetRLMDA	0.983 (± 0.003)	0.983 (± 0.003)
	MAMFGAT	0.952 (± 0.004)	0.944 (± 0.006)
	MHXGMDA	0.933 (± 0.005)	0.903 (± 0.011)
	MHMDA	0.972 (± 0.104)	0.981 (± 0.052)

Best performance values are shown in bold

As shown in Fig. 3, integrating multiple miRNA similarity networks further improved SimNetRLMDA’s predictive performance. The best results were obtained with an embedding size of 512 and 100–400 training epochs, yielding AUROC and AUPRC values of approximately 0.980 for miRNANet_SB and 0.983 for miRNANet_WSB. Even under less favorable configurations, performance remained stable, with AUROC and AUPRC values consistently above 0.967. Figure S2 (Supplementary Materials) shows the ROC and PR curves for the optimal results for miRNANet_WSB.

SimNetRLMDA outperformed MAMFGAT, MHXGMDA, and MHMDA across all configurations for both miRNANet_SB and miRNANet_WSB (Fig. 3, Table 3). Although MHMDA benefited from integrating multiple networks and showed improved performance for miRNANet_WSB, it remained inferior to SimNetRLMDA’s best results. Overall, Tables 2 and 3 indicate that while performance gains from network integration are modest, they are consistent for SimNetRLMDA and not observed for competing methods.

To assess whether the performance improvements of SimNetRLMDA over competing methods are statistically significant, we compared SimNetRLMDA with the strongest deep learning-based baseline method, MAMFGAT, using 10-fold cross-validation on miRNANet_WSB. Fold-wise AUROC and AUPRC values were analyzed using Wilcoxon signed-rank test. SimNetRLMDA achieved AUROC = 0.985 ± 0.004 and AUPRC = 0.983 ± 0.009, outperforming MAMFGAT (AUROC = 0.953 ± 0.005; AUPRC = 0.944 ± 0.010). Statistical testing confirmed that these improvements were highly significant for both AUROC (p = 1.953 × 10⁻³) and AUPRC (p = 1.953 × 10⁻³). These findings provide additional evidence that the superior performance of SimNetRLMDA is robust and reproducible across different cross-validation folds rather than arising from stochastic variation.

To further assess the generalization ability of SimNetRLMDA on a larger and more recent dataset, we additionally evaluated the proposed method using known disease–miRNA associations from HMDD v4. This dataset contains 27,189 associations involving 568 diseases and 1,387 miRNAs, substantially expanding the coverage compared to HMDD v2. The experiments were conducted using the best-performing miRNA similarity network configuration (miRNANet_WSB), following the same evaluation protocol as in the main experiments. As shown in Figure S3 (Supplementary Materials), SimNetRLMDA achieves slightly improved AUROC and AUPRC on HMDD v4 compared to HMDD v2, demonstrating that the proposed model generalizes well to larger and more comprehensive disease–miRNA association datasets.

Evaluation under disease-wise and miRNA-wise holdout settings

To further evaluate the generalization capability of SimNetRLMDA under stricter cold-start scenarios, we performed additional disease-wise and miRNA-wise cross-validation experiments. Unlike conventional pair-wise cross-validation, where diseases and miRNAs appearing in the test set may also appear during training through different association pairs, these protocols evaluate the model’s ability to generalize to entirely unseen diseases or miRNAs. In the disease-wise setting, all associations involving diseases assigned to the test fold were excluded from training. Similarly, in the miRNA-wise setting, all associations involving miRNAs in the test fold were withheld during training. These evaluation settings therefore represent substantially more challenging prediction tasks.

The results are summarized in Table 4. As expected, predictive performance decreased compared with the pair-wise setting due to the increased difficulty of cold-start prediction. Nevertheless, SimNetRLMDA maintained strong predictive performance across all evaluated networks. For example with miRNANet_WSB, the model achieved AUROC/AUPRC values of 0.855 ± 0.020 / 0.861 ± 0.022 under disease-wise evaluation and 0.957 ± 0.013 / 0.962 ± 0.007 under miRNA-wise evaluation.

Table 4.

Performance comparison of SimNetRLMDA under disease-wise and miRNA-wise cross-validation settings across different miRNA similarity networks

miRNA similarity network	Cross-validation setting	AUROC	AUPRC
miRNANetB	Disease-wise	0.809 (± 0.046)	0.801 (± 0.047)
	miRNA-wise	0.970 (± 0.011)	0.973 (± 0.010)
miRNANetS	Disease-wise	0.801 (± 0.026)	0.814 (± 0.025)
	miRNA-wise	0.959 (± 0.009)	0.963 (± 0.008)
miRNANetSB	Disease-wise	0.840 (± 0.026)	0.846 (± 0.020)
	miRNA-wise	0.977 (± 0.005)	0.979 (± 0.004)
miRNANetWSB	Disease-wise	0.855 (± 0.020)	0.861 (± 0.022)
	miRNA-wise	0.957 (± 0.013)	0.962 (± 0.007)

Results are reported as mean ± standard deviation over five folds

Notably, integrated miRNA similarity networks (miRNANet_SB and miRNANet_WSB) consistently outperformed single-source networks under the disease-wise evaluation setting, indicating that multi-source similarity integration improves robustness when generalizing to previously unseen diseases. In contrast, under the miRNA-wise setting, integrated networks achieved performance comparable to single-source networks but did not consistently provide further improvements. This observation suggests that the benefit of multi-source integration may depend on the specific cold-start scenario. In particular, predicting associations for unseen miRNAs may rely more strongly on the intrinsic quality of individual miRNA embeddings than on broader network integration. The larger performance reduction observed in disease-wise evaluation compared with miRNA-wise evaluation further suggests that predicting associations for entirely unseen diseases remains the more challenging scenario, likely due to greater heterogeneity in disease phenotypes and disease similarity structures.

Overall, these results demonstrate that the proposed association-independent representation learning framework retains meaningful predictive capability under strict cold-start conditions and is not limited to recovering missing edges among previously observed entities.

Hyperparameter sensitivity analysis

To evaluate the robustness of SimNetRLMDA and to justify the choice of key hyperparameters, we conducted a systematic sensitivity analysis and ablation study focusing on three critical components: (i) the number of largest neighbors (kLN) used in similarity network construction, (ii) the number of attention heads in the graph attention network (GAT), and (iii) the dropout ratio used for regularization. All experiments were performed using five-fold cross-validation, and model performance was assessed using AUROC and AUPRC. The reported results correspond to the mean performance across folds, while error bars represent standard deviations, allowing the stability of each configuration to be evaluated. The final parameter settings were selected based on overall performance trends, model simplicity, and stability rather than small fluctuations in individual measurements.

Sensitivity to the number of largest neighbors (kLN)

In the construction of disease and miRNA similarity networks, we employed a k-largest-neighbor strategy to retain the most informative similarity edges. While kLN was set to 5 in the main experiments, we additionally evaluated kLN values of 10 and 15 to assess sensitivity. As shown in Fig. 4a–b, the predictive performance of SimNetRLMDA remains highly stable across all tested kLN values. Increasing kLN beyond 5 does not yield consistent performance improvements and, in some cases, slightly increases variability.

Fig. 4.

Hyperparameter sensitivity analysis of SimNetRLMDA. AUROC and AUPRC under varying a–b k-largest neighbors (kLN), c–d numbers of attention heads (H), and e–f dropout ratio (p). Error bars denote standard deviations across five-fold cross-validation

From a biological perspective, disease similarity networks are expected to exhibit relatively sparse modular structures, where each disease is strongly associated with only a limited number of phenotypically or mechanistically related diseases. Retaining only the top five largest neighbors therefore helps preserve the most biologically meaningful relationships while reducing the inclusion of weak or noisy similarity edges caused by broad or nonspecific phenotypic overlap. This design is also consistent with the disease-module principle, which suggests that diseases sharing similar molecular mechanisms tend to form localized clusters rather than densely connected global networks.

Overall, these results indicate that kLN = 5 provides an effective balance between network connectivity and specificity, improving the quality of representation learning while avoiding noise introduced by weaker similarity relationships.

Effect of the number of attention heads

We further investigated the impact of the number of attention heads (H) in the GAT encoder by evaluating H ∈ {2, 4, 8}. The results, presented in Fig. 4c–d, show that SimNetRLMDA achieves consistently high AUROC and AUPRC values across different settings. Although the overall differences are relatively small, H = 4 provides a favorable trade-off between predictive performance, representation diversity and model complexity. Increasing the number of heads beyond 4 does not lead to notable performance gains, while fewer heads slightly reduce performance stability. These observations support the use of H = 4 in the final model configuration.

Effect of dropout ratio

To assess the influence of regularization, we evaluated dropout ratios (p) ranging from 0.0 to 0.5 in increments of 0.1. As illustrated in Fig. 4e–f, the model exhibits strong robustness across a wide range of dropout values, with only marginal fluctuations in AUROC and AUPRC. Moderate dropout values (0.2–0.4) achieve performance comparable to or slightly better than no dropout, indicating that the model is not overly sensitive to this parameter. A dropout ratio of 0.5 was selected in the main experiments to provide effective regularization without compromising predictive accuracy.

Overall, these analyses indicate that SimNetRLMDA is not highly dependent on narrowly tuned hyperparameter settings and maintains robust predictive performance across a range of reasonable configurations.

Ablation study

To elucidate the contribution of individual components and the underlying working mechanism of SimNetRLMDA, we conducted a series of ablation studies in which key modules were systematically removed or replaced while keeping the remaining components unchanged. These include encoder-level ablation (replacing GAT with GCN, Node2Vec, and DeepWalk), network-level ablation on miRNA similarity sources, and disease representation ablation by substituting similarity-based embeddings with text-based embeddings. Overall, the ablation results consistently demonstrate that the association-independent similarity networks provide biologically meaningful prior information and that the proposed GAT-based representation learning framework effectively exploits this information for disease–miRNA association prediction.

Ablation study on representation learning methods

To assess the contribution of the graph attention mechanism to representation learning, we performed an encoder-level ablation study in which the Graph Attention Network (GAT) in SimNetRLMDA was replaced with alternative graph embedding methods while keeping all other components unchanged. Specifically, we evaluated Graph Convolutional Network (GCN), Node2Vec, and DeepWalk as substitutes for GAT, which were proposed to use as encoders in GCNCF [33], NCMD [34] and DWMDA [35], resulting in three ablated variants that differ only in the node representation learning module.

The comparison results are summarized in Table 5, with detailed performance under different embedding dimensions and training epochs provided in Figure S4. Across all evaluated settings, SimNetRLMDA with the GAT encoder consistently outperformed its GCN-, Node2Vec-, and DeepWalk-based counterparts in terms of both AUROC and AUPRC. Replacing GAT with GCN resulted in a notable performance decrease, while the use of Node2Vec and DeepWalk led to further degradation, indicating limited capacity to model complex neighborhood importance.

Table 5.

Encoder-level ablation study comparing different representation learning methods on miRNANet_WSB

Methods	AUROC	AUPRC
SimNetRLMDA (GAT)	0.983 (± 0.003)	0.983 (± 0.003)
SimNetRLMDA (GCN)	0.962 (± 0.018)	0.942 (± 0.018)
SimNetRLMDA (Node2Vec)	0.946 (± 0.003)	0.947 (± 0.003)
SimNetRLMDA (DeepWalk)	0.947 (± 0.003)	0.947 (±0.003)

Best performance values are shown in bold

These results demonstrate that the attention mechanism in GAT plays a critical role in learning informative miRNA and disease representations by adaptively weighting neighboring nodes, thereby capturing heterogeneous and non-uniform similarity contributions that are not adequately modeled by convolution-based or random-walk-based encoders.

Ablation study on miRNA and disease similarity networks

To investigate the contribution of individual network components and to assess whether integrating multiple similarity networks introduces noise, we conducted a comprehensive ablation study on both miRNA and disease similarity networks.

We evaluated SimNetRLMDA on all available single miRNA similarity networks (miRNANet_S, miRNANet_B, and miRNANet_W) as well as their combinations (miRNANet_SB, miRNANet_WS, miRNANet_WB, and miRNANet_WSB). For each configuration, performance was averaged across all tested embedding sizes and training epochs. The results are summarized in Fig. 5.

Fig. 5.

Ablation study on miRNA similarity networks. Average AUROC a and AUPRC b of SimNetRLMDA across different single and integrated miRNA similarity networks, computed over all tested embedding sizes and training epochs. Single networks include miRNANet_S, miRNANet_B, and miRNANet_W, while integrated networks include miRNANet_SB, miRNANet_WS, miRNANet_WB, and miRNANet_WSB. The results show a consistent positive trend from single to integrated networks, indicating complementary contributions of individual similarity sources without performance degradation

All individual networks already achieve strong predictive performance (AUROC ≥ 0.972), indicating that biologically meaningful miRNA functional information can be captured directly from independent miRNA–target interaction datasets without relying on known disease–miRNA associations. This observation supports the role of the proposed similarity networks as effective biological priors for representation learning. Moreover, a consistent positive trend is observed when integrating multiple similarity networks, with the best performance achieved by the fully integrated miRNANet_WSB network. Importantly, none of the integrated configurations results in performance degradation, suggesting that the individual similarity sources provide complementary rather than redundant or noisy information. These findings indicate that combining multiple association-independent similarity networks enables SimNetRLMDA to learn more robust and biologically informative miRNA representations.

To evaluate the role of disease similarity modeling, we compared SimNetRLMDA using disease embeddings learned from the MeSH-based disease similarity network via GAT with a variant that replaces these embeddings with text-based representations learned using Word2Vec [53]. Disease embeddings of dimensions 128, 256, and 512 were generated using Word2Vec and used as input to the prediction model.

As shown in Table 6, the model using disease similarity–based GAT embeddings substantially outperforms the Word2Vec-based variant in both AUROC and AUPRC. This result highlights the importance of explicitly modeling disease similarity through a structured network rather than relying solely on textual representations.

Table 6.

Contribution of the disease similarity network

Methods	AUROC	AUPRC
SimNetRLMDA (Disease similarity-based embeddings by GAT)	0.983 (± 0.003)	0.983 (± 0.003)
SimNetRLMDA (Text-based disease embeddings by Word2Vec)	0.939 (± 0.008)	0.930 (± 0.011)

Performance comparison between SimNetRLMDA with disease similarity–based GAT embeddings and text-based disease embeddings learned using Word2Vec

Overall, these ablation studies demonstrate that both the miRNA similarity networks and the disease similarity network contribute meaningfully and synergistically to the predictive performance of SimNetRLMDA.

Prediction of novel disease-miRNA associations

In the previous section, we demonstrated that SimNetRLMDA achieved optimal performance with an embedding size of 512 for miRNANet_WSB (Fig. 3c-d). Here, we applied this configuration to predict novel disease-miRNA associations, validated using independent disease-miRNA datasets and literature. The model was trained on known disease-miRNA associations (positive set) and an equal number of randomly selected disease-miRNA pairs (negative set). Predictions were generated for the remaining pairs, with the top 20 highest-scoring pairs per disease selected as promising associations for further validation.

Novel associations supported by public databases and literature

We validated the top 20 predictions for each disease using two up-to-date disease-miRNA association databases, i.e., dbDEMC v3 [54] and HMDD v3 [55]. A prediction was considered validated if the miRNA-disease pair was reported in either database. Table S1 show a list of 516 validated associations between 144 diseases and 72 miRNAs by the two databases. Interestingly, 16 of the associations are found for diseases and miRNAs which have not had any known disease-miRNA associations (Table 7). Note that, those could not be found by methods like MHXGMDA [31] and MAMFGAT [32], since the input matrices they constructed totally depends on diseases and miRNAs having at least one known disease-miRNA associations.

Table 7.

Validated novel disease-miRNA associations for diseases and miRNAs which have not had any known disease-miRNA associations

OMIM ID	Disease	miRNA	Supported Datasets
153,700	macular dystrophy	hsa-miR-6810	HMDDv3
154,700	marfan syndrome	hsa-miR-26a	HMDDv3
161,550	nasopharyngeal carcinoma	hsa-miR-26a	HMDDv3
166,760	otitis media	hsa-miR-26a	HMDDv3
176,670	hutchinson-gilford progeria syndrome	hsa-miR-26a	HMDDv3
184,700	polycystic ovary syndrome 1	hsa-miR-26a	HMDDv3
189,800	preeclampsia/eclampsia 1	hsa-miR-26a	HMDDv3
246,470	leukemia	hsa-miR-4708	dbDEMC
277,700	werner syndrome	hsa-miR-26a	HMDDv3
605,218	systemic lupus erythematosus	hsa-miR-26a	HMDDv3
605,364	psoriasis susceptibility 6	hsa-miR-26a	HMDDv3
607,154	allergic rhinitis	hsa-miR-26a	HMDDv3
607,893	ovarian cancer	hsa-miR-26a	HMDDv3
608,266	parathyroid carcinoma	hsa-miR-6512	dbDEMC
608,266	parathyroid carcinoma	hsa-miR-6810	dbDEMC
608,901	coronary heart disease	hsa-miR-26a	HMDDv3

Table 7 lists 16 novel disease-miRNA associations predicted by SimNetRLMDA and validated using HMDD v3 and dbDEMC v3 databases, involving 14 unique diseases (e.g., macular dystrophy, leukemia, ovarian cancer) and three miRNAs (hsa-miR-26a, hsa-miR-4708, hsa-miR-6810), none of which had prior associations in the HMDD v2 training dataset. Notably, hsa-miR-26a is linked to 11 distinct diseases, suggesting a broad and pleiotropic regulatory role, while others like hsa-miR-6810 and hsa-miR-6512, show more disease-specific associations, including parathyroid carcinoma.

Importantly, these validated predictions are not only statistically robust but also biologically meaningful. Several of the newly identified associations involve both diseases and miRNAs that were previously unseen during training, demonstrating the model’s ability to uncover genuinely novel disease–miRNA relationships rather than rediscovering known patterns. For example, hsa-miR-26a was independently validated across multiple disease contexts, including ovarian cancer, systemic lupus erythematosus, and coronary heart disease, supporting its functional relevance in diverse pathological processes.

Validated primarily by HMDD v3 (12 associations) and dbDEMC v3 (three associations), these findings highlight SimNetRLMDA’s ability to predict disease–miRNA relationships without relying on prior disease–miRNA associations, unlike methods like MHXGMDA and MAMFGAT, which rely on known associations. This capability is enabled by constructing miRNA and disease similarity networks from independent biological evidence, thereby reducing data leakage and improving generalizability. Collectively, these findings underscore SimNetRLMDA’s potential to uncover novel biomarkers and therapeutic targets across a wide spectrum of diseases, ranging from rare genetic disorders to common complex conditions, contributing to precision medicine and mechanistic disease research.

Due to limitations in the curation quality and update frequency of public disease-miRNA association databases, we further validated the top 20 predicted miRNAs per disease by mining PubMed for evidence not captured in dbDEMC v3 or HMDD v3, using the easyPubMed R package [56]. Specifically, we queried PubMed with “(disease[Title/Abstract]) AND (miRNA[Title/Abstract])” for each disease-miRNA pair, considering a prediction validated if at least one supporting study was identified. Table S3 lists 64 disease-miRNA pairs with at least one supporting study (NuStudy > 0).

miRNA-target gene interactions underlying disease-miRNA associations

In this study, we constructed miRNA functional similarity networks based on miRNA-target gene interactions to predict disease-miRNA associations. Our method’s superior performance, driven by effective representation learning of diseases and miRNAs on these networks, outperforms existing approaches. To elucidate why miRNA-target interactions enhance prediction performance, we visualized connections between miRNAs and diseases via shared target genes. We collected disease-gene associations from OMIM [47] and miRNA-target interactions from miRNA-target databases (see Materials and Methods). Figure 6 illustrates eight of the 64 literature-supported disease-miRNA associations linked through target genes. Most interactions between miRNAs and shared target genes are supported by both validated (miRTarBase) and predicted (TargetScan) datasets, except for the validated interaction between hsa-let-7b and TP53 (7157) (Fig. 6b) and the predicted interaction between hsa-miR-204 and PRF1 (5551) (Fig. 6d). This suggests that while most validated miRNA-gene interactions can be predicted, some cannot (e.g., Fig. 6b), and certain novel interactions can be predicted despite lacking validation (e.g., Fig. 6d). Thus, validated and predicted interactions complement each other, strengthening the prediction of disease-miRNA associations.

Fig. 6.

Visualization of disease-miRNA associations via shared target genes. a Gastric cancer (613659) and hsa-miR-630. b Nasopharyngeal carcinoma (607107) and hsa-let-7d. c Nasopharyngeal carcinoma (607107) and hsa-miR-30e. d Lymphoma (605027) and hsa-miR-204. e Myeloma (254500) and hsa-miR-204. f Myeloma (254500) and hsa-miR-15b. g Psoriatic arthritis (607507) and hsa-miR-30e. h Autism (613436) and hsa-miR-590. Note: Due to the large number of miRNA-target interactions, only interactions between miRNAs and shared target genes are shown

The coexistence of experimentally validated and computationally predicted miRNA–gene interactions indicates that integrating heterogeneous miRNA–target data enhances functional coverage. Together, these results demonstrate that SimNetRLMDA not only achieves high predictive accuracy but also captures biologically interpretable patterns that may facilitate the discovery of novel biomarkers and therapeutic targets.

Discussion

In this study, we present SimNetRLMDA, an association-independent graph representation learning strategy that significantly enhances the prediction of disease-associated microRNAs by integrating multiple miRNA functional similarity networks with a disease similarity network. By employing graph attention networks (GATs) for representation learning and a Multi-Layer Perceptron (MLP) for association prediction, SimNetRLMDA addresses critical limitations of existing methods, notably data leakage and limited generalizability, particularly for diseases or miRNAs without prior known associations. Our approach constructs miRNA similarity networks from miRNA–target datasets (TargetScan, miRTarBase, and miRWalk), fully independent of disease–miRNA associations, ensuring robustness and generalizability. The disease similarity network, built using MeSH vocabulary, further strengthens the model by capturing phenotypic relationships.

Our experimental results demonstrate that SimNetRLMDA consistently outperforms state-of-the-art methods, including network-based approaches (RWRHMDA, MHMDA) and deep learning-based methods (MHXGMDA, MAMFGAT), across both single and integrated miRNA similarity networks. Achieving AUROC and AUPRC scores of 0.980–0.983 on integrated networks (miRNANet_SB and miRNANet_WSB), SimNetRLMDA highlights the advantage of combining association-independent similarity network construction with representation learning. The integration of multiple miRNA similarity networks further improves performance, as evidenced by superior metrics on miRNANet_WSB compared to single networks. Importantly, SimNetRLMDA’s ability to predict 516 novel disease–miRNA associations, including 16 associations involving diseases and miRNAs without prior known links, validated by independent databases (HMDD v3, dbDEMC v3) and literature, underscores its potential to uncover biologically meaningful relationships.

To further validate the robustness of SimNetRLMDA and address concerns regarding overfitting and parameter sensitivity, we conducted extensive hyperparameter sensitivity analyses and ablation studies. Sensitivity analyses demonstrated that the model’s performance remains stable across a wide range of key hyperparameters, including embedding dimensionality, the number of GAT attention heads, dropout ratios, and the number of largest neighbors used in similarity network construction. From a biological perspective, the selected kLN = 5 promotes sparse and modular disease similarity structures by retaining only the strongest phenotypic relationships, thereby reducing noise from weak or nonspecific similarity edges. This design is consistent with the disease-module principle, where diseases sharing molecular mechanisms tend to form localized clusters rather than densely connected global networks. In addition, systematic ablation experiments – covering encoder-level replacements (GAT vs. GCN, Node2Vec, and DeepWalk), network-level ablation of miRNA similarity sources, and disease representation variants – confirmed that each major component contributes meaningfully to the final performance, with the full SimNetRLMDA configuration consistently achieving the best results. These analyses provide strong empirical evidence that the observed performance gains are not driven by specific parameter choices or data leakage, but by the proposed learning strategy.

The success of SimNetRLMDA stems from several key innovations. First, constructing miRNA similarity networks based on shared target genes, rather than known disease–miRNA associations, explicitly prevents information leakage, a significant issue in methods such as MHXGMDA and MAMFGAT. Second, the use of GATs effectively captures heterogeneous neighborhood importance within similarity networks, improving representation quality beyond uniform aggregation schemes. Third, integrating multiple miRNA–target datasets provides a more comprehensive and balanced view of miRNA functional relationships, mitigating biases introduced by individual data sources. Fourth, incorporating a disease similarity network aligns with the disease module principle, enabling the model to leverage phenotypic similarities in a principled manner. Finally, additional disease-wise and miRNA-wise holdout experiments demonstrated that SimNetRLMDA retains strong predictive performance under strict cold-start settings involving previously unseen diseases or miRNAs, further supporting the generalizability of the proposed association-independent representation learning strategy.

Recent studies further demonstrate the rapid advancement of graph-learning-based frameworks for disease–miRNA association prediction. For example, ADHGMDA [57] introduced adaptive dynamic heterogeneous graph attention learning with virtual nodes and multi-task optimization strategies to address graph sparsity and cold-start prediction problems. BADMDA [58] combined multiple kernel learning, drop aggregation, and bi-directional attention mechanisms to improve heterogeneous feature integration and mitigate over-smoothing. HybridGNN [59] integrated GCN, GAT, and matrix factorization modules with multi-source similarity networks to enhance representation learning and scalability, while APMF [60] employed multi-granularity transformer contrastive learning and feature reconstruction strategies to capture complex semantic relationships from heterogeneous biological data. Collectively, these studies highlight the increasing importance of advanced graph neural networks, attention mechanisms, contrastive learning, and multi-source biological data integration for improving disease–miRNA association prediction. In contrast to many recent methods that directly incorporate known disease–miRNA associations into heterogeneous graph construction or representation learning, SimNetRLMDA focuses primarily on an association-independent representation learning strategy to reduce potential information leakage and improve generalizability to unseen diseases and miRNAs.

From a computational perspective, SimNetRLMDA is efficient and scalable. The representation learning stage applies lightweight two-layer weighted GAT encoders separately to disease and miRNA similarity networks, whose sizes depend only on the number of diseases or miRNAs rather than the full disease–miRNA association space. This modular design keeps the number of trainable parameters modest and allows training time to scale linearly with the number of nodes and retained similarity edges. In addition, the k-largest-neighbor filtering strategy substantially reduces graph density, limiting unnecessary message passing over weak similarity edges and improving computational efficiency. The model achieves stable convergence within a moderate number of training epochs on standard GPU hardware, while the downstream MLP classifier introduces minimal computational overhead. Once embeddings are learned, the downstream MLP classifier is compact, and inference for disease–miRNA association prediction is fast. Compared with approaches that jointly optimize large heterogeneous networks or repeatedly reconstruct similarity matrices from association data, SimNetRLMDA offers a favorable balance between predictive performance and computational cost.

Despite the strong performance of SimNetRLMDA, several directions remain for future research. Recent studies on lncRNA–disease association prediction, such as GMFLDA [61], neighborhood-regularized matrix factorization [62], and expression-based matrix factorization frameworks [63], further demonstrate the effectiveness of graph neural networks and representation learning techniques across different non-coding RNA prediction tasks. More broadly, recent AI-driven biomedical prediction frameworks, including MPOGAN [64], circRDRP [65], STExplore [66], and CLCDA [67], highlight the growing importance of graph learning, multi-modal biological data integration, and collaborative deep learning for modeling complex biological systems. Inspired by these advances, future extensions of SimNetRLMDA could generalize the proposed framework to other RNA–disease association prediction tasks, including lncRNA- and circRNA-related diseases. In addition, incorporating multi-modal biological data, such as gene expression or epigenetic profiles, and exploring more advanced graph neural network architectures may further enhance model expressiveness and biological interpretability. Furthermore, like most existing disease–miRNA association prediction methods, our framework relies on randomly sampled unknown disease–miRNA pairs as negative examples. Because currently unlabeled pairs may contain undiscovered true associations, the prediction task is more appropriately interpreted within a positive–unlabeled learning setting rather than a strict binary classification framework. Although such label incompleteness generally increases task difficulty rather than artificially inflating predictive performance, future work could explore more conservative negative sampling strategies or dedicated positive–unlabeled learning approaches to further improve robustness and biological reliability. Finally, experimental validation of predicted associations remains an important direction to confirm their biological relevance.

Conclusions

SimNetRLMDA introduces an association-independent representation learning strategy for disease–miRNA association prediction by learning disease and miRNA embeddings exclusively from external similarity networks and incorporating known disease–miRNA associations only during the final supervised prediction stage. This design effectively reduces the risk of information leakage while improving generalizability to diseases and miRNAs without previously known associations. Extensive benchmarking, ablation studies, statistical analyses, and disease-wise and miRNA-wise holdout experiments demonstrate that the proposed framework consistently achieves state-of-the-art predictive performance while remaining robust across different parameter settings and similarity networks.

Beyond its predictive accuracy, SimNetRLMDA provides a scalable and biologically interpretable framework that successfully identifies novel disease–miRNA associations supported by independent databases and published literature. By emphasizing association-independent representation learning, this work offers a practical strategy for developing more reliable graph-based prediction models in biomedical informatics. The framework may facilitate biomarker discovery, improve understanding of disease mechanisms, and support future precision medicine applications. To promote reproducibility and further research, the source code and datasets are publicly available at https://github.com/hauldhut/SimNetRLMDA.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

The author (DHL) contributed to all aspects of this study.

Funding

This research is funded by Hanoi University of Science and Technology (HUST) under project number T2025-PC-016.

Data availability

The source code and the preprocessed datasets used in this study are publicly available at https://github.com/hauldhut/SimNetRLMDA. The original datasets used to construct the similarity networks and evaluate the model were obtained from publicly available resources: TargetScan: https://www.targetscan.org/. miRTarBase: https://mirtarbase.cuhk.edu.cn/. miRWalk: http://mirwalk.umm.uni-heidelberg.de/. HMDD: http://www.cuilab.cn/hmdd. dbDEMC: https://www.biosino.org/dbDEMC.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

The author declares that there exists no conflict of interest regarding the publication of this article.

Competing interests

The authors declare no competing interests.