Enhanced drug-disease association prediction through representation learning on similarity networks

Abstract

Drug repositioning has emerged as a promising strategy for accelerating therapeutic discovery by identifying novel indications for existing drugs. Recent graph representation learning methods have shown encouraging performance for drug–disease association prediction; however, many existing approaches directly utilize heterogeneous drug–disease networks during representation learning, potentially introducing label leakage and limiting generalizability. In this study, we propose similarity network-based representation learning for drug repositioning (SimNetRLDR), a similarity network-based representation learning framework for drug repositioning. The proposed method independently learns drug and disease embeddings from homogeneous similarity networks using a weighted graph attention network encoder. The learned representations are subsequently integrated and used for downstream drug–disease association prediction through an extreme gradient boosting (XGBoost) classifier. Comprehensive experiments were conducted on benchmark datasets under single and integrated/multiplex disease similarity network settings. SimNetRLDR consistently outperformed existing methods, achieving superior area under the receiver operating characteristic curve, area under the precision–recall curve, F1-score, and accuracy with strong robustness across cross-validation folds. Additional robustness evaluations using external dataset, drug-wise and disease-wise cold-start settings further demonstrated the generalizability of the proposed framework, particularly for unseen drugs. Hyperparameter sensitivity analysis demonstrated stable performance across different neighborhood sizes and attention head numbers. Component-wise ablation studies further confirmed the effectiveness of the weighted graph attention encoder and the decoupled XGBoost classifier design. To evaluate biological and clinical relevance, we analyzed predicted associations supported by shared Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways and manually curated evidence from ClinicalTrials.gov. After rigorous evidence filtering, 12 predicted drug–disease associations showed plausible clinical support, including Sulindac–Breast Neoplasms, Methotrexate–Schizophrenia, and Liothyronine–Breast Neoplasms. Overall, these findings demonstrate that SimNetRLDR provides an effective, robust, and biologically meaningful framework for computational drug repositioning.

Introduction

Drug repositioning, the process of discovering new therapeutic applications for existing or experimental drugs, provides a cost-efficient alternative to conventional drug development, which is often protracted and costly [1]. Historically, clinical insights have facilitated notable repositioning successes. For example, bevacizumab, originally developed to target angiogenesis in metastatic colon cancer and non-small cell lung cancer, has been repurposed for exudative macular degeneration by inhibiting abnormal retinal vascularization [2]. Likewise, sildenafil is now used for erectile dysfunction and pulmonary hypertension, thalidomide treats severe erythema nodosum leprosum, and retinoids are applied in acute promyelocytic leukemia [3, 4]. Further examples of successful drug repositioning are documented by Ashburn et al. and Sardana et al. [5, 6]. To enhance scalability, computational approaches, including matrix-based, network-based, and machine learning-based methods, have been developed for in silico prediction of drug–disease associations [7–10].

Matrix factorization-based approaches constitute another important class of computational methods for drug–disease association prediction. These methods infer missing associations by decomposing large-scale drug–disease interaction matrices into lower-dimensional latent representations that capture underlying structural patterns [11]. Representative approaches include non-negative matrix factorization [12, 13], non-negative matrix tri-factorization [14, 15], and collective matrix factorization [16, 17]. Such methods are particularly effective because they can incorporate prior biological knowledge, including drug–target interactions, pathway information, and side-effect profiles, through regularization terms or structural constraints. More recently, tensor factorization approaches have further extended this framework by enabling multidimensional integration of heterogeneous biomedical relationships, such as drug–disease–gene associations, thereby improving predictive performance under sparse and noisy data conditions [18]. Despite their effectiveness, most matrix-based approaches primarily exploit global interaction patterns and often do not explicitly model similarity network topology, which may limit their ability to capture localized structural relationships among pharmacologically similar drugs and biologically related diseases.

Network-based approaches are highly effective for drug repositioning, as they leverage known interactions among drugs, diseases, and associated proteins or genes to build predictive models [19–21]. These methods typically operate on the premise that drugs with similar properties can address diseases sharing common pathological mechanisms or clinical manifestations, utilizing heterogeneous networks that connect drug and disease similarity networks through established drug–disease associations. For instance, Wu et al. employed network clustering to detect drug–disease modules within a heterogeneous network [22], whereas MBiRW [23] and TP-NRWRH [24] implemented variants of random walk with restart (RWR) algorithm on heterogeneous networks to identify novel drug indications. Similarly, RLSDR utilized semi-supervised learning to prioritize candidate diseases [21]. However, a significant limitation of these approaches is their reliance on a single phenotype-based disease similarity network, often derived from MimMiner, which uses MeSH terms to annotate Online Mendelian Inheritance in Man (OMIM) diseases [25, 26]. This approach neglects diverse disease relationships, such as those based on molecular or ontological similarities, thereby constraining prediction accuracy [27]. While earlier methods depended on MeSH vocabularies [25] or shared gene profiles [28], recent advancements, such as the human phenotype ontology (HPO), have facilitated more robust semantic similarity measures [29, 30]. Acknowledging the multidimensional nature of diseases—encompassing phenotypic, molecular, and ontological aspects—integrating these diverse perspectives can enhance disease relationship modeling. Recently, we introduced MHDR [31], a method that constructs multiple disease similarity networks and integrates them into multiplex and multiplex-heterogeneous networks using known drug-disease associations. By adapting the RWR algorithm to rank candidate diseases, MHDR outperformed methods relying on single disease similarity networks.

Recently, deep learning approaches have gained traction in drug repositioning research [32, 33]. A key strength of deep learning lies in its capacity to generate robust representations. For instance, methods like DDAGDL [34] and RGLDR [35] utilize advanced techniques, such as graph attention networks (GAT) and regulation-aware learning, within complex heterogeneous information networks (HINs). These networks integrate drug–protein, protein–disease, and drug–disease associations, along with annotation data, including SMILES for drugs [36], MeSH terms for diseases, and protein sequences, to derive high-quality representations of drugs and diseases, thereby improving prediction accuracy. Moreover, multimodal and contrastive learning frameworks have further advanced computational drug repositioning. For example, AMDGT [37] employs a dual graph transformer architecture together with an attention-aware multimodality interaction module to integrate similarity information and biochemical features. Similarly, GCGB [38] incorporates heterogeneous graph contrastive learning and adaptive gradient balancing to alleviate label sparsity and optimization imbalance between auxiliary contrastive objectives and the primary drug–disease association prediction task. These methods demonstrate the growing importance of multimodal biomedical representation learning and cross-view feature integration in drug repositioning research.

These approaches combine the strengths of network-based and deep learning methodologies by applying deep learning to networks comprising drugs, diseases, and proteins. However, most existing heterogeneous graph-based methods primarily focus on integrating heterogeneous associations and multimodal annotations, while often underutilizing direct similarity relationships among drugs or among diseases. Consequently, they may not fully exploit the classical network-based assumption that similar drugs tend to treat phenotypically or mechanistically related diseases. Additionally, incorporating known drug–disease associations into heterogeneous networks during representation learning may introduce potential label leakage or optimistic bias, as these learned representations are subsequently used as inputs in the classification phase. Furthermore, many recent multimodal frameworks require extensive heterogeneous biomedical resources, such as molecular graphs, omics profiles, biomedical literature, or large-scale knowledge graphs, which may limit their applicability and reproducibility across standard benchmark settings.

Motivated by these challenges, we propose SimNetRLDR (similarity network-based representation learning for drug repositioning), a similarity-network-based framework for drug repositioning that systematically integrates heterogeneous drug and disease similarities through RWR over integrated and multiplex network structures. Unlike many existing heterogeneous graph approaches that directly incorporate known drug–disease associations during representation learning, SimNetRLDR independently learns drug and disease embeddings from homogeneous similarity networks using a weighted GAT encoder, thereby reducing the risk of explicit label leakage during embedding construction. In addition, unlike existing approaches that typically rely on single-view similarity representations or limited biological integration, SimNetRLDR evaluates multiple disease network configurations, including single, integrated, and multiplex similarity networks, to better capture complementary disease relationships. This design enables a comprehensive assessment of how network topology influences predictive performance and robustness.

The major contributions of this study are 3-fold. First, we develop a unified computational framework capable of exploiting heterogeneous biological similarities while preserving network-specific information. Second, we comprehensively evaluate SimNetRLDR against recent state-of-the-art methods under diverse network settings, including external validation and stricter drug-wise and disease-wise cold-start evaluations, to assess robustness and generalizability. Third, we perform rigorous downstream validation through shared KEGG pathway analysis and manually curated ClinicalTrials.gov evidence using stricter therapeutic relevance criteria. These analyses not only confirm the predictive utility of SimNetRLDR but also provide biologically plausible candidate drug–disease associations for future experimental investigation. Overall, the proposed framework provides an effective and interpretable strategy for prioritizing candidate drug–disease associations in computational drug repositioning.

Materials and methods

In this section, we outline the construction of drug and disease similarity networks. Subsequently, we detail the process of learning representations for drugs and diseases from these networks, enabling the prediction of novel drug–disease associations (Fig. 1).

Figure 1.

Overview of SimNetRLDR. Initially, similarity networks for drugs and diseases are constructed, comprising: (a) DrSimNet, a drug similarity network derived from chemical structure similarities computed between pairs of drugs; (b) DiSimNet_M, a disease similarity network based on MeSH, a disease vocabulary database; (c) DiSimNet_P, a disease similarity network built using known disease-associated genes and their interactions within a protein-protein interaction network. Next, (d) representations of drugs and diseases are generated using a graph attention network (GAT) encoder. Finally, (e) novel drug-disease associations are predicted using an XGBoost classifier, with top-scoring pairs validated through molecular and clinical evidence.

Network construction

Construction of drug similarity network

We developed a drug similarity network based on the assumption that drugs with comparable chemical structures exhibit similar therapeutic effects. Specifically, we calculated pairwise chemical structure similarities among 7838 drugs from the KEGG database [39–41] using the SIMCOMP tool [42]. The drug similarity graph was constructed by retaining all drug pairs with positive SIMCOMP similarity scores (i.e. similarity score > 0), without applying additional sparsification strategies such as threshold-based filtering or fixed k-largest-neighbor selection. As a result, the final drug similarity network, referred to as DrSimNet, comprised 887,883 weighted interactions among the 7838 drugs (Fig. 1a).

Construction of disease similarity networks

MeSH-based disease similarity network

Initially, we built a disease similarity network using the MeSH disease vocabulary database, which annotates diseases in OMIM medical records [25, 26]. Each matrix element reflects the phenotypic similarity between pairs of diseases, normalized to the range [0,1]. To create a reliable and sparse network, we selected the five largest neighbors (kLN) with the highest similarity scores for each disease, yielding a phenotypic disease similarity network, termed DiSimNet_M, with 19,729 interactions across 5080 phenotypes (Fig. 1b). Setting kLN = 5 ensures a balance between network connectivity and specificity, emphasizing the strongest associations.

PPI-based disease similarity network

Next, we developed a disease similarity network utilizing known disease-associated genes and a protein interaction network [43]. Leveraging 3222 diseases from OMIM and the functional protein-protein interaction (PPI) network [44], we determined the similarity between diseases d_i and d_j by comparing their respective gene/protein sets P₁ and P₂​:

$$\mathit{simDis}\left(d_i,d_j\right)=\mathit{simGeneSet}\left(P_1,P_2\right)=\frac{{\sum }_{p_i\in P_1}\max \left(\mathit{sim}\right(p_i,P_2\left)\right)+{\sum }_{p_j\in P_2}\max \left(\mathit{sim}\right(p_j,P_1\left)\right)}{\left|P_1\right|+\left|P_2\right|}.$$

This resulted in a disease similarity network, namely DiSimNet_P, with 80,696 interactions among 3222 diseases (Fig. 1c).

Integration of disease similarity networks

To evaluate the advantages of integrating multiple disease similarity networks, we merged these networks into a unified network, DiSimNet_MP, using a per-edge average approach. For each pair of diseases i and j, the integrated edge weight was computed as:

$${\overline{w}}_{ij}=\frac{1}{M}\sum _{k=1}^M{\left(w_{ij}\right)}_k,$$

where M represents the number of networks containing an interaction between diseases i and j, and ${\left(w_{ij}\right)}_k$ ​ denotes the edge weight in network k. The per-edge average method was selected to balance the contributions of each network while maintaining the relative strength of interactions. The resulting DiSimNet_MP network comprises 6261 diseases and 99,688 interactions. Table 1 provides a summary of the statistics for the three disease similarity networks. Additionally, these individual disease similarity networks were combined into a multiplex disease similarity network, with each network serving as a layer, for comparison with the MHDR method [31].

Table 1.

Summary of constructed disease similarity networks.

Disease similarity network	Number of diseases	Number of interactions
DiSimNetM	5080	19,729
DiSimNetP	3222	80,696
DiSimNetMP	6261	99,688

Benchmark datasets

To evaluate the proposed framework, we employed two widely used benchmark datasets for drug–disease association prediction. The primary benchmark was the F-dataset compiled by Gottlieb et al. [45], which contains 1933 experimentally validated drug–disease associations involving 593 drugs and 313 diseases. This dataset has been widely adopted as a gold-standard benchmark for evaluating computational drug repositioning methods.

To further assess robustness and generalizability, we additionally evaluated the proposed method on the C-dataset introduced by Luo et al. [23], which consists of 2532 validated associations among 663 drugs and 409 diseases. This dataset was constructed as an expanded benchmark to support large-scale evaluation of drug–disease association prediction methods.

In addition, external validation was conducted using the D-dataset (also referred to as the DN-dataset in some studies), which originates from the DrugNet study proposed by Martínez et al. [46], which consists of 1008 validated associations among 1490 drugs and 4517 diseases. This dataset was used to evaluate the generalization capability of the trained model on associations not included in the primary training benchmark.

It should be noted that the dataset names (F-, C-, and D-/DN-dataset) do not correspond to specific biological concepts or chemical properties. Instead, these names represent historical benchmark identifiers adopted from previous studies to facilitate standardized comparison across computational drug repositioning methods [47].

Drug and disease identifiers were standardized and mapped across all similarity network resources through a unified preprocessing pipeline to ensure consistency between association matrices and similarity networks. Only entities with complete mapping information across all required data sources were retained.

Representation learning from similarity networks

We used weighted GATs to learn low-dimensional vector representations (embeddings) of drugs and diseases from their respective similarity networks (Fig. 1d).

Each similarity network was represented as an undirected, weighted graph:

$$G=(V,E,W),$$

where:

• $V=\{v_1,v_2,\dots ,v_N\}$ is the set of N nodes (drugs or diseases).

• $E \subseteq V\times V\mathrm{ }$ is the set of undirected edges representing similarity relationships.

• $W=\left\{w_{ij}\mathrm{ }\right|\mathrm{ }\left(v_i,v_j\right)\mathrm{ }\in \mathrm{ }E\}$ is the set of scalar edge weights $w_{ij}\in {\mathbb{R}}^{+}$ indicating similarity scores between nodes $v_i$ and $v_j$.

The graph was converted into a PyTorch Geometric [48] data object consisting of:

• Node features: $X\in {\mathbb{R}}^{N\times d_{\text{in}}}$​, initialized as constant vectors of dimension $d_{\text{in}}=d_{\text{embed}}$​ (the embedding size).

• Edge indices: edge_index $\in {\mathbb{Z}}^{2\times 2\left|E\right|}$, an undirected edge is represented as two directed ones.

• Edge weights: w  $\in {\mathbb{R}}^{2\left|E\right|}$ containing the similarity scores for each directed edge.

Weighted graph attention layer

We implemented a custom WeightedGATConv layer, extending the original GAT [49] formulation by incorporating scalar edge weights into the attention mechanism. Let:

• $h_{\mathrm{i}}\in {\mathbb{R}}^{d_{\text{in}}}$ ​ ​be the input feature vector of node $v_i$.

• H be the number of attention heads.

• $W^{\left(k\right)}\in {\mathbb{R}}^{d_{\text{hid}}\times d_{\text{in}}}$ ​ be the learnable weight matrix for head k.

• $a^{\left(k\right)}\in {\mathbb{R}}^{2d_{\text{hid}}}$ be the learnable attention vector for head k.

For each head k and edge $\left(v_i,v_j\right)\mathrm{ }\in \mathrm{ }E$:

• Feature transformation: ${z_i^{\left(k\right)}=W}^{\left(k\right)}{h}_{\mathrm{i}}$

• Unnormalized attention coefficient: $e_{ij}^{\left(k\right)}=\mathit{LeakyReLU}\left(a^{{\left(k\right)}^{\mathrm{{\rm T}}}}\right[z_i^{\left(k\right)}\left|\right|z_j^{\left(k\right)}]·w_{\text{ij}}$), where ‖ denotes vector concatenation and $w_{\text{ij}}$​ is the scalar edge weight.

  Normalization: (softmax over neighbors $\mathcal{N}(i$)): ${\alpha }_{ij}^{\left(k\right)}=\frac{\exp \left(e_{ij}^{\left(k\right)}\right)}{\sum _{v_{\mathrm{t}}\in \mathcal{N}\left(i\right)}\exp (e_{it}^{\left(k\right)})}$.

• Neighborhood aggregation: $h_i^{\left(k\right)\mathrm{\prime }}=\mathrm{\sigma }\left(\sum _{v_{\mathrm{j}}\in \mathcal{N}\left(i\right)}{\alpha }_{ij}^{\left(k\right)}{z}_j^{\left(k\right)}\right)$

  where σ(⋅) is a non-linear activation function (i.e. exponential linear unit [ELU] in our implementation).

The outputs from all heads were concatenated for intermediate layers and averaged for the final layer.

GAT encoder

Our GAT encoder (Fig. 2) comprised two WeightedGATConv layers:

Figure 2.

GAT encoder.

• First layer: H = 4 heads, hidden size $d_{\text{hid}}$​, ELU activation, dropout p = 0.5.

• Second layer: H = 1 head (no concatenation), output size $d_{\text{embed}}$.

The forward pass for a given input graph (X, edge_index, w) was:

$$Z={ \text{GAT}}_2(\text{Dropout}(\text{ELU}({\text{GAT}}_1(X, \text{edge}\_\text{index},\mathbf{w})))),$$

where $Z\in {\mathbb{R}}^{N\times d_{\text{embed}}}$​ is the matrix of learned embeddings.

Unsupervised training via graph reconstruction

The model was trained to preserve graph structure by distinguishing existing edges from randomly generated non-edges.

For each epoch t ∈{1,…,T}:

• Positive samples: For each observed edge $\left(v_i,v_j\right)\mathrm{ }\in \mathrm{ }E$, compute: $s_{ij}=z_i^{\mathrm{{\rm T}}}{\mathrm{z}}_j$ and positive loss:

  $${\mathcal{L}}_{\text{pos}}=-\frac{1}{\left|E\right|}\sum _{\left(v_i,v_j\right)\mathrm{ }\in \mathrm{ }E}\log \mathrm{ }\sigma \left(s_{ij}\right).$$
• Negative samples: Randomly select |$E_{\mathit{neg}}$| node pairs $\left(v_u,v_v\right)\mathrm{ }\notin \mathrm{ }E$. For these pairs, compute negative loss:

  $${\mathcal{L}}_{\text{neg}}=-\frac{1}{\left|E_{\mathrm{neg}}\right|}\sum _{\left(v_u,v_v\right)\mathrm{ }\in \mathrm{ }{E}_{\mathit{neg}}}\log (1-\sigma (s_{uv}\left)\right).$$

Finally, the total loss is: $\mathcal{L}={\mathcal{L}}_{\text{pos}}+{\mathcal{L}}_{\text{neg}}$.

Optimization was performed using the Adam optimizer with learning rate η = 0.01.

This process was applied separately to drug and disease similarity networks, yielding embedding matrices $Z_{ }^{\mathrm{drug}}$ and $Z_{ }^{\mathrm{disease}}$.

Drug–disease association prediction

For each drug–disease pair (dr_i, d_j), we constructed a feature vector:

$$x_{ij}^{ }=\left[Z_{i,:}^{\mathrm{drug}}\left|\right|Z_{j,:}^{\mathrm{disease}}\right]\in {\mathbb{R}}^{2d_{\text{embed}}}$$

where $Z_{i,:}^{\mathrm{drug}}$ and $Z_{j,:}^{\mathrm{disease}}$ are the embedding vectors of drug dr_i and disease d_j, respectively.

Negative samples were generated by randomly selecting unknown drug–disease pairs from the complement of the known association matrix, under the assumption commonly adopted in drug–disease association prediction that unlabeled pairs are treated as candidate negatives. To avoid class imbalance, the number of negative samples was set equal to the number of positive associations, resulting in a 1:1 positive-to-negative sampling ratio.

We used extreme gradient boosting (XGBoost) [50] for binary classification. XGBoost builds an ensemble of decision trees in sequence, each correcting the residuals of the previous trees. The objective function is:

$$\mathcal{L}(\Phi )=\sum _{i=1}^n\ell (y_i,{\widehat{y}}_i)+\sum _{k=1}^K\mathrm{\Omega }\left(f_k\right),$$

where:

• n is number of training samples (instances).

• ℓ(⋅,⋅) is the logistic loss for binary classification.

• $y_i$ ∈{0,1} is the true label.

• ${\widehat{y}}_i$ ​=${\sum }_{k=1}^K{f}_k\left(s_i\right)$ is the predicted probability. $s_i$ represents a constructed feature vector of a drug–disease pair.

• $\mathrm{\Omega }\left(f_k\right)=\gamma T+\frac{1}{2}\lambda {\Vert w\Vert }^2$ is the complexity penalty for tree $f_k$​ with T leaves and leaf weights w.

In our implementation, we set hyperparameters (number of trees K = 100, Maximum depth of each decision tree = 5, learning rate = 0.1, evaluation metric = log loss) for the XGBoost classifier.

Performance evaluation

We evaluated model performance using 10-fold stratified cross-validation, ensuring that the class distribution of positive and negative samples was preserved in each fold.

For each benchmark dataset, all known positive drug–disease associations were retained. Negative samples were generated by randomly selecting unknown drug–disease pairs from the complement of the known association matrix, using a 1:1 positive-to-negative sampling ratio. The resulting dataset was partitioned into 10 folds, with 90% of samples used for training and 10% for testing in each iteration.

Model performance was assessed using four evaluation metrics: Area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), F1-score, and accuracy. AUROC and AUPRC were used to evaluate ranking-based predictive performance, while F1-score and accuracy were included to assess classification effectiveness and the balance between precision and recall. Final performance was reported as the mean ± standard deviation of all metrics across the 10 cross-validation folds.

Results

Hyperparameter sensitivity analysis

To evaluate the robustness of SimNetRLDR and provide greater transparency regarding hyperparameter selection, we conducted a series of sensitivity analyses on key model and network construction parameters. Specifically, we examined the effects of the number of largest neighbors (kLN) used in constructing the MeSH-based disease similarity network, and the number of attention heads in the weighted GAT encoder applied during training. The hyperparameter analysis was conducted to evaluate the stability and robustness of SimNetRLDR under different parameter settings rather than to perform exhaustive hyperparameter optimization. A limited range of commonly used values was examined for the kLN and the number of attention heads. The final parameter configuration was selected based on both predictive performance and computational efficiency. The same hyperparameters were subsequently applied consistently across all experiments, including external validation and cold-start evaluations.

Effect of the kLN

To evaluate the sensitivity of SimNetRLDR to the sparsification level of the MeSH-based disease similarity network, we varied the number of retained largest neighbors (kLN) from 5 to 15 and assessed predictive performance using AUROC and AUPRC (Fig. 3a and b). The results show that SimNetRLDR achieved highly stable performance across all tested values, with only negligible variations in both AUROC and AUPRC. This indicates that the proposed framework is robust to moderate changes in local neighborhood size and is not strongly dependent on a specific network sparsification setting. Notably, the best performance was observed at kLN = 5, while larger values did not yield measurable improvement. Since smaller kLN values preserve a sparser network structure and reduce computational complexity while maintaining comparable predictive accuracy, kLN = 5 was selected as the default setting for all subsequent experiments. This result also suggests that preserving only the strongest local phenotypic similarity relationships is sufficient for effective disease representation learning.

Figure 3.

Sensitivity analysis of SimNetRLDR with respect to the number of largest neighbors (kLN) used in constructing the MeSH-based disease similarity network. (a) AUROC and (b) AUPRC obtained under different kLN values (5, 10, and 15). The results indicate that SimNetRLDR maintains stable predictive performance across varying neighborhood sizes, with kLN = 5 yielding the best overall performance and therefore selected as the default setting for subsequent experiments.

Effect of the number of attention heads

To evaluate the sensitivity of SimNetRLDR to architectural hyperparameters of the weighted GAT encoder, we examined the effects of varying the number of attention heads on predictive performance (Fig. 4). For the number of attention heads, values of 2, 4, and 8 were tested. As shown in Fig. 4a and b, predictive performance remained highly stable across all tested settings, with only minor fluctuations in AUROC and AUPRC. The best overall performance was achieved with eight attention heads, while four attention heads produced nearly identical results. Since increasing the number of attention heads substantially increases computational cost and memory consumption, four attention heads were retained as the default configuration, providing an effective balance between predictive accuracy and computational efficiency.

Figure 4.

Hyperparameter sensitivity analysis of SimNetRLDR with respect to weighted GAT encoder settings. (a) AUROC and (b) AUPRC under different numbers of attention heads (2, 4, and 8). The results demonstrate that SimNetRLDR maintains stable predictive performance across different architectural settings, supporting the robustness of the proposed framework. The default configuration of four attention heads was selected to balance predictive accuracy, and computational efficiency.

These results demonstrate that SimNetRLDR is robust to moderate variations in key encoder hyperparameters, further supporting the stability and reliability of the proposed framework.

Representation learning from similarity networks improves the prediction performance

To rigorously evaluate the effectiveness of similarity-network-based representation learning for drug–disease association prediction, we compared SimNetRLDR with representative state-of-the-art methods from two major methodological categories: deep learning-based approaches operating on HINs and conventional network-based methods based on random walk strategies. To ensure fair and reproducible comparison, all methods were evaluated under the same experimental framework using the gold-standard F-dataset, identical 10-fold cross-validation splits, and the same equal-size negative sampling strategy. Performance was assessed using AUROC, AUPRC, F1-score, and accuracy across multiple disease similarity network configurations.

Comparison with deep learning-based methods

To demonstrate that representation learning on similarity networks outperforms deep learning-based approaches operating on HINs, we compared SimNetRLDR with two recent state-of-the-art methods, DDAGDL [34] and RGLDR [35]. Specifically, DDAGDL constructs a HIN integrating drugs, diseases, and proteins through known drug–disease, drug–protein, and protein–disease associations. It first employs an autoencoder to generate initial representations from SMILES data for drugs, MeSH annotations for diseases, and protein sequence information, followed by GATs [49] to refine node representations, with XGBoost used for downstream association prediction. Similarly, RGLDR constructs a HIN comprising drugs, diseases, and proteins, applies metapath-guided random walks (e.g. drug → protein → disease) to generate regulation graphs, and uses regulation-aware graph representation learning to derive embeddings for prediction using XGBoost. In addition, we expanded the benchmark comparison to include two recent state-of-the-art multimodal and contrastive learning-based drug repositioning methods, namely AMDGT [37] and GCGB [38]. Specifically, AMDGT introduces a dual graph transformer together with an attention-aware multimodality interaction module to integrate similarity information with biochemical and chemical features more effectively than conventional feature concatenation approaches. GCGB further incorporates heterogeneous graph contrastive learning, fusion-view representation learning, and adaptive gradient balancing to alleviate label sparsity and optimization imbalance between the auxiliary contrastive objective and the primary drug–disease association prediction task.

In contrast, SimNetRLDR learns drug and disease representations independently from homogeneous similarity networks without directly incorporating known drug–disease associations during embedding construction. This design enables the model to exploit intrinsic similarity structures while reducing direct label leakage during representation learning. The detailed performance of SimNetRLDR across all combinations of training epochs and embedding sizes is presented in Supplementary Figs. 1 and 2 for DiSimNet_M and DiSimNet_P, respectively. For DiSimNet_M, the best performance was obtained with an embedding size of 512 and 400 epochs, whereas DiSimNet_P achieved optimal results with an embedding size of 512 and 200 epochs (Table 2). Figure 5 further illustrates the ROC and precision–recall curves across all 10 cross-validation folds for both disease similarity networks in the best configurations. The substantial overlap among the curves indicates highly stable predictive behavior across folds, confirming the robustness and reproducibility of SimNetRLDR. Despite these optimal configurations, the heatmaps (Supplementary Figs. 1 and 2) reveal consistently strong performance across all tested parameter settings, demonstrating the robustness of the learned representations.

Table 2.

Performance comparison between SimNetRLDR and representative deep learning-based methods using F-dataset.

Methods	AUROC	AUPRC	F1	ACC
SimNetRLDR
DiSimNetM	0.979 (±0.004)	0.980 (±0.006)	0.943 (±0.008)	0.944 (±0.008)
DiSimNetP	0.974 (±0.013)	0.977 (±0.012)	0.943 (±0.016)	0.943 (±0.015)
DDAGDL	0.848 (±0.030)	0.860 (±0.034)	0.777 (±0.029)	0.774 (±0.026)
RGLDR	0.829 (±0.038)	0.828 (±0.042)	0.763 (±0.038)	0.766 (±0.035)
GCGB	0.968 (±0.006)	0.863 (±0.028)	0.923 (±0.010)	0.922 (±0.009)
AMDGT	0.960 (±0.005)	0.962 (±0.003)	0.893 (±0.005)	0.891 (±0.003)

SimNetRLDR was evaluated using two single disease similarity networks: DiSimNet_M and DiSimNet_P. Performance was compared against DDAGDL and RGLDR using AUROC, AUPRC, F1-score, and Accuracy. Results are reported as mean ± standard deviation over 10-fold cross-validation. Superior performance is highlighted in bold.

Figure 5.

ROC and precision–recall curves of SimNetRLDR across 10-fold cross-validation for single disease similarity networks under the optimal hyperparameter settings on F-dataset. (a) ROC curves and (b) Precision–recall curves for DiSimNet_M using embedding size = 512 and epochs = 400. (c) ROC curves and (d) precision–recall curves for DiSimNet_P using embedding size = 512 and epochs = 200. The substantial overlap among curves across folds indicates stable and robust predictive performance of SimNetRLDR under both disease similarity network configurations.

Table 2 summarizes the comparative results. SimNetRLDR consistently outperformed both DDAGDL and RGLDR across all evaluation metrics under both disease similarity network configurations (DiSimNet_M and DiSimNet_P), including AUROC, AUPRC, F1-score, and accuracy. More specifically, using the MeSH-based disease similarity network (DiSimNet_M), SimNetRLDR achieved the strongest overall performance with an AUROC of 0.979 ± 0.004, AUPRC of 0.980 ± 0.006, F1-score of 0.943 ± 0.008, and Accuracy of 0.944 ± 0.008. Using the PPI-based disease similarity network (DiSimNet_P), similarly strong performance was obtained, with an AUROC of 0.974 ± 0.013, AUPRC of 0.977 ± 0.012, F1-score of 0.943 ± 0.016, and Accuracy of 0.943 ± 0.015. These results indicate that SimNetRLDR maintains stable predictive performance across distinct disease similarity formulations. In comparison, DDAGDL achieved an AUROC of 0.848 ± 0.030 and AUPRC of 0.860 ± 0.034, while RGLDR showed lower predictive capability, with AUROC and AUPRC values of 0.829 ± 0.038 and 0.828 ± 0.042, respectively. Similar performance gaps were observed for F1-score and Accuracy, further confirming the superiority of the proposed framework. In addition, despite the advanced multimodal and contrastive architectures adopted by AMDGT and GCGB, SimNetRLDR consistently achieved superior or competitive performance.

The superior performance of SimNetRLDR may be attributed to its decoupled representation learning strategy. By learning embeddings solely from similarity networks rather than directly embedding known drug–disease associations within a heterogeneous graph, SimNetRLDR more effectively captures intrinsic structural relationships among pharmacologically similar drugs and biologically related diseases while mitigating explicit label leakage during representation learning. These results suggest that similarity-network-based representation learning provides a more effective and generalizable framework for drug–disease association prediction than current heterogeneous-network-based deep learning approaches.

Comparison with network-based methods

To evaluate whether similarity-network-based representation learning offers advantages over conventional network-based approaches, we compared SimNetRLDR with two representative random walk-based methods, TP-NRWRH [24] and MHDR [31]. Both methods operate on heterogeneous drug–disease networks in which drug and disease similarity networks are explicitly connected through known drug–disease associations.

Specifically, TP-NRWRH applies a two-pass RWR algorithm on a heterogeneous network integrating the drug similarity network with a single disease similarity network, while MHDR extends this framework to multiplex-heterogeneous networks by incorporating multiple disease similarity layers. In contrast, SimNetRLDR learns latent drug and disease representations independently from homogeneous similarity networks before downstream association prediction, thereby avoiding direct embedding of known drug–disease associations during representation learning.

Table 3 summarizes the comparative results across both single and integrated/multiplex disease similarity network configurations. For the single disease similarity networks, the relative performance differences between the two methods further highlight the sensitivity of TP-NRWRH to network construction. Under the DiSimNet_M setting, TP-NRWRH achieved relatively strong predictive performance, with AUROC and AUPRC values of 0.951 ± 0.065 and 0.959 ± 0.040, respectively, indicating that this network provides sufficient topological information for effective random-walk-based prediction. However, its performance declined substantially when using DiSimNet_P, where AUROC and AUPRC dropped to 0.820 ± 0.208 and 0.861 ± 0.026, respectively, accompanied by markedly increased variance, suggesting reduced robustness and stability under this network configuration. In contrast, SimNetRLDR maintained consistently high performance across both single-network settings, indicating greater resilience to variations in disease similarity representation. These results suggest that while TP-NRWRH is strongly influenced by the structural properties of the input similarity network, SimNetRLDR is better able to extract stable predictive features from diverse disease similarity sources.

Table 3.

Performance comparison between SimNetRLDR and network-based methods on F-dataset.

Disease similarity network		Method	AUROC	AUPRC	F1	Accuracy
Single disease similarity network	DiSimNetM	SimNetRLDR	0.979 (±0.004)	0.980 (±0.006)	0.943 (±0.008)	0.944 (±0.008)
		TP-NRWRH	0.951 (±0.065)	0.959 (±0.040)	0.896 (±0.067)	0.896 (±0.067)
	DiSimNetP	SimNetRLDR	0.974 (±0.013)	0.977 (±0.012)	0.943 (±0.016)	0.943 (±0.015)
		TP-NRWRH	0.820 (±0.208)	0.861 (±0.026)	0.771 (±0.208)	0.771 (±0.208)
Integrated/multiplex disease similarity network	DiSimNetMP	SimNetRLDR	0.982 (±0.005)	0.983 (±0.003)	0.940 (±0.009)	0.941 (±0.009)
		MHDR	0.949 (±0.093)	0.966 (±0.047)	0.899 (±0.093)	0.899 (±0.093)
	DiSimNetMPH	SimNetRLDR	0.982 (±0.006)	0.984 (±0.005)	0.941 (±0.015)	0.942 (±0.014)
		MHDR	0.961 (±0.100)	0.970 (±0.045)	0.913 (±0.101)	0.913 (±0.101)

SimNetRLDR was compared against TP-NRWRH and MHDR using both single disease similarity networks (DiSimNet_M and DiSimNet_P) and integrated/multiplex disease similarity networks (DiSimNet_MP and DiSimNet_MPH). Performance was assessed using AUROC, AUPRC, F1-score, and Accuracy. Results are reported as mean ± standard deviation over 10-fold cross-validation. Superior performance is highlighted in bold.

For integrated and multiplex disease similarity networks, as MHDR works on multiplex-heterogeneous networks by incorporating multiple disease similarity layers, thus, to ensure a fair comparison with MHDR, we created an integrated disease similarity network, DiSimNet_MP, by combining DiSimNet_M and DiSimNet_P (see Materials and Methods section). Additionally, we constructed a HPO-based disease similarity network, DiSimNet_H, comprising 34,476 associations among 6,521 diseases, as described in MHDR [31]. We then integrated the three individual disease similarity networks (DiSimNet_M, DiSimNet_P, and DiSimNet_H) to form another integrated network, DiSimNet_MPH, containing 133,141 associations among 7496 diseases for SimNetRLDR, alongside a corresponding multiplex disease similarity network for MHDR. The detailed hyperparameter performance landscapes for SimNetRLDR on the integrated disease similarity networks are presented in Supplementary Figs. 3 and 4. Across all tested configurations, SimNetRLDR demonstrated highly stable predictive performance, indicating robustness to hyperparameter selection. For DiSimNet_MP, optimal performance was obtained with an embedding size of 512 and 200 epochs, while DiSimNet_MPH achieved its best performance with an embedding size of 512 and 100 epochs (Table 3). Figure 6 further presents ROC and Precision–Recall curves across all 10 cross-validation folds under the optimal hyperparameter settings. The strong overlap among folds confirms the reproducibility and stability of the proposed framework.

Figure 6.

ROC and precision–recall curves of SimNetRLDR across 10-fold cross-validation for integrated disease similarity networks under the optimal hyperparameter settings on F-dataset. (a) ROC curves and (b) Precision–recall curves for DiSimNet_MP using embedding size = 512 and epochs = 200. (c) ROC curves and (d) Precision–recall curves for DiSimNet_MPH using embedding size = 512 and epochs = 100. The strong overlap among curves across folds demonstrates stable and robust predictive performance of SimNetRLDR under integrated and multiplex disease similarity network configurations.

Table 3 shows that SimNetRLDR surpassed MHDR across all evaluated metrics. Using the integrated disease similarity network DiSimNet_MP, SimNetRLDR achieved an AUROC of 0.982 ± 0.005 and AUPRC of 0.983 ± 0.003, compared with 0.949 ± 0.093 and 0.966 ± 0.047 for MHDR. Under the more comprehensive DiSimNet_MPH configuration, SimNetRLDR maintained superior performance, reaching an AUROC of 0.982 ± 0.006 and AUPRC of 0.984 ± 0.005, outperforming MHDR by 2.1% and 1.4%, respectively.

Compared with the single-network settings, the integrated and multiplex networks reveal clear differences in methodological sensitivity. As observed previously, TP-NRWRH showed strong dependence on network choice, with performance dropping substantially from AUROC 0.951 ± 0.065 under DiSimNet_M to 0.820 ± 0.208 under DiSimNet_P. In contrast, MHDR exhibited notable recovery and improvement under the multiplex configurations, with AUROC increasing to 0.949–0.961 and substantially improved F1-score and Accuracy. This indicates that multi-source disease similarity fusion alleviates information sparsity and reduces the instability associated with relying on a single similarity representation.

By comparison, SimNetRLDR remained consistently robust across all disease similarity configurations. Its AUROC varied only marginally from 0.974 to 0.979 in the single-network settings to 0.982 in the integrated settings, while maintaining similarly stable AUPRC, F1-score, and Accuracy values. This limited variation highlights the ability of the graph representation learning framework to capture latent structural features regardless of network construction strategy. Nevertheless, the slight performance advantage under DiSimNet_MPH suggests that preserving disease similarity heterogeneity across multiple network layers provides additional structural cues that can be more effectively leveraged than direct network integration.

Overall, these findings indicate that although both methods benefit from richer disease similarity representations, SimNetRLDR demonstrates superior robustness and generalizability across all network configurations. The stronger dependence of MHDR on multiplex structures further emphasizes the advantage of SimNetRLDR in learning stable predictive representations from diverse disease similarity networks.

To further validate the robustness of SimNetRLDR, we additionally evaluated all methods on the independent C-dataset benchmark. As shown in Supplementary Tables 1 and 2, SimNetRLDR consistently achieved superior predictive performance compared with baseline methods under both single-network and multiplex disease similarity settings, confirming the generalizability of the proposed framework across benchmark datasets.

The superior performance of SimNetRLDR over TP-NRWRH and MHDR highlights the advantage of representation learning over diffusion-based random walk strategies. While random walk methods rely on iterative information propagation through predefined heterogeneous graph structures, SimNetRLDR learns expressive latent representations directly from similarity topology through weighted graph attention, enabling more effective capture of local structural relationships. Moreover, by decoupling embedding construction from known drug–disease associations, SimNetRLDR mitigates explicit label leakage and achieves stronger generalization across diverse disease similarity formulations.

Component-wide ablation study

To identify the key architectural components responsible for the performance of SimNetRLDR, we conducted a component-wise ablation study by systematically replacing the graph encoder and downstream classifier with alternative methods. Specifically, we evaluated the effects of substituting the proposed weighted GAT encoder with other graph representation learning architectures and replacing the XGBoost classifier with several commonly used classification models. These experiments were designed to assess the individual contributions of each component and to verify whether the observed performance gains arise from the proposed framework as a whole rather than from any single methodological choice.

Ablation study on graph encoders

To evaluate the contribution of the weighted graph attention mechanism to SimNetRLDR, we replaced the original weighted GAT encoder with two alternative graph representation learning architectures: graph transformer network (GTN) and graph convolutional network (GCN). As shown in Table 4, the original weighted GAT-based SimNetRLDR consistently achieved the best performance across all evaluation metrics, attaining an AUROC of 0.979 ± 0.004, AUPRC of 0.980 ± 0.006, F1-score of 0.943 ± 0.008, and accuracy of 0.944 ± 0.008.

Table 4.

Ablation study on graph encoder architectures within SimNetRLDR.

Methods	AUROC	AUPRC	F1	ACC
SimNetRLDR (GAT)	0.979 (±0.004)	0.980 (±0.006)	0.943 (±0.008)	0.944 (±0.008)
SimNetRLDR (GTN)	0.960 (±0.012)	0.962 (±0.016)	0.903 (±0.018)	0.904 (±0.019)
SimNetRLDR (GCN)	0.842 (±0.029)	0.850 (±0.035)	0.766 (±0.031)	0.777 (±0.027)

Performance comparison of SimNetRLDR using different graph encoders, including the proposed weighted graph attention network (GAT), graph transformer network (GTN), and graph convolutional network (GCN), evaluated using AUROC, AUPRC, F1-score, and Accuracy. Results are reported as mean ± standard deviation over 10-fold cross-validation. Superior performance is highlighted in bold.

Replacing weighted GAT with GTN resulted in a moderate performance decrease across all metrics (AUROC: 0.960 ± 0.012, AUPRC: 0.962 ± 0.016), indicating that although transformer-based graph modeling can capture global structural dependencies, it is less effective than weighted graph attention for exploiting local similarity relationships in drug and disease similarity networks. The performance degradation was substantially larger when weighted GAT was replaced by GCN, with AUROC and AUPRC dropping to 0.842 ± 0.029 and 0.850 ± 0.035, respectively.

These results demonstrate that the weighted GAT encoder is a critical component of SimNetRLDR. Its ability to adaptively assign attention coefficients to neighboring nodes while explicitly incorporating edge similarity weights enables more effective representation learning than both convolution-based and transformer-based alternatives. This confirms that the proposed weighted attention mechanism plays a central role in the superior predictive performance of SimNetRLDR.

Ablation study on classification methods

To assess the contribution of the downstream classifier to the overall performance of SimNetRLDR, we replaced the original XGBoost classifier with three alternative classification methods: Random forest (RF), multi-layer perceptron (MLP), and k-nearest neighbors (kNN). As shown in Table 5, the original XGBoost-based SimNetRLDR achieved the best performance across all evaluation metrics, obtaining an AUROC of 0.979 ± 0.007, AUPRC of 0.981 ± 0.009, accuracy of 0.943 ± 0.012, and F1-score of 0.944 ± 0.011.

Table 5.

Ablation study on downstream classification methods within SimNetRLDR.

Methods	AUROC	AUPRC	ACC	F1
SimNetRLDR (XGB)	0.979 (±0.007)	0.981 (±0.009)	0.943 (±0.012)	0.944 (±0.011)
SimNetRLDR (RF)	0.976 (±0.006)	0.979 (±0.006)	0.938 (±0.013)	0.940 (±0.012)
SimNetRLDR (MLP)	0.911 (±0.014)	0.904 (±0.020)	0.841 (±0.022)	0.838 (±0.021)
SimNetRLDR (kNN)	0.901 (±0.015)	0.877 (±0.019)	0.816 (±0.015)	0.793 (±0.019)

Performance comparison of SimNetRLDR using different classifiers, including XGBoost (XGB), random forest (RF), multi-layer perceptron (MLP), and k-nearest neighbors (kNN), evaluated using AUROC, AUPRC, accuracy, and F1-score. Results are reported as mean ± standard deviation over 10-fold cross-validation. Superior performance is highlighted in bold.

RF produced highly competitive results, with only a slight decrease in performance (AUROC: 0.976 ± 0.006, AUPRC: 0.979 ± 0.006), indicating that ensemble tree-based classifiers are well-suited for leveraging the learned drug–disease embedding representations. In contrast, substantially lower performance was observed when XGBoost was replaced by MLP or kNN. The MLP classifier achieved an AUROC of 0.911 ± 0.014, while kNN showed the weakest performance, with AUROC and AUPRC decreasing to 0.901 ± 0.015 and 0.877 ± 0.019, respectively.

These results suggest that the learned embeddings are most effectively exploited by tree-based ensemble methods, likely due to their ability to capture complex nonlinear interactions between drug and disease representations while maintaining strong generalization. The superior performance of XGBoost confirms that it is an appropriate and effective downstream classifier for the proposed framework.

Robustness and generalizability assessment

Sensitivity to negative sampling ratios

To evaluate the robustness of SimNetRLDR under more realistic class imbalance conditions, we assessed predictive performance using progressively increasing negative-to-positive sampling ratios ranging from 1:1 to 10:1. As shown in Fig. 7, SimNetRLDR maintained highly stable AUROC performance across all tested ratios, with values consistently remaining above 0.97 (Fig. 7a). This indicates that the model preserves strong discriminative capability even as the number of negative samples increases substantially.

Figure 7.

Predictive performance of SimNetRLDR under different negative sampling ratios. Performance was evaluated using negative-to-positive sampling ratios ranging from 1:1 to 10:1 under 10-fold cross-validation. (a) AUROC and (b) AUPRC across varying class imbalance settings. Error bars represent standard deviation across folds. The results demonstrate that SimNetRLDR maintains stable discriminative performance under increasingly imbalanced sampling conditions, although precision–recall performance gradually decreases as expected with increasing class imbalance.

In contrast, AUPRC exhibited a gradual decline as class imbalance increased, decreasing from approximately 0.98 at the balanced 1:1 ratio to 0.88 at the 10:1 ratio (Fig. 7b). This behavior is expected because AUPRC is inherently sensitive to class imbalance and becomes more challenging to optimize as the proportion of negative samples grows. Notably, despite this decline, SimNetRLDR maintained relatively strong precision–recall performance even under severe imbalance conditions, demonstrating robust ranking capability in more realistic prediction settings.

These results suggest that while balanced sampling provides favorable evaluation conditions, the proposed framework remains resilient when exposed to increasingly imbalanced negative sampling schemes that better reflect real-world drug–disease association sparsity. This analysis further supports the robustness and practical applicability of SimNetRLDR for large-scale drug repositioning tasks.

Cold-start evaluation for generalizability assessment

To further assess the robustness and generalizability of SimNetRLDR under stricter and more realistic prediction scenarios, we additionally conducted drug-wise and disease-wise cold-start evaluations. Unlike conventional pair-wise cross-validation, these settings prevent the model from observing any associations involving the test drugs or test diseases during training, thereby reducing potential information overlap and providing a more challenging evaluation framework.

As shown in Table 6, SimNetRLDR maintained strong predictive performance under the drug cold-start setting. Using the PPI-based disease similarity network (DiSimNet_P), the model achieved an AUROC of 0.963 ± 0.015 and AUPRC of 0.964 ± 0.021, while the MeSH-based network (DiSimNet_M) yielded AUROC and AUPRC values of 0.928 ± 0.019 and 0.933 ± 0.017, respectively. These results indicate that the learned drug similarity representations generalize well to previously unseen drugs when sufficient disease similarity information is available.

Table 6.

Performance of SimNetRLDR under drug-wise and disease-wise cold-start evaluation settings.

Evaluation setting	Disease similarity network	AUROC	AUPRC	F1-score	Accuracy
Drug cold-start	DiSimNetM	0.928 (±0.019)	0.933 (±0.017)	0.883 (±0.023)	0.884 (±0.025)
	DiSimNetP	0.963 (±0.015)	0.964 (±0.021)	0.939 (±0.025)	0.940 (±0.022)
Disease cold-start	DiSimNetM	0.756 (±0.096)	0.779 (±0.068)	0.734 (±0.042)	0.665 (±0.096)
	DiSimNetP	0.672 (±0.121)	0.667 (±0.104)	0.725 (±0.039)	0.653 (±0.080)

Drug cold-start evaluation excludes all associations involving test drugs from training, whereas disease cold-start evaluation excludes all associations involving test diseases. Results are reported as mean ± standard deviation across 10 folds.

In contrast, performance decreased more substantially under the disease cold-start setting. The best performance was obtained using DiSimNet_M, with AUROC and AUPRC values of 0.756 ± 0.096 and 0.779 ± 0.068, respectively, whereas DiSimNet_P achieved AUROC and AUPRC values of 0.672 ± 0.121 and 0.667 ± 0.104. This reduction likely reflects the greater heterogeneity and sparsity of disease similarity relationships compared with drug similarity networks, making representation transfer to previously unseen diseases substantially more challenging.

Importantly, although performance under the cold-start settings was lower than that observed in conventional pair-wise cross-validation, SimNetRLDR still demonstrated meaningful predictive capability, particularly in the drug cold-start scenario. These findings suggest that the proposed decoupled similarity-network representation learning framework captures transferable topological and biological information beyond simple memorization of known associations.

Overall, the cold-start experiments provide a more rigorous assessment of model generalizability and partially address concerns regarding optimistic estimates arising from randomly sampled negative pairs in conventional pair-wise evaluation settings.

External validation on an independent dataset

To further assess the robustness and generalizability of SimNetRLDR beyond internal cross-validation, we performed external validation using an independent benchmark dataset, namely the D-dataset (DN-dataset) proposed by Martínez et al. [46]. The original D-dataset contains 1008 experimentally validated drug–disease associations involving 1490 drugs and 4517 diseases. SimNetRLDR was trained exclusively on the F-dataset and subsequently evaluated on associations derived from the D-dataset. After removing associations overlapping with the F-dataset training set, 599 unique experimentally validated drug–disease associations remained for external testing. To compute AUROC, AUPRC, F1-score, and Accuracy, an equal number of randomly selected unlabeled drug–disease pairs were used as negative samples, resulting in a balanced external test set containing 599 positive and 599 negative pairs.

The external validation results are summarized in Table 7. SimNetRLDR achieved satisfactory predictive performance under both disease similarity network configurations. Notably, the PPI-based disease similarity network (DiSimNet_P) consistently outperformed the MeSH-based disease similarity network (DiSimNet_M) across all evaluation metrics. Specifically, DiSimNet_P achieved an AUROC of 0.814, AUPRC of 0.829, Accuracy of 0.621, and F1-score of 0.712, compared with 0.745, 0.770, 0.595, and 0.648, respectively, for DiSimNet_M.

Table 7.

External validation results on the independent D-dataset using DiSimNet_M and DiSimNet_P networks.

Disease similarity network	AUROC	AUPRC	ACC	F1
DiSimNetM	0.745	0.770	0.595	0.648
DiSimNetP	0.814	0.829	0.621	0.712

These findings suggest that disease similarity inferred from protein–protein interaction information captures biologically meaningful relationships that generalize more effectively across independent datasets than ontology-based similarity derived from MeSH descriptors. Although classification accuracy remains moderate, this is expected given the distributional differences between benchmark datasets and the inherent uncertainty of unlabeled drug–disease pairs, some of which may correspond to undiscovered true associations. Overall, the external validation confirms the robustness and transferability of SimNetRLDR and highlights the importance of biologically informed disease similarity modeling for drug–disease association prediction.

Prediction of novel drug-disease associations

In the prior section, we showed that SimNetRLDR achieved optimal performance with an embedding size of 512 and 100 epochs for DiSimNet_MP. Here, we utilized this configuration to predict novel drug-disease associations and assessed their practical significance using molecular and clinical evidence. To evaluate the biological relevance of newly predicted associations, the model was trained on known drug–disease associations (positive set) and an equal number of randomly selected drug–disease pairs (negative set). Predictions were then made on the remaining drug–disease pairs, with the top 100 highest-scoring pairs identified as promising drug–disease associations for further evidence analysis.

These top-ranked candidate associations were subsequently examined for supporting biological evidence based on shared KEGG pathways between drug target genes and disease-associated genes. Following previous pathway-based drug repositioning studies [31, 51–54], we adopted a pathway-overlap-based validation strategy rather than statistical pathway enrichment analysis. Specifically, a candidate association was considered biologically supported if the drug targets and disease-associated genes shared at least one curated KEGG pathway. Using this criterion, 23 out of the top 100 predicted drug–disease associations were found to have pathway-supported evidence (Supplementary Table 3), while the remaining 77 candidate pairs did not exhibit shared KEGG pathway overlap under the adopted annotation resources.

The pathway-supported associations were further manually evaluated using ClinicalTrials.gov records using the “clintrialx” R package. After manual evaluation of all ClinicalTrials.gov records using strict evidence criteria that excluded studies reporting only incidental co-occurrence, symptom management, unrelated comorbidities, or adverse-event monitoring, 12 of these 23 associations were considered to have plausible clinical validation (Table 8). Representative examples are discussed below. Figure 8 illustrates these associations through shared KEGG pathways.

Table 8.

A list of 12 out of 23 promising drug-disease associations supported by shared pathways and clinical trials.

Drug ID (KEGG ID)	Drug name	Disease ID (MIM ID)	Disease name	Shared pathways (KEGG ID)	ClinicalTrials.gov (NCT ID)
D00120	Sulindac	114480	Breast Neoplasms	map04064, map04370, map04611, map04625, map04657, map04668, map04726, map04921, map04923, map05163, map05165, map05167, map05200, map05206, map05222	NCT00037609, NCT00039520, NCT04542135, NCT01761877, NCT00245024
D00120	Sulindac	276300	Neoplastic Syndromes, Hereditary	map05200	NCT01187901, NCT00319007, NCT01483144, NCT00927485, NCT01245816
D00141	Indomethacin	114480	Breast Neoplasms	map04064, map04370, map04611, map04625, map04657, map04668, map04726, map04921, map04923, map05163, map05165, map05167, map05200, map05206, map05222	NCT02950259, NCT05035407, NCT01612728
D00141	Indomethacin	276300	Neoplastic Syndromes, Hereditary	map05200	NCT01471639
D00142	Methotrexate	181500	Schizophrenia	map00670	NCT02074319, NCT03369795
D00332	Gabapentin	600807	Asthma	map04010, map04261, map05410, map05414	NCT00153283, NCT04278404
D00426	Risperidone	167870	Anxiety Disorders	map04728	NCT00167479, NCT00231335, NCT05737511, NCT01342120, NCT06611592, NCT03227562, NCT03790085, NCT00277654, NCT02593734, NCT00854919, NCT02374567, NCT00086112, NCT04478526, NCT00389493, NCT00206765, NCT07060872, NCT05401019, NCT00457106, NCT00215241
D00506	Phenobarbital	181500	Schizophrenia	map04080, map04726	NCT02374567
D00554	Ethinylestradiol	125853	Diabetes Mellitus, Type 2	map04915, map04917, map04919, map05200, map05205, map05224	NCT00504816, NCT01458210, NCT02845219, NCT06942936, NCT01324505, NCT06513104, NCT05153564, NCT01508858, NCT00254800, NCT01077505
D01136	Tranexamic acid	600807	Asthma	map04080, map05164	NCT04278404
D08128	Liothyronine	114480	Breast Neoplasms	map04919	NCT03787303, NCT06820554, NCT01360177, NCT00725946, NCT02126449, NCT07044791
D08560	Tamsulosin	608622	Hypertension	map04022, map04270	NCT00223717

Figure 8.

Visualization of drug-disease associations via shared KEGG pathways. (a) Sulindac and breast neoplasms; (b) Sulindac and neoplastic syndromes, hereditary; (c) Indomethacin and neoplastic syndromes, hereditary; (d) Methotrexate and schizophrenia; (e) Gabapentin and asthma; (f) Risperidone and anxiety disorders; (g) Phenobarbital and schizophrenia; (h) Ethinylestradiol and diabetes mellitus, Type 2; (i) Tranexamic acid and asthma, (j) Liothyronine and breast neoplasms; and (k) Tamsulosin and hypertension. Edge weight is proportional to the number of shared genes between a drug or disease and a pathway. For (a), pathways are 1: NF-kappa B signaling pathway, 2: VEGF signaling pathway, 3: Platelet activation, 4: C-type lectin receptor signaling pathway, 5: IL-17 signaling pathway, 6: TNF signaling pathway, 7: Serotonergic synapse, 8: Oxytocin signaling pathway, 9: Regulation of lipolysis in adipocytes, 10: Human cytomegalovirus infection, 11: Human papillomavirus infection, 12: Kaposi sarcoma-associated herpesvirus infection, 13: Pathways in cancer, 14: MicroRNAs in cancer, and 15: Small cell lung cancer.

• Sulindac (D00120) and Breast Neoplasms (114480): Sulindac, a nonsteroidal anti-inflammatory drug, is associated with breast cancer through 15 KEGG pathways, including NF-kappa B signaling (map04064), TNF signaling (map04668), and pathways in cancer (map05200), indicating roles in inflammation modulation and tumor progression. Clinical evidence is supported by trials such as NCT00037609, NCT00039520, NCT04542135, and NCT01761877, which investigated sulindac for breast cancer prevention, recurrence reduction, or therapeutic intervention.

• Sulindac (D00120) and Neoplastic Syndromes, Hereditary (276300): This association is supported by pathways related to colorectal tumorigenesis and inflammatory signaling. Trials including NCT01187901, NCT00319007, and NCT01483144 evaluated sulindac in patients with familial adenomatous polyposis (FAP) and hereditary cancer predisposition syndromes, supporting its chemopreventive relevance.

• Indomethacin (D00141) and Breast Neoplasms (114480): Indomethacin is linked through inflammatory signaling pathways relevant to breast tumor progression. Clinical evidence includes NCT02950259 and NCT05035407, where indomethacin was used as part of therapeutic regimens or supportive interventions in breast cancer settings.

• Methotrexate (D00142) and Schizophrenia (181500): Methotrexate is connected through the one-carbon pool by folate pathway (map00670), which may influence neuroinflammatory and neurotransmitter-related mechanisms implicated in schizophrenia. Clinical trials NCT02074319 and NCT03369795 directly evaluated methotrexate as an adjunctive therapy for schizophrenia, supporting therapeutic relevance.

• Gabapentin (D00332) and Asthma (600807): Gabapentin is linked via MAPK signaling (map04010) and related neuroimmune pathways. Although evidence is limited, NCT00153283 directly investigated gabapentin in asthma therapy, suggesting potential relevance beyond incidental mention.

• Ethinyl Estradiol (D00554) and Type 2 Diabetes Mellitus (125853): This association is supported by estrogen signaling pathways (map04915), which may influence insulin sensitivity and glucose metabolism. Trials including NCT00504816, NCT02845219, and NCT01324505 assessed pharmacokinetic interactions of ethinyl estradiol in participants with type 2 diabetes or with antidiabetic agents, indicating clinically relevant metabolic interaction.

• Tranexamic Acid (D01136) and Asthma (600807): Clinical evidence is limited but plausible, primarily based on NCT04278404, where tranexamic acid was included among monitored drugs in pediatric asthma pharmacokinetic studies. This suggests preliminary relevance warranting further investigation.

• Liothyronine (D08128) and Breast Neoplasms (114480): Liothyronine is linked through thyroid hormone signaling (map04919), a pathway implicated in tumor metabolism and proliferation. Multiple trials, including NCT03787303, and NCT01360177, evaluated thyroid hormone manipulation in breast cancer imaging or treatment contexts, supporting mechanistic plausibility.

Overall, these findings demonstrate that our framework not only achieves strong predictive performance (AUROC/AUPRC), but also identifies biologically plausible drug–disease associations with emerging clinical support. At the same time, the manual curation highlights the importance of rigorous evidence assessment when translating computational predictions into actionable drug repositioning hypotheses.

Conclusions and discussion

In this study, we proposed SimNetRLDR, a similarity network-based representation learning framework for computational drug repositioning. Unlike conventional heterogeneous graph approaches that directly incorporate known drug–disease associations into the representation learning process, SimNetRLDR independently learns drug and disease embeddings from homogeneous similarity networks using a weighted GAT encoder, followed by downstream association prediction using XGBoost. This design reduces the risk of explicit label leakage during embedding construction, while enabling more effective exploitation of intrinsic topological and biological relationships within similarity networks.

Comprehensive experiments on two benchmark datasets demonstrated that SimNetRLDR consistently outperformed existing state-of-the-art methods across multiple disease similarity network configurations. In addition to conventional baselines, we further incorporated comparisons with recent multimodal and contrastive learning-based approaches. Despite their more complex architectures and reliance on heterogeneous auxiliary biomedical modalities, SimNetRLDR achieved consistently superior or competitive performance. These findings indicate that similarity-network-based representation learning alone can provide highly informative representations for drug–disease association prediction without requiring extensive multimodal integration. The proposed framework achieved strong and stable predictive performance under both single disease similarity networks and integrated/multiplex disease similarity settings. In particular, integrated and multiplex disease similarity networks generally provided superior predictive performance compared with single-network settings, indicating that combining complementary disease similarity information can enhance representation quality and improve downstream drug–disease association prediction.

Beyond standard pair-wise cross-validation, we additionally conducted stricter drug-wise and disease-wise cold-start evaluations to assess the generalizability of the proposed framework under more realistic unseen-entity prediction scenarios. SimNetRLDR maintained strong predictive performance under drug cold-start settings, with AUROC values up to 0.963 and AUPRC values up to 0.964 using the PPI-based disease similarity network. Although performance decreased under disease cold-start settings, the framework still achieved meaningful predictive capability. These results suggest that SimNetRLDR generalizes more effectively to previously unseen drugs than to entirely unseen diseases, likely because disease similarity networks exhibit greater biological heterogeneity and sparsity. External validation on an independent dataset further confirmed the robustness and generalizability of the proposed framework.

Extensive sensitivity analyses demonstrated that SimNetRLDR remains highly stable across different neighborhood sizes, attention head numbers, embedding dimensions, and training epochs. The hyperparameter experiments additionally showed that relatively small neighborhood sizes (e.g. kLN = 5) were sufficient to preserve informative local similarity structures while avoiding the introduction of noisy long-range relationships, whereas four attention heads provided a favorable balance between representational diversity and computational efficiency. Component-wise ablation studies additionally showed that the weighted GAT encoder consistently outperformed alternative graph encoders such as GCN and GTN, while XGBoost achieved better downstream classification performance than RF, MLP, and kNN classifiers. These findings suggest that adaptive attention-based aggregation over weighted similarity networks plays a key role in capturing informative biological patterns relevant to drug repositioning.

To further investigate robustness under realistic class imbalance conditions, we evaluated SimNetRLDR using progressively increasing negative-to-positive sampling ratios ranging from 1:1 to 10:1. The results showed that AUROC remained highly stable across imbalance settings, whereas AUPRC gradually decreased as class imbalance increased, consistent with the known sensitivity of precision–recall metrics to skewed label distributions. These findings indicate that the proposed framework remains robust under substantially more challenging evaluation conditions while also highlighting the inherent difficulty of drug–disease association prediction under extreme sparsity.

To further evaluate practical relevance, we investigated high-confidence predicted drug–disease associations using shared KEGG pathway analysis and manually curated ClinicalTrials.gov evidence. After rigorous manual filtering to exclude incidental co-occurrences, symptom-management-only studies, unrelated comorbidities, and adverse-event monitoring records, 12 predicted associations were found to have plausible clinical support. Representative examples included Sulindac for breast neoplasms and hereditary neoplastic syndromes, Methotrexate for schizophrenia, Gabapentin for asthma, Ethinyl Estradiol for type 2 diabetes mellitus, Tranexamic Acid for asthma, and Liothyronine for breast neoplasms. These findings suggest that SimNetRLDR is capable of identifying biologically meaningful and clinically relevant drug repositioning candidates.

Despite these promising results, several limitations should be acknowledged. First, although SimNetRLDR reduces direct label leakage by excluding known drug–disease associations during representation learning, indirect or second-order information leakage may still exist because similarity networks themselves can implicitly encode prior biomedical knowledge related to known therapeutic associations. Second, negative samples were generated from unknown drug–disease pairs using random sampling strategies, and some unlabeled pairs may correspond to undiscovered true associations. Therefore, the current formulation should more appropriately be viewed as a positive–unlabeled (PU) learning problem rather than a strict supervised binary classification task. Although robustness experiments under increasing class imbalance ratios and cold-start evaluations demonstrated stable predictive behavior, more realistic strategies such as ranking, semi-supervised learning (e.g. PU-learning) frameworks, or hard-negative mining approaches may further improve reliability and real-world applicability in future studies. Third, the current framework primarily relies on predefined biological similarity measures and does not yet fully exploit rich semantic information available from biomedical literature, knowledge graphs, or multimodal molecular data. In addition, although the current study provides pathway-level and clinical-trial-supported evidence, experimental validation through in vitro, in vivo, or prospective clinical studies remains necessary to fully establish translational relevance.

Future work may address these limitations by incorporating advanced representation learning techniques. Recent studies have demonstrated that large language models (LLMs) can generate rich biomedical semantic embeddings by integrating diverse textual and knowledge graph information, substantially improving molecular relationship prediction tasks [55]. Incorporating LLM-derived drug and disease embeddings into SimNetRLDR could provide more expressive representations and better capture latent biomedical semantics beyond handcrafted similarity metrics. Additionally, contrastive learning offers a promising direction for aligning heterogeneous biological views while preserving complementary information. Multi-view contrastive frameworks have recently shown strong performance in drug interaction prediction by learning robust cross-modal representations [56]. Integrating contrastive objectives into our framework could enhance consistency across similarity layers and improve generalization for sparse or noisy biological networks. Furthermore, recent advances in biomedical representation learning highlight several promising directions for extending SimNetRLDR. Zhang et al. [57] demonstrated that 3D spatial multi-view molecular representations can capture fine-grained structural and conformational characteristics of compounds that are not explicitly represented by similarity-network-derived embeddings. While SimNetRLDR currently relies on biological and topological similarity information, incorporating learned 3D molecular representations may provide complementary chemical and spatial information for improved drug characterization. In addition, Wu et al. [58] showed that metapath-enhanced multisource graph fusion can effectively model complex semantic relationships across heterogeneous biomedical entities. Compared with such metapath-guided heterogeneous graph approaches, SimNetRLDR adopts a similarity-network-based strategy that emphasizes homogeneous biological relationships and representation learning from integrated or multiplex similarity networks. Future research may combine the strengths of both paradigms by integrating metapath-guided semantic information, heterogeneous graph fusion, and similarity-network-based embeddings within a unified framework, potentially further improving predictive performance, biological interpretability, and generalization ability.

Overall, the results demonstrate that SimNetRLDR provides an effective and robust framework for drug–disease association prediction by leveraging similarity network-based representation learning. The proposed approach offers a promising direction for identifying biologically plausible drug repositioning candidates and may contribute to accelerating drug discovery and precision medicine research.

Author contributions

The author contributed to all aspects of this study.

Supplementary material

Supplementary material is available at Biology Methods and Protocols online.

Conflicts of interest

None declared.

Funding

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

Data availability

The data and source code are hosted on Github (https://github.com/hauldhut/SimNetRLDR).