DIAM: Adaptive Drug Repositioning with Decoupled Biological Mechanism and Instance-Aware Modulation

1 Introduction

Over the past few decades, the traditional model of drug research and development (R&D) has faced significant challenges, characterized by high costs, lengthy development cycles, and high failure rates [Huang et al., 2025]. Statistics indicate that only about 11% of the candidate drugs that enter human clinical trials ultimately receive regulatory approval [DiMasi et al., 2003; Kola and Landis, 2004]. In this context, drug repositioning, which identifies new indications for existing drugs, has emerged as a pivotal strategy to mitigate failure risks and accelerate therapeutic discovery [Lei et al., 2024]. To support this strategy, computational methods, particularly deep learning, have been widely adopted to predict potential DDAs.

∗Corresponding Author.

Despite the promising performance of recent deep learning models, the dominant computational paradigm in DDA prediction still suffers from two fundamental limitations, arising from a mismatch between simplified computational assumptions and complex biological realities: (1) Prevalent methods predominantly rely on random negative sampling [Huang et al., 2024; Liu et al., 2024], which indiscriminately treats unobserved drug-disease pairs as uniform negatives. Since random samples are often topologically distant from true associations, they fail to provide sufficient gradient penalties for ambiguous candidates, reducing the training process to a trivial discrimination against easy negatives. This deficiency causes the resulting decision boundary to remain insufficiently compact, compromising the model's capability to discern fine-grained discrepancies between verified interactions and ambiguous candidates. (2) Existing approaches typically employ a static, monolithic fusion logic (e.g., a fixed dot-product) to integrate heterogeneous biological features [Liu et al., 2024; Wang et al., 2023; Yuan et al., 2024]. This one-size-fits-all strategy fails to accommodate instancespecific heterogeneity, overlooking that the dominant informative view varies dynamically between global structural propagation and local molecular interactions across different drug-disease pairs. Consequently, this rigidity leads to suboptimal representation learning for complex biological associations.

To address these challenges, we propose DIAM, a novel framework that rethinks DDA prediction from the perspective of decoupled biological mechanisms. The overall architecture of DIAM is depicted in Figure 1. Specifically, DIAM first constructs a Dual-Encoder Representation Learning backbone. We employ Heterogeneous Graph Transformer (HGT) and Hypergraph Convolutional Networks (HGCN) to capture heterogeneous structural patterns and high-order correlations, respectively. Distinct from previous approaches that rely on implicit feature propagation, we introduce a DualStream Biological Mechanism Decoupling module to explicitly quantify association evidence into topological propagation signals and target-mediated interaction signals. These decoupled signals serve two critical functions. First, we utilize them to calculate a biological plausibility score, which guides the mining of hard negative samples and forces the model to learn robust decision boundaries. Second, we design an Adaptive Residual Gating to perform instance-aware modulation. This module uses prior confidence projection to dynamically modulate the fusion of the learned representations. Additionally, we integrate a contrastive learning objective to align the semantic spaces of the dual-encoder representations. Finally, the modulated representations are fed into a predictor to infer association probabilities. The main contributions of this study are summarized as follows.

• We propose DIAM, a novel framework that orchestrates a Dual-Encoder backbone with a Dual-Stream Mechanism Decoupling module to synthesize the biological semantics underlying both global structural and local molecular views.
• We introduce two mechanism-driven strategies leveraging the decoupled signals: (1) Hard Negative Sampling, which utilizes biological plausibility scores to identify ambiguous candidates and compel the model to learn finer-grained discriminative features; and (2) Adaptive Residual Gating for instance-aware modulation, which dynamically rectifies the fusion of learned representations.
• Extensive experiments on three benchmark datasets demonstrate that DIAM outperforms the state-of-the-art methods.

2 RelatedWork

2.1 Network Diffusion-BasedMethods

Network diffusion-based methods infer associations by simulating information propagation across heterogeneous biological graphs. Representative works, such as TL-HGBI [Wang et al., 2014] and MBiRW [Luo et al., 2016], utilize update algorithm or random walk algorithms to predict potential drug-disease pairs. Despite offering strong interpretability, these methods are heavily reliant on the network's existing edges. This dependency hinders their ability to capture complex drug-disease associations, particularly with sparse data, limiting their predictive performance.

2.2 Machine Learning-BasedMethods

Machine learning-based methods treat drug-disease association prediction as a classification or regression problem, making predictions by learning patterns from the feature vectors of drugs and diseases. Early works exhibit diverse strategies: PREDICT [Gottlieb et al., 2011] utilizes logistic regression, whereas FBLM [Ding et al., 2020] adopts a fuzzy SVM to model drug-disease relationships. Subsequently, SFRLDDA [Zhao et al., 2023] employs random forests, while DDASKF [Gao et al., 2022] leverages Laplacian Regularized Least Squares (LapRLS) to integrate heterogeneous similarities. However, machine learning-based methods often rely on features extracted by shallow learning frameworks, making it difficult to capture the deep, abstract associations hidden between drugs and diseases, thereby hindering further improvement of prediction performance [Aittokallio, 2022].

2.3 Deep Learning-BasedMethods

Deep learning-based methods leverage complex neural architectures to extract nonlinear representations from biological networks. For example, deepDR [Zeng et al., 2019] employs multimodal deep autoencoders to integrate multiple drug-related network information. In another case, LAGCN [Yu et al., 2021] applies a Graph Convolutional Network (GCN) to learn node embeddings from a heterogeneous network with a hierarchical attention mechanism. Beyond these, REDDA [Gu et al., 2022] learns node representations on a five-element heterogeneous network containing drugs, proteins, genes, pathways, and diseases by combining GNN with multiple attention mechanisms. More recently, AMDGT [Liu et al., 2024] utilizes a dual-graph transformer to learn features from both similarity networks and three-element heterogeneous networks.

3 Methodology

3.1 Network Construction

Association Networks

In this work, three types of associations are considered: drugdisease associations, drug-protein associations, and diseaseprotein associations. These are denoted Ard ∈ RNr×Nd , Arp ∈ RNr×Np , Adp ∈ RNd×Np , where Nr, Nd, and Np represent the number of drugs, diseases, and proteins. Taking Ard as an example, if drug ri and disease dj are confirmed to be associated, the corresponding element Ard ij is assigned a value of 1; otherwise, it is assigned a value of 0.

Similarity networks

For the drug similarity network Sr , we employ the RDKit tool to process the SMILES strings of drugs and compute the fingerprint similarity for each drug, denoted as Srf [Steinbeck et al., 2003]. Furthermore, the Gaussian Interaction Profile (GIP) kernel similarity based on known DDAs is incorporated to supplement the fingerprint similarity when it is insufficient. Specifically, the GIP similarity Srg (ri, rj) between drug ri and rj is calculated as follows:

where IP(ri) represents the corresponding column of drug ri in Ard , and γ ′r is set to 1. The final drug similarity matrix Sr is defined as:

Similarly, the disease similarity network Sd is computed based on Medical Subject Headings (MeSH) descriptors

[Van Driel et al., 2006] and GIP similarity. The protein similarity network integrates sequence similarity and GIP kernel similarity, where sequence similarity is calculated using the cosine similarity measure to compare 3-mer feature vectors. It is worth noting that the GIP similarity for proteins can be calculated from both drug and disease perspectives. Therefore, the two GIP matrices are averaged to derive the final GIP similarity.

3.2 Dual-Encoder Representation Learning Backbone

Heterogeneous Graph Transformer

We employ HGT to learn the interaction patterns between different types of nodes in the heterogeneous graph [Hu et al., 2020; Mei et al., 2022; Ni et al., 2023]. HGT aggregates messages from neighbors N(t) using attention weights specific to the meta-relation ⟨τ(s), ϕ(e), τ(t)⟩. In particular, the intermediate representation at the l-th layer is computed as:

Subsequently, the node embedding is updated via a typespecific linear projection and a residual connection:

Through this process, we obtain contextually enriched representations for drug node Hhgt r and disease nodes Hhgt d . These representations will serve as key inputs for the subsequent association prediction task.

Hypergraph Convolutional Network

To capture high-order correlations beyond pairwise connections, we construct homogeneous hypergraphs Gr and Gd using the K-Nearest Neighbors (KNN) strategy [Huang et al., 2022; Ouyang et al., 2024; Wang et al., 2024]. Using the drug hypergraph as an example, for each drug node v ∈ Vr, a hyperedge ev is formed by grouping v with its K most similar neighbors based on Sr . Since each node generates a unique hyperedge, this results in a square incidence matrix H, defined as:

In the constructed homogeneous hypergraphs, we employ HGCN to encode group-level dependencies. Its feature propagation rule at the l - th layer is defined as:

where X(l) and θ(l) denote the node feature matrix and learnable filter at the l - th layer. Dv, De, and W represent the diagonal matrices of node degrees, hyperedge degrees, and hyperedge weights, respectively. By applying HGCN to the drug hypergraph Hr and the disease hypergraph Hd, we finally obtain high-order embeddings Hhg r and Hhg d .

3.3 Dual-Stream Biological Mechanism Decoupling

Global Structural Propagation View

This global structural propagation view quantifies the topological propagation signal by modeling the global reachability. The core idea is that the association potential of a pair (ri, dj) can be quantified by leveraging the similarity information of its topological neighbors.

Formally, let Nr(j) = {k | Ard kj = 1 ∧ k ̸= i} denote the set of drugs treating disease dj, and Nd(i) = {l | Ard il = 1 ∧ l ̸= j} denote the set of diseases treated by drug ri. We define the topological propagation signal Ctopo(ri, dj) as a two-side weighted aggregation:

where the hyperparameter α ∈ [0, 1] is used to balance the relative contributions of the drug-side and disease-side information.

Local Molecular Interaction View

Complementing the global topology, this view explicitly models the molecular interaction signal driven by shared protein targets. We first define the importance I(t) for each protein target t based on its connectivity degree distributions in Arp and Adp , followed by Min-Max normalization to obtain I norm(t). The specific formula is as follows:

where ϵ1 is a small constant 10-8 to avoid division by zero. Finally, the molecular interaction signal Ctarget(ri, dj) is derived by aggregating the normalized importance of all shared targets that bridge drug ri and disease dj:

3.4 Mechanism-Driven Learning Strategies

Prior Confidence Projection

For the purpose of integrating the prior information with implicit deep representations, we propose the Prior Confidence Projection (PCP) module. Given the normalized topological propagation signal ˆ Ctopo(ri, dj) and the target-mediated interaction signal ˆ Ctarget(ri, dj), PCP projects these mechanism indicators into a unified latent confidence scalar βij:

Hard Negative Sampling

To enforce a finer-grained decision boundary, we design a Hard Negative Sampling strategy that identifies unobserved pairs exhibiting high biological plausibility. We first quantify this plausibility by synthesizing the decoupled signals, weighted by the projected confidence βij, to generate the biological plausibility score W:

Consequently, the hard negative pool Nhard is constructed by filtering unobserved pairs that exceed a plausibility threshold θneg:

During training, negative samples are drawn from Nhard to penalize ambiguous candidates and sharpen the model's discriminative capability.

Adaptive Residual Gating

To overcome the limitations of monolithic fusion strategies and perform instance-aware signal rectification, we propose the Adaptive Residual Gating (ARG) module. ARG adopts a residual enhancement formulation modulated by the projected confidence βij .

Formally, given the base interaction embedding zij = hhgt ri ⊙ hhgt dj , the final rectified representation zfinal ij is computed as:

where λ controls the modulation intensity. This residual gain selectively amplifies biologically plausible signals, thereby enhancing the expressiveness of embeddings.

3.5 Model Optimization

Contrastive Learning

To jointly optimize the heterogeneous graph and the homogeneous hypergraph views, we adopt the InfoNCE loss [Oord et al., 2018] for dual-view contrastive learning. Considering drug nodes as an example, let hhgt ri and hhg ri denote the embedding vectors of drug ri in the two views. The cross-view contrastive loss is designed to maximize the mutual information across views:

where Pri and Nri are the positive and negative sample sets for ri, and τ is the temperature parameter. γc is the balancing coefficient for the cross-view loss.

Complementarily, the intra-view contrastive loss for drug nodes is formulated as follows:

The total contrastive loss for drugs is defined as Lr = L crossr +L intrar. Similarly, the contrastive loss for diseases Ld can be derived.

Joint Optimization Objective

We employ a Multi-Layer Perceptron (MLP) as the decoder, optimizing the supervised task via the cross-entropy loss Lcls. To simultaneously leverage the benefits of self-supervised alignment and supervised association mining, the overall objective function is defined as:

where η ∈ [0, 1] is a hyperparameter that balances the contribution of the contrastive learning task and the classification task to the model's optimization.

4 Experiments

4.1 Experimental Setup

Datasets

We evaluate our proposed DIAM on three public benchmark datasets, each comprising three types of biological entities: drugs, diseases, and proteins, along with their association networks. Table 1 summarizes the dataset statistics. Cdataset[Luo et al., 2016], sourced from DrugBank [Knox et al., 2024] and Online Mendelian Inheritance in Man (OMIM) [Hamosh et al., 2005], contains 663 drugs, 409 diseases, and 2532 drug-disease associations. Bdataset[Zhang et al., 2018], extracted from the Comparative Toxicogenomics Database (CTD) [Davis et al., 2023], includes 269 drugs, 598 diseases, and 18416 drug-disease associations. Fdataset[Gottlieb et al., 2011], obtained from DrugBank and OMIM, comprises 593 drugs, 313 diseases, and 1933 drug-disease associations.

Evaluation metrics

To ensure a comprehensive and fair performance evaluation, we compare DIAM against all baselines using six standard metrics: Area Under the Receiver Operating Characteristic Curve (AUC), Area Under the Precision-Recall Curve (AUPR), Accuracy (ACC), Precision, Recall, and F1-Score. The parameter settings for all baseline models are kept consistent with their papers. All results are reported as the mean ± standard deviation with different random seeds to ensure stability and reproducibility.

Baselines

We compare DIAM with seven state-of-the-art methods, including DRHGCN [Cai et al., 2021], DRWBNCF [Meng et al., 2022], SMGCL [Gao et al., 2023], AMDGT [Liu et al., 2024], MilGNet [Gu et al., 2025], DFDRNN [Zhu et al., 2025], SIGDR [Cui et al., 2025].

4.2 Performance Comparison

We conduct 10-CV on benchmark datasets and compare DIAM with seven state-of-the-art baseline methods. As shown in Table 2, DIAM achieves optimal performance on the majority of evaluation metrics across all three datasets, consistently attaining the best results in AUC, AUPR, Precision, Recall, and F1-Score. For instance, on the Cdataset, DIAM achieves an AUC of 0.9948 and an AUPR of 0.9954, representing relative improvements of 2.61% and 2.49%, respectively, over the second-best model. Similarly, on the Fdataset, it attains an AUC of 0.9935 and an AUPR of 0.9945, outperforming the strongest baselines by 3.18% in AUC and 3.00% in AUPR. On the more challenging Bdataset, DIAM ranks first across all six metrics, achieving an AUC of 0.9861 and an AUPR of 0.9890, which surpass the second-best model by 5.85% and 6.43%, respectively, demonstrating strong robustness even in complex scenarios.

Notably, on both the Cdataset and Fdataset, although DFDRNN attains the highest ACC, its AUPR and F1-Score are markedly lower. This suggests that its high ACC may stem from correctly classifying a large number of negative samples, while it performs poorly in identifying the critical yet scarce positive samples. In contrast, DIAM's superior AUPR and F1-Score demonstrate its robust capability to effectively identify the positive samples. This superiority is quantitatively evident, with DIAM surpassing the second-best baseline by over 2.49% in AUPR and 4.71% in F1-Score on both datasets.

In summary, the outstanding performance of DIAM underscores the effectiveness of our mechanism-adaptive paradigm. Unlike conventional approaches that depend on random negative sampling, DIAM employs hard negative sampling to enhance the model's discriminative capacity. For embedding fusion, DIAM incorporates Adaptive Residual Gating (ARG) for instance-aware modulation. These targeted designs enable the model to more effectively uncover drugdisease associations, ultimately leading to superior and robust predictive performance across benchmark datasets.

4.3 Ablation Study

To verify the contribution of each component in DIAM, we compare the full model against eight variants with AUC as the primary metric. The variants are defined as follows: V1 (w/o Decoupling) removes the decoupling module; V2 (w/o Local) and V3 (w/o Global) exclude molecular and structural views, respectively; V4 (w/o HNS) uses random sampling; V5 (w/o ARG) replaces ARG with a standard dot product; V6 (w/o PCP) replaces the PCP with a fixed weight; V7 (w/o CL) removes the contrastive learning objective; and V8 (w/o HGT) replaces the HGT encoder with GAT. The results are shown in Figure 2.

Impact of Mechanism Decoupling (V1-V3). V1 suffers the sharpest degradation across all datasets, confirming that decoupled signals are foundational to the model's performance. V2 and V3 further validate that biological signal dominance varies dynamically across datasets.

Efficacy of Sampling and Modulation (V4-V6). The significant drop in V4 confirms that mining hard negatives is essential for sharpening decision boundaries against ambiguous candidates. Moreover, V5 and V6 demonstrate that instanceaware gating effectively enhances the expressiveness of embeddings.

Backbone and Auxiliary Tasks (V7, V8). V7 and V8 confirm the value of cross-view alignment and HGT's superiority in modeling multi-relational dependencies, respectively.

4.4 Parameter sensitivity analysis

We perform a parameter sensitivity analysis on three key hyperparameters of DIAM using grid search. The model is trained for 300 epochs with a consistent learning rate of 5e3 across all datasets. Optimal values are selected based on AUC. The detailed sensitivity analysis for these three parameters is presented in Figure 3.

Impact of embedding dimension (d). The embedding dimension d determines the capacity of the node feature representations. It defines both the output dimension of the HGT

and the input dimension of the final MLP predictor. We evaluate its impact by testing values from the set {32, 64, 128, 256, 512}. The optimal dimension is 128 for Cdataset and 64 for both Bdataset and Fdataset.

Impact of the structural propagation coefficient (α). The coefficient α balances the relative contributions of the drug-side and disease-side information when computing Ctopo. It is evaluated over {0.1, 0.3, 0.5, 0.7, 0.9}, where the optimal values are 0.3, 0.7, and 0.5 for Cdataset, Bdataset, and Fdataset, respectively.

Impact of the negative sampling threshold (θneg). The negative sampling threshold θneg is set to a value that filters the top percentile of unobserved pairs as hard negatives. We assess its impact over the range {50, 60, 70, 80, 90}. We select 90 as the optimal threshold, since further increases drastically shrink the negative pool and heighten the risk of overfitting on a small subset of samples.

4.5 Effectiveness of Adaptive Residual Gating

To validate the superiority of the proposed ARG module, we compare DIAM against two representative fusion strategy: -concat, which directly concatenates drug and disease embeddings, and -attention, which employs a standard attention mechanism to adaptively assign weights.

As shown in Table 3, DIAM consistently achieves the highest AUC across all datasets. The primary limitation of -concat and -attention is that they fail to incorporate biological semantics for instance-specific modulation. In contrast, DIAM leverages the decoupled signals from both the global structural and local molecular views to guide the fusion, ensuring that the final embedding is adaptively rectified by the biological signals for each drug-disease pair.

4.6 Case Study

To assess the model's practical application, we conduct a case study on Alzheimer's Disease (AD). AD is a progressive neurodegenerative disorder with limited therapeutic options [Livingston et al., 2024]. Drug repositioning provides an important avenue for exploring new treatment strategies.

In this study, we rank the candidate drugs from high to low based on their prediction scores and select the top 10 for in-depth analysis. As shown in Table 4, 8 of these top 10 candidates have demonstrated therapeutic potential for AD, according to existing literature evidence.

Tacrolimus ranks first in our predictions. Recent preclinical studies have confirmed it as a strong candidate drug for AD [PMID: 40412360]. As a calcineurin inhibitor (CNI), it targets key AD features like neuroinflammation and glutamatergic disruptions. Preclinical studies have demonstrated its neuroprotective effects: in an acute Aβ1-42 model, tacrolimus prevents cognitive impairment and neurodegeneration. Furthermore, in chronic APP/PS1 mice, it improves social interaction deficits, partially reduces microgliosis, and restores impaired glutamate release and calcium levels.

Progesterone ranks second in our list. Existing literature demonstrates its multifaceted neuroprotective potential, such as upregulating BDNF and activating the PI3K/Akt signaling pathway [PMID: 39663597]. However, as this literature also points out, current clinical research conclusions regarding the direct treatment of AD with Progesterone are inconsistent and have not reached a consensus, suggesting its complex mechanisms of action require further investigation. Overall, the fact that 80% of the top-ranked candidates are supported by recent studies validates DIAM's reliability in prioritizing biologically significant candidates for drug repositioning.

5 Conclusions

In this paper, we present DIAM, a mechanism-adaptive framework that rectifies the mismatch between static computational paradigms and dynamic biological realities in drug repositioning. Specifically, we design a Dual-Stream Biological Mechanism Decoupling module to explicitly quantify distinct signals from global structural propagation and local molecular interaction views. Leveraging these decoupled signals, we implement Hard Negative Sampling to enforce finer-grained decision boundaries, and an Adaptive Residual Gating module to perform instance-aware modulation. Extensive experiments on three benchmark datasets demonstrate DIAM's consistent superiority over state-of-the-art methods, while case studies on Alzheimer's Disease validate its reliability in prioritizing biologically significant candidates. In future work, we will extend the depth of this paradigm by integrating multi-omics modalities to construct a comprehensive multi-view biological landscape. Code and datasets are available at https://github.com/Eyre68/DIAM.

Acknowledgements

This work was supported in part by the National Major Science and Technology Projects of China (2024ZD0523903), and the National Natural Science Foundation of China (72342014).