www.nature.

com/scientificreports

OPEN Multi-layer graph attention neural

networks for accurate drug-target
interaction mapping
Qianwen Lu2, Zhiheng Zhou3,4 & Qi Wang1
In the crucial process of drug discovery and repurposing, precise prediction of drug-target interactions
(DTIs) is paramount. This study introduces a novel DTI prediction approach—Multi-Layer Graph
Attention Neural Network (MLGANN), through a groundbreaking computational framework that
effectively harnesses multi-source information to enhance prediction accuracy. MLGANN not only
strides forward in constructing a multi-layer DTI network by capturing both direct interactions between
drugs and targets as well as their multi-level information but also amalgamates Graph Convolutional
Networks (GCN) with a self-attention mechanism to comprehensively integrate diverse data sources.
This method exhibited significant performance surpassing existing approaches in comparative
experiments, underscoring its immense potential in elevating the efficiency and accuracy of DTI
predictions. More importantly, this study accentuates the significance of considering multi-source data
information and network heterogeneity in the drug discovery process, offering new perspectives and
tools for future pharmaceutical research.

Drug-Target Interaction (DTI) prediction is a computational task aimed at identifying potential interactions
between chemical compounds (drugs) and biological molecules (targets, typically proteins). Prediction of
DTI is a crucial component in modern drug research and development. Accurate DTI prediction can provide
valuable clues for drug design and expedite drug repurposing, with approximately 75% of drugs being eligible
for repurposing1. Traditional in vitro experimental tests can validate DTI, but this approach is both expensive
and time-consuming. In recent years, computer-assisted methods for DTI prediction have garnered widespread
attention, including matrix factorization2, kernel-based methods3, graph embedding techniques4, and more.
The introduction of these methods narrows the search space, reduces the workload of in vitro experiments, and
accelerates the drug development process.
Although the importance of DTI prediction is widely recognized, most current research in this area primarily
relies on the chemical structure of drugs and the protein sequences of targets for DTI prediction5–7. This
approach neglects other crucial multi-source information about drugs and targets, such as the physicochemical
properties of drugs and the relationships between targets and diseases8,9, which are equally important for DTI
prediction. With deep learning demonstrating excellent performance across various domains, some researchers
have also applied it to DTI prediction and have achieved promising results10–12. However, methods based on
neural networks or machine learning, when integrating information about drugs and targets, can only utilize the
information from individual drugs or targets3,10. The result of this approach is a loss of the unique information
regarding the similarity between drugs and targets. In contrast, network-based methods can model the
similarity between drugs and targets and various features as a DTI network. This network is used to represent
multidimensional information about drugs and targets and is applied to DTI prediction.
Due to the superior performance of graph neural network (GNN) methods in network data analysis, some
researchers have proposed GNN-based models for DTI prediction on heterogeneous DTI networks13–15.
However, most of the work separately considers drug and target networks, neglecting the information regarding
their interactions. Additionally, some models treat heterogeneous DTI networks as homogeneous graphs for
neighbor aggregation13, which results in the inability to capture the heterogeneous information within the DTI
network, leading to suboptimal network node representations.
To address the above issues, in this paper, we propose a multi-layer graph attention neural network (MLGANN)
to capture multi-source information on drugs and targets for DTI prediction. Specifically, we leverage collected
multi-source information about drugs and targets to construct a multi-layer DTI network, which is then utilized

1College of Science, China Agricultural University, Beijing 100083, China. 2SDU-ANU Joint Science College,
Shandong University, Weihai 264209, Shandong, China. 3Academy of Mathematics and Systems Science, Chinese
Academy of Sciences, Beijing 100190, China. 4School of Mathematical Sciences, University of Chinese Academy of
Sciences, Beijing 100190, China. email: [email protected]

Scientific Reports | (2024) 14:26119 | https://doi.org/10.1038/s41598-024-75742-1 1

www.nature.com/scientificreports/

for DTI prediction. The multi-layer DTI network not only encompasses interaction information between
drugs and targets but also captures multi-level information about drugs and targets. Furthermore, we design
a graph neural network-based model that employs Graph Convolutional Networks (GCN) to capture multi-
source information about drugs and targets within the multi-layer DTI network. Subsequently, a self-attention
mechanism layer is employed to integrate the multi-source information and obtain representations for drugs
and targets. Finally, through comparisons with multiple models on DTI datasets, the results demonstrate that
the proposed approach outperforms state-of-the-art methods.
In sum, the major contributions of our proposed method can be summarized as follows:

• We construct a multi-layer DTI network for DTI prediction, which not only encapsulates multi-level in-
formation of drugs and targets but also captures the interactions between these levels. This comprehensive
integration enhances the accuracy of DTI predictions.
• We design a multi-layer graph attention neural network (MLGANN) that effectively captures multi-source
drug and target information within the multi-layer DTI network, as well as the interactions between them,
further improving DTI prediction outcomes.
• Through comparative analysis with multiple models on the DTI dataset, our experimental results demon-
strate the effectiveness and superiority of the proposed method in addressing the DTI prediction problem.

Related work
Drug-target interaction (DTI) prediction is a critical task in drug discovery and development, where the goal
is to identify potential interactions between drugs and target proteins. Over the years, various computational
approaches have been developed to address this problem, ranging from similarity-based methods to more
advanced graph-based approaches.
Traditional methods for DTI prediction often rely on similarity measures, where similar drugs are assumed
to interact with similar targets. These methods typically compute drug-drug and target-target similarity
matrices, which are then used in conjunction with machine learning models to predict interactions16. However,
these approaches have limitations, such as their inability to integrate multi-source information about drugs and
targets. For example, methods like DTINet and deepDTnet fall into this category but struggle with capturing the
full complexity of drug-target relationships.
Recent advancements have seen the application of knowledge graphs to DTI prediction17, where biological
knowledge is structured as a graph and used to infer new interactions. Methods such as ComplEx18 and KGE_
NFM19 have been developed to leverage knowledge graph embeddings for DTI prediction. While these methods
provide a way to incorporate diverse biological information, they often require significant domain-specific
knowledge to construct the knowledge graph, which can limit their applicability.
Graph neural networks (GNNs) have gained popularity for their ability to model complex relationships in
network data. Several GNN-based methods have been proposed for DTI prediction, including IMCHGAN15,
SGCL-DTI14, and MHGNN20. These methods treat drug-target pairs as nodes in a heterogeneous network and
use GNNs to learn embeddings that capture the intricate relationships between drugs and targets. For instance,
SGCL-DTI employs a supervised graph co-contrastive learning approach, while MHGNN uses heterogeneous
graph attention networks to enhance prediction accuracy.
In addition to the methods mentioned above, several recent works have further advanced the field.
SGCLDGA21 and NGCN22 are notable examples. SGCLDGA enhances DTI prediction by incorporating graph
neural networks and contrastive learning that focus on graph structures. NGCN, on the other hand, leverages
integrated heterogeneous network to extract relevant biological properties and association information while
maintaining the topology information.
Furthermore, methods like TripletMultiDTI23, DeepTraSynergy24, and HGTDR25 represent the cutting edge
of DTI prediction. TripletMultiDTI uses a triplet loss to improve the distinction between interacting and non-
interacting pairs. DeepTraSynergy is designed to predict drug synergy, which is closely related to DTI, using
deep learning techniques. HGTDR employs heterogeneous graph transformers to advance drug repurposing
efforts by effectively modeling the multi-relational nature of drug-target interactions.
While existing methods have made significant strides in DTI prediction, they often face challenges in
integrating multi-source information or dealing with the heterogeneity of drug-target interactions. To address
these challenges, our proposed Multi-Layer Graph Attention Neural Network (MLGANN) not only captures
multi-level interaction information between drugs and targets but also integrates diverse data sources through a
novel multi-layer attention mechanism. This approach allows for a more comprehensive and accurate prediction
of DTIs, as demonstrated in our experimental results.

Methods
Multi-layer DTI network
n
Give a drug set D = {di}i=1 d
and a target set T = {tj }nj=1
t
, the similarity between drugs (targets) can be
md
assessed in various perspectives, which are represented by a set of matrices {AD,k }k=1 l=1 ),where
, ({AT,l }m t
D,k
A ∈R nd ×nd T,l
, (A ∈ R nt ×nt
) and md(mt) is the number of similarity types for drugs (targets). Let the binary
matrix AY ∈ {0, 1}nd×nt indicate the interactions between drugs in D and targets in T, where AYij = 1 denotes
that di and tj interact with each other, and AYij = 0 otherwise. A multi-layer DTI network GM = (V M , E M ) as
md
shown in Fig. 1 for D and T consists of {AD,k }k=1 l=1 and A , with its adjacency matrices represented
, {AT,l }m t Y

as follows:

Scientific Reports | (2024) 14:26119 | https://doi.org/10.1038/s41598-024-75742-1 2

www.nature.com/scientificreports/

Fig. 1. The figure illustrates a multiplex layer drug-target interaction (DTI) network, which integrates
multi-level information from drugs (D) and targets (T) across several layers. On the left, different layers of
drug associations are shown (labeled as AD,1, AD,2, AD,3), representing various relationships among drugs
D1, D2, D3 and D4. On the right, target associations are depicted in similar layers (AT,1 and AT,2) with targets
T1, T2 and T3. The central part of the diagram displays the interaction between drugs and targets, where a
multi-layer network structure is used to capture the complex interplay between different layers of information.
This multi-layered approach enables for more comprehensive DTI prediction by considering both intra-layer
and inter-layer interactions.

 
AD,1 I I · · · AY AY
 I AD,2 I ··· A Y
AY 
 
 
 I I AD,3 ··· A Y
AY 
AM =  .. ... ... .
. ...  .(1)
 . ··· . 
 Y Y T,mt −1 
 A A AY ··· A I 
AY AY AY ··· I AT,mt

where ||V M || = N = (nd × md + nt × mt).

Multi-layer attention graph neural network

We propose a model called the Multi-Layer Graph Attention Neural Network (MLGANN) for DTI prediction.
In the multi-layer network of DTI, apart from the interaction between drugs and targets, there is also interaction
information among various properties within drugs and targets themselves. Therefore, we utilize the designed
MLGANN to capture both the interaction information between drugs and targets and the multi-source
information within drugs and targets.

Multi-layer neighbor aggregation

Let X ∈ RN ×f represent the initial features of nodes in the multi-layer DTI network, where f denotes the
dimension of the embedding space. We apply graph neural networks to learn embeddings for drugs and targets
on the multi-layer DTI network. Specifically, in our model, we employ Graph Convolutional Networks (GCN),
as they are both simple and effective. These embeddings can be refined by applying P layers of GCN across the
entire multi-layer DTI network:
 1 1

X (p) = σ D̂− 2 ÂD̂− 2 X (p−1)W (p) , p = 1, 2, . . . , P,(2)

where X (0) = X, Â = AM + I M , and AM is the adjacency matrix of the multi-layer DTI network, I M is an
identity matrix of the same size as A , D̂ is a diagonal matrix with D̂ij = j=1 Âij , W (p) ∈ Rf ×f is a trainable
M

weight matrix, σ is a nonlinear activation function ReLU.

Scientific Reports | (2024) 14:26119 | https://doi.org/10.1038/s41598-024-75742-1 3

www.nature.com/scientificreports/

For a node v ∈ GM (representing either a drug or a target), Eq. (2) updates the embedding of that node as
follows:
 
 1
(p)
xv = σ W (p)
x (p−1)
, u ∈ {v ∪ Nv ∪ Cv }(3)
u
αvu u

where αvu is the normalized weight, Nv is the set of neighbors of node v in layer of GM , and Cv is the set of nodes
that correspond to the same drug/target as node v. Therefore, MLGANN not only aggregates the neighbors of
node v in layer of GM (similar to what GCN does) but also embeds nodes corresponding to the same drug/target
in different layers of GM . This allows information to be transmitted across different layers of GM . By leveraging
information from different layers of GM , MLGANN can learn better representations for each node, especially
for nodes with limited interactions in a particular layer of GM . This is the main distinction between MLGANN
and existing other network embedding methods.

Multi-layer attention pooling

We concatenate all representations learned by the P-layer GCN to obtain the final node embedding:
 
D,k(0) D,k(1) D,k(P )
ziD,k = zi , zi , · · · , zi
  (4)
T,l(0) T,l(1) T,l(P )
zjT,l = zj , zj , · · · , zj ,

D,k(p)
where ziD,k denotes final embedding of ith drug in k layer of GM and zi denotes GM ’s kth layer embedding
T,l(p)
of ith drug in GCN pth layer, zjT,l represents final embedding of jth target in lth layer of GM , zj denotes GM

’s lth layer embedding of jth drug in GCN pth layer.

To obtain the final representations of drugs and targets, we have designed a self-attention mechanism to aggregate
the representation vectors of drugs and targets across different layers for DTI prediction in the GM graph. The
computer process is as follows:
   
eD,k
i = q D · LeakyReLU W D ziD,k , eT,l T
j = q · LeakyReLU W zj
T T,l

eD,k
md
 eT,l
mt
 (5)
αik = m i D,k′ , ziD = αik ziD,k , βjl = m i T,l′ , zjT = βjl zjT,l ,
d t
k ′ =1 ei k=1 l′ =1 ej l=1

f′ f′ ′ ′ ′ ′
where ziD ∈ R and ziD ∈ R are the final′ representations
′
of drugs and targets, W D ∈ Rf ×f and W T ∈ Rf ×f
are trainable parameter matrices, q D ∈ Rf and q T ∈ Rf are trainable vectors.

DTI prediction
Let GY be the DTI network derived from the adjacency matrix AY . For an edge ditj in GY , where ziD and zjT
are final representation vectors of drug di and target tj , respectively. we sample a non-existing edge dutv in GY ,
where zuD and zvT are final representation vectors of drug du and target tv , respectively. We consider DTP ditj as
a positive sample and dutv as a negative sample. Therefore, we design the loss function based on cross-entropy
as follows:
     
L = − log σ < ziD , zjT > − log σ − < zuD , zvT > (6)
where σ is a nonlinear activation function Sigmoid, < ·, · > is the inner product in Euclidean space.

Experiments
Experimental setup
Datasets We collected data on DTIs from DrugBank (Version5.1.8), as well as relevant data on drug chemical
structural similarity, drug-side effect relationships, drug-disease relationships, target interaction relationships,
and target-disease relationships. All of this data is available on DrugBank. By taking the intersection of these
datasets, we obtained a total of 201 drugs, 252 targets, and 907 interactions between them. We used 607 of
these interactions for training and 300 interactions for testing. The statistical information for these datasets
is summarized in Table 1, where D denotes drug, T denotes target, D-D-1 denotes drug chemical structural
similarity, D-D-2 denotes similarity based on drug-side effect relationships, D-D-3 denotes similarity based on
drug-disease relationships, T-T-1 denotes target interaction relationships, T-T-2 denotes similarity based on
target-disease relationships.
Baselines The baselines can be categorized into three categories: (1) Similarity-based methods, mainly
including DTINet12, deepDTnet26 and NEDTP4; (2) KG-based models which use TransE27, ComplEx18 and KGE_
NFM19; (3) Graph neural network (GNN)-based methods, such as DTI-MGNN8, SGCL-DTI14, MHGNN20 and
IMCHGAN15.

Scientific Reports | (2024) 14:26119 | https://doi.org/10.1038/s41598-024-75742-1 4

www.nature.com/scientificreports/

D T DTI D-D-1 D-D-2 D-D-3 T-T-1 T-T-2

201 252 907 5142 1263 1598 12123 23545

Table 1. Statistics for the drug target interaction dataset.

Models AUC Precision Recall F1 score AUPR

DTNet 0.6811±0.0050 0.6503±0.0042 0.5837±0.0063 0.6287±0.0055 0.6924±0.0037
Similarity-based deepDTNet 0.6704±0.0078 0.6619±0.0099 0.5736±0.0079 0.5878±0.0108 0.6544±0.0049
NEDTP 0.8027± 0.0041 0.7662± 0.0072 0.7018±0.0037 0.7240±0.0061 0.8191±0.0031
TransE 0.6245±0.0045 0.5535±0.0044 0.4980±0.0068 0.5122±0.0073 0.6674±0.0055
KG-based KGE-NFM 0.7584±0.0033 0.6973±0.0030 0.6901±0.0046 0.6784±0.0065 0.7688±0.0041
ComplEx 0.7061±0.0083 0.6229±0.0057 0.6417±0.0032 0.6527±0.0044 0.7477±0.0067
DTI-MGNN 0.7408±0.0069 0.6894±0.0057 0.6427±0.0073 0.6339±0.0045 0.7540±0.0058
SGCL-DTI 0.6567±0.0167 0.6639±0.0112 0.5422±0.0089 0.5847±0.0095 0.7111±0.0098
IMCHGAN 0.6864±0.0071 0.6675±0.0074 0.6149±0.0056 0.6482±0.0044 0.7001±0.0063
GNN-based
MHGNN 0.7980±0.0056 0.7531±0.0035 0.7618± 0.0066 0.7252± 0.0052 0.8263± 0.0036
MLGANN 0.8699±0.0023 0.8122±0.0037 0.7989±0.0041 0.8018±0.0035 0.8552±0.0026
P-value ***P<0.001 ***P<0.001 **P<0.01 ***P<0.001 **P<0.01

Table 2. Comparison of our model with baseline on dataset DTI. The highest and the second highest results
over each measurement are in bold and italics, respectively. Best model indicate significant improvement at P-
value level when compared to the second highest performance, using one-sided paired t-test. **, *** represent
significant levels of 1% and 0.1% respectively.

Parameter Configuration The hidden node embedding dimension of our models is set to 256 and the output
embedding dimension is set to 64. We stack two convolutional layers. For the model optimization, our model
will be trained 200 epochs with a learning rate of 0.01 and a batch size of 256. For HAN and MAGNN, the
dimension of the attention vector in the feature fusion seted to 128. Other parameters are set to the same as these
in MLGANN. For fair comparison, all these baselines adopt the same training set and test set as MLGANN. All
models use Adam optimizer. In the experiments, all models are conducted with DGL platform and PyTorch
framework.
Evaluation Metrics We conduct the link prediction task to validate the performance of all models. For
training and testing stage, positive sample consists of a DTI, and the corresponding negative samples consist of
an unknown DTI. We use the precision score, the recall score, the F1 score, the area under the curve (AUC) and
the area under the precision-recall curve (AUPR) as evaluation metrics.

Overall performance
The experimental results for DTI prediction are shown in Table 2, where all results are the averaged across 5
repeated runs. As we can see, MLGANN significantly outperforms all baseline methods. It shows improvements
of 0.0672, 0.0460, 0.0371, and 0.0766 over metrics of AUC, the precision score, the recall score, and the F1 score,
demonstrating the superiority of our proposed model. Compared to similarity-based methods, MLGANN can
harness multi-source data for drugs and targets, capturing various feature information of drugs and targets in the
multi-layer DTI network for prediction. The end-to-end modeling approach also proves advantageous for DTI
prediction. However, similarity-based methods rely on strong assumptions in modeling, leading to suboptimal
final prediction results. Furthermore, two-stage models separate representation learning from the prediction

Models AUC Precision Recall F1 score AUPR

w/o D-D-1 0.7832±0.0037 0.7478±0.0062 0.7118±0.0043 0.7281±0.0047 0.8069±0.0034
w/o D-D-2 0.8393±0.0082 0.7973±0.0067 0.7525±0.0056 0.7490±0.0077 0.8413±0.0050
w/o D-D-3 0.8078±0.0057 0.7775±0.0033 0.7402±0.0048 0.7415±0.0072 0.8098±0.0066
w/o T-T-1 0.8114±0.0058 0.7726±0.0044 0.7567±0.0039 0.7385±0.0051 0.8284±0.0053
w/o T-T-2 0.8235±0.0077 0.8004±0.0031 0.7531±0.0062 0.7306±0.0068 0.8316±0.0047
Ours 0.8699±0.0023 0.8122±0.0037 0.7989±0.0041 0.8018±0.0035 0.8552±0.0039
P-value **P<0.01 **P<0.01 **P<0.01 ***P<0.001 **P<0.01

Table 3. Results of ablation study (where w/o = without). Best model indicate significant improvement at
P-value level when compared to the second highest performance, using one-sided paired t-test. The best results
are marked in bold (the higher the better). **, *** represent significant levels of 1% and 0.1% respectively.

Scientific Reports | (2024) 14:26119 | https://doi.org/10.1038/s41598-024-75742-1 5

www.nature.com/scientificreports/

Fig. 2. Impacts of different hidden feature dimensions.

task, resulting in learned representation vectors that contradict DTI prediction outcomes. Compared to KG-
based methods, MLGANN can learn better representations of drugs and targets.
Models based on GNNs approach DTI prediction by modeling the multi-source data of drugs and targets as
a heterogeneous network and using metapaths and GCN for the task. SGCL-DTI constructs two heterogeneous
networks, a topological network and a semantic network, with DTPs as nodes. The edges between nodes in the
network are determined by the similarity of DTPs, as DTPs that share common drugs or targets have higher
similarity than those without shared drugs or targets. Therefore, the topological and semantic networks exhibit
significant overlap. Additionally, SGCL-DTI has access to labels of DTPs during training. DTI-MGNN constructs
a DTP graph with DTPs as nodes, separating the representation learning of drugs and targets from the DTI
prediction task. By comparing MLGANN with MHGNN, we can conclude that directly applying heterogeneous
graph embedding methods to DTI prediction tasks does not yield optimal DTI prediction results.

Ablation study
To clearly demonstrate the effectiveness of the multi-layer graph we constructed, we conducted ablation
experiments by progressively removing each layer in the multi-layer DTI network. The results are presented in
Table 3. The experimental findings indicate that the results obtained with the multi-layer DTI network, utilizing
3 layers for drugs and 2 layers for targets, outperform the results obtained by removing any single layer within
the multi-layer DTI network. Furthermore, the removal of D-D-1, D-D-3, and T-T-1 has a significant impact on
the experimental results, suggesting that information related to drug chemical structural similarity, similarity
based on drug-disease relationship, and target interaction relationship are more important than similarity based
on drug-disease relationship and target-disease relationships.

Scientific Reports | (2024) 14:26119 | https://doi.org/10.1038/s41598-024-75742-1 6

www.nature.com/scientificreports/

Fig. 3. The experiment result of robust analysis.

Parameter analysis
To analyze the performance of DTI prediction, we altered the dimension of the hidden layer from 32 to 512. The
results for different metrics are shown in Fig. 2. From Fig. 2, we observe that the values of AUC, precision, and
F1 score initially increase and then decrease with the increment of feature dimensions, with AUC and precision
peaking at a dimension of 256, and F1 score peaking at a dimension of 128. The recall score reaches its peak
result at the dimension of 256. This indicates that too small a dimension of the hidden layer cannot capture the
characteristic information of the data, while too large a dimension causes the model to be prone to overfitting,
affecting the model’s generalization performance.

Robust analysis
To analyze the robustness of model, we randomly remove 10–50% train data. The results for different metrics are
shown in Fig. 3. From Fig. 3, we observe that the values of AUC, F1 score, and AUPR decrease with the increment
of remove ratio. However, among the three models, MLGANN not only maintains the highest performance in
all metrics, but also has the smallest decline. This shows that MLGANN has good robustness in DTI prediction
tasks.

Conclusion
In this study, we have successfully developed the Multi-Layer Graph Attention Neural Network (MLGANN)
aimed at elucidating the intricate structures and nuanced semantics inherent in multi-layer Drug-Target
Interaction (DTI) networks. MLGANN leverages GCN to capture multi-source information about drugs and
targets, and further employs a self-attention mechanism to obtain representations for drugs and targets. We
conducted model training and validation using data obtained from DrugBank, and the results demonstrate the
effectiveness of MLGANN for DTI prediction. Looking ahead, we are poised to expand our research horizons
by incorporating additional datasets, thereby enriching our training and validation frameworks. This strategic
augmentation is anticipated to further refine the predictive capabilities of our model, making it a more potent
tool in the realm of drug discovery and development.

Data availability
The data on drug-target interactions underpinning this study were sourced from DrugBank (Version 5.1.8). For
more details, please visit https://go.drugbank.com/.

Received: 2 February 2024; Accepted: 8 October 2024

References
1. Nosengo, N. et al. Can you teach old drugs new tricks? Nature 534, 314–316 (2016).
2. Ezzat, A., Zhao, P., Wu, M., Li, X.-L. & Kwoh, C.-K. Drug-target interaction prediction with graph regularized matrix factorization.
IEEE/ACM Trans. Comput. Biol. Bioinform. 14, 646–656 (2016).
3. Ding, Y., Tang, J. & Guo, F. Identification of drug-target interactions via dual Laplacian regularized least squares with multiple
kernel fusion. Knowl.-Based Syst. 204, 106254 (2020).
4. An, Q. & Yu, L. A heterogeneous network embedding framework for predicting similarity-based drug-target interactions. Brief.
Bioinform. 22, bbab275 (2021).
5. Gao, K. Y. et al. Interpretable drug target prediction using deep neural representation. In IJCAI Vol. 2018, 3371–3377 (2018).
6. Huang, K., Xiao, C., Glass, L. M. & Sun, J. Moltrans: Molecular interaction transformer for drug-target interaction prediction.
Bioinformatics 37, 830–836 (2021).
7. Chu, Y. et al. DTI-MLCD: Predicting drug-target interactions using multi-label learning with community detection method. Brief.
Bioinform. 22, bbaa205 (2021).
8. Liu, B., Pliakos, K., Vens, C. & Tsoumakas, G. Drug-target interaction prediction via an ensemble of weighted nearest neighbors
with interaction recovery. Appl. Intell. 52, 1–23 (2022).
9. Pliakos, K., Vens, C. & Tsoumakas, G. Predicting drug-target interactions with multi-label classification and label partitioning.
IEEE/ACM Trans. Comput. Biol. Bioinform. 18, 1596–1607 (2019).
10. Olayan, R. S., Ashoor, H. & Bajic, V. B. DDR: Efficient computational method to predict drug-target interactions using graph
mining and machine learning approaches. Bioinformatics 34, 1164–1173 (2018).

Scientific Reports | (2024) 14:26119 | https://doi.org/10.1038/s41598-024-75742-1 7

www.nature.com/scientificreports/

11. Zhang, Z.-C. et al. A graph regularized generalized matrix factorization model for predicting links in biomedical bipartite
networks. Bioinformatics 36, 3474–3481 (2020).
12. Luo, Y. et al. A network integration approach for drug-target interaction prediction and computational drug repositioning from
heterogeneous information. Nat. Commun. 8, 573 (2017).
13. Wan, F., Hong, L., Xiao, A., Jiang, T. & Zeng, J. NeoDTI: Neural integration of neighbor information from a heterogeneous network
for discovering new drug-target interactions. Bioinformatics 35, 104–111 (2019).
14. Li, Y., Qiao, G., Gao, X. & Wang, G. Supervised graph co-contrastive learning for drug-target interaction prediction. Bioinformatics
38, 2847–2854 (2022).
15. Li, J., Wang, J., Lv, H., Zhang, Z. & Wang, Z. IMCHGAN: Inductive matrix completion with heterogeneous graph attention
networks for drug-target interactions prediction. IEEE/ACM Trans. Comput. Biol. Bioinform. 19, 655–665 (2021).
16. Adasme, M. F., Parisi, D., Sveshnikova, A. & Schroeder, M. Structure-based drug repositioning: Potential and limits. In Seminars
in cancer biology Vol. 68 192–198 (Elsevier, 2021).
17. Mohamed, S. K., Nounu, A. & Nováček, V. Drug target discovery using knowledge graph embeddings. In Proceedings of the 34th
ACM/SIGAPP Symposium on Applied Computing, 11–18 (2019).
18. Trouillon, T., Welbl, J., Riedel, S., Gaussier, É. & Bouchard, G. Complex embeddings for simple link prediction. In International
Conference on Machine Learning, 2071–2080 (PMLR, 2016).
19. Ye, Q. et al. A unified drug-target interaction prediction framework based on knowledge graph and recommendation system. Nat.
Commun. 12, 6775 (2021).
20. Li, M., Cai, X., Xu, S. & Ji, H. Metapath-aggregated heterogeneous graph neural network for drug-target interaction prediction.
Brief. Bioinform. 24, bbac578 (2023).
21. Fan, Y. et al. SGCLDGA: Unveiling drug-gene associations through simple graph contrastive learning. Brief. Bioinform. 25,
bbae231 (2024).
22. Cao, J. et al. NGCN: Drug-target interaction prediction by integrating information and feature learning from heterogeneous
network. J. Cell Mol. Med. 28, e18224 (2024).
23. Dehghan, A., Razzaghi, P., Abbasi, K. & Gharaghani, S. TripletMultiDTI: Multimodal representation learning in drug-target
interaction prediction with triplet loss function. Expert Syst. Appl. 232, 120754 (2023).
24. Rafiei, F. et al. DeepTraSynergy: Drug combinations using multimodal deep learning with transformers. Bioinformatics 39, btad438
(2023).
25. Gharizadeh, A., Abbasi, K., Ghareyazi, A., Mofrad, M. R. & Rabiee, H. R. HGTDR: Advancing drug repurposing with heterogeneous
graph transformers. Preprint at arXiv:2405.08031 (2024).
26. Zeng, X. et al. Target identification among known drugs by deep learning from heterogeneous networks. Chem. Sci. 11, 1775–1797
(2020).
27. Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J. & Yakhnenko, O. Translating embeddings for modeling multi-relational data.
Adv. Neural Inf. Processi. Syst. 26 (2013).

Author contributions
Q.L. conceived the project, developed the prediction method, designed the experiments, implemented the ex-
periments, and wrote the paper. Z.Z. conceived the project, designed the experiments, analyzed the result, and
wrote the paper. Q.W. conceived the project, analyzed the result, and revised the paper. All authors have re-
viewed and approved the manuscript.

Declarations

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to Q.W.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives
4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in
any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide
a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have
permission under this licence to share adapted material derived from this article or parts of it. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence
and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to
obtain permission directly from the copyright holder. To view a copy of this licence, visit ​h​t​t​p​:​//​ ​c​r​e​a​t​i​v​e​co
​ ​m​m​o​
n​s​.​o​r​g/​ ​l​i​c​e​n​s​e​s​/b
​ ​y​-​n​c​-​n​d​/​4​.​0​/​​​​.​ ​​

© The Author(s) 2024

Scientific Reports | (2024) 14:26119 | https://doi.org/10.1038/s41598-024-75742-1 8