Drug–target interaction prediction by integrating heterogeneous information with mutual attention network

Abstract

Background

Identification of drug–target interactions is an indispensable part of drug discovery. While conventional shallow machine learning and recent deep learning methods based on chemogenomic properties of drugs and target proteins have pushed this prediction performance improvement to a new level, these methods are still difficult to adapt to novel structures. Alternatively, large-scale biological and pharmacological data provide new ways to accelerate drug–target interaction prediction.

Methods

Here, we propose DrugMAN, a deep learning model for predicting drug–target interaction by integrating multiplex heterogeneous functional networks with a mutual attention network (MAN). DrugMAN uses a graph attention network-based integration algorithm to learn network-specific low-dimensional features for drugs and target proteins by integrating four drug networks and seven gene/protein networks collected by a certain screening conditions, respectively. DrugMAN then captures interaction information between drug and target representations by a mutual attention network to improve drug–target prediction.

Results

DrugMAN achieved the best performance compared with cheminformation-based methods SVM, RF, DeepPurpose and network-based deep learing methods DTINet and NeoDT in four different scenarios, especially in real-world scenarios. Compared with SVM, RF, deepurpose, DTINet, and NeoDT, DrugMAN showed the smallest decrease in AUROC, AUPRC, and F1-Score from warm-start to Both-cold scenarios. This result is attributed to DrugMAN’s learning from heterogeneous data and indicates that DrugMAN has a good generalization ability. Taking together, DrugMAN spotlights heterogeneous information to mine drug–target interactions and can be a powerful tool for drug discovery and drug repurposing.

Introduction

Elucidating mechanistic actions of drugs is one of the critical tasks in drug discovery, necessary to identify on-target drugs and new therapeutic targets, avoid unwanted off-target effects, and improve the success rate in clinical trials [1]. Identification of interactions between drugs and their biological targets is one of the core step of exploring drug mechanisms of action. In these years, both experimental and computational approaches are frequently employed to study drug–target interactions. For instance, direct biochemical assays label the protein or small molecule of interest and directly detect the binding affinity [2]. Molecular dynamics (MD) and docking simulations model potential ligand-target binding configurations with low-energy states [3, 4]. Machine learning (ML) and artificial intelligence (AI) models learn molecular representation from chemical structures and capture the complex nonlinear relationships between drugs and targets [5,6,7,8,9]. Till now, there have been many specialist databases providing data on drug–target interactions, such as DrugBank [10], BindingDB [11], ChEMBL [12], and Comparative Toxicogenomics Database (CTD) [13]. These approaches or technologies highlight chemogenomic information to formalize binding interactions between chemicals and targets and neglect other biological information for drugs and protein targets.

Different from chemogenomic-based models, network-based models encode drug and target representations by integrating heterogeneous information from multiplex functional interaction networks, such as inducible gene expression, drug side effects, related diseases, and genetic associations [14]. Intuitively, network-based methods that integrate more information for both drugs and targets could be adept at mining drug–target interactions. In recent years, various network-based methods have been developed based on heterogeneous biological networks to mine drug–target interactions and achieved promising results. For example, DTINet combines Random walk with restart (RWR) and diffusion component analysis (DCA) to learn low-dimensional drug and target representations from heterogeneous networks and predict drug–target interactions using inductive matrix completion [15]. NeoDTI integrates different networks and automatically learns topology-preserving representations of drugs and targets to facilitate drug–target prediction [16].

Despite these promising effects, two challenges remain for network-based methods. (i) An excellent graph embedding method to learn drug and target features. Biological interaction networks from real-world biomedical high-through data inevitably have varied false positives and -negatives while preserving meaningful functional links. A solid graph embedding method should be scalable in both the size and quantity of input networks and learn low-dimensional node features that can reflect the functional and topological properties of all heterogeneous networks. (ii) An embedded module to connect the interaction information between drug and target representations. The network-specific features characterize drug and target proteins in the existing biological big data resources. It should be emphasized to learn the interaction patterns between drugs and targets bridged by these underlying functional media. A simple concatenation of drug and target features is quite inefficient. This problem is similar to the natural language processing (NLP) problem of integrating multiple word embedding into a sentence representation, which has been broken through by the attention mechanism [17].

To address these limitations, we here present a novel deep learning model, termed DrugMAN, which extracts accurate network-specific features of drugs and target proteins by a scalable graph attention network-based integration algorithm and captures interaction patterns between drugs and targets by a mutual attention network to improve drug–target prediction. We have evaluated the performance of DrugMAN and other five baseline models in real-world applications and found that DrugMAN outperforms both chemoinformatics (SVM, RF and DeepPurpose) and network-based methods (DTINet and NeoDTI) for predicting drug–target interactions under different distributions of test and training datasets. DrugMAN shows a good performance in learning and mining the potential drug–target interaction patterns from heterogeneous information and improves the drug–target interaction prediction. DrugMAN promises to be a powerful tool for drug discovery and drug repositioning.

Methods

DrugMAN architecture

Extract drug and target representations from heterogeneous networks

We adopt BIONIC (Biological Network Integration using Convolutions), a scalable deep learning framework for network integration to learn the accurate and comprehensive representations of drugs and protein targets from different types of drug and target networks, respectively [18]. Each network is represented by its adjacency matrix $A$ where $A_{ij} = A_{ij}> 0$ if node i and node j share an edge and $A_{ij} = A_{ij} = 0$ otherwise. BIONIC encodes each input network using three graph attention networks (GAT) to sequentially extract the three-order neighbors of each node. Each GAT encoder has 10 heads with a hidden dimension of 68 per head. The GAT block formulation is then given by:

$$H_{d}^{l + 1} = \sigma \left( {GAT\left( {A,W_{g}^{\left( l \right)} ,b_{g}^{\left( l \right)} ,H_{d}^{\left( l \right)} } \right)} \right)$$

(1)

where $W_{g}^{\left( l \right)}$ and $b_{g}^{\left( l \right)}$ are the layer-specific learnable weight matrix and bias vector of GAT, $A$ is the adjacency matrix for network g, and $H_{d}^{l}$ is the $l$ th hidden node representation with $H_{d}^{0} = H_{P}$, $H_{p}$ is the initial node feature of one-hot encoded so that each node is uniquely identified. $\sigma$ is a nonlinear function (here is LeakyReLU).

The final network-specific features learned by the GAT block can retain both local and global features of the network. The network-specific node features for each network are combined through a weighted and stochastically masked summation to produce combined node features $H_{combined}$. Then, BIONIC maps $H_{combined}$ to a low-dimensional space F through a learned linear transformation. In F, each row corresponds to a node with learned features of 512 dimensions. BIONIC reconstructs the network $\hat{A} = F \cdot F^{T}$ and minimize the discrepancy between the reconstruction and input networks to obtain a high-quality F. Finally, we use the BIONIC framework to learn drug and target representations $F_{d}$ and $F_{t}$ from four types of drug networks and seven gene/protein networks, respectively.

Mutual attention network

We further use the self-attention framework to capture pairwise interactions between drug and target features. As input to the mutual attention network, the network-specific drug and target representations $F_{d}$ and $F_{t}$ are combined into a new matrix $F_{dt}$:

$$F_{dt} = concat\left( {F_{d} ,F_{t} , {\text{axis }} = { }0} \right)$$

(2)

$F_{dt}$ has a shape of $2 \times 512$, where the first and second rows are the drug and target representation with 512-dimensional features, respectively. $F_{dt}$ is fed into sequential transformer encoder units [17] to effectively learn the interrelated information between drug and target features:

$$F_{dt}^{l + 1} = \sigma \left( {atten\left( {F_{dt}^{l} , W^{l} ,b^{l} } \right)} \right)$$

(3)

where each layer $l$ corresponds to a transformer encoder unit and consists of a self-attention layer and feed-forward neural network layer. W and b are learnable weight matrices and bias vectors in the $l$ th transformer encoder unit. $F_{dt}^{l}$ is the $l$th hidden feature matrix and $F_{dt}^{0}$ is the initial matrix from the network encoder. σ is the activation function ReLU.

In $F_{dt}$, the updated drug and target representations $F_{d}$ and $F_{t}$ are directly concatenated into the joint drug–target pair representation $F_{pair}$ with a dimension of 1024. $F_{pair}$ is then inputted into a classification layer which is a fully connected linear layer connected by a sigmoid output function:

$$F_{pair}^{l + 1} = \sigma \left( {W^{l} F_{pair}^{l} + b^{l} } \right))$$

(4)

$$p = sigmoid\left( {W_{o} F_{pair} + b_{o} } \right){ }$$

(5)

$W^{l}$ and $b^{l}$ are learnable weight matrix and bias vector in linear layers. ${\text{W}}_{o}$ and $b_{o}$ are learnable weight matrix and bias vector in the sigmoid layer. $p$ represents the drug–target interaction probability. The binary classification is optimized by minimizing the cross-entropy function as follows:

$$loss = - \mathop \sum \limits_{i} (y_{i} log\left( {p_{i} } \right) + \left( {1 - y_{i} } \right)log\left( {1 - p_{i} } \right))$$

(6)

where $y_{i}$ is the ground-truth label of the ith drug–target pair, $p_{i}$ is its output probability by the model.

Experimental setting

Construction of drug–target interaction dataset

The success of mechanism-based drug discovery depends on the definition of the drug target. To improve the accuracy and reliability of drug discovery, we selected the known drug–target pairs that have been rigorously validated through experiments or supported by extensive literature as the gold-standard data. Drug–target interaction data are collected from five public sources including, Drugbank [10], map of Molecular Targets of Approved drugs (MTA) [19], CTD [13], ChEMBL [12] and BindingDB [11]. MTA is a manually curated dataset that contains 1578 US FDA-approved drugs and 893 human and pathogen-derived biomolecules, of which 667 human-genome-derived proteins targeted by 1194 drugs for human diseases are used in the present analysis. We first collect all drug–target pairs from the Drugbank and MTA. Drugs in the form of inorganic salts are removed. Then we collate drug–target pairs in CTD, ChEMBL and BindingDB corresponding to drugs in Drugbank and MTA. All these data are combined and further filtered by the kinetic constants Ki, Kd, IC50 and EC50. Finally, we choose thresholds of ≤ 10³ nM to obtain 20,565 drug–target binding data, corresponding to 5135 drugs and 2894 protein targets. All drug compound SMILES and amino acid sequences of targets are obtained from PubChem and UniProt, respectively. The PubChem CID and the gene Entrez ID are used as the unique identifiers for drugs and targets, respectively. When datasets are used for evaluating models in the prediction of drug–target interactions, they are balanced with validated positive interactions and an equal number of negative samples randomly obtained from unseen pairs. The dataset is randomly divided into training, validation and test sets with a 7:1:2 ratio.

Construction of cold start dataset

We set up drug cold-start, target cold-start and both cold-start scenarios to evaluate the performance of the model in the real-world drug–target interaction prediction. For cold-start scenarios, each dataset is also randomly divided into training, validation and test sets with a 7:1:2 ratio similar to warm-start.

Evaluation criteria

We curate drug–target interaction data that meet the rigorous standards from five common sources including DrugBank, map of Molecular Targets of Approved drugs (MTA), CTD, ChEMBL, and BindingDB. To train the model, the dataset is randomly divided into training, validation, and test sets with a 7:1:2 ratio. The training set is used to fit the model, and the test set is used to evaluate the model's performance. It should be noted that in real-world scenarios, drug–target pairs that need to be predicted are often unseen and dissimilar to any pairs in the training data. To evaluate the performance of the model in real-world applications, we set three different scenarios to simulate the real-world prediction for drug–target interactions: (i) scenario of drug cold-start (drug-cold), in which drugs from the test set are absent in the training data, (ii) scenario of target cold-start (target-cold), where targets from the test set are not present in the training data, and (iii) both cold-start (both-cold), where both drugs and targets from the test set are missing in the training data.

The area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), and F1-score are used as the major metrics to evaluate the model classification performance. For each experiment, the best-performing model is the one with the best AUROC on the validation set. We conduct five independent runs with different random seeds for each dataset split. The average value of AUROC, AUPRC, and F1-score of five runs are used as indicators for model evaluation. To evaluate the performance of DrugMAN prediction, we introduced the AUROC, AUPRC and F1-Score evaluation metrics. Since the drug–target interaction prediction is regarded as a binary classification problem in the study, the sample can assign positive and negative label, and the prediction value can assign true and false label. We can compute True Positive Rate (TPR), False Positive Rate (FPR), Precision and Recall as follows:

$$TPR = \frac{TP}{{TP + FN}}$$

(7)

$$FPR = \frac{FP}{{TN + FP}},$$

(8)

$$Precision = \frac{TP}{{TP + FP}}$$

(9)

$$Recall = \frac{FP}{{TP + FN}}$$

(10)

where $TP$ is the number of positive samples correctly predicted, $FP$ is the number of positive samples incorrectly predicted, $TN$ is the number of negative samples correctly predicted, and $FN$ is the number of negative samples incorrectly predicted. The receiver operating characteristic curve (ROC) is obtained by taking TPR and FPR as the vertical and horizontal axes, respectively and the AUROC is the area value under the ROC curve. The PR is obtained by taking Presision and Recall as the vertical and horizontal axes, respectively, and AUPRC is the area under PR curve. The higher score of the AUROC and AUPRC values means a more accurate prediction model. The F1-Score is a weighted average of precision and recall and is calculated as follows:

$$F1 = 2 \times \frac{Precision \times Recall}{{Precision + Recall}}$$

(11)

Precision reflects the discrimination ability of the model to negative samples. The higher the Precision, the stronger the discrimination ability of the model to negative samples. Recall reflects the ability of the model to identify positive samples. The higher the Recall, the stronger the ability of the model to identify positive samples. The F1-Score is a synthesis of both, and a higher F1-Score indicates a more robust model.

Baseline

We compare the performance of DrugMAN with that of the five baseline models including SVM, RF, DeepPurpose [6], DTINet [15] and NeoDTI [16]. Two shallow machine learning methods, SVM and RF, are applied using the drug–target pair representation concatenated by the drug ECFP4 fingerprint and target protein AAC features. DeepPurpose models drug–target interaction using CNN to encode drug molecular graphs and protein sequences. The learned drug and protein representation vectors are combined with a simple concatenation and processed by a binary classification layer [6]. DTINet is a network-based model by combining restarted random walk and diffusion component analysis to learn low-dimensional feature representations of drugs and targets from heterogeneous networks and predicts drug–target interaction using inductive matrix completion [15]. NeoDTI is an end-to-end network-based model to integrates diverse heterogeneous networks and automatically learns topology-preserving representations of drugs and targets to predict drug–target interactions [16].

Result

DrugMAN architecture

DrugMAN contains two main networks. The first network intends to learn accurate and comprehensive representations for drugs and protein targets from heterogeneous drug and gene/protein networks by the network integration algorithm BIONIC (Biological Network Integration using Convolutions), which outperforms the existing state-of-the-art network embedding methods [18]. The core idea of BIONIC is to characterize the neighborhoods of network nodes through sequential graph attention networks (GAT) to learn network-specific integrated features. The second network captures the relevant information in a drug–target pair and learns the predictive score for drug–target interaction. We treat the drug and target in a drug–target pair as words in a sentence. The drug and target representations are processed by the mutual attention network that utilizes a series of transformer encoders to apprehend interaction information between the drug and target. The updated drug and target features are concatenated to form the drug–target pair representation, which is then run through a sequence of fully connected classification layers to obtain the predictive score, indicating the probability of drug–target interaction (Fig. 1).

Fig. 1

DrugMAN framework. DrugMAN contains two main parts. The first part encodes network-specific drug and target features from heterogeneous drug and gene/protein networks through sequential graph attention networks (GAT). The combined drug and target features (F_dt) are fed into the second part to learn the updated F_dt by the five transformer encoders. The updated F_dt captures interaction information between the drug and the target. Then the drug and target features in the updated F_dt are concatenated to drug–target pair representation (F_pair), which is input to the fully connected classification layer to calculate the drug–target binding probability score

Evaluation of DrugMAN and baselines

We first compare the performance of DrugMAN with three chemoinformatic baseline models, support vector machine (SVM), random forest (RF), and DeepPurpose. As shown in Fig. 2, DrugMAN consistently outperforms the three chemoinformatic baselines in terms of AUROC, AUPRC, and F1-score. To discern whether the superiority of DrugMAN over these chemoinformatic methods is due to its additional heterogeneous network information, two classical network-based models DTINet and NeoDTI are introduced with the same network data as DrugMAN. We find that DTINet and NeoDTI only perform better than SVM but underperform both RF and DeepPurpose. The results indicate integrating more information cannot necessarily guarantee better performance in predicting drug–target interactions for network-based models compared to chemoinformatic methods [20]. More importantly, the comparison between network-based methods demonstrates that DrugMAN yields superior performance than the two state-of-the-art models. Specifically, it outperforms DTINet and NeoDTI by 16.6% and 7.1% in AUROC, 12.8% and 8.1% in AUPRC, and 18.3% and 8.3% in F1-score, respectively. The results show that DrugMAN better integrates data from different types of networks and captures relevant information in drug–target pairs.

Fig. 2

Comparison of DrugMAN to state-of-the-art methods. We compare the prediction performance of DrugMAN to that of five baselines, including three chemoinformatics models SVM, RF and DeepPurpose, and two network-based models DTINet and NeoDTI. The rows from top to bottom correspond to four scenarios: warm-start, drug-cold, target-cold and both cold, respectively. The columns from left to right correspond to three metrics: the receiver operating characteristic curve, precision-recall curve and F1 Score. The box plots show the median as the center lines and the mean as green triangles for five random runs. The minima and lower percentile represent the worst and second-worst scores. The maxima and upper percentile indicate the best and second-best scores

We further evaluate the stability of DrugMAN in three different real-world scenarios. As expected, compared to the normal random (warm-start) condition, the performance of all models drops significantly due to less information overlap between training and test data in cold-start scenarios. Even so, DrugMAN still achieves the best performance against other state-of-the-art baselines including chemoinformatic and network-based models in drug-cold (AUROC = 0.910, AUPRC = 0.921 and F1-score = 0.835), target-cold (AUROC = 0.922, AUPRC = 0.931 and F1-score = 0.846) and both-cold (AUROC = 0.850, AUPRC = 0.861 and F1-score = 0.776). To our surprise, DrugMAN showed the smallest decrease in three evaluation metrics (AUROC, AUPRC and F1-Score) compared with other methods under the warm-start and both-cold scenarios, which was attributed to its learning from heterogeneous data. This result demonstrates the notable strengths of DrugMAN in learning from heterogeneous networks. The results confirm the good generalization ability of DrugMAN in the drug–target interaction prediction.

Evaluation of different network embedding methods

The network integration algorithm is the essential part of DrugMAN for extracting drug and target features from heterogeneous networks. Although BIONIC has been proven to perform better than other established network integration methods in various benchmarks as a whole, for a fair comparison in drug–target interaction prediction task, we compare BIONIC with two classical network integration approaches, deepNF a deep learning multi-modal autoencoder [21] and multi-node2vec a multi-network extension of the node2vec model [22] by directly replacing BIONIC in the DrugMAN framework. When DrugMAN uses the drug and target representations learned from deepNF and multi-node2vec, we observe an overall drop in performance in terms of AUROC, AUPRC, and F1-score across all scenarios (Fig. 3). Especially, we observe that the advantage of BIONIC is even more remarkable in cold-start conditions compared to the random split. For example, in the random testing, with the assistance of BIONIC, DrugMAN outperforms models with substitution of multi-node2vec by 2.8%, 2.6% and 3.8% in AUROC, AUPRC and F1-score, respectively. In the both-cold testing, this discrepancy has increased to 15.3%, 15.4% and 10.4% in AUROC, AUPRC and F1-score, respectively. The results indicate that BIONIC captures more sophisticated functional and topological information from drug and target heterogeneous networks to power the drug–target interaction prediction compared to the established network integration methods.

Fig. 3

Evaluation of different network embedding methods. In DrugMAN, the drug and target embedding module BIONIC is replaced by two other network integration methods: DeepNF and Multi-node2vec. The vertical bars represent the mean value of five random runs, and the black lines are error bars indicating the standard deviation. The dots indicate performance scores in each random run

Evaluation of integrated and single networks

To evaluate the impact of each single network on the model predicting ability, for each run we use one single network to produce the drug or target features and fix all other settings in DrugMAN. We find that all single network-based models (including four drug networks and seven protein target networks) show poorer performance compared to the primary DrugMAN (Supplementary Tables 1 and 2), confirming the significance of integrating heterogeneous networks for drug–target interaction prediction. For input drug networks, the best performance is observed for the drug structure similarity network-based model with AUROC of 0.949, 0.896, 0.913 and 0.834, AUPRC of 0.953, 0.910, 0.923 and 0.844 and F1-score of 0.883, 0.823, 0.837 and 0.764, in random, drug-cold, target-cold and both-cold testing, respectively (Supplementary Table 1). This is reasonable as the structure information is the basis for drug binding to the target. Consistently, in all single protein network-based models, the protein sequence similarity network-based model performs better than or as well as other single network models across all scenarios (Supplementary Table 2). Based on the prominent contribution of structural information to DrugMAN, we further examine the performance of the model with only the drug structure similarity network and the protein sequence similarity network as input (DrugMAN_STR). Encouragingly, DrugMAN_STR outperforms those state-of-the-art structural models including SVM, RF and DeepPurpose, indicating the DrugMAN framework can capture more information from the structural similarity data related to drug–target interactions compared to the established methods. Moreover, compared to DrugMAN_STR, we can observe that DrugMAN has significant performance improvements with the introduction of non-structural heterogeneous information in all scenarios (Fig. 4). The discrepancy between DrugMAN_STR and the primary DrugMAN can to some extent reflect the contribution of non-structural information to the performance. For example, in the both-cold scenario, DrugMAN outperforms DrugMAN_STR by 7.6%, 6.2% and 6.7% in AUROC, AUPRC and F1-score, respectively (Fig. 4). These results demonstrate the strength of DrugMAN in generalizing prediction performance across different conditions by integrating various pharmacological and biological information.

Fig. 4

Evaluation of DrugMAN based on the drug structure similarity network and the protein sequence similarity network. DrugMAN with only the drug structure similarity network and the protein sequence similarity network as input (DrugMAN_STR) outperforms three state-of-the-art chemoinformatic models but underperforms DrugMAN. The box plots show the median as the center lines and the mean as green triangles. The minima and lower percentile represent the worst and second-worst scores. The maxima and upper percentile indicate the best and second-best scores

Ablation study

We perform an ablation study to investigate the impact of the mutual attention network on DrugMAN. As shown in Fig. 5, the introduction of the mutual attention network has significantly improved the performance of DrugMAN in all scenarios. The results indicate that the mutual attention network can capture the pairwise interaction information for drug–target interaction prediction.

Fig. 5

Ablation study for DrugMAN without attention mechanisms. Performance comparison of DrugMAN with and without attention mechanisms (No attention) in different scenarios. The vertical bars represent the mean of five random runs, and the black lines are error bars indicating the standard deviation. The dots indicate performance scores in each random run

Discussion and conclusion

In this work, we develop a new framework, DrugMAN, to integrate drug and protein target information from multiplex biological networks to mine drug–target interactions. DrugMAN achieves superior performance over both state-of-the-art chemoinformatics and network-based models. Especially, the stability of DrugMAN performance has been confirmed in different real-world scenarios. DrugMAN’s effectiveness can be attributed to two intrinsic advantages: the sophisticated network embedding module for learning suitable drug and target features from heterogeneous data and the mutual attention block to capture interaction information between drugs and targets.

The main challenge of network embedding is how to encode accurate node features from heterogeneous networks with high-dimensional, incomplete and noisy traits. DrugMAN takes BIONIC, the latest deep learning-based network integration algorithm, which first characterizes the topology of each individual network by applying a GAT algorithm, and then formalizes a low-dimensional representation by combining the features learned from each individual network to approximate the initial networks. The network-specific integrated features learned by BIONIC can reflect both functional and topological properties of heterogeneous networks and excel other unsupervised methods in a range of downstream tasks [18]. We demonstrate that DrugMAN can achieve substantial improvement over the state-of-the-art network-based methods for drug–target interaction prediction (Fig. 2). Consistently, we here compare BIONIC with two classical network integration approaches, deepNF and multi-node2vec in the DrugMAN framework (Fig. 3), indicating that BIONIC is more adaptive for drug–target interaction prediction compared to the established network integration methods. In addition, we find when DrugMAN uses single drug or target networks as input, the predicting performance is greatly inferior compared to DrugMAN with multiple networks (Supplementary Tables 1 and 2), indicating BIONIC can produce accurate integrated drug and target features from heterogeneous networks for drug–target interaction prediction.

Most existing drug–target prediction models learn drug and target representations using their separate encoders and ignore mutual impacts between the targets and drugs [23, 24]. The pairwise interaction information between drugs and targets is explicitly important for drug–target interaction prediction [25, 26]. We use a mutual attention network that utilizes a series of transformer encoders to capture interaction patterns between drugs and targets. We demonstrate that the introduction of the mutual attention block in the DrugMAN architecture significantly improves the performance of DrugMAN in all scenarios (Fig. 5). In summary, DrugMAN can provide a powerful and useful tool to facilitate drug discovery and drug repositioning. However, DrugMAN still has some limitations in predicting drug–target interactions based on chemical structure information. In this study, we only used the one-dimensional structural information of drugs and targets, and more information in the three-dimensional structure of drugs and targets has not been mined. However, to the best of our knowledge, there are currently existing drug–target interaction prediction models that use deep learning methods to extract the characteristics of drug and target based on the 3D structure of drug and target. For example, the GTAE-VF model with good performance is based on the 3D structure of drugs and targets, and the deep learning method of Graph transformer is used to extract the features of drugs and targets [27]. This is where we need to consider improving our future work.