Datasets

To evaluate the performance of a drug-target interaction prediction algorithm based on a drug-target bipartite network and graph representation learning, we tested our model on Luo et al. dataset [14] as well as on a newly constructed dataset. Detailed data for the above two datasets were provided in Additional file 1.

Luo’s dataset contains a total of four biological entities (drugs, proteins, diseases, and side effects). These entities constitute a total of seven interaction networks that GSRF-DTI uses to construct homogeneous networks of drugs and targets for subsequent DTI prediction.

Among them, drug entities were collected from the DrugBank database (Version 3.0) [30], protein entities were collected from the HPRD database (Release 9) [31], disease entities were collected from the Comparative Toxicogenomics database [32], and side effects entities were collected from the SIDER database (Version 2) [33]. A detailed description of the number of various biological entities and their interactions is provided in Additional file 2: S1.1. In addition, we used the chemical structure of the drugs and the similarity information of the protein sequence. The chemical structures of drugs were downloaded from the DrugBank database (Version 3.0), and the protein sequence was downloaded from the integrated medicinal genomic database of Sophic [34].

The newly constructed dataset similarly contains four types of entities (drugs, proteins, diseases, and side effects). Inspired by [14] and [9], the sources of drug-disease, drug-side effect, and protein-disease association information were the same as those of Luo’s dataset. Then, chemical structures of the drugs were obtained from the DRUG and COMPOUND Sections in the KEGG LIGAND database [35]. Amino acid sequences of the target proteins were obtained from the KEGG GENES database. Finally, known drug-protein interactions were obtained from KEGG BRITE, BRENDA [36], SuperTarget [37], and DrugBank databases; we focused only on regulatory interactions between enzymes and compounds.

Since different data sources use different identifiers for the same entity, we utilized BioKG [38] to parse drugs and proteins identities of different data sources into unified DrugBank ids and UniProt ids, respectively. Based on the unified IDs, we only retained drugs that have disease, side effect, and chemical structure information simultaneously as well as proteins that have disease and amino acid sequence information simultaneously. A detailed description of the number of various biological entities and their interactions is provided in Additional file 2: S1.2.

Data generation

The evaluation dataset was generated in the same way as the method used in EEG-DTI [39]. In terms of expansion, for Luo’s dataset, this dataset contains 708 drugs and 1512 targets; the DTP-NET consisted of 1,070,496 \((708 \times 1512)\) nodes. We labeled the nodes corresponding to the 1923 known drug-target interactions as 1, considering them as positive samples. Next, we randomly selected nodes with an equal number of positive and negative samples from the remaining 1,068,573 nodes labeled as 0. For the newly constructed dataset, there are 151 drugs and 285 proteins, so the DTP-NET constructed based on this has a total of 43,035 \((151\times 285)\) nodes, of which 481 are known drug-target interactions.

To minimize the influence of data variability on the results, we used fivefold cross-validation to evaluate our model. Positive samples and negative samples were divided into 5 parts. Then, one positive part and one negative part were selected as the test sets every time, and the remaining parts were successively selected as the training set. Finally, the average value of the five results was calculated as the final evaluation metric.

Performance evaluation on Luo dataset

To comprehensively evaluate the performance of GSRF-DTI, we used the area under the receiver operating characteristic curve (AUROC) and the area under the exact recall curve (AUPR) as evaluation index, similar to previous work [15, 17, 39]. We compared GSRF-DTI with five state-of-the-art DTI prediction approaches on Luo’s dataset, including EEG-DTI [39], GCN-DTI [15], BLMNII [10], NRLMF [40], and DTI-NET [14]. The introduction of approaches proposed above is provided in Additional file 2: S2.1. In addition, the specific parameter settings in GSRF-DTI were shown in Additional file 2: S2.2.

The comparative results are shown in Table 1. GSRF-DTI consistently outperformed the other five baseline methods, with AUROC and AUPR values of up to 97.78% and 98.04%, respectively. These AUROC and AUPR values were 2.59% and 1.94% higher, respectively, compared to EEG-DTI, which ranked as the third-best approach. The potential reason is that GSRF-DTI additionally considers the association relationship between DTPs. Compared to GCN-DTI, GSRF-DTI achieved a 4.27% higher AUROC and a 3.32% higher AUPR. GSRF-DTI may have a greater consideration of the multiple interrelated interactions between biological entities in the process of drug and target feature representation learning. The visual representation of the results is shown in Fig. 4a.

Table 1 AUROC and AUPR results of DTI prediction from the different methods on Luo’s datasetFig. 4

AUROC and AUPR performance of different methods. a AUROC and AUPR performance of different methods on Luo’s dataset. b AUROC and AUPR performance of different of different methods on the newly constructed dataset

Performance evaluation on the newly constructed dataset

For further evaluation, we constructed a new dataset, implemented GSRF-DTI algorithm, and compared it with other five methods. The experimental results are shown in Table 2.

Table 2 AUROC and AUPR values of DTI prediction from the different methods on the newly constructed dataset

Comparing the results in Table 1, it is evident that the performance of models based on deep learning, such as GSRF-DTI, DTI-MGNN, EEG-DTI, and GCN-DTI, in predicting DTI on the newly constructed dataset has slightly decreased. This can be attributed to the fact that deep learning algorithms generally perform better on large sample problems. The results in Table 2 show that the AUROC and AUPR values of GSRF-DTI reach as high as 96.66% and 96.85% respectively, both outperforming the other five baseline methods, thus further emphasizing the effectiveness of our proposed method GSRF-DTI in identifying DTIs. The visual representation of the results is shown in Fig. 3b.

Sensitivity analysisThe effects of learning_rate

The learning rate is an important parameter in supervised learning and deep learning, which determines the step size of the weight change and affects the convergence of the objective function. To obtain the optimal model, we set different learning rates to train GSRF-DTI. The detailed results are shown in Additional file 2: S3.1, and the average AUROC and AUPR values of fivefold cross-validation are shown in Fig. 5.

Fig. 5

AUROC and AUPR values of the different learning rates

From Fig. 5, we can conclude that AUROC and AUPR values of the model prediction results were the highest when the learning rate was 0.001. We continuously compared the effects of different values of learning rate on the model performance and finally set learning rate to 0.001.

The effects of the aggregate function

The aggregation function directly affects the feature representation of nodes in the DTP-NET, which indirectly affects DTI prediction. The GraphSAGE algorithm usually has three commonly used aggregation functions: the mean aggregator, pooling aggregate, and LSTM aggregator. To assess the impact of different aggregation functions on model performance, we calculated the evaluation index of the model prediction results under each aggregation function. The detailed results are shown in Additional file 2: S3.2, and the average AUROC and AUPR values of fivefold cross-validation are shown in Fig. 6. Note that we set learning_rate to 0.001 at this point.

Fig. 6

AUROC and AUPR values from the different aggregators

Figure 6 shows that the performance of the LSTM aggregator was slightly better. However, the time complexity of the LSTM aggregator was much larger than that of the other two aggregators. Considering the time complexity and model performance, we chose the mean aggregator as the aggregation function of our GSRF-DTI model.

The effects of the classifier

The classical algorithms in machine learning are logistic regression (LR) [41], support vector machine (SVM) [42], and random forest (RF) [29], which are commonly used to work out binary classification tasks. To obtain better model performance, we took the node label \(_\) and the node features \(_\) sampled and aggregated by the GraphSAGE algorithm as the input of the above classification algorithms to train the model and predict the DTIs. To avoid the occasionality of the results, we randomly divide the dataset 50 times, namely, 75% as the training set and 25% as the validation set. The average of the results was used as the final index value, as shown in Table 3.

Table 3 AUROC and AUPR values from the different classification algorithms

The experimental data showed that RF performed the best in classification; therefore, RF was determined as the classification method of our model for DTI identification.

Ablation experimentThe effects of different types of network information

One of the innovative points of the GSRF-DTI proposed in this paper is the integration of seven types of network information. In order to evaluate the importance of each type of network on DTI prediction, we designed five ablation experiments on Luo’s dataset. The experimental findings are summarized in Table 4.

Table 4 Evaluation metric values of the ablation experiments

From Table 4, it is evident that the model performed best when utilizing all networks. When protein sequence information was removed, the model exhibited the poorest performance in predicting DTI. This suggests that among the seven types of information considered, protein sequence information was the most important for accurately identifying DTI. When removing drug chemical structures or side effects, the AUROC and AUPR decreased, but only slightly. However, when removing diseases, AUROC and AUPR experienced a significant decrease. This could be attributed to diseases being associated with both drugs and proteins, whereas chemical structures and side effects are exclusively related to drugs. Finally, if side effects and diseases were removed simultaneously, the AUROC and AUPR decreased compared to when diseases or side effects were removed individually.

All the ablation experiments conducted above provide evidence that the integration of the seven types of information can enhance the identification performance of DTI.

Evaluation of the effectiveness of the GraphSAGE algorithm

We propose that the GSRF-DTI model is suitable for feature representation learning on large-scale networks. To further illustrate the effectiveness of the GraphSAGE algorithm, we designed three comparative experiments. Specifically, in the basic experiment, we identified DTIs limited to drug and target-related information, that is, the initial features \(_\) of DTPs obtained by concatenating the corresponding drug feature \(_\) and target feature \(_\) after the action of the Deepwalk were directly used as the input of the three classification algorithms. In the contrast experiment, the association between DTPs was further considered, that is, the DTP features \(_\) after the GraphSAGE that was applied to the DTP-NET were used as the input of the three classification algorithms. We evaluated the model performance by calculating the following six evaluation indicators: accuracy, precision, recall, F1 score, AUROC, and AUPR. The detailed results are shown in Additional file 2: S4, and the average values of fivefold cross-validation are shown in Table 5.

Table 5 The six evaluation metric values of the comparative experiments

Table 5 depicts that further considering the association information between DTPs was far better than the experimental results that only considered the drug/target association information. The model prediction results of the three classical classification algorithms combined with GraphSAGE were higher than those of each classification algorithm alone. Thus, the effectiveness of the GraphSAGE algorithm was more strongly illustrated. In this paper, GraphSAGE combined with RF was finally applied to predict DTIs. Table 5 also depicts that the AUROC and AUPR values of the model (Deepwalk + GraphSAGE + RF) prediction were the highest. Specifically, GSRF-DTI achieved an AUROC score of 0.9818 and an AUPR score of 0.9839. To compare the results of the basic experiment and the contrast experiment more intuitively, ROC and PR curves were drawn and shown in Fig. 7.

Fig. 7

ROC curves and PR curves of the comparative experiments

Based on the results in Fig. 7, the area under all the solid lines exceeded that under the dotted lines. As previously mentioned, the classification algorithm combined with the GraphSAGE algorithm exhibited slightly superiority, indicating the effectiveness of GraphSAGE for inductive representation learning of large graphs. It also shows that the consideration of the interaction information between DTPs played an important role in DTI prediction.