Predicting drug-target interactions using matrix factorization with self-paced learning and dual similarity information.

This study introduced a novel DTI prediction model, self-paced learning with dual similarity information and MF method (SPLDMF), to predict unknown DTIs. Five matrices S t , S d , P t , P d , and Y represented target similarity, drug similarity, drug topological feature similarity, target topological feature similarity, and known DTI, respectively. The task was to explore how to use known information to predict unknown DTIs. Then, four scenarios based on DTI were created to more comprehensively display the performance of the model (Fig.  1 ). To describe these four scenarios, we utilized five drugs (i.e., D 1 to D 5 ) and four targets (i.e., T 1 to T 4 ) as an example. Then, the D 1 – T 1 interaction pair on the orange background can represent four scenarios depending on the conditions: (1) known drug-known target (scenario 1 in Fig.  1 a); (2) known drug-new target (scenario 2 in Fig.  1 b); (3) new drug-known target (scenario 3 in Fig.  1 c); and (4) new drug-new target (scenario 4 in Fig.  1 d) Figure 1. Four scenarios of DTI predictions. The pair with orange background represents (a) known drug-known target; (b) known drug-new target; (c) new drug-known target; and (d) new drug-new target. Four scenarios of DTI predictions. The pair with orange background represents (a) known drug-known target; (b) known drug-new target; (c) new drug-known target; and (d) new drug-new target. Figure 2. Process of our proposed model. Process of our proposed model. In the protocol, definitions reference to a “known drug” means that the experimental drug has at least one interaction with the targets (e.g., D 1 in Figs  1 a and  2 b, respectively). Similarly, “known target” means that the experimental target has at least one interaction with drugs. In contrast, “new drug” denotes that the experimental drug has no known interactions with the targets (e.g., D 1 in Fig.  1 c and d). Similarly, “new target” means that the experimental target has no existing interaction with drugs. The focus of this study is to use the SPLDMF method to improve the DTI prediction ability of the model. Specifically, the algorithm assigns scores to drug-target pairs to estimate the likelihood of their interaction, and the higher the score is, the more likely the drug and target will interact. Suppose N d known drugs are represented by a matrix D ), then D = { d 1 , d 2 , … , d N d } . Assuming N t known targets, a set of known targets T can be represented as T = { t 1 , t 2 , … , t N t } . Let { S d ⁢ P d } represent similarity matrices related to drugs, and the dimension of S d and P d is N d × N d . Similarly, if { S t ⁢ P t } are the similarity matrices involving targets, then the dimension of S t and P t is N t × N t . Let Y be an N d × N t adjacency matrix, which can be expressed as the DTI. When Y i ⁢ j = 1, the drug d i interacts with the target t j ; when Y i ⁢ j = 0, no interaction between drug d i and the target t j is observed. Our goal was to reconstruct F , which was an N d × N t score matrix. When the score F i ⁢ j of F is higher, it meant that the drug d i more likely interacted with the target t j . In this study, the attributive and topological properties of the drug and the target were used. The drug and target attributive features referred to the drug structure and the amino acid sequence of the target protein, respectively. Yamanishi et al. [ 28 ] also collected a dataset including the attributive feature similarity of the drug and the target. The structural data of all network nodes were referred to as topological features. Drug-drug topological feature similarity and target-target topological feature similarity were measured using the Node2vec method and the cosine similarity method, respectively, to extract the topological features of drugs and targets from the DTI network [ 43 ]. The DTI matrix Y ∈ R N d × N t was obtained from the dataset. Then, a weightless and undirected network graph G = ( V , E ) was constructed based on the DTI matrix Y , where V denotes the set of nodes, | V | = N d + N t , where | V | denotes the number of nodes. E denotes the set of edges, | E | = ∑ i ∈ N d ∑ j ∈ N t Y i ⁢ j , where | E | denotes the number of edges. When Y ⁢ ( i , j ) = 1, an edge exists such that V i and V j are connected; when Y ⁢ ( i , j ) = 0, no edge exists, and V i and V j are not connected. Then, a second-order random walk was performed on the network graph G using the Node2vec method to obtain the topological features of drugs and targets. Moreover, we obtained the d-dimensional topological features of the drug and target using the Node2vec method. Next, we calculated the drug-drug and target-target topological feature similarity. We used the cosine similarity to calculate the topological feature similarity, and the cosine similarity between drugs represented the similarity of two drug vectors in the topological feature space. Likewise, the cosine similarity of target-target topological features was predicted as the similarity of two target vectors in the topological feature space. The topological feature vectors of two drugs d i and d j are denoted as x i and x j , both of which are d-dimensional topological features. Finally, the drug-drug topological feature similarity was measured with the help of cosine similarity using the sampling vertex sequence: (1) 𝑆𝑖𝑚 𝑑𝑡𝑝 = x i ⁢ x j T || x i || ⋅ || x j || For ease of description, the drug-drug topological feature similarity matrix can be represented as P d ∈ R N d × N d , where P d ⁢ ( i , j ) denotes the topological feature similarity between the i -th and the j -th drugs. Correspondingly, the target-target topological feature similarity matrix is represented by P t ∈ R N t × N t , where P t ⁢ ( i , j ) denotes the topological feature similarity between the i -th and the j -th targets. The goal of MF was to factorize the identified DTI matrix Y into two low-rank matrices A and B . The dimensionality of A and B are matrices of N d × r and N t × r , respectively, where r denotes the dimensionality of the feature space, A denotes the potential feature representation of the drug, and B denotes the potential feature representation of the target. As the DTI matrix Y can be factorized into A and B , the inner product of A and B is approximately equal to the DTI score, and Y is represented as: (2) Y ≈ A ⁢ B T First, A and B were calculated to obtain Y . Subsequently, the squared error of Eq. ( 2 ) was minimized to obtain: (3) argmin A , B ⁢ || Y - A ⁢ B T || F 2 where | | ⋅ | | F 2 is the Frobenius norm. Solving for Eq. ( 3 ) might directly lead to overfitting during training. Therefore, the L 2 regularization term was added to solve the aforementioned problem. Then, Eq. ( 3 ) was rewritten as: (4) argmin A , B ⁢ || W ⊙ ( Y - A ⁢ B T ) || F 2 + λ l ⁢ ( || A || F 2 + || B || F 2 ) where λ l represents the regularization parameter. Based on the idea that drugs with a higher degree of similarity tend to act on a similar set of targets, and vice versa, we integrated drug-related similarity matrices S d and P d and target-related similarity matrices S t and P t into the model to more accurately discover potential DTIs. Based on a previous study [ 33 ], the inner product of the corresponding two drug feature vectors and two target feature vectors was used to approximate the drug similarity and target similarity matrices, respectively. The detailed decomposition process was as follows: (5) S d ≈ A ⁢ A T   S t ≈ B ⁢ B T   P d ≈ A ⁢ A T   P t ≈ B ⁢ B T Therefore, we added the drug similarity matrix S d , the target similarity matrix S t , the drug topological feature matrix P d , and the target topological feature P t into Eq. ( 5 ). The new equation was as follows: argmin A , B ⁢ || W ⊙ ( Y - A ⁢ B T ) || F 2 + λ l ⁢ ( || A || F 2 + || B || F 2 ) + λ d ⁢ || S d - A ⁢ A T || F 2 (6) + λ t ⁢ || S t - B ⁢ B T || F 2 + λ m ⁢ || P d - A ⁢ A T || F 2 + λ n ⁢ || P t - B ⁢ B || F 2 where λ d , λ t , λ m , and λ n are the regularization parameters. The objective function of the most recent MF-based approaches for DTI prediction is nonconvex. As a result, the optimized objective function can be easily trapped in local minima, particularly when dealing with enhanced noise and a large amount of missing data. Many studies showed that SPL could alleviate the model falling into a bad local optimal solution because of its training strategy of selecting samples from easy to complex [ 44 , 45 ]. Thus, we integrated the SPL algorithm into the MF model to improve its strength. Consequently, Eq. ( 3.3 ) could be modified as: argmin A , B ⁢ || W ⊙ ( Y - A ⁢ B T ) || F 2 + λ 1 ⁢ ( || A || F 2 + || B || F 2 ) + λ d ⁢ || S d - A ⁢ A T || F 2 (7) + λ t ⁢ || S t - B ⁢ B T || F 2 + λ m ⁢ || P d - A ⁢ A T || F 2 + λ n ⁢ || P t - B ⁢ B T || F 2 + γ 2 W + γ ⁢ k where k and λ denote the model age and the weights assigned to the selected samples, respectively. According to Zhao et al. [ 44 ], the optimal W i , j ∗ was calculated using Eq. ( 8 ) when A and B were fixed. (8) W i ⁢ j ∗ = { 1 if  ⁢ l i ⁢ j ⩽ 1 ( k + 1 / γ ) 2 0 if  ⁢ l i ⁢ j ⩾ 1 k 2 γ ⁢ ( 1 l i ⁢ j - k ) otherwise where l i ⁢ j = [ ( Y - A ⁢ B T ) i ⁢ j ] 2 . When l i ⁢ j ⩽ 1 / ( k + 1 / γ ) 2 , the corresponding weight was 1, implying that the sample was taken as a simple sample and selected by the model during the training; when l i ⁢ j ⩾ 1 k 2 , the sample was considered a difficult sample and was temporarily not selected by the model; in other cases, the sample was assigned a non-zero weight and was considered an easy sample. The alternative search strategy (ASS) was used to calculate A and B to overcome the problem of the potential feature vectors of the drug and the target to not easily solved as they tended to couple together. The potential feature vector of the drug was represented by a i , which is a row vector of matrix A . Furthermore, the potential feature vector of the target was represented by b j , which is a row vector of matrix B . The objective function was transformed as in Eq. ( 9 ) to implement the ASS algorithm. (9) L = ∑ i = 1 N d ∑ j = 1 N t W i ⁢ j ⁢ ( Y i ⁢ j - a i ⁢ b j T ) 2 + λ l ⁢ ( ∑ i = 1 N d || a i || 2 + ∑ j = 1 N t || b j || 2 ) + λ d ⁢ ∑ i = 1 N d ∑ p = 1 N d ( S d ⁢ ( d i , d p ) - a i ⁢ a p T ) 2 + λ t ⁢ ∑ j = 1 N t ∑ q = 1 N t ( S t ⁢ ( t j , t q ) - b j ⁢ b q T ) 2 ⁢ λ m ⁢ ∑ i = 1 N d ∑ p = 1 N d ( P d ⁢ ( d i , d p ) - a i ⁢ a p T ) 2 + λ n ⁢ ∑ j = 1 N t ∑ q = 1 N t ( P t ⁢ ( t j , t q ) - b j ⁢ b q T ) 2 We fixed B and computed the partial derivative of L with respect to a i to minimize L . Afterward, A was updated by ∂ ⁡ L ∂ ⁡ a i = 0. The updated equation obtained after derivation was as enumerated by the equation: (10) a i = ∑ j = 1 N t W i ⁢ j ⁢ Y i ⁢ j ⁢ b j + λ d ⁢ ∑ p = 1 N d S d ⁢ ( d i , d p ) ⁢ a p + λ m ⁢ ∑ p = 1 N d P d ⁢ ( d i , d p ) ⁢ a p ∑ j = 1 N t W i ⁢ j ⁢ b j T ⁢ b j + λ l ⁢ I k + λ d ⁢ ∑ p = 1 N d a p T ⁢ a p + λ m ⁢ ∑ p 1 N d a p T ⁢ a p Similarly, we fixed A and computed the partial derivative of L with respect to b j . Then, B was updated using ∂ ⁡ L ∂ ⁡ b i = 0. The updated equation obtained after derivation was as enumerated by the equation: (11) b j = ∑ i = 1 N d W i ⁢ j ⁢ Y i ⁢ j ⁢ a i + λ t ⁢ ∑ q = 1 N t S t ⁢ ( t j , t q ) ⁢ b q + λ n ⁢ ∑ q = 1 N t P t ⁢ ( t j , t q ) ⁢ b q ∑ i = 1 N d W i ⁢ j ⁢ a i T ⁢ a i + λ l ⁢ I k + λ t ⁢ ∑ q = 1 N t b q T ⁢ b q + λ n ⁢ ∑ q = 1 N t b q T ⁢ b q where l k in Eqs ( 10 ) and ( 11 ) is the identity matrix. Algorithms 1 and 2 explain the process of assessing individual parameters. The potential drug characteristic representation A and the potential target characteristic representation B were obtained after several iterations using Eqs ( 10 ) and ( 11 ). We obtained the DTI prediction matrix F by reconstructing the DTI matrix Y , and the calculation procedure was as enumerated by the equation: (12) F = A ⁢ B T Algorithm 1: Pseudocode of parameter estimation for MF Input:   Y : true drug-target interaction matrix; W : weight matrix; S d , S t : drug and target similarity matrices;   P d , P t : drug and target topological feature matrix; r : feature space; λ l , λ d , λ t , λ m , λ n : regularization parameters Ouput:  drug potential representation A , target potential representation B and score matrix F 1: initial A and B randomly; 2: repeat 3:  Update A using Eq. ( 10 ); 4:  Update B using Eq. ( 11 ); 5:  Update F using Eq. ( 12 ); 6: until The drugs (compounds) and targets (small molecules) could be determined based on the prediction result, that is, the scoring and ranking of matrix F . The workflow of the whole method is shown in Fig.  2 . Algorithm 2: Pseudocode of parameter estimation for SPLDMF Input:   Y : true drug-target interaction matrix; S d , S t : drug and target similarity matrices;   P d , P t : drug and target topological feature matrix; r : feature space; λ l , λ d , λ t , λ m , λ n : regularization parameters;   μ > 1: step size; k 0 ; k 𝑒𝑛𝑑 Ouput:  score matrix F 1: initial solve the MF problem with all the observation equally weighted to obtain A 0 and B 0 , calculate t ← 0 , k ← k 𝑒𝑛𝑑 2: while k > k 𝑒𝑛𝑑 do 3:  Update W using Eq. ( 8 ); 4:  Update A and B using Algorithm 1; 5:  Update F using Eq. ( 12 ); 6:  Compute currentd l i ⁢ j ⁢ ( Y - F ) ; 7:  t ← t + 1 , k ← k / μ ; 8: end while

Compared with other methods, the performance of the proposed model was assessed by simulating experiments under different missing rates and noise ratios. Then, compared with the performance of the advanced model, the performance was tested using four application scenarios. Further, two realistic and challenging extended datasets were selected for experimental comparison. We used four matrices such as root-mean-squared error (RMSE), mean absolute error (MAE), area under the receiver operating characteristic curve (AUC), and precision-recall curve (AUPR) to evaluate the effectiveness of SPLDMF. Simulation experiments were carried out to test the robustness of the model under different missing rates and noise ratios. We compared the proposed SPLDMF with two popular DTI prediction methods: MF and SVD. According to the studies by Xia et al. [ 33 ], Zheng et al. [ 46 ], and Zhao et al. [ 44 ], a matrix Y ′ following Gaussian distribution N ⁢ ( 0 , 1 ) was developed randomly using n = 300, m = 200, and r = 3. We set three missing ratios (50%, 50%, and 90%) and five noise ratios (5%, 10%, 20%, 25%, and 40%) to verify the validity and robustness of the models. We determined that the noise property of Y ′ was uniform noise in the range [ - 20 , 20 ] . Based on a previous study [ 47 ], the conversion between matrices Y ′ and Y was possible, and Y ′ with a well-fitting effect could help explore new DTIs. RMSE and MAE criteria were used for evaluating the performance of the three methods, where 𝑅𝑀𝑆𝐸 = 1 m ⁢ n ⁢ || Y ′ - A ⁢ B T || F , 𝑀𝐴𝐸 = 1 m ⁢ n ⁢ || Y ′ - A ⁢ B T || 1 , and m , n are the rows and columns of the matrix Y , respectively. We performed 30 replicate experiments for each method, and the performance of each method was qualified based on the average of the experimental results (Table  2 ). SPLDMF achieved the best RMSE and MAE performance in each case by comparing the three methods with three missing ratio levels and five noise ratio levels. For instance, when missing ratio = 10% and noise ratio = 10%, the RMSE and MAE of SPLDMF reached 0.886 and 0.296, respectively, which were much better compared with the values of MF (1.472 and 0.667, respectively) and SVD (1.970 and 0.935, respectively). The predictive performance of the models decreased as the deletion rate increased. The proposed SPLDMF imposed more regularization constraints on the self-similarity of drugs and targets, allowing more similar DTIs to be accurately predicted. Table  2 demonstrates that the best performance of our method could be obtained at all three data missing ratios. Additionally, the prediction error of all models increased with the increase in the noise ratio. However, the proposed SPLDMF was capable of adaptively weighting both clean and noisy samples due to the introduction of the SPL strategy. This learning strategy enabled the model to avoid falling into bad local optima and had better robustness to mitigate the effects of noise. Overall, the results of the simulation experiments revealed that the SPLDMF outperformed the MF and SVD methods under noise and missing data conditions. Table 2 Performance comparison of MF, SVD, and SPLDMF on synthetic data in terms of MAE and RMSE Missing_ratio (%) Noise_ratio (%) MAE RMSE CMF SVD SPLDMF CMF SVD SPLDMF 10 5 0.497 (0.040) 0.755 (0.048) 0.218 (0.005) 1.340 (0.040) 1.804 (0.048) 0.743 (0.029) 10 0.667 (0.026) 0.935 (0.034) 0.296 (0.009) 1.472 (0.038) 1.970 (0.034) 0.886 (0.031) 20 0.864 (0.023) 1.159 (0.048) 0.432 (0.016) 1.635 (0.038) 2.164 (0.048) 1.078 (0.042) 25 0.930 (0.022) 1.426 (0.045) 0.514 (0.020) 1.694 (0.026) 2.230 (0.045) 1.189 (0.043) 40 1.113 (0.025) 1.481 (0.049) 0.872 (0.032) 1.833 (0.036) 2.411 (0.049) 1.710 (0.060) 50 5 0.681 (0.039) 0.795 (0.045) 0.259 (0.008) 1.776 (0.047) 2.153 (0.045) 0.889 (0.047) 10 0.911 (0.033) 1.018 (0.038) 0.351 (0.013) 1.970 (0.052) 2.346 (0.038) 1.075 (0.053) 20 1.151 (0.022) 1.297 (0.031) 0.552 (0.022) 2.231 (0.037) 2.565 (0.031) 1.393 (0.063) 25 1.228 (0.032) 1.411 (0.038) 0.659 (0.032) 2.289 (0.051) 2.650 (0.038) 1.554 (0.085) 40 1.462 (0.026) 1.734 (0.042) 1.094 (0.037) 2.469 (0.042) 2.889 (0.042) 2.157 (0.078) 90 5 0.656 (0.027) 0.775 (0.078) 0.402 (0.019) 2.453 (0.072) 2.571 (0.078) 0.996 (0.059) 10 1.247 (0.035) 1.138 (0.034) 0.497 (0.017) 3.315 (0.080) 2.881 (0.034) 1.337 (0.072) 20 2.027 (0.037) 1.514 (0.032) 0.890 (0.042) 4.262 (0.064) 3.186 (0.032) 2.171 (0.120) 25 2.307 (0.036) 1.683 (0.037) 1.110 (0.045) 4.559 (0.083) 3.322 (0.037) 2.520 (0.128) 40 2.940 (0.052) 2.138 (0.044) 1.846 (0.079) 5.129 (0.085) 3.649 (0.044) 3.540 (0.148) Performance comparison of MF, SVD, and SPLDMF on synthetic data in terms of MAE and RMSE We used the same dataset and cross-validation technique to compare our method with state-of-the-art methods (i.e., 5-time-10-fold cross-validation using Yamanishi’s benchmark dataset in four different applications scenarios) to validate the performance of the model. Three cross-validation settings were used to better evaluate the model in these four scenarios: (1) CVP, which was based on the cross-validation of drug-target pairs; (2) CVR, which was based on cross-validation on rows; (3) CVC, which was based on cross-validation on columns; and (4) CV4S, which was based on random cross-validation. Table  3 depicts the application scenario as well as the optimal potential feature dimensionality settings in our experiments. We employed the CVP settings to predict known drug-known target interactions (i.e., scenario 1, named CVPS). Figure  3 illustrates the model’s AUPR and AUC values for several potential features. The findings revealed that a higher potential feature dimensionality was more consistent AUPR and AUC values. In the CVP scenario, the GPCR dataset also reached the optimal feature dimensionality at r = 80 (Fig.  3 a). We used the CVR settings (i.e., scenario 3, named CVRS) for predicting a new drug-known target interaction. The model’s AUPR and AUC values were calculated for various potential features. Table 3 Application scenarios and dataset settings and optimal feature dimensionality CVPS CVCS CVRS CV4S Dataset settings CVP CVC CVR CVP/CVC/CVR Best feature dimension 80 100 100 100 Application scenarios and dataset settings and optimal feature dimensionality Figure 3. Performance comparison of SPLDMF and other advanced models, and the influence and change of r on AUC and AUPR in different scenarios. (a) Changes in AUC and AUPR under different feature dimensions under CVPS. (b) Changes in AUC and AUPR under different feature dimensions under CVRS. (c) Variation in AUC and AUPR under different feature dimensions under CVCS. (d) Performance comparison of SPLDMF and other advanced models under the GPCR dataset in four scenarios. Performance comparison of SPLDMF and other advanced models, and the influence and change of r on AUC and AUPR in different scenarios. (a) Changes in AUC and AUPR under different feature dimensions under CVPS. (b) Changes in AUC and AUPR under different feature dimensions under CVRS. (c) Variation in AUC and AUPR under different feature dimensions under CVCS. (d) Performance comparison of SPLDMF and other advanced models under the GPCR dataset in four scenarios. The values are the average findings of 30 runs. The best results are shown in bold, and the values in parentheses are standard deviations. The value was found to be the highest at r = 100. In the CVR scenario, the GPCR dataset also achieved the optimal feature dimensionality at r = 100 (Fig.  3 c). The CVC configuration was applied (i.e., scenario 2, named CVCS) for predicting new target-known drug interactions. Figure  3 c illustrates the model’s AUPR and AUC values for several potential feature dimensionalities. The experimental findings revealed that the AUC curves in the CVC scenario differed significantly from those in the CVP and CVR scenarios, particularly with the possible feature dimensionality r = 70 (a variation amplitude of more than 0.2). In the CVC scenario, the GPCR dataset also had the best feature dimensionality at r = 100. The fourth of the four scenarios (CV4S, new drug-new target) was the most difficult for DTI prediction. Since this sort of cross-validation was random and the training datasets and test datasets were also generated randomly, the test dataset might contain samples of fresh medications and fresh targets to aid in the inclusion of drug-target combinations in the new drug-new target category ( D 1 – T 1 pairs in Fig.  1 d). In the CVP situation, we performed 50 times of 5-time-10-fold cross-validation tests based on GPCR data. The optimal AUPR and AUC results were 0.651 ± 0.050 and 0.910 ± 0.012, respectively. The detailed calculation procedure was demonstrated in the code. Table 4 Comparison of the matrices from the major algorithms in CVPS, CVRS, CVCS, and CV4S scenarios based on the GPCR dataset Scenario Method AUC AUPR CVPS NRLMF 0.969 ± 0.004 0.749 ± 0.015 DNILMF 0.975 ± 0.003 0.812 ± 0.009 SPLCMF 0.976 ± 0.012 0.779 ± 0.015 SPLDMF 0.982 ± 0.004 0.815 ± 0.015 CVRS NRLMF 0.895 ± 0.011 0.364 ± 0.023 DNILMF 0.967 ± 0.006 0.781 ± 0.050 SPLCMF 0.967 ± 0.002 0.784 ± 0.023 SPLDMF 0.971 ± 0.012 0.792 ± 0.050 CVCS NRLMF 0.930 ± 0.012 0.556 ± 0.038 DNILMF 0.933 ± 0.009 0.684 ± 0.036 SPLCMF 0.931 ± 0.010 0.675 ± 0.015 SPLDMF 0.941 ± 0.023 0.710 ± 0.050 CV4S NRLMF 0.706 ± 0.008 0.385 ± 0.006 DNILMF 0.897 ± 0.004 0.633 ± 0.025 SPLCMF 0.856 ± 0.008 0.645 ± 0.025 SPLDMF 0.910 ± 0.012 0.651 ± 0.050 Comparison of the matrices from the major algorithms in CVPS, CVRS, CVCS, and CV4S scenarios based on the GPCR dataset Table 5 Top 10 drug-target relationship prediction scores and their validation Rank Drug name Target name Score Databases Literature 1 Verapamil SCN4A 0.983 C, D, K [ 49 , 50 ] 2 Clozapine DD5R 0.978 D [ 51 ] 3 Mirtazapine 5HR1A 0.902 D [ 52 ] 4 Diethylstilbestrol ESR1 0.896 C, D, K [ 53 , 54 ] 5 Norehindrone ESR1 0.894 – – 6 Methysergide 5HR1D 0.893 C, D, K [ 55 ] 7 Flunitrazepam GARSA1 0.891 C, K [ 56 ] 8 Clozapine ADRA1A 0.886 C, D [ 57 , 58 ] 9 Loxapine 5HR2B 0.879 C, D, K [ 59 ] 10 Isoflurane GABRA1 0.876 D [ 60 ] Top 10 drug-target relationship prediction scores and their validation Table 6 Comparison of the matrices from DNILMF, SPLCMF, and SPLDMF algorithms in four scenarios based on the Kuang and Hao datasets Dataset Scenario AUC AUPR DNILMF SPLCMF SPLDMF DNILMF SPLCMF SPLDMF Kuang CVP 0.941 0.933 0.949 0.649 0.733 0.842 CVR 0.803 0.831 0.840 0.602 0.491 0.710 CVC 0.862 0.886 0.888 0.643 0.456 0.731 CV4S 0.897 0.826 0.903 0.633 0.435 0.742 Hao CVP 0.943 0.935 0.943 0.748 0.721 0.816 CVR 0.811 0.792 0.843 0.736 0.740 0.741 CVC 0.852 0.868 0.881 0.683 0.710 0.726 CV4S 0.901 0.816 0.912 0.621 0.593 0.735 Comparison of the matrices from DNILMF, SPLCMF, and SPLDMF algorithms in four scenarios based on the Kuang and Hao datasets We conducted sufficient comparative experiments for the aforementioned four scenarios to verify the effectiveness of the proposed method. Specifically, we compared SPLDMF with three other state-of-the-art methods, and the results are depicted in Table  4 . The results indicated that the AUC and AUPR of SPLDMF were currently the best among the comparison methods. Our method could deal with noisy data more robustly due to the introduction of the SPL strategy, thus achieving better performance. The result showed that SPLDMF under all scenarios outperformed NRLMF and DNILMF in AUC and AUPR, suggesting that the proposed SPLDMF was more robust when using ligand-based methods to anticipate the interactions between ligands and target proteins. Our method outperformed in all scenarios compared with SPLCMF, which also used SPL strategy. An insightful explanation was that we leveraged more drug-drug and target-target similarities to improve predictive capacity for unknown outcomes. The result also demonstrated that SPLDMF had an improvement of 0.054 and 0.006 in AUC and AUPR, respectively, in the most difficult scenario CV4S, compared with SPLCMF. The prediction matrix was scored using Eq. ( 12 ). We took the top 10 DTI pairs with the prediction scores after synthesizing the DTI prediction scores of NR, GPCR, IC, and E. Data validation was performed using ChEMBL, DrugBank, and KEGG databases, labeled C, D, and K, respectively. We validated the partial prediction results based on previous studies. The fifth and sixth columns of Table  5 list the database used for data validation and the studies referred to for the validation method, respectively. Table  5 lists the top 10 predicted DTIs. The most anticipated interaction was between DB00661 (verapamil) and P35499 (SCN4A) with a predicted high score of 0.983. This predicted relationship was found in the three databases C, D, and K. Furthermore, they were also reported in previous studies (Shafi et al., 2022; Stee et al., 2020). Except for the fifth item, other predictions were found in relevant reports in the database and literature, which verified these predictions to a certain extent. The fifth pair, the relationship between norethindrone (DB00717) and ESR1 ( P03372 ), had no relevant reports in the current database and literature. According to the FDA, the drug norethindrone (DB00717), similar to the drug diethylstilbestrol (DB00255), is a progestin used for contraception, the prevention of endometrial hyperplasia in hormone replacement therapy, and the treatment of other hormone-mediated diseases such as endometriosis. Diethylstilbestrol is also used to treat diseases such as breast and prostate cancer, but it is listed as a known carcinogen. The predicted results indicated that norethindrone has the same target (ESR1) as diethylstilbestrol. Besides its proven contraceptive use, norethindrone may also be used to treat breast cancer, prostate cancer, and other diseases based on the target principle. We verified our speculation through the KEGG pathway analysis experiment. Besides simulated data and common benchmark datasets, the proposed SPLDMF was also tested with additional expanded datasets (prepared by Kuang [ 39 ] and Hao [ 31 ]) to fully verify the effectiveness of the suggested model on various datasets. A total of 3681 known interactions, 786 drugs, and 809 targets were detected in the Kuang dataset. Moreover, 3688 known interactions, 829 drugs, and 733 targets were detected in the Hao dataset. Table  6 depicts the performance comparison of SPLDMF and other methods on the expanded dataset, indicating that SPLDMF achieved the best prediction performance on both augmented datasets. This was mainly attributed to the fact that the SPL strategy improved the generalization performance of the model, enabling it to perform more robustly on noisy data. Meanwhile, the use of more feature similarity also enhanced the prediction accuracy, which was conducive to the discovery of potential DTIs.

Yamanishi [ 28 ], Kuang [ 39 ], and Hao [ 31 ] datasets are three critical databases used for validating the proposed DTI-related algorithm. The Yamanishi dataset is called a benchmark database, which contains drug-target relationships from databases such as KEGG BRITE [ 40 ], BRENDA [ 41 ], SuperTarget [ 27 ], and DrugBank [ 23 ], target protein sequence from KEGG Gene Database [ 40 ], and drug compounds from KEGG Drug and Compound Database [ 40 ]. Moreover, the Yamanishi database is categorized into four datasets: NR, GPCR, IC, and E. It contained 445 drugs and 664 targets in E, 210 drugs and 204 targets in IC, 223 drugs and 95 targets in GPCR, and 54 drugs and 26 targets in NR. The details of the dataset are depicted in Table  1 . The Kuang dataset had 3681 known interaction pairs [ 39 ], including 786 drugs and 809 targets (Table  1 ). The Hao dataset comprised 829 drugs, 733 targets, and 3688 identified interaction pairs [ 31 ] (Table  1 ). Table 1 Summary of four benchmark and two expanded datasets Dataset No of drugs No of targets No of interactions Sparsity E 445 664 2,926 0.010 IC 210 204 1,476 0.034 GPCR 223 95 635 0.030 NR 54 26 90 0.064 Kuang 786 809 3681 0.006 Hao 829 733 3688 0.006 Summary of four benchmark and two expanded datasets For targeted analysis and prediction, we ensured that each drug contained at least one FDA-approved ATC code in the dataset.

Several computational-based methods, including similarity-based methods, standard machine learning methods, and MF-based methods, have been developed in recent years to achieve efficient and accurate DTI prediction. A recent study by Shi et al. [ 48 ] revealed that MF-based methods had the best prediction accuracy. Existing MF-based methods, however, might easily fall into bad local minima due to noise and missing data, as well as the nonconvex pattern of MF models. Meanwhile, the lack of prior information made it challenging for the model to accurately predict more potential associations. Therefore, we proposed a DTI prediction model based on an SPL strategy and incorporated more similarity information. The novelty of SPLDMF might be attributed to a combination of several factors. First, introducing the SPL strategy enabled the model to avoid falling into a bad local optimum solution and thus had stronger robustness. The proposed SPLDMF had better prediction performance when the data were affected by noise. Moreover, we employed more prior similarity information to improve the feature extraction capability of the model, thus enabling the model to observe more potential DTIs accurately. Extensive experiments on synthetic data and four benchmark datasets were performed to assess the validity of the proposed SPLDMF method, which was then compared with three state-of-the-art DTI prediction methods. Two extended datasets were also used to verify the validity of each method. Comprehensive analysis results demonstrated that our proposed SPLDMF outperformed other state-of-the-art approaches. SPLDMF, for example, was more robust for noisy and missing data based on synthetic data. Furthermore, it outperformed all four scenarios and two expanded datasets in terms of common machine learning evaluation matrices. The prediction results revealed that 9 of the top 10 DTI pairs were found in the database and literature, and they were proven or considered effective. An unproven DTI pair (DB00717- P03372 ) was also preliminarily proven using pathway enrichment experiments. These results suggested that SPLDMF might provide a useful tool for predicting new DTIs and redirecting the use of existing drugs.

Predicting drug-target interaction (DTI) is a crucial phase in drug discovery (DD) [ 1 ] and drug repositioning (DR) [ 2 ] for discovering novel targets of existing drugs [ 3 , 4 , 5 ]. The traditional methods for new DD are time-consuming and have a high failure rate; therefore, traditional new drug development is not a good choice [ 3 , 6 ]. Various computer prediction methods have been proposed in recent years to improve the efficiency of new drug research and discovery, thus increasing the development efficiency and reducing expenditure to a certain extent. According to previous works [ 7 , 8 , 9 ], the current methods are mainly categorized into three groups [ 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 ]: (1) molecular docking (MD) methods, (2) ligand-based methods, and (3) chemical genomics methods. The MD methods involve simulation experiments based on the 3D structure drug and protein [ 11 , 18 ]. However, the simulation of the 3D structure of massive ligands and targets, as well as their massive simulation calculation using MD-based methods, requires a lot of time and computing equipment [ 19 , 20 ]. The ligand-based methods assume that drugs with similar functions have similar functional properties and may also have corresponding targets. They predict the drug target using ligand similarity. However, this approach suffers from unpredictable targets without known ligands. On the contrary, errors in chemical structure and physiological effects beyond structural relationships (e.g., the metabolites may be active molecules) may limit its use in drug repurposing. The chemical method facilitates rapid and large-scale DTI predictions to generate drug candidates and targets, making it the most efficient method in drug research [ 21 , 22 ]. Adopting this method for DTI prediction has become a prominent research issue with the continuous increase in drug-related data and the launch of a large number of databases, such as DrugBank [ 23 ], KEGG [ 24 ], PubChem [ 25 ], BRENDA [ 26 ], and SuperTarget [ 27 ]. Recently, chemical genomics-based computational approaches for DTI prediction have advanced rapidly. They are mainly categorized into three groups: classification-based methods, network diffusion (network propagation), and matrix factorization (MF). The classification-based methods treat a DR prediction task as a binary classification task that whether has an association between drug and target. These methods are not yet proof with wet experimental. In 2008, Yamanishi et al. [ 28 ] established a bipartite network technique to predict DTIs for four target classes: G protein-coupled receptors, by combining chemical and genomic spaces (GPCRs), nuclear receptor (NR), ion channel (IC), and enzyme (E). Yamanishi’s dataset [ 28 ] is regarded as the gold standard by many researchers; several newly developed algorithms based on it have displayed better performance. Based on this benchmark dataset, Bleakley et al. [ 29 ] suggested a novel supervised inference method for predicting unknown DTIs based on benchmark datasets, namely, a kernel-based support vector machine (KN-SVM) model. In recent years, the MF methods are widely used in many DR prediction works, which combines two low-rank matrices to factorize the matrix. Liu et al. [ 30 ] proposed a neighborhood regularized logistic MF model. Hao et al. [ 31 ] designed a logistic MF based on a dual network (DNILMF) approach to predict DTIs. Yang et al. [ 32 ] performed the nonlinear MF technique and the negative sampling technique for DR prediction. SPLCMF, a collaborative MF method combined with self-paced learning (SPL), is an efficient DTI prediction method proposed by Xia et al. [ 33 ]. Yang et al. [ 34 ] developed an MF method based on multi-similarities bilinear MF for DR prediction. Ding et al. [ 35 ] developed a multiple kernel-based triple collaborative MF method to predict DTIs. Wang et al. [ 36 ] used a neighborhood regularized logistic MF method based on extracted features from a neural tangent kernel to predict DTIs. These previous studies showed the feasibility of MF used in DR prediction tasks, but it still had two challenges. (1) The model easily fell into a bad local optimal solution due to the high noise and high missing rate in the data. (2) Single similarity information makes the learning power of the model insufficient in terms of identifying the potential associations accurately. To cope with the aforementioned challenges, we propose a model named Self-Paced Learning with Dual similarity information and Matrix Factorization (SPLDMF), which combines the self-paced learning method into MF. Furthermore, more similarity information related to drugs and targets is integrated into the model to improve the prediction performance. First, many previous works demonstrate that SPL has the superiority of relieving the problem of bad local optimum, especially when data is sparse [ 37 , 38 ]. Inspired by the human learning process, the core idea of SPL is to automatically include more samples from simple to complex for training in a purely self-paced manner. Thus, we make improvement of MF based on the SPL mechanism to adapt for the data with high noise and high missing rate. Then, the SPLDMF method also incorporates more data into our model to improve its capacity for learning, which can predict the potential relationship more accurately. Experimental results on five benchmark datasets and two extend datasets demonstrate the effectiveness of our approach in predicting drug-target interactions. Our model obtains the best results on each dataset we tested, such as AUC and AUPR of SPLDMF achieve 0.982 and 0.815, outperforming state-of-the-art models among similar methods to our knowledge