Drug target association prediction method based on sequence attribute extraction and graph neural network

Abstract

The application discloses a drug-target correlation prediction method based on sequence attribute extraction and a graph neural network, and relates to the field of drug-target prediction.The method adopts a prediction model SaeGraphDTI, and the overall architecture comprises a sequence attribute extractor, a graph encoder and a graph decoder;initial sequences of drugs and targets are input into the sequence attribute extractor, and sequences with inconsistent lengths are extracted into attribute sequences with consistent lengths and position alignment;then, the feature sequences of the drugs and the targets are input into the graph encoder, node information is updated correspondingly to obtain feature expression based on network topology;finally, the graph decoder is input, the probability of existence of a specific drug-target edge is calculated by the graph decoder to make a prediction.The application combines the sequence attributes of drugs and targets and the topological attributes of the correlation network, and exhibits the effectiveness of feature extraction and update in a complex relationship network, and progress is achieved in most key indicators.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug-target interaction prediction, and particularly relates to a drug-target correlation prediction method based on sequence attribute extraction and graph neural network. BACKGROUND

[0002] Drugs can enhance or inhibit the expression of related targets through drug-target interaction (DTI), thereby affecting the progress of related diseases. Discovering potential DTI is a key step in drug research and development and drug repositioning, however, traditional DTI discovery is usually accompanied by a large number of experiments, which consumes a large amount of funds and time, and meanwhile, millions of drugs and potential targets also make it impractical.

[0003] DTI prediction based on machine learning is generally regarded as a binary classification problem, the drug-target combination with interaction, i.e. DTI, is regarded as a positive example, and the drug-target combination without interaction is regarded as a negative example. DTI prediction based on machine learning has low cost and high efficiency, can significantly narrow the search range of drugs and targets, and can greatly shorten the pre-development time required for drug research, so it has attracted widespread attention.

[0004] Classical machine learning methods have been widely used in the field of DTI prediction, such as random forest (RF) and support vector machine (SVM). In recent years, with the rapid development of deep learning, DTI prediction based on deep learning is also rapidly developing and has shown good results. Some people proposed DeepConvDTI, which uses multiple groups of convolution kernels of different sizes to perform convolution on protein sequences of different lengths to capture the local residues of general proteins and detect the binding sites of DTI. Some people proposed DeepEmbedding-DTI, which respectively encodes drugs and proteins through GNN and BiLSTM, and then inputs them into the transformer to learn the embedding vector of the protein sequence. Some people proposed GraphDTA, which represents drugs as molecular graphs, uses multiple GNNs to obtain drug features, uses multiple CNNs to obtain target features, and finally uses FCN for prediction. Some people proposed SSPGraphCPI, which combines drug two-dimensional molecular graphs and one-dimensional sequences to obtain high-dimensional features of drugs, and finally predicts DTI.

[0005] To capture the interaction between drugs and proteins, people no longer focus on individual drugs or targets, but begin to integrate the existing DTI correlation features into their respective methods. Someone proposed DTI TR, which applies self-attention mechanism to drug and protein feature extraction, and uses cross-attention to capture the correlation features of drug-protein combination for DTI prediction. Someone proposed MolTrans, which decomposes drugs and proteins into a set of explicit substructure sequences through a frequent continuous subsequence mining algorithm and uses Transformer for feature extraction. The extracted drug and target features are mapped into a two-dimensional binding graph, and 2D-CNN and FCN are used for final prediction. Someone proposed DeepAffinity, which uses a recurrent neural network-based autoencoder for drug and target feature learning, followed by an attention mechanism to learn the binding features between drugs and proteins, and finally inputs into a convolutional neural network to obtain the final output. Someone proposed HyperAttentionDTI, which uses CNN to extract drug and protein features, and uses HyperAttention mechanism to capture the interaction between drugs and targets. Someone proposed MCN-CPI, which requires information about the binding site between proteins and compounds in addition to using the molecular fingerprint of the drug and the molecular sequence of the protein as input. The binding site features are extracted using 3D-CNN in three-dimensional space as a supplement to the original features. Someone proposed MCANet, which uses a multi-head cross-attention mechanism based on shared weights to extract the interaction features between drugs and proteins, and uses PolyLoss loss function to alleviate the overfitting problem and class imbalance problem in the original dataset.

[0006] With the development of graph neural networks, people began to apply the topological relationship between drugs and targets in the entire biological information network to DTI prediction. Someone proposed GENNIUS, which is a fast DTI prediction model based on DTI network. By calculating the properties of drugs and targets as original features, combining topological networks, and using multi-layer SAGEConv for final prediction. Someone proposed SGCL-DTI, which combines the complex relationships between drugs, targets, diseases and side effects to construct node meta-paths and perform initial feature acquisition and update based on this graph. Then based on the existing topological graph and semantic graph, the features are updated again, the updated DTI is spliced and FCN is used for final prediction.

[0007] Although the above models have achieved good results, in these models, how to obtain key information from the basic expression sequence is still a very important and difficult work; in addition, the drug-target, drug-drug, target-target forms a complex relationship network, and the feature acquisition and processing of the drug and the target in the above model are often limited to itself or the DTI inside, and the topological information in the whole relationship network is ignored. SUMMARY

[0008] In view of the problems and deficiencies of the prior art described above, the problem to be solved by the present application is to propose a drug target association prediction method based on sequence attribute extraction and graph neural network, which can discover potential drug-target interaction relationships in the existing drug-target interaction network.

[0009] In order to solve the above problems, the technical scheme adopted by the present application is as follows:

[0010] The present application is a drug target association prediction method based on sequence attribute extraction and graph neural network, which is characterized by the following steps: the method uses a DTI prediction model SaeGraphDTI based on sequence attribute extraction and graph neural network, and the overall architecture of SaeGraphDTI is divided into three main parts: sequence attribute extractor, graph encoder and graph decoder.

[0011] Firstly, the initial sequence of the drug and the target is input into the sequence attribute extractor, and the related attributes of the drug sequence and the target sequence are extracted by applying the sequence attribute extractor; the drug and target feature acquisition method based on sequence attribute extraction extracts the sequences of different lengths into attribute sequences of consistent length and position alignment.

[0012] Subsequently, the feature sequence of the drug and the target is input into the graph encoder, and the graph encoder based on this network updates the node information correspondingly to obtain the feature expression based on the network topology; the graph neural network based on adaptive connection reasonably selects the edges by dynamically judging the connectivity between nodes.

[0013] Finally, the obtained node features and their position relationships are input into the graph decoder, and the graph decoder calculates the probability of the existence of a specific drug-target edge to predict the DTI.

[0014] The drug target association prediction method based on sequence attribute extraction and graph neural network according to the present application is further preferably: using Davis, E, GPCR and IC four public data sets to train and test SaeGraphDTI; the Davis data set contains the selectivity determination of kinase proteins and related inhibitors and their respective dissociation constants K dThe value, which ranges from 0.016 to 10,000, reflects the size of the affinity value of the drug molecule to the target protein; K d The smaller the value, the stronger the correlation, and K d Greater than or equal to 10,000 represents no correlation; the K d value in the Davis dataset is logarithmically transformed, and the calculation method is as follows:

[0015]

[0016] The threshold value is set to 5.0, and the drug-target combination less than 5.0 is regarded as a positive example, and the rest is regarded as a negative example to construct a binary classification dataset;

[0017] The proteins constituting the target are divided into three main parts of enzymes, G protein-coupled receptors and ion channels, wherein the enzyme E and the ion channel IC come from the KEGG BRITE database, the DrugBank database, the SuperTarget database and the BRENDA database, the G protein-coupled receptor GPCR comes from the GLASS database, all DTIs in the dataset are regarded as positive examples, and the rest is regarded as negative examples, and the data distribution of each dataset is as follows:

[0018]

[0019] The drug and the target are expressed in the form of a sequence, and the lengths of the drugs and the targets are unified respectively; the sequence attribute extractor based on one-dimensional convolution of a variable convolution kernel obtains the most significant output of the sequence under different convolution kernels, while maintaining the output order of each convolution kernel, so that the output of each sequence is a position corresponding attribute list; the output of the drug and the target after passing through the sequence attribute extractor is set to the same dimension, so as to be processed by the subsequent graph encoder and decoder.

[0020] The drug and the target are expressed in the form of a sequence, and the lengths of the drugs and the targets are unified respectively; the sequence attribute extractor based on one-dimensional convolution of a variable convolution kernel obtains the most significant output of the sequence under different convolution kernels, while maintaining the output order of each convolution kernel, so that the output of each sequence is a position corresponding attribute list; the output of the drug and the target after passing through the sequence attribute extractor is set to the same dimension, so as to be processed by the subsequent graph encoder and decoder.

[0021] Firstly, each element in the original drug sequence and the original target sequence is encoded, so that the original sequence can be represented as an encoded sequence, wherein the range of drug encoding is an integer from 1 to 32, and the range of target encoding is an integer from 1 to 22; subsequently, all drug or target encoding sequences are made consistent in length by adding 0 at the end or cutting the end sequence; finally, the fixed-length encoding sequence is input into the embedding layer to generate the corresponding drug embedding matrix and the target embedding matrix wherein, L d is the drug sequence length, L t is the target sequence length, Dim d is the drug embedding dimension, Dim t is the target embedding dimension.

[0022] The further preferred technical solution of the drug target association prediction method based on sequence attribute extraction and graph neural network is that the convolutional neural network CNN obtains local features by moving a fixed size convolution kernel on an image or a sequence, and the size of the convolution kernel determines the receptive field of the current convolution; in order to make the sequence attribute extractor have a wider initial receptive field, after encoding the sequence embedding, first use two CNNs with fixed size convolution kernels to extract local features from the sequence, and each CNN is connected with an average pooling layer and a Relu activation function layer; then, use multiple CNNs with different convolution kernel sizes but consistent output channel numbers for processing, in order to obtain the most obvious features of the current channel under the current convolution kernel size, the maximum convolution result is taken as the output here;

[0023] Next, the outputs of all convolution kernels are spliced into a one-dimensional tensor with a length of N kernel_size ×N out_channel , wherein N kernel_size is the number of convolution kernel sizes, and N out_channel is the number of channels;

[0024] Finally, it is input into the FC layer, and the output of each original sequence is a one-dimensional feature vector with a length of N out_channel .

[0025] The attribute extractor converts the outputs of different dimensions with inconsistent convolution kernel sizes into outputs of the same dimension, that is, a one-dimensional feature sequence with consistent length, which is taken as the final output result of the sequence attribute extractor; at the same time, by taking the maximum value under different convolution kernel size dimensions, the appropriate convolution kernel size at each position is obtained, so as to locate the key subsequence with the appropriate length most related to the result.

[0026] The drug target correlation prediction method based on sequence attribute extraction and graph neural network provided by the application further preferably has the following technical scheme: the original data set is represented in the form of a D-T network, wherein the drugs and the targets constitute all nodes, and the drug-target combinations with DTI constitute all edges; a D-D network and a T-T network are constructed by a calculation method based on topological similarity, and are spliced with the D-T network to construct a D-D-T-T network covering all drugs and targets, each node of the network is connected to the remaining nodes by jumping, and the distance between nodes is represented by the number of jumps between nodes; a graph encoder module based on the network uses a graph neural network method to propagate and update node information, and adds topological information to the nodes; a graph decoder accepts the node features coded by the encoder, calculates the probability of the existence of edges between specific nodes, and regards the drug-target combinations with higher probability as positive examples.

[0027] The drug target correlation prediction method based on sequence attribute extraction and graph neural network provided by the application further preferably has the following technical scheme: the approximate network construction method based on topological similarity is as follows:

[0028] All drug sets are represented as D={d1, d2, …, d m}, all target sets are represented as T={t1, t2, …, t n}, and all drug-target interaction combinations are represented as E, wherein m is the number of all drugs, and n is the number of all targets.

[0029] The drugs and the targets are regarded as nodes, and the interaction relationship between the two is regarded as an edge, so that the nodes and edges constitute a drug-target interaction network, and the corresponding interaction matrix is represented as A and is as follows:

[0030]

[0031] All drug-target interaction combinations participating in the construction of the approximate network are defined as B, and the number of elements of B is less than the number of all interaction combinations A.

[0032] The approximate matrix between different drugs is defined as DS, and the calculation method of DS is as follows:

[0033]

[0034] Wherein, DR is the Pearson similarity, and DC is the cosine similarity, and the calculation is as follows:

[0035]

[0036] After obtaining the drug approximate matrix DS, for any two drugs di and drug d j , if DS[d i , d j ]>a, it is considered that there is high similarity, that is, there is an edge between nodes d i and d j in the drug-drug approximate network;

[0037] Meanwhile, for the remaining unconnected nodes, the first n maximum approximate nodes are selected as the neighbor nodes and the corresponding edges are added; finally, the network obtained according to the above method is the D-D network;

[0038] The approximate matrix between different target points is defined as TS, and the calculation method of TS is as follows:

[0039]

[0040] Wherein, TR is the Pearson similarity, and TC is the cosine similarity, and the calculation is as follows:

[0041]

[0042] After obtaining the target point approximate matrix TS, for any two target points t i and target point t j , if TS[t i , t j ]>b, it is considered that there is high similarity, that is, there is an edge between nodes t i and t j in the target-target approximate network; meanwhile, for the remaining unconnected nodes, the first m maximum approximate nodes are selected as the neighbor nodes and the corresponding edges are added; finally, the network obtained according to the above method is the T-T network;

[0043] The D-D network, the T-T network and the existing D-T network are spliced to obtain the D-D-T-T network.

[0044] The drug target correlation prediction method based on sequence attribute extraction and graph neural network provided by the application further preferably has the technical scheme that the graph encoder is stacked by a plurality of GATConv layers, SAGEConv layers and FC layers, and based on the topological relationship in the D-T network, the corresponding node information is updated in a message passing manner;

[0045] The drug and target feature vectors processed by the sequence attribute extractor have the same dimension, the drug features and the target features are placed in the same graph neural network for processing; the D-D-T-T network is represented as: G=(V, E), wherein V is a node set, containing drug nodes and target nodes, E is an edge set, and each node vi There exists a corresponding feature vector h. i GATConv learns the importance of a node's neighboring nodes, uses the learned importance weights to perform a weighted sum, and then updates the node itself; for any node v i Its feature vector at layer l+1 Represented as:

[0046]

[0047] Where N(i) is the node v i neighboring nodes, W l Let α be the weight matrix of the l-th layer. ij For neighbor node v j For node v i The importance weight is calculated as follows:

[0048]

[0049] SAGEConv samples the neighboring nodes of a node and aggregates the sampled neighboring nodes through a sink function to update the node itself; for any node v i Its feature vector at layer l+1 Represented as:

[0050]

[0051] Where σ is the activation function, The aggregated features of the sampled neighbor nodes are calculated as follows:

[0052]

[0053] Where aggregate is an aggregation function;

[0054] Since the graph encoder updates based on local iteration, it depends on the existing network structure. Therefore, before each training round, it is determined whether each edge should be retained, i.e., the dynamic connection of graph nodes. The decision to retain an edge depends on its weight. All edge weights are between 0 and 1. If the weight of an edge is greater than 0.5, it is considered an important edge and is retained; otherwise, it is discarded. If an edge is determined to be retained, its two endpoints exchange information; otherwise, each endpoint retains only its own information.

[0055] During node updates, the graph encoder brings the original features from the previous layer into the new layer to enhance the preservation of the original features.

[0056] The graph decoder calculates the inner product of the feature vectors of the two end points on all edges according to the probability of the existence of the node, and takes the result as the initial value of the probability of the existence of the edge, and then performs a Sigmod processing on the probability values of all edges, and takes the processed result as the probability of the existence of each edge, that is, the probability of the existence of the corresponding drug node and target node on the edge.

[0057] Compared with the prior art, the method has the following beneficial effects:

[0058] The model used in the method combines the sequence attributes of drugs and targets and the topological attributes of the associated network, and exhibits effectiveness in feature extraction and updating in a complex relationship network.

[0059] The application is based on a sequence attribute extractor and a graph neural network, extracts drug sequence and target sequence information through the sequence attribute extractor, updates drug information and target information through the graph encoder, and calculates the probability of the existence of drug-target edges through the graph decoder to predict potential drug-target interactions. The model constructed by the application achieves the best results in most key indicators compared with other most advanced methods on four public data sets, and the good effect of the model in predicting potential drug-target interactions is verified through experiments. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is the overall architecture diagram of the SaeGraphDTI model;

[0061] Figure 2 is the architecture diagram of the sequence attribute extractor;

[0062] Figure 3 is the architecture diagram of the graph encoder. DETAILED DESCRIPTION

[0063] The specific technical solutions of the application are further described below to facilitate those skilled in the art to further understand the application without limiting its rights.

[0064] Example 1, drug-target association prediction method based on sequence attribute extraction and graph neural network

[0065] 1.1 Dataset

[0066] The SaeGraphDTI model is proposed by training and testing four public data sets, namely Davis, E, GPCR and IC. The Davis data set contains the selectivity assay of kinase proteins and related inhibitors and their respective dissociation constant (K d ) values, which range from 0.016 to 10,000, reflecting the size of the affinity value of the drug molecule to the target protein. K dThe smaller, the stronger the correlation, K d Equal to or greater than 10000 represents no correlation. Since K d The value distribution is too large, we will take the logarithmic transformation, the calculation method is as follows:

[0067]

[0068] Set its threshold to 5.0, and consider the drug-target combination less than 5.0 as positive examples, and the rest as negative examples, to construct a binary classification dataset. Dataset E, GPCR and IC come from the public literature of MCANet. It divides the proteins that constitute the target into three main parts: enzyme (Enzyme), G protein-coupled receptor (GPCR) and ion channel (Ion Channel), among which enzymes and ion channels come from KEGG BRITE database, DrugBank database, SuperTarget database and BRENDA database. G protein-coupled receptors come from GLASS database. All DTIs in the dataset are considered as positive examples, and the rest are considered as negative examples. The data distribution of each dataset is shown in Table 1.

[0069] Table 1. Data distribution of the dataset

[0070]

[0071] 1.2 Method

[0072] The overall architecture of the SaeGraphDTI model proposed in the present application is shown in Figure 1 The model can be divided into three main parts: sequence attribute extractor, graph encoder and graph decoder. First, the initial sequence of the drug and the target is input into the sequence attribute extractor (Attribute extractor) to obtain the feature sequence with uniform length and attribute alignment. Then, the feature sequence of the drug and the target is input into the graph encoder (Graph encoder) to obtain the feature expression based on topological information. Finally, it is input into the graph decoder (Graph decoder) to calculate the possibility of the existence of a specific drug-target edge to infer the potential DTI.

[0073] 1.2.1 Sequence attribute extractor

[0074] The drug and the target are expressed in the form of sequence, and the length of each is unified. The sequence attribute extractor based on one-dimensional convolution of variable convolution kernel can obtain the most significant output of the sequence under different convolution kernels, while maintaining the output order of each convolution kernel, so that the output of each sequence is a list of position corresponding attributes. The relevant structure is as follows: Figure 2The drug and target are set to the same dimension by the output of the sequence attribute extractor in our implementation, so that subsequent graph encoder and decoder processes can be performed.

[0075] 1.2.1.1 Embedding layer

[0076] A suitable expression can enable the model to obtain more accurate information, therefore, the SMILES (Simplified molecular input line entry system) string is used as the serialized expression of the drug, and the amino acid sequence is used as the serialized expression of the target protein.

[0077] Firstly, each element in the original drug sequence and the original target sequence is encoded, so that the original sequence can be represented as an encoded sequence, wherein the drug code ranges from 1 to 32, and the target code ranges from 1 to 22. Subsequently, all drug or target code sequences are made consistent in length by adding 0 at the end or cutting the end sequence. Finally, the fixed-length encoded sequence is input into the embedding layer to generate the corresponding drug embedding matrix and the target embedding matrix wherein L d is the length of the drug sequence, L t is the length of the target sequence, Dim d is the drug embedding dimension, and Dim t is the target embedding dimension.

[0078] 1.2.1.2 Variable convolution layer

[0079] CNN can obtain local features by moving a fixed-size convolution kernel on an image or a sequence, and the size of the convolution kernel determines the receptive field of the current convolution. In order to make the sequence attribute extractor have a wider initial receptive field, after the sequence embedding, firstly, two CNNs with fixed-size convolution kernels are used to extract local features from the sequence, and an average pooling layer and a Relu activation function layer are connected after each CNN. Subsequently, multiple CNNs with different convolution kernel sizes but consistent output channel numbers are used for processing, and the maximum convolution result is taken as the output here in order to obtain the most obvious features of the current channel under the current convolution kernel size. Then, the outputs of all convolution kernels are spliced into a one-dimensional tensor with a length of N kernel_size ×N out_channel , wherein N kernel_size is the number of convolution kernel sizes, and N out_channel is the number of channels. Finally, it is input into the FC layer, and the output of each original sequence is a one-dimensional feature vector with a length of N out_channel .

[0080] Based on the structure of the attribute extractor, different dimensions of the output with inconsistent convolution kernel size can be converted into the same dimension of the output, i.e., a one-dimensional feature sequence with consistent length, which is used as the final output result of the sequence attribute extractor. At the same time, by taking the maximum value under different convolution kernel size dimensions, the appropriate convolution kernel size at each position can be obtained, so as to locate the key sub-sequence with the appropriate length most related to the result.

[0081] 1.2.2 Graph encoder and graph decoder

[0082] The original data set is represented in the form of a D-T (Drug-Target) network, in which drugs and targets constitute all nodes, and drug-target combinations with DTI constitute all edges. By constructing D-D (Drug-Drug) and T-T (Target-Target) networks based on topological similarity calculation, and splicing them with the existing D-T network, a D-D-T-T (Drug-Drug-Target-Target) network covering all drugs and targets can be constructed, and each node of the network can be connected to the remaining nodes through jumping, and the distance between nodes is represented as the number of jumps between nodes.

[0083] The graph encoder module based on this network uses a graph neural network method to propagate and update node information, thereby adding topological information to the node. The graph decoder accepts the node features encoded by the encoder and calculates the probability of the existence of edges between specific nodes, regarding the drug-target combination with a higher probability as a positive example.

[0084] 1.2.2.1 Approximate network construction method based on topological similarity

[0085] The set of all drugs is represented as D = {d1, d2, …, d m}, the set of all targets is represented as T = {t1, t2, …, t n}, and the set of all drug-target interaction combinations is represented as E, where m is the number of all drugs and n is the number of all targets. If the drugs and targets are regarded as nodes and the interaction relationship between them is regarded as edges, these nodes and edges constitute a drug-target interaction network, and the corresponding interaction matrix is represented as A, as follows:

[0086]

[0087] Since the construction of drug-drug approximate network and target-target approximate network is based on the existing drug-target interaction combination, if all interaction combinations are involved in the calculation, it may lead to the model effect higher than the actual real effect. Therefore, the set of all drug-target interaction combinations involved in the construction of approximate network is defined as B, the number of elements of which is less than the number of all interaction combinations A.

[0088] The approximate matrix between different drugs is defined as DS, and the calculation method is as follows:

[0089]

[0090] Wherein, DR is Pearson similarity, DC is cosine similarity, and the calculation is as follows:

[0091]

[0092] After obtaining the drug approximate matrix DS, for any two drugs d i and drug d j , if DS[d i ,d j ]>α, it is considered that there is a higher similarity, that is, there is an edge between node d i and node d j in the drug-drug approximate network. At the same time, for the remaining unconnected nodes, the first n largest approximate nodes are selected as their neighbor nodes and the corresponding edges are added. Finally, the network obtained according to the above method is the D-D network.

[0093] Similarly, the approximate matrix between different targets is defined as TS, and the calculation method is as follows:

[0094]

[0095] Wherein, TR is Pearson similarity, TC is cosine similarity, and the calculation is as follows:

[0096]

[0097] After obtaining the target approximate matrix TS, for any two targets t i and target t j , if TS[t i ,t j ]>β, it is considered that there is a higher similarity, that is, there is an edge between node t i and node t jThere are edges between them. Meanwhile, for the remaining unconnected nodes, the top m nodes with the highest approximation are selected as their neighbors, and corresponding edges are added. Finally, the network obtained using the above method is the TT network.

[0098] By splicing the DD network, TT network and the existing DT network, the DDTT network is obtained.

[0099] 1.2.2.2 Graphic Encoder

[0100] The graph encoder consists of multiple stacked GATConv, SAGEConv, and FC layers. Based on the topology of the existing DT network, it updates the corresponding node information using a "message passing" method, such as... Figure 3 As shown in (A).

[0101] The drug and target feature vectors processed by the sequence attribute extractor have the same dimension, so they can be processed in the same graph neural network. The DDTT network can be represented as G = (V, E), where V is the set of nodes containing drug and target nodes, and E is the set of edges, with each node v... i There exists a corresponding feature vector h. i .

[0102] GATConv by The proposed approach involves learning the importance of a node's neighboring nodes, then using these learned importance weights to perform a weighted summation, thereby updating the node itself. For any node v... i Its feature vector at layer l+1 It can be represented as:

[0103]

[0104] Where N(i) is the node v i neighboring nodes, W l Let α be the weight matrix of the l-th layer. ij For neighbor node v j For node v i The importance weight is calculated as follows:

[0105]

[0106] SAGEConv is designed for large-scale graph structures. Its core idea is to sample the neighboring nodes of a node and aggregate these sampled neighboring nodes using a sink function, thereby updating the node itself. For any node v... i Its feature vector at layer l+1 It can be represented as:

[0107]

[0108] wherein σ is an activation function, aggregate features of sampled neighbor nodes, which is calculated as follows:

[0109]

[0110] wherein aggregate is an aggregation function, such as average aggregation, maximum aggregation, LSTM aggregation, etc.

[0111] Since the graph encoder is updated based on local iteration, it depends on the existing network structure, so it is necessary to determine whether each edge is retained before each round of training begins, i.e., the dynamic connection of the graph nodes. Whether a certain edge is retained depends on the weight of this edge. The weights of all edges are between 0 and 1. If the weight of a certain edge is greater than 0.5, it is considered to be an important edge and is retained, otherwise it is discarded. If a certain edge is determined to be retained, the two endpoints of the edge exchange information, otherwise each endpoint only retains its own information, as shown in Figure 3 (B).

[0112] At the same time, the traditional graph neural network will cause the difference between node features to decrease with the increase of network layers, i.e., the problem of over-smoothing. Therefore, in this model, the original features of the previous layer are brought into the new layer to enhance the preservation effect of the original features. The related update method is shown in Figure 3 (C).

[0113] 1.2.2.3 Graph decoder

[0114] The graph decoder can calculate the probability of the existence of a specific edge according to the node. Specifically, it calculates the inner product of the feature vectors of the two endpoints of all edges, and takes this result as the initial value of the probability of this edge. Then, the probability values of all edges are processed by Sigmod, and the processed results are taken as the probability of the existence of each edge, i.e., the probability of the interaction between the drug node and the target node on the edge.

[0115] 2. Experiment

[0116] 2.1 Experimental setup

[0117] 2.1.1 Hyperparameter setting

[0118] Appropriate hyperparameters, such as convolution kernel size, edge threshold, learning rate, and training rounds, can improve the model's ability to acquire key features, improve model stability, and improve model performance.

[0119] In the sequence expression of drugs and targets, the maximum length of the drug is set to 128, and the maximum length of the target is set to 2000. For the part with insufficient length, it is filled with the number 0.

[0120] In the sequence attribute extractor, since the length of the drug sequence is much smaller than that of the target sequence, we set different hyperparameters for drugs and targets: the convolution kernel of the drug sequence attribute extractor is set to 3, 6, 9, and 12, while the convolution kernel of the target Seqfilter is set to 3, 6, 9, 12, 15, 18, 21, 24, 27, and 30.

[0121] In the approximate network construction, the drug approximate edge takes the top 10% of the total drug approximation value. For the unconnected drug nodes, it is connected to at most one similar drug node. The target approximate edge takes the top 10% of the total target approximation value. For the unconnected target nodes, it is connected to at most one similar target node.

[0122] In the graph encoder, the number of layers of the graph neural network part is set to 6 layers, and the dropout of all nodes is set to 0.2.

[0123] In the training process, the model is optimized using the Adam optimizer, with the initial learning rate set to 0.001, the training rounds set to 10000 rounds, and the learning rate decay set to 0.8 every 1000 rounds. However, the iteration time on each dataset does not exceed 15 minutes.

[0124] 2.1.1 Benchmark model

[0125] In order to test SaeGraphDTI, we selected the classical machine learning models RF and SVM and the widely cited DTI prediction model in recent years for comparison testing. All benchmark model parameters are set to the corresponding default settings.

[0126] MolTrans: Apply Transformer to drug and protein feature embedding, and create interaction matrix based on respective embedding sequence. Use multiple two-dimensional convolution and fully connected layer for prediction.

[0127] HpyerAttentionDTI: Use multiple one-dimensional convolution for feature extraction of drugs and proteins, then use HpyerAttention mechanism to obtain the binding features between the two, and finally use multiple fully connected layers for prediction.

[0128] MCANet: Use multi-head cross-attention mechanism based on shared weights to extract interaction features between drugs and proteins, and use PolyLoss loss function to alleviate the overfitting problem and class imbalance problem in the original dataset.

[0129] 2.1.2 Experimental Environment and Training Settings

[0130] The model is implemented based on Pytorch and Pytorch_geometric and trained on the following hardware configuration:

[0131] • CPU: Intel i5-12400F, 2.50GHz

[0132] • GPU: NVIDIA RTX2060Super, 8GB

[0133] • Memory: 32GB DDR4

[0134] Five-fold cross-validation is adopted in the experiment. All DTIs are divided into five equal parts, and each time four parts are selected for network construction and model training, and the remaining one is only used for testing. The best result is taken for each test, and the average value of each test result is taken as the final result of the experiment.

[0135] 2.2 Results and Analysis

[0136] 2.2.1 Evaluation Indicators

[0137] This model is a binary classification model. In order to evaluate the classification performance of the model, the following evaluation indicators are used. Among them, TP, FP, TN and FN represent the number of true positive, false positive, true negative and false negative samples, respectively.

[0138] Accuracy: The proportion of correctly predicted samples to all samples, calculated as follows:

[0139]

[0140] Precision: The proportion of actual positive examples to predicted positive examples, calculated as follows:

[0141]

[0142] Recall: The proportion of predicted positive examples to actual positive examples, calculated as follows:

[0143]

[0144] F1-Score, the harmonic mean of Precision and Recall, calculated as follows:

[0145]

[0146] AUROC (Area Under the Receiver Operating Characteristic Curve): the area under the ROC curve, in the binary classification problem, the closer the AUROC value is to 1, the better the model is in predicting positive examples.

[0147] AUPRC (Area Under the Precision-Recall Curve): the area under the precision-recall curve, suitable for imbalanced data. The closer the AUPRC value is to 1, the better the model can maintain high precision and recall at various thresholds.

[0148] 2.2.2 Effect analysis

[0149] First, the SaeGraphDTI model of the present application and the benchmark model are trained and tested on the E, GPCR, and IC three data sets. The three data sets are balanced data sets, the number of positive examples is close to or equal to the number of negative examples, but the ratio of the number of drugs to the number of targets in each data set is not consistent. Experiments show that the SaeGraphDTI model of the present application can obtain good results on each data set.

[0150] On the data set E, compared with the best experimental indicators of all models, the SaeGraphDTI model of the present application achieved 13.7% accuracy, 13.5% precision, 10.4% F1 score, 13.4% AUC, and 16.8% AUPR improvement, but the recall rate decreased by 1%. Overall, the SaeGraphDTI model basically exceeds all benchmark models on this data set, and the relevant data is shown in Table 2.

[0151] Table 2. Experimental results of SaeGraphDTI and benchmark models on data set E

[0152]

[0153]

[0154] The model in this embodiment is from:

[0155] RF model: Breiman L. Random forests [J]. Machine learning, 2001, 45: 5-32.

[0156] SVM model: Burges C J C. A tutorial on support vector machines for pattern recognition [J]. Data mining and knowledge discovery, 1998, 2(2): 121-167.

[0157] MolTrans model: Huang K, Xiao C, Glass L M, et al. MolTrans: molecular interaction transformer for drug-target interaction prediction [J]. Bioinformatics, 2021, 37(6): 830-836.

[0158] HpyerAttentionDTI model: Zhao Q, Zhao H, Zheng K, et al. HyperAttentionDTI: improving drug-protein interaction prediction by sequence-based deep learning with attention mechanism [J]. Bioinformatics, 2022, 38(3): 655-662.

[0159] MCANet model: Bian J, Zhang X, Zhang X, et al. MCANet: shared-weight-based Multihead Cross Attention network for drug-target interaction prediction [J]. Briefings in Bioinformatics, 2023, 24(2): bbad082.

[0160] On the dataset GPCR, compared with the best experimental indicators of all models, the SaeGraphDTI model of the application achieved 0.3% accuracy, 0.4% precision, 0.2% F1 score, 1.3% AUC, 0.2% AUPR improvement, but a 3.9% decrease in recall rate. Overall, the SaeGraphDTI model basically outperformed all benchmark models on this dataset, and the relevant data is shown in Table 3.

[0161] Table 3. Experimental results of SaeGraphDTI and benchmark models on dataset GPCR

[0162]

[0163]

[0164] On the dataset IC, compared with the best experimental indicators of all benchmark models, the SaeGraphDTI model achieved 9.9% accuracy, 7.6% precision, 10.7% recall, 9.5% F1 score, 8% AUC, and 8.9% AUPR improvement. Overall, the SaeGraphDTI model outperformed all benchmark models on the dataset E, and the relevant data is shown in Table 4.

[0165] Table 4. Experimental results of SaeGraphDTI and benchmark models on dataset IC

[0166]

[0167] Subsequently, the SaeGraphDTI model and the benchmark models were trained and tested on the Davis dataset, which is more challenging than the other three datasets because it is an imbalanced dataset, in which the ratio of positive examples to negative examples is about 1:3. Compared with the best experimental indicators of all benchmark models, the SaeGraphDTI model achieved an accuracy of 0.7%, a precision of 2.7%, an F1 score of 0.6%, an AUC of 0.2%, an AUPR of 1%, but a recall of 3.3%. Overall, the SaeGraphDTI model basically outperformed all benchmark models on the Davis dataset, and the relevant data is shown in Table 5.

[0168] Table 5. Experimental results of SaeGraphDTI and benchmark models on the Davis dataset

[0169]

[0170] When the dataset is represented as a graph structure, the correlation value of the node degree of each dataset can be calculated, and it is found that the average node degree of the three datasets E, GPCR, and IC is between 4.4 and 10.5, but in the Davis dataset, each drug is connected to an average of 107.6 target points, and each target point is connected to an average of 19.3 drugs, which is significantly higher than other datasets. Excessive connection density can cause relevant information to spread throughout the graph, thereby reducing the distinction between each drug or target point, so the relevant indicators of the SaeGraphDTI model and other benchmark models on the Davis dataset are slightly lower than those of the E dataset and the IC dataset. At the same time, the number of drugs and target points in the GPCR dataset and the Davis dataset has a large gap, which also affects the balance of the model's judgment of various classes. SaeGraphDTI limits the propagation range of node information while obtaining topological information through the graph encoder and node dynamic connection, so it achieves better results compared to the benchmark models.

[0171] In order to verify the inference ability of SaeGraphDTI on potential DTI, part of the original DTI was deleted, and the remaining data was used to predict the deleted DTI, which was used as the result of this case analysis, and the verification process is as follows:

[0172] First, two target points, target point YANK1 and target point PAK2, were randomly selected from the Davis dataset, and 10 known relationships (5 associated drugs and 5 non-associated drugs) of each target point were randomly deleted, thereby obtaining the drug-target set to be tested and the corresponding reduced dataset.

[0173] Then, based on the reduced dataset, the similarity of nodes of the same type is recalculated, and then the association network is constructed. All parameters remain in the default configuration.

[0174] Then, the constructed association network is input into the model, and the corresponding training and testing are performed until the model converges stably. The training and testing data are from the reduced dataset, and all parameters remain in the default configuration.

[0175] Finally, the index of the drug-target combination to be tested is input into the model to obtain the probability of the existence of the association of the combination. The probability value is between 0 and 1, and the larger the value, the higher the association. Values greater than 0.5 are considered to be associated, i.e., there is a DTI, and values less than 0.5 are considered to be unrelated, i.e., there is no DTI. In this experiment, the prediction results of the target YANK1 and the target PAK2 and the related test drugs are shown in Table 6.

[0176] Table 6. Prediction results of SaeGraphDTI for target YANK1 and target PAK2

[0177]

[0178] This experiment shows that the SaeGraphDTI model has achieved 90% accuracy in identifying the potential DTI of the two randomly selected target YANK1 and PAK2, verifying the good effect of the model in predicting potential DTI.

[0179] 3. Conclusion

[0180] Experiments show that SaeGraphDTI is superior to the most advanced model in most indicators under multiple datasets.

Claims

1. A method for drug target association prediction based on sequence attribute extraction and graph neural network, characterized in that, The method adopts a DTI prediction model SaeGraphDTI based on sequence attribute extraction and graph neural network, and the overall architecture of SaeGraphDTI is divided into three parts: a sequence attribute extractor, a graph encoder, and a graph decoder; First, the initial sequences of drugs and targets are input into the sequence attribute extractor, and the relevant attributes of the drug sequences and target sequences are extracted by applying the sequence attribute extractor; The sequence attribute extraction-based drug and target feature acquisition method extracts sequences of different lengths into attribute sequences of consistent length and position alignment; Subsequently, the feature sequences of the drugs and targets are input into the graph encoder, and the graph encoder updates the node information correspondingly to obtain feature expressions based on network topology; The graph neural network based on adaptive connection dynamically determines the connectivity between nodes and reasonably selects edges; Finally, the obtained node features and their positional relationships are input into the graph decoder; the graph decoder calculates the probability of existence of a specific drug-target edge to make a prediction of DTI; The original data set is represented in the form of a D-T network, in which drugs and targets constitute all nodes, and drug-target combinations with DTI constitute all edges; a D-D network and a T-T network are constructed based on a topological similarity calculation method, and are spliced with the D-T network to construct a D-D-T-T network covering all drugs and targets; each node of the network is connected to the remaining nodes through jumping, and the distance between nodes is represented as the number of jumps between nodes; the graph encoder module based on this network uses a graph neural network method to propagate and update node information, and adds topological information to the nodes; the graph decoder accepts the node features encoded by the encoder, and calculates the probability of existence of an edge between specific nodes, regarding drug-target combinations with higher probability as positive examples; The approximate network construction method based on topological similarity is as follows: Let the set of all drugs be represented as D = {d1, d2,..., d m}, the set of all targets be represented as T = {t1, t2,..., t n}, and the set of all drug-target interaction combinations be represented as E, where m is the number of all drugs and n is the number of all targets. Regarding drugs and targets as nodes and the interaction relationship between them as edges, these nodes and edges constitute a drug-target interaction network, and the corresponding interaction matrix is represented as A, as follows: The set of all drug-target interaction combinations of the approximate network based on topological similarity is defined as B, and the number of elements is less than the number of all interaction combinations A; The approximate matrix between different drugs is defined as DS, and the calculation method is as follows: Wherein, DR is the Pearson similarity, and DC is the cosine similarity, and the calculation is as follows: After obtaining the drug similarity matrix DS, for any two drugs d i and d j , if DS[d i ,d j ]>a, it is considered that there is a higher similarity, i.e. there is an edge between node d i and node d j in the drug-drug similarity network; meanwhile, for the remaining unconnected nodes, the first n largest similarity nodes are selected as their neighbor nodes and the corresponding edges are added; finally, the network obtained according to the above method is the D-D network. The approximate matrix between different targets is defined as TS, and the calculation method is as follows: Wherein, TR is the Pearson similarity, and TC is the cosine similarity, and the calculation is as follows: After obtaining the target point approximate matrix TS, for any two target points t i and t j , if TS[t i ,t j ]>β, it is considered that there is a higher similarity, that is, there is an edge between the nodes t i and t j in the target point-target point approximate network; meanwhile, for the remaining unconnected nodes, the first m largest approximate nodes are selected as the neighbor nodes and the corresponding edges are added; finally, the network obtained according to the above method is the T-T network; The D-D network, the T-T network, and the existing D-T network are spliced to obtain the D-D-T-T network.

2. The method of claim 1, wherein: SaeGraphDTI is trained and tested using four published datasets: Davis, E, GPCR, and IC; the Davis dataset contains the selectivity assays of kinase proteins and their respective dissociation constants K d values, which range from 0.016 to 10,000, reflect the magnitude of the drug molecule's affinity for the target protein; K d values, which range from 0.016 to 10,000, reflect the magnitude of the drug molecule's affinity for the target protein; K d values, which range from 0.016 to 10,000, reflect the magnitude of the drug molecule's affinity for the target protein; K d The K d values in the Davis dataset are log-transformed, calculated as follows: The threshold is set to 5.0, and drug-target combinations less than 5.0 are regarded as positive examples, and the rest are regarded as negative examples to construct a binary classification data set; The proteins constituting the target points are divided into three parts of enzymes, G protein-coupled receptors and ion channels, wherein the enzymes E and the ion channels IC are from the KEGG BRITE database, the DrugBank database, the SuperTarget database and the BRENDA database, the G protein-coupled receptor GPCR is from the GLASS database, all DTIs in the data set are regarded as positive examples, and the rest are regarded as negative examples, and the data distribution of each data set is as follows:

3. The method of claim 1, wherein: The drugs and target points are expressed in the form of sequences, and the lengths of the drugs and target points are unified respectively; the sequence attribute extractor based on one-dimensional convolution of a variable convolution kernel obtains the most significant output of the sequence under different convolution kernel conditions, while maintaining the output order of each convolution kernel, so that the output of each sequence is a position corresponding attribute list; The output of the drug and the target point after the sequence attribute extractor is set to the same dimension, so as to be processed by the subsequent graph encoder and decoder.

4. The method according to claim 1 or 3, characterized in that: The SMILES string is used as the serialized expression of the drug, and the amino acid sequence is used as the serialized expression of the target protein; First, each element in the original drug sequence and the original target sequence is encoded, so that the original sequence can be represented as an encoded sequence, wherein the drug code ranges from 1 to 32, and the target code ranges from 1 to 22; then, all drug or target code sequences are made consistent in length by adding 0 at the end or cutting the end sequence; finally, the fixed-length encoded sequence is input into the embedding layer to generate the corresponding drug embedding matrix and the target embedding matrix wherein L d is the length of the drug sequence, L t is the length of the target sequence, Dim d is the drug embedding dimension, and Dim t is the target embedding dimension.

5. The method of claim 1 or 3, wherein: The convolutional neural network CNN obtains local features by moving a fixed size convolution kernel on an image or a sequence, and the size of the convolution kernel determines the receptive field of the current convolution; in order to make the sequence attribute extractor have a wider initial receptive field, after the encoding sequence embedding, firstly, two CNNs with fixed size convolution kernels are used to extract local features from the sequence, and each CNN is connected with an average pooling layer and a Relu activation function layer; then, a plurality of CNNs with different convolution kernel sizes but consistent output channel numbers are used for processing; in order to obtain the most obvious features of the current channel under the current convolution kernel size, the maximum convolution result is taken as the output here; Next, the outputs of all the convolution kernels are spliced into a one-dimensional tensor of length N kernel_size x N out_channel , where N kernel_size is the number of dimensions of the convolution kernel, and N out_channel is the number of channels; Finally, it is input to the FC layer, and the output of each original sequence is a one-dimensional feature vector with length N out_channel . The sequence attribute extractor converts the outputs of different dimensions with different convolution kernel sizes into outputs of the same dimension, that is, a one-dimensional feature sequence with consistent length, which is taken as the final output result of the sequence attribute extractor; at the same time, by taking the maximum value under different convolution kernel size dimensions, the appropriate convolution kernel size at each position is obtained, so as to locate the key subsequence with the appropriate length most related to the result.

6. The method of claim 1, wherein: The graph encoder is stacked by a plurality of GATConv layers, SAGEConv layers and FC layers, and based on the topological relationship in the D-T network, the corresponding node information is updated in a "message passing" manner; The drug and target feature vectors processed by the sequence attribute extractor have the same dimension, and the drug features and target features are placed in the same graph neural network for processing; the D-D-T-T network is represented as: G=(V, E), where V is a node set containing drug nodes and target nodes, and E is an edge set i There is a corresponding feature vector h i ; GATConv learns the importance of the neighbor nodes of the node, and updates the node itself by weighted summation using the learned importance weight; for any node v i , the feature vector of the (l+1)th layer is represented as: where N(i) is the neighbor node of node v i , W l is the weight matrix of the lth layer, and a ij is the importance weight of the neighbor node v j to node v i , which is calculated as follows: SAGEConv samples the neighbor nodes of a node and updates the node itself by aggregating the sampled neighbor nodes through a pooling function; for any node v i a feature vector at the (l+1)th layer is expressed as: where σ is an activation function, is the aggregated features of the sampled neighbor nodes, which is calculated as follows: Wherein, aggregate is an aggregation function; Since the graph encoder is updated based on local iteration, it depends on the existing network structure, so whether each edge is retained is judged before each round of training starts, that is, the dynamic connection of the graph node; whether an edge is retained depends on the weight of the edge, and the weight of all edges is between 0 and 1; if the weight of an edge is greater than 0.5, the edge is considered as an important edge and is retained, otherwise the edge is discarded; if an edge is judged to be retained, the two endpoints of the edge exchange information, otherwise each endpoint only retains its own information; In the node updating process of the graph encoder, the original features of the previous layer are brought into the new layer to enhance the retention effect of the original features. The graph decoder calculates the probability of existence of a specific edge according to the node, calculates the inner product of the feature vectors of the two end points on all edges, and takes the result as the initial value of the probability of the edge, and then performs a Sigmod process on the probability values of all edges, and takes the processed result as the probability of existence of each edge, i.e. the probability of interaction between the corresponding drug node and target node on the edge.

Images