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

Abstract

The invention 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. According to the method, a prediction model SaeGraphDTI is adopted, and the overall architecture comprises a sequence attribute extractor, a graph encoder and a graph decoder; inputting the initial sequences of the drugs and the target spots into a sequence attribute extractor, and extracting the sequences with inconsistent lengths into attribute sequences with consistent lengths and aligned positions; inputting the feature sequences of the drugs and the target spots into a graph encoder, and correspondingly updating node information to obtain feature expression based on a network topology structure; and finally, inputting into a graph decoder, and calculating the probability of existence of a specific drug-target edge through the graph decoder to make a prediction. According to the method, the sequence attributes of the drugs and the targets and the topological attributes of the associated network are combined, the effectiveness of feature extraction and updating in the complex relation network is shown, and progress is made in most key indexes.

Description

Technical Field

[0001] The present invention relates to the technical field of drug-target interaction prediction, and in particular to a drug-target association prediction method based on sequence attribute extraction and graph neural network. Background Art

[0002] Drugs can enhance or inhibit the expression of related targets through drug-target interaction, thereby affecting the progression of related diseases. Discovering potential DTIs is a key step in drug development and drug repurposing. However, traditional DTI discovery is usually accompanied by a large number of experiments, which requires a lot of money and time. At the same time, the millions of drugs and potential targets make it impractical.

[0003] DTI prediction based on machine learning is generally regarded as a binary classification problem. Drug-target combinations with interactions, namely DTI, are regarded as positive examples, and drug-target combinations without interactions are regarded as negative examples. DTI prediction based on machine learning has low cost and high efficiency, which can significantly narrow the search scope of drugs and targets and greatly shorten the initial time required for drug development. Therefore, it has attracted widespread attention.

[0004] Classic 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 depth and learning has also developed rapidly and has shown good results. Some people have proposed DeepConvDTI, which uses multiple groups of convolution kernels of different sizes to perform convolutions on protein sequences of different lengths to capture local residues of generalized proteins and detect binding sites of DTI. Some people have proposed DeepEmbedding-DTI, which encodes drugs and proteins through GNN and BiLSTM respectively, and then inputs them into transformer to learn the embedding vector of protein sequences. Some people have 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 have proposed SSGraphCPI, which combines the two-dimensional molecular graph of drugs with one-dimensional sequences to obtain high-dimensional features of drugs and finally predict DTI.

[0005] In order to capture the interaction between drugs and proteins, people no longer focus on individual drugs or targets, but begin to incorporate the existing DTI association features into their respective methods. Some people proposed DTITR, which applies the self-attention mechanism to drug feature extraction and protein feature extraction, and uses cross attention to capture the association features of drug-protein combinations to predict DTI. Some people 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, maps the extracted drug and target features into a two-dimensional binding graph, and uses 2D-CNN and FCN for final prediction. Some people proposed DeepAffinity, which uses a recurrent neural network-based autoencoder for drug and target feature learning, and then uses the attention mechanism to learn the binding features between drugs and proteins, and finally inputs them into a convolutional neural network to obtain the final output. Some people proposed HyperAttentionDTI, which uses CNN to extract drug and protein features and uses the HyperAttention mechanism to capture the interaction between drugs and targets. Someone proposed MCN-CPI, which, in addition to using the molecular fingerprint of the drug and the protein molecular sequence as input, also requires information about the binding site between the protein and the compound, and uses 3D-CNN in three-dimensional space to extract the binding site features 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 the PolyLoss loss function to alleviate the overfitting and class imbalance problems in the original dataset.

[0006] With the development of graph neural networks, people began to apply the topological relationship of drugs and targets in the entire bioinformatics network to DTI prediction. Someone proposed GENNIUS, a fast DTI prediction model based on the DTI network, which calculates the various properties of drugs and targets as original features, combines the topological network, and uses multi-layer SAGEConv for final prediction. Someone proposed SGCL-DTI, which combines the complex relationship between drugs, targets, diseases and side effects, constructs node meta-paths and performs initial feature acquisition and update based on this graph, then updates features again based on the existing topological graph and semantic graph, splices the updated DTI and uses FCN for final prediction.

[0007] Although the above models have achieved good results, in these models, how to obtain key information from basic expression sequences is still a very important and difficult task; in addition, drug-target, drug-drug, target-target constitute a complex relationship network, and the feature acquisition and processing of drugs and targets in the above models are often limited to themselves or the DTI in which they are located, while ignoring their topological information in the entire relationship network. Summary of the invention

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

[0009] In order to solve the above problems, the present invention adopts the following technical solutions:

[0010] The present invention 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 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 main parts: a sequence attribute extractor, a graph encoder and a graph decoder;

[0011] First, the initial sequences of drugs and targets are input into the sequence attribute extractor, and the relevant attributes of drug sequences and target sequences are extracted by applying the sequence attribute extractor; based on the drug and target feature acquisition method extracted from sequence attributes, sequences with inconsistent lengths are extracted into attribute sequences with consistent lengths and aligned positions;

[0012] Subsequently, the feature sequences of drugs and targets are input into the graph encoder. The graph encoder based on this network updates the node information accordingly to obtain feature expressions based on the network topology structure. The graph neural network based on adaptive connections dynamically judges the connectivity between nodes and makes reasonable choices for edges.

[0013] Finally, the obtained node features and their positional relationships are input into the graph decoder, and the probability of the existence of a specific drug-target edge is calculated by the graph decoder to make a prediction on DTI.

[0014] The drug target association prediction method based on sequence attribute extraction and graph neural network described in the present invention has a further preferred technical solution: using four public data sets of Davis, E, GPCR and IC to train and test SaeGraphDTI; the Davis data set contains the selective determination of kinase proteins and related inhibitors and their respective dissociation constants K dThe value ranges from 0.016 to 10,000, which reflects the affinity of the drug molecule to the target protein; K d The smaller the K, the stronger the correlation. d If it is greater than or equal to 10,000, it means there is no correlation; d The values ​​are logarithmically transformed and calculated as follows:

[0015]

[0016] The threshold is set to 5.0, and drug-target combinations with a value less than 5.0 are considered positive examples, and the rest are considered negative examples to construct a binary classification dataset;

[0017] The proteins that constitute the target are divided into three main parts: enzymes, G protein-coupled receptors and ion channels. The enzyme E and ion channel IC are from the KEGG BRITE database, DrugBank database, SuperTarget database and BRENDA database, and 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. The data distribution of each data set is as follows:

[0018]

[0019] The drug-target association prediction method based on sequence attribute extraction and graph neural network described in the present invention has a further preferred technical solution: the drugs and targets are expressed in the form of sequences, and their respective lengths are unified; a sequence attribute extractor based on one-dimensional convolution with 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 list of attributes corresponding to the position; the outputs of the drugs and targets through the sequence attribute extractor are set to the same dimension for subsequent graph encoder and decoder processing.

[0020] The drug-target association prediction method based on sequence attribute extraction and graph neural network described in the present invention has a further preferred technical solution of: using a SMILES string as a serialized expression of the drug, and using an amino acid sequence as a serialized expression of the target protein;

[0021] First, encode each element in the original drug sequence and the original target sequence so that the original sequence can be represented as a coding sequence, where the drug coding range is an integer from 1 to 32, and the target coding range is an integer from 1 to 22; then, by adding 0 at the end or trimming the terminal sequence, the length of all drug or target coding sequences is made consistent; finally, the fixed-length coding sequence is input into the embedding layer to generate the corresponding drug embedding matrix Target embedding matrix Among them, L d is the drug sequence length, L t is the target sequence length, Dim d is the drug embedding dimension, Dim t Embedding dimension for the target.

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

[0023] Next, concatenate the outputs of all convolution kernels into a network of length N kernel_size ×N out_channel A one-dimensional tensor, where N kernel_size is the number of convolution kernel sizes, 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 sequence of length N. out_channel The one-dimensional feature vector of

[0025] The attribute extractor converts the outputs of different dimensions with inconsistent numbers of convolution kernel sizes into outputs of the same dimension, that is, outputs a one-dimensional feature sequence with consistent output 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 is obtained, so as to locate the key subsequence of suitable length that is most relevant to this result.

[0026] The drug-target association prediction method based on sequence attribute extraction and graph neural network described in the present invention has a further preferred technical solution: the original data set is represented in the form of a DT network, in which drugs and targets constitute all nodes, and drug-target combinations with DTI constitute all edges; a DD network and a TT network are constructed by a calculation method based on topological similarity, and they are spliced ​​with the DT network to construct a DDTT 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 this network uses a graph neural network method to propagate and update node information, and add topological information to the inside of the node; a graph decoder accepts node features encoded by the encoder, and calculates the probability of edges existing between specific nodes, and regards drug-target combinations with higher probabilities as positive examples.

[0027] The drug target association prediction method based on sequence attribute extraction and graph neural network described in the present invention has a further preferred technical solution: the approximate network construction method based on topological similarity is as follows:

[0028] The set of all drugs is represented as D = {d1, d2, ..., d m}, the set of all targets is represented by T = {t1, t2, …, t n}, the set of all drug-target interaction combinations is denoted as E, where m is the number of all drugs and n is the number of all targets;

[0029] If drugs and targets are regarded as nodes and the interaction between them is regarded as edges, then these nodes and edges constitute a drug-target interaction network, and the corresponding interaction matrix is ​​represented as A, which is expressed as follows:

[0030]

[0031] The set of all drug-target interaction combinations involved in constructing the approximate network is defined as B, whose number of elements is less than the number of all interaction combinations A;

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

[0033]

[0034] Among them, DR is Pearson similarity and DC is cosine similarity, which are calculated as follows:

[0035]

[0036] After obtaining the drug approximation matrix DS, for any two drugs di and drugs j , if DS[d i ,d j ]>α, it is considered that there is a high similarity, that is, the node d in the drug-drug approximate network i and node d j There are edges between them;

[0037] At the same time, for the remaining unconnected nodes, select the first n nodes with the largest similarity as their neighbor nodes and add corresponding edges; finally, the network obtained according to the above method is the DD network;

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

[0039]

[0040] Among them, TR is Pearson similarity and TC is cosine similarity, which are calculated as follows:

[0041]

[0042] After obtaining the target approximation matrix TS, for any two target points t i and target t j , if TS[t i ,t j ]>β, it is considered that there is a high similarity, that is, the node t in the target-target approximation network i and node t j There are edges between them; at the same time, for the remaining unconnected nodes, the first m nodes with the largest similarity are selected as their neighbor nodes and the corresponding edges are added; finally, the network obtained according to the above method is the TT network;

[0043] By splicing the DD network, TT network and the existing DT network, we get the DDTT network.

[0044] The drug target association prediction method based on sequence attribute extraction and graph neural network described in the present invention has a further preferred technical solution: the graph encoder is composed of a plurality of GATConv layers, SAGEConv layers and FC layers stacked together, and based on the topological relationship in the DT 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 target features are placed in the same graph neural network for processing. The DDTT network is represented as: G = (V, E), where V is a node set, including drug nodes and target nodes, and E is an edge set. Each node vi There exists a corresponding characteristic eigenvector h i ; GATConv learns the importance of the node's neighbor nodes, and uses the learned importance weights for weighted summation to update the node itself; for any node v i , whose feature vector at the l+1th layer It is expressed as:

[0046]

[0047] Where N(i) is the node v i Neighbor nodes, W l is the weight matrix of the lth layer, α ij is the 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 the sink function to update the node itself; for any node v i , whose feature vector at the l+1th layer It is expressed as:

[0050]

[0051] Among them, σ is the activation function, is the aggregate feature of the sampled neighbor nodes, which is calculated as follows:

[0052]

[0053] Among them, aggregate is the aggregation function;

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

[0055] The graph encoder brings the original features of the previous layer into the new layer during the node update process to enhance the retention of the original features;

[0056] The graph decoder calculates the probability of the existence of a specific edge based on the node, calculates the inner product of the two endpoint feature vectors on all edges, and uses this result as the initial value of the probability of this edge. The probability values ​​of all edges are then Sigmod processed, and the processed results are regarded as the probability of the existence of each edge, that is, the probability that the drug node corresponding to the edge interacts with the target node.

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

[0058] The model adopted by the method of the present invention combines the sequence properties of drugs and targets and the topological properties of the association network, demonstrating its effectiveness in feature extraction and updating in complex relationship networks.

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

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

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

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

[0063] The specific technical solutions of the present invention are further described below to help those skilled in the art further understand the present invention without limiting the rights thereof.

[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 was proposed using four public datasets for training and testing, namely Davis, E, GPCR, and IC. The Davis dataset contains the selectivity determination of kinase proteins and related inhibitors and their respective dissociation constants (K d ) value, which ranges from 0.016 to 10,000, and reflects the affinity of the drug molecule to the target protein. dThe smaller the K, the stronger the correlation. d If it is greater than or equal to 10,000, it means there is no correlation. d The value distribution is too different, so we perform logarithmic transformation on it. The calculation method is as follows:

[0067]

[0068] The threshold is set to 5.0, and drug-target combinations less than 5.0 are considered positive examples, and the rest are considered negative examples to construct a binary classification dataset. Datasets E, GPCR, and IC are from the public literature of MCANet. It divides the proteins that constitute the target into three main parts: enzymes, G protein-coupled receptors (GPCRs), and ion channels. Enzymes and ion channels are from the KEGG BRITE database, DrugBank database, SuperTarget database, and BRENDA database. G protein-coupled receptors are from the GLASS database. All DTIs in the dataset are considered positive examples, and the rest are considered negative examples. The data distribution of each dataset is shown in Table 1.

[0069] Table 1. Data distribution of datasets

[0070]

[0071] 1.2 Methods

[0072] The overall architecture of the SaeGraphDTI model proposed in this invention is as follows: Figure 1 As shown in Figure 1, this model can be divided into three main parts: sequence attribute extractor, graph encoder, and graph decoder. First, the initial sequence of drugs and targets is input into the sequence attribute extractor to obtain feature sequences with uniform length and aligned attributes. Subsequently, the feature sequences of drugs and targets are input into the graph encoder to obtain feature expressions based on topological information. Finally, it is input into the 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] Drugs and targets are expressed in the form of sequences, and their lengths are unified. The sequence attribute extractor based on one-dimensional convolution with variable convolution kernels 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 attributes corresponding to the position. The relevant structure is as follows: Figure 2As shown in Figure 2, in our case, the outputs of drugs and targets after the sequence attribute extractor are set to the same dimension for subsequent graph encoder and decoder processing.

[0075] 1.2.1.1 Embedding Layer

[0076] Appropriate expressions can enable the model to obtain more accurate information. Therefore, SMILES (Simplified molecular input line entry system) strings are used as serialized expressions of drugs, and amino acid sequences are used as serialized expressions of target proteins.

[0077] First, encode each element in the original drug sequence and the original target sequence so that the original sequence can be represented as a coding sequence, where the drug code ranges from 1 to 32 integers and the target code ranges from 1 to 22 integers. Then, by adding 0 at the end or trimming the end sequence, the length of all drug or target coding sequences is made consistent. Finally, the fixed-length coding sequence is input into the embedding layer to generate the corresponding drug embedding matrix. Target embedding matrix Among them, L d is the drug sequence length, L t is the target sequence length, Dim d is the drug embedding dimension, Dim t Embedding dimension for the target.

[0078] 1.2.1.2 Variable Convolutional Layer

[0079] CNN can obtain local features by moving a fixed-size convolution kernel on an image or sequence, and the size of the convolution kernel determines the receptive range of the convolution layer. In order to give the sequence attribute extractor a wider initial receptive range, after the encoded sequence is embedded, two CNNs with fixed-size convolution kernels are first used to extract local features of the sequence. Each CNN is followed by an average pooling layer and a Relu activation function layer. Subsequently, multiple CNNs with different convolution kernel sizes but the same number of output channels 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 used as the output here. Next, the outputs of all convolution kernels are concatenated into a length of N. kernel_size ×N out_channel A one-dimensional tensor, where N kernel_size is the number of convolution kernel sizes, 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 length of N out_channel The one-dimensional feature vector of .

[0080] The attribute extractor based on this structure can convert the outputs of different dimensions with inconsistent numbers of convolution kernel sizes into outputs of the same dimension, that is, output 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 subsequence of suitable length that is most relevant to this result.

[0081] 1.2.2 Graph Encoder and Graph Decoder

[0082] The original data set is represented in the form of a DT (Drug-Target) network, where drugs and targets constitute all nodes, and drug-target combinations with DTI constitute all edges. By constructing a DD (Drug-Drug) network and a TT (Target-Target) network based on the calculation method of topological similarity, and splicing it with the existing DT network, a DDTT (Drug-Drug-Target-Target) network covering all drugs and targets can be constructed. Each node of the network can be connected to the remaining nodes by jumping, and the distance between nodes is represented by 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 inside of 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, treating drug-target combinations with higher probabilities as positive examples.

[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 by T = {t1, t2, …, t n}, 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 drugs and targets are regarded as nodes and the interaction relationship between them is regarded as edges, then these nodes and edges constitute a drug-target interaction network, and the corresponding interaction matrix is ​​represented as A, which is expressed as follows:

[0086]

[0087] Since the construction of drug-drug approximate network and target-target approximate network is based on the calculation of existing drug-target interaction combinations with interactions, if all interaction combinations are involved in the calculation, this may cause the model effect to be higher than the actual true effect. Therefore, the set of all drug-target interaction combinations involved in the construction of the approximate network is defined as B, and the number of its elements is less than the number of all interaction combinations A.

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

[0089]

[0090] Among them, DR is Pearson similarity and DC is cosine similarity, which are calculated as follows:

[0091]

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

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

[0094]

[0095] Among them, TR is Pearson similarity and TC is cosine similarity, which are calculated as follows:

[0096]

[0097] After obtaining the target approximation matrix TS, for any two target points t i and target t j , if TS[t i ,t j ]>β, it is considered that there is a high similarity, that is, the node t in the target-target approximation network i and node t jAt the same time, for the remaining unconnected nodes, select the first m nodes with the largest similarity as their neighbor nodes and add corresponding edges. Finally, the network obtained according to the above method is the TT network.

[0098] By splicing the DD network, TT network and the existing DT network, we get the DDTT network.

[0099] 1.2.2.2 Graph Encoder

[0100] The graph encoder is composed of multiple GATConv layers, SAGEConv layers and FC layers. Based on the topological relationship in the existing DT network, the corresponding node information is updated in a "message passing" manner, such as Figure 3 (A) shown.

[0101] The drug and target feature vectors processed by the sequence attribute extractor have the same dimension, so they can be placed in the same graph neural network for processing. The DDTT network is represented as G = (V, E), where V is a node set, including drug nodes and target nodes, and E is an edge set. Each node v i There exists a corresponding characteristic eigenvector h i .

[0102] GATConv is composed of et al. proposed that the idea is to learn the importance of the node's neighbor nodes, use the learned importance weights for weighted summation, and then update the node itself. i , whose feature vector at the l+1th layer It can be expressed as:

[0103]

[0104] Where N(i) is the node v i Neighbor nodes, W l is the weight matrix of the lth layer, α ij is the neighbor node v j For node v i The importance weight is calculated as follows:

[0105]

[0106] SAGEConv is designed to be applied to large-scale graph structures. Its core idea is to sample the neighbor nodes of a node and aggregate the sampled neighbor nodes through a sink function to update the node itself. i , whose feature vector at the l+1th layer It can be expressed as:

[0107]

[0108] Among them, σ is the activation function, is the aggregate feature of the sampled neighbor nodes, which is calculated as follows:

[0109]

[0110] Here, 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 will rely on the existing network structure. Therefore, before each round of training, it is determined whether each edge is retained, that is, the dynamic connection of the graph nodes. The decision of whether an edge is retained depends on the weight corresponding to this edge. The weights of all edges are between 0 and 1. If the weight of an edge is greater than 0.5, it is considered to be an important edge and retained, otherwise it is discarded. If an edge is determined to be retained, the corresponding two endpoints exchange information, otherwise each endpoint only retains its own information, such as Figure 3 (B) shown.

[0112] At the same time, the traditional graph neural network will cause the difference of node features to decrease as the number of network layers increases, that is, the over-smoothing problem. Therefore, in this model, the original features of the previous layer are brought into the new layer to enhance the retention of the original features. The relevant update methods are as follows: Figure 3 (C) shown.

[0113] 1.2.2.3 Image Decoder

[0114] The graph decoder can calculate the probability of the existence of a specific edge based on the node. Specifically, it calculates the inner product of the two endpoint feature vectors on all edges, and uses this result as the initial value of the probability of this edge. Then, the probability values ​​of all edges are Sigmod processed, and the processed results are regarded as the probability of the existence of each edge, that is, the probability that the drug node corresponding to the edge interacts with the target node.

[0115] 2. Experiment

[0116] 2.1 Experimental Setup

[0117] 2.1.1 Hyperparameter Settings

[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 serialized expression of drugs and targets, the maximum length of drugs is set to 128, and the maximum length of targets is set to 2000. For the insufficient length, the number 0 is used to fill.

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

[0121] In the construction of the approximate network, the drug approximate edge takes the top 10% of all drug approximate values. For unconnected drug nodes, they are connected to at most one most approximate drug node. The target approximate edge takes the top 10% of all target approximate values. For unconnected target nodes, they are connected to at most one most approximate target node.

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

[0123] During the training process, the Adam optimizer was used to optimize the model, with the initial learning rate set to 0.001, the number of training rounds to 10,000, and the learning rate decayed to 0.8 every 1,000 rounds. Despite this, the iteration time on each data set did not exceed 15 minutes.

[0124] 2.1.1 Baseline Model

[0125] In order to test SaeGraphDTI, the classic machine learning models RF and SVM were selected for comparative testing with the DTI prediction models that have been widely cited in recent years. All benchmark model parameter settings were the corresponding default settings.

[0126] MolTrans: applies Transformer to feature embedding of drugs and proteins and creates interaction matrices based on their respective embedding sequences, using multiple 2D convolutional and fully connected layers for prediction.

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

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

[0129] 2.1.2 Experimental environment and training settings

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

[0131] ·CPU: Intel i5-12400F, 2.50GHz

[0132] ·GPU: NVIDIA RTX2060Super, 8GB

[0133] Memory: 32GB DDR4

[0134] The experiment adopted five-fold cross validation. All DTIs were evenly divided into five parts, four of which were selected for network construction and model training each time, and the remaining one was only used for testing. The best result was taken for each test, and the average of each test result was 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. 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 ratio of actual positive examples to predicted positive examples. It is calculated as follows:

[0141]

[0142] Recall: The ratio of predicted positive examples to actual positive examples. It is calculated as follows:

[0143]

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

[0145]

[0146] AUROC (Area Under the Receiver Operating Characteristic Curve): The area under the ROC curve. In a 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, which is applicable to the case of data imbalance. The closer the AUPRC value is to 1, the higher the precision and recall of the model can be maintained at various thresholds.

[0148] 2.2.2 Effect Analysis

[0149] First, the SaeGraphDTI model of the present invention and the benchmark model were trained and tested on three datasets: E, GPCR, and IC. These three datasets are balanced datasets, and the number of positive examples is close to or equal to the number of negative examples. However, the ratio of the number of drugs to the number of targets in each dataset is inconsistent. The experiments show that the SaeGraphDTI model of the present invention can obtain better results on each dataset.

[0150] On dataset E, compared with the best experimental indicators of all models, the SaeGraphDTI model of the present invention 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 surpassed all benchmark models on this dataset, and the relevant data are shown in Table 2.

[0151] Table 2. Experimental results of SaeGraphDTI and baseline models on dataset E

[0152]

[0153]

[0154] The model in this embodiment is derived from:

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

[0156] SVM model: Burges CJ C. Tutorial on support vector machines for pattern recognition[J].Data mining and knowledge discovery,1998,2(2):121-167.(Burges CJ CA 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 LM, et al. MolTrans: molecular interaction transformer for drug–target interaction prediction[J]. Bioinformatics, 2021, 37(6): 830-836.

[0158] HyperAttentionDTI 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 CrossAttention network for drug–target interaction prediction[J]. Briefings in Bioinformatics, 2023, 24(2): bbad082.

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

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

[0162]

[0163]

[0164] On dataset IC, the SaeGraphDTI model achieves 9.9% accuracy, 7.6% precision, 10.7% recall, 9.5% F1 score, 8% AUC, and 8.9% AUPR improvements compared to the best experimental indicators of all baseline models. Overall, the SaeGraphDTI model surpasses all baseline models on dataset E, and the relevant data are shown in Table 4.

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

[0166]

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

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

[0169]

[0170] By representing the data set as a graph structure, the relevant values ​​of the node degree under each data set can be calculated. It is found that the average node degree under the three data sets of E, GPCR, and IC is between 4.4 and 10.5. However, in the Davis data set, each drug is connected to an average of 107.6 targets, and each target is connected to an average of 19.3 drugs, which is significantly higher than other data sets. Too high a connection density may cause the spread of relevant information to the entire graph, thereby reducing the distinction between each drug or target. Therefore, the relevant indicators of the SaeGraphDTI model and other benchmark models under the Davis data set are slightly lower than those of the data sets E and IC. At the same time, there is a large gap between the number of drugs and targets in the GPCR data set and the Davis data set, which also affects the balance of the model's judgment of various categories. SaeGraphDTI dynamically connects to nodes through a graph encoder, which limits the propagation range of node information while obtaining topological information, so it achieves better results than the benchmark model.

[0171] In order to verify the inference ability of SaeGraphDTI for potential DTI, some original DTIs were deleted, and the deleted DTIs were predicted based on the remaining data as the result of this case analysis. The verification process is as follows:

[0172] First, two targets were randomly selected from the Davis dataset: target YANK1 and target PAK2, and 10 known confirmed relationships of each target (5 associated drugs and 5 unassociated drugs) were randomly deleted, thus obtaining the drug-target set to be tested and the corresponding deleted dataset.

[0173] Then, based on this pruned data set, the similarity of nodes of the same type is recalculated to construct the association network. All parameters are kept in default configuration.

[0174] Then, the constructed association network is input into the model, and corresponding training and testing are performed until the model converges and stabilizes. Among them, the training and test data are all from the deleted data set, and all parameters are kept 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 association of this combination. The probability value is between 0 and 1. The larger the value, the higher the association. Values ​​greater than 0.5 are considered to be associated, that is, there is DTI, and values ​​less than 0.5 are considered to be unassociated, that is, there is no DTI. In this experiment, the prediction results of target YANK1 and target PAK2 and related test drugs are shown in Table 6.

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

[0177]

[0178] This experiment showed that the SaeGraphDTI model achieved 90% accuracy in identifying the potential DTIs of two randomly selected targets, YANK1 and PAK2, verifying the good effect of this model in predicting potential DTIs.

[0179] 3. Conclusion

[0180] Experiments show that SaeGraphDTI outperforms the state-of-the-art models on most metrics under multiple datasets.

Claims

1. A drug target association prediction method based on sequence attribute extraction and graph neural network, characterized in that: This method adopts the DTI prediction model SaeGraphDTI based on sequence attribute extraction and graph neural network. The overall architecture of SaeGraphDTI is divided into three main parts: sequence attribute extractor, graph encoder and graph decoder; First, the initial sequences of drugs and targets are input into the sequence attribute extractor, and the relevant attributes of drug sequences and target sequences are extracted by applying the sequence attribute extractor; The drug and target feature acquisition method based on sequence attribute extraction extracts sequences with inconsistent lengths into attribute sequences with consistent lengths and aligned positions; Subsequently, the characteristic sequences of drugs and targets are input into the graph encoder, and the graph encoder based on this network updates the node information accordingly to obtain the characteristic expression based on the network topology structure; Based on the adaptive connection graph neural network, the connectivity between nodes is dynamically judged to make reasonable choices for edges. Finally, the obtained node features and their positional relationships are input into the graph decoder; the graph decoder calculates the probability of the existence of a specific drug-target edge to make a prediction for DTI.

2. The method according to claim 1, characterized in that: SaeGraphDTI was trained and tested using four public datasets: Davis, E, GPCR, and IC; the Davis dataset contains the selectivity determination of kinase proteins and related inhibitors and their respective dissociation constants K d The value ranges from 0.016 to 10,000, which reflects the affinity of the drug molecule to the target protein; K d The smaller the K, the stronger the correlation. d If it is greater than or equal to 10,000, it means there is no correlation; d The values ​​are logarithmically transformed and calculated as follows: The threshold is set to 5.0, and drug-target combinations with a value less than 5.0 are considered positive examples, and the rest are considered negative examples to construct a binary classification dataset; The proteins that constitute the target are divided into three main parts: enzymes, G protein-coupled receptors and ion channels. The enzyme E and ion channel IC are from the KEGG BRITE database, DrugBank database, SuperTarget database and BRENDA database, and 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. The data distribution of each data set is as follows:

3. The method according to claim 1, characterized in that: Drugs and targets are expressed in the form of sequences, and their lengths are unified. The sequence attribute extractor based on one-dimensional convolution with variable convolution kernels 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 list of attributes corresponding to the position. The outputs of the sequence attribute extractor for drugs and targets are set to the same dimension for subsequent graph encoder and decoder processing.

4. The method according to claim 1 or 3, characterized in that: Use SMILES string as the serialized expression of the drug and amino acid sequence as the serialized expression of the target protein; First, encode each element in the original drug sequence and the original target sequence so that the original sequence can be represented as a coding sequence, where the drug coding range is an integer from 1 to 32, and the target coding range is an integer from 1 to 22; then, by adding 0 at the end or trimming the terminal sequence, the length of all drug or target coding sequences is made consistent; finally, the fixed-length coding sequence is input into the embedding layer to generate the corresponding drug embedding matrix Target embedding matrix Among them, L d is the drug sequence length, L t is the target sequence length, Dim d is the drug embedding dimension, Dim t Embedding dimension for the target.

5. The method according to claim 1 or 3, characterized in that: Convolutional neural network (CNN) obtains local features by moving a fixed-size convolution kernel on an image or sequence, and the size of the convolution kernel determines the receptive range of the convolution at this layer. In order to give the sequence attribute extractor a wider initial receptive range, after the encoding sequence is embedded, two CNNs with fixed-size convolution kernels are first used to extract local features of the sequence. Each CNN is followed by an average pooling layer and a Relu activation function layer. Subsequently, multiple CNNs with different convolution kernel sizes but the same number of output channels 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 used as the output here. Next, concatenate the outputs of all convolution kernels into a network of length N kernel_size ×N out_channel A one-dimensional tensor of N kernel_size is the number of convolution kernel sizes, 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 sequence of length N. out_channel The one-dimensional feature vector of The attribute extractor converts the outputs of different dimensions with inconsistent numbers of convolution kernel sizes into outputs of the same dimension, that is, outputs a one-dimensional feature sequence with consistent output 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 is obtained, so as to locate the key subsequence of suitable length that is most relevant to this result.

6. The method according to claim 1 or 3, characterized in that: The original dataset is represented in the form of a DT network, in which drugs and targets constitute all nodes, and drug-target combinations with DTI constitute all edges; DD networks and TT networks are constructed by a calculation method based on topological similarity, and they are spliced ​​with the DT network to construct a DDTT 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; the graph encoder module based on this network uses a graph neural network method to propagate and update node information, and add topological information to the nodes; the graph decoder accepts the node features encoded by the encoder, and calculates the probability of the existence of edges between specific nodes, and regards drug-target combinations with higher probabilities as positive examples.

7. The method according to claim 6, characterized in that The method for constructing an approximate network based on topological similarity is as follows: The set of all drugs is represented as D = {d1, d2, ..., d m }, the set of all targets is represented by T = {t1, t2, …, t n }, the set of all drug-target interaction combinations is denoted as E, where m is the number of all drugs and n is the number of all targets; If drugs and targets are regarded as nodes and the interaction between them is regarded as edges, then these nodes and edges constitute a drug-target interaction network, and the corresponding interaction matrix is ​​represented as A, which is expressed as follows: The set of all drug-target interaction combinations involved in constructing the approximate network is defined as B, whose 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 DS calculation method is as follows: Among them, DR is Pearson similarity and DC is cosine similarity, which are calculated as follows: After obtaining the drug approximation matrix DS, for any two drugs d i and drugs j , if DS[d i ,d j ]>α, it is considered that there is a high similarity, that is, the node d in the drug-drug approximate network i and node d j There are edges between them; at the same time, for the remaining unconnected nodes, select the first n nodes with the largest similarity as their neighbor nodes and add corresponding edges; finally, the network obtained according to the above method is the DD network; The approximate matrix between different targets is defined as TS, and the TS calculation method is as follows: Among them, TR is Pearson similarity and TC is cosine similarity, which are calculated as follows: After obtaining the target approximation matrix TS, for any two target points t i and target t j , if TS[t i ,t j ]>β, it is considered that there is a high similarity, that is, the node t in the target-target approximation network i and node t j There are edges between them; at the same time, for the remaining unconnected nodes, the first m nodes with the largest similarity are selected as their neighbor nodes and the corresponding edges are added; finally, the network obtained according to the above method is the TT network; By splicing the DD network, TT network and the existing DT network, we get the DDTT network.

8. The method according to claim 1, characterized in that: The graph encoder is composed of multiple GATConv layers, SAGEConv layers and FC layers stacked together, and updates the corresponding node information in a "message passing" manner based on the topological relationship in the DT network; The drug and target feature vectors processed by the sequence attribute extractor have the same dimension. The drug features and target features are placed in the same graph neural network for processing. The DDTT network is represented as: G = (V, E), where V is a node set, including drug nodes and target nodes, and E is an edge set. Each node v i There exists a corresponding characteristic eigenvector h i ; GATConv learns the importance of the node's neighbor nodes, and uses the learned importance weights for weighted summation to update the node itself; for any node v i , whose feature vector at the l+1th layer It is expressed as: Where N(i) is the node v i Neighbor nodes, W l is the weight matrix of the lth layer, α ij is the neighbor node v j For node v i The importance weight is calculated as follows: SAGEConv samples the neighboring nodes of a node and aggregates the sampled neighboring nodes through the sink function to update the node itself; for any node i, its feature vector at the l+1 layer It is expressed as: Among them, σ is the activation function, is the aggregate feature of the sampled neighbor nodes, which is calculated as follows: Among them, aggregate is the aggregation function; Since the graph encoder is updated based on local iteration, it will be dependent on the existing network structure. Therefore, before each round of training, it is determined whether each edge is retained, that is, the dynamic connection of the graph nodes; whether an edge is retained depends on the weight corresponding to the edge. The weights of all edges are between 0 and 1. If the weight of an edge is greater than 0.5, it is considered to be an important edge and retained, otherwise it is discarded; if an edge is determined to be retained, the corresponding two endpoints exchange information, otherwise each endpoint only retains its own information; The graph encoder brings the original features of the previous layer into the new layer during the node update process to enhance the retention of the original features; The graph decoder calculates the probability of the existence of a specific edge based on the node, calculates the inner product of the two endpoint feature vectors on all edges, and uses this result as the initial value of the probability of this edge. The probability values ​​of all edges are then Sigmod processed, and the processed results are regarded as the probability of the existence of each edge, that is, the probability that the drug node corresponding to the edge interacts with the target node.

Images