Drug-target affinity prediction method based on multi-scale convolutional neural network

Abstract

The invention discloses a drug-target affinity prediction method based on a multi-scale convolutional neural network, and the method comprises the steps: obtaining a sequence of a target protein and an SMILES character string of a drug small molecule, obtaining an SMILES word vector matrix and a protein word vector matrix based on word vector coding, constructing the multi-scale convolutional neural network, carrying out the learning of the SMILES word vector matrix, and carrying out the prediction of the drug-target affinity. The method comprises the following steps: constructing a protein word vector matrix, acquiring chemical and structural characteristics of drug small molecules, constructing a one-dimensional convolutional neural network, learning the protein word vector matrix, acquiring structural characteristics of target protein, and fusing the chemical and structural characteristics of the drug small molecules and the structural characteristics of the target protein through splicing operation. And expressing the fused features as drug-target features, constructing an affinity prediction model, and learning the drug-target features to obtain an affinity prediction value between the drug and the protein. Compared with a graph neural network and a pre-training model, the drug-target affinity can be better predicted, and the prediction capability is better than that of a deep learning method.

Description

Technical Field

[0001] The present invention relates to the field of target binding affinity prediction, and in particular to a drug-target affinity prediction method based on a multi-scale convolutional neural network. Background Art

[0002] With the development of artificial intelligence technology, especially the continuous advancement of deep learning technology, the method of using deep learning models to predict drug-target binding affinity has been widely recognized. Artificial intelligence methods can help screen potential drug candidates and accelerate the drug discovery process.

[0003] DeepDTA proposes a method to predict the binding affinity of protein-ligand interactions using a deep learning model using only the sequences of proteins and ligands (1D representation). To this end, the sequences of proteins and SMILES (Simplified Molecular Input Line Entry System) representations of compounds are used instead of external features or the 3D structure of the binding complex. Representations are learned from the original protein sequence and SMILES strings using a convolutional neural network, and these representations are combined and input into the fully connected layer block. The WideDTA model is an extension of DeepDTA. In terms of original feature selection, not only the SMILES code of the drug molecule and the amino acid sequence of the target protein are used, but also the protein domains and motifs and the maximum common substructure words of the molecule are additionally used. And the differences in the results obtained by using two, three or four of these features at the same time are tested. However, these models represent drugs as strings, which is not a natural way to represent molecules. When using string representation, the structural information of the molecule is lost, which may affect the predictive ability of the model and the functional relevance of the learned latent space.

[0004] Nguyen et al. proposed a graph-based model GraphDTA, which encodes drugs as an undirected graph with feature maps and adjacency matrices. Graph convolutional networks (GCN), graph attention networks (GAT), and graph isomorphism networks (GIN) are designed to extract features from drugs. This method uses 1D convolution blocks to extract protein feature encoders, but crudely using 1D convolutional layers or global pooling operations to aggregate features may result in the loss of a lot of useful information.

[0005] GraphCL-DTA's graph contrastive learning approach for drug-target binding affinity prediction. In GraphCL-DTA, a molecular graph contrastive learning framework for learning drug representations is designed to preserve the semantic information of molecular graphs. With this graph contrastive learning framework, more essential and effective drug representations can be learned. GraphscoreDTA develops a novel graph neural network strategy combined with Vina distance optimization term (GraphscoreDTA) for predicting protein-ligand binding affinity, which is the first time that graph neural networks, bidirectional information transfer mechanisms, and physics-based distance terms are combined. Unlike other methods, GraphscoreDTA can not only effectively capture the mutual information of protein-ligand pairs, but also highlight the important atoms of the ligand and the residues of the protein.

[0006] However, the complex graph structure makes the interpretability of the model an obstacle for experts in professional fields to adopt such models. With unique ethical and regulatory requirements, the demand for interpretable deep models in the biomedical field is growing. Recent studies have explored methods based on attention mechanisms to address the interpretability problem.

[0007] DrugVQA proposed a question-answering model for drug-target interaction tasks, in which a sequential attention mechanism was used to capture dependencies of a dynamic convolutional neural network (CNN). From another perspective, MATT_DTI designed a multi-head attention model that treats drug representations as queries and protein representations as keys and values. ColdDTA uses data augmentation and attention-based feature fusion to improve the generalization ability of predicting drug-target binding affinity. Although the attention mechanism has been widely studied for building pairwise long-range dependencies. However, the relatively limited well-annotated medical data makes it difficult for this mechanism to extract diverse global features, resulting in attention collapse, that is, the attention maps become similar or even identical.

[0008] In summary, existing technologies have problems such as difficult to understand models and low predictive ability. Summary of the invention

[0009] The present invention provides a drug-target affinity prediction method based on a multi-scale convolutional neural network to overcome the above technical problems.

[0010] Drug-target affinity prediction methods based on multi-scale convolutional neural networks, including:

[0011] Step 1: Get the sequence of the target protein and the SMILES string of the drug small molecule.

[0012] Step 2: Encode the SMILES string of the drug small molecule based on the word vector encoding to obtain the SMILES word vector matrix, and construct a multi-scale convolutional neural network. The multi-scale convolutional neural network is used to learn the SMILES word vector matrix and obtain the chemical and structural characteristics of the drug small molecule. The multi-scale convolutional neural network consists of N convolutional layers with different convolution kernel sizes.

[0013] Step 3: After encoding the sequence of the target protein based on the word vector encoding, a protein word vector matrix is ​​obtained, and a one-dimensional convolutional neural network is constructed. The one-dimensional convolutional neural network is used to learn the protein word vector matrix and obtain the structural characteristics of the target protein.

[0014] Step 4: The chemical and structural features of the drug small molecule and the structural features of the target protein are fused through a splicing operation, the fused features are represented as drug-target features, and an affinity prediction model is constructed. The affinity prediction model is used to learn the drug-target features and obtain the affinity prediction value between the drug and the protein.

[0015] Preferably, the word vector encoding-based encoding of the SMILES string of the drug small molecule to obtain the SMILES word vector matrix includes randomly initializing a V*d matrix as the word vector matrix according to the size V of the vocabulary and the embedding dimension d of each word, and for each character in the SMILES string, converting it into a corresponding index, wherein the index is the position of the character in the vocabulary, and searching the corresponding embedding vector from the word vector matrix according to the index of the character, obtaining the embedding vector of each character in the SMILES string, and constructing the SMILES word vector matrix according to the embedding vector of each character.

[0016] Preferably, the affinity prediction model includes three fully connected layers and two Dropout layers, the three fully connected layers are a first fully connected layer, a second fully connected layer and a third fully connected layer, the two Dropout layers include a first Dropout layer and a second Dropout layer, the first fully connected layer is used to accept drug-target features and pass them as input data to a hidden layer, wherein the first Dropout layer, the second fully connected layer and the second Dropout layer serve as hidden layers, and the hidden layers perform nonlinear transformation and feature extraction on the input data by applying an activation function; the third fully connected layer uses a ReLU activation function to predict the affinity prediction value between the drug and the protein and outputs it.

[0017] Preferably, the N convolutional layers with different convolution kernel sizes are 9 convolutional layers with different convolution kernel sizes, wherein the 9 different convolution kernel sizes include 1×em_dim, 2×em_dim, 4×em_dim, 8×em_dim, 10×em_dim, 12×em_dim, 16×em_dim, 32×em_dim, and 64×em_dim, wherein em_dim represents the dimension of the SMILES word vector matrix.

[0018] Preferably, obtaining the predicted affinity value between the drug and the protein comprises evaluating the predicted result according to formula (1),

[0019]

[0020] Among them, y i ,y j It represents the predicted value of drug-protein affinity obtained by the affinity prediction model. represents the actual affinity value of the drug to the protein, z represents a constant, and when y i >y j ,h is equal to 1, when y i =y j , h is equal to 0.5, otherwise the h function value is 0.

[0021] The present invention provides a drug-target affinity prediction method based on a multi-scale convolutional neural network. First, the drug small molecule string is treated as a text containing a biological language, and then a word vector encoding is performed. This encoding form is simple and easy to understand, and retains semantic information to the greatest extent. Many studies have shown that when the amount of data is limited, the convolutional neural network will obtain a better learning effect, and the mature visualization method makes the convolutional neural network easier to understand. Therefore, the word vector encoding of the drug small molecule is regarded as an "image", and the word vector "image" is processed using a multi-scale large convolution kernel two-dimensional convolutional neural network, which makes full use of the ability of the convolutional neural network to efficiently process image data. In the two-dimensional convolutional neural network, the word vector "image" dimension is used as the size of the convolution kernel. The advantage of this is that it replaces the pooling layer usually used in the convolutional neural network, reducing the information lost due to the pooling operation. The present invention not only predicts drug-target affinity better than the graph neural network and the pre-training model, but also has better prediction ability than the deep learning method. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0023] Figure 1 is a flow chart of the method of the present invention;

[0024] Figure 2 is a diagram of a small molecule of the drug of the present invention;

[0025] Figure 3 It is a multi-scale convolution model of drug small molecules of the present invention;

[0026] Figure 4 It is a one-dimensional convolutional neural network model of the present invention;

[0027] Figure 5 is the affinity prediction model of the present invention;

[0028] Figure 6 It is the prediction model structure of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] Figure 1 It is a flow chart of the method of the present invention, such as Figure 1 As shown, the method of this embodiment may include:

[0031] Step 1: Get the sequence of the target protein and the SMILES string of the drug small molecule.

[0032] Step 2: Encode the SMILES string of the drug small molecule based on the word vector encoding to obtain the SMILES word vector matrix, and construct a multi-scale convolutional neural network. The multi-scale convolutional neural network is used to learn the SMILES word vector matrix and obtain the chemical and structural characteristics of the drug small molecule. The multi-scale convolutional neural network consists of N convolutional layers with different convolution kernel sizes.

[0033] Step 3: After encoding the sequence of the target protein based on the word vector encoding, a protein word vector matrix is ​​obtained, and a one-dimensional convolutional neural network is constructed. The one-dimensional convolutional neural network is used to learn the protein word vector matrix and obtain the structural characteristics of the target protein.

[0034] Step 4: The chemical and structural features of the drug small molecule and the structural features of the target protein are fused through a splicing operation, the fused features are represented as drug-target features, and an affinity prediction model is constructed. The affinity prediction model is used to learn the drug-target features and obtain the affinity prediction value between the drug and the protein.

[0035] Based on the above scheme, the present invention provides a drug-target affinity prediction method based on a multi-scale convolutional neural network. First, the drug small molecule string is treated as a text containing biological language, and then word vector encoding is performed. This encoding form is simple and easy to understand, and retains semantic information to the greatest extent. Many studies have shown that when the amount of data is limited, the convolutional neural network will obtain a better learning effect, and the mature visualization method makes the convolutional neural network easier to understand. Therefore, the word vector encoding of the drug small molecule is regarded as an "image", and the word vector "image" is processed using a multi-scale large convolution kernel two-dimensional convolutional neural network, which makes full use of the ability of the convolutional neural network to efficiently process image data. In the two-dimensional convolutional neural network, the word vector "image" dimension is used as the size of the convolution kernel. The advantage of this is that it replaces the pooling layer usually used in the convolutional neural network and reduces the information lost due to the pooling operation. The present invention not only predicts drug-target affinity better than the graph neural network and the pre-trained model, but also has better prediction ability than the deep learning method.

[0036] Specifically, this embodiment provides the details of the prediction method, including:

[0037] First, a mathematical definition of the problem is given: The purpose of the present invention is to design a method that can effectively predict drug-target binding affinity. The prediction of drug small molecule and protein binding affinity belongs to a regression problem. The goal is to predict the binding affinity score (CI) between a protein and a drug small molecule given a string sequence. In order to better understand the method of the present invention, a mathematical definition of the problem is given.

[0038] Input: protein sequence and drug SMILES string.

[0039] Output: Binding affinity score, evaluated by Mean Square Error (MSE) and Concordance Index (CI). Mean Square Error is a common metric for evaluating the performance of regression models. It measures the average squared difference between the predicted value and the actual value. The smaller the MSE, the better the model fits the data. The consistency index, also known as the C-index or consistency index, is a metric used to evaluate the performance of survival analysis models. It measures the accuracy of the model in ranking samples according to the predicted survival time. The C-index ranges from 0 to 1, where 1 indicates perfect consistency and 0.5 indicates random prediction. The closer the C-index is to 1, the better the predictive ability of the model.

[0040] The data sets used in this embodiment are: the commonly used public data sets Davis and KIBA are used as the training set, validation set and test set of drug-target binding affinity, as shown in Table 1.

[0041] Table 1 Dataset

[0042]

[0043] In order to fairly evaluate the prediction performance of the proposed model, the present invention uses two commonly used data sets in the benchmark methods SimBoost, DeepDTA, and GraphDTA: Davis and KIBA, and the two data sets solve the problem of data heterogeneity; and according to the atomic sequence length distribution in the two benchmark data sets, the label encoding lengths of proteins and drug molecules in Davis are 1200 and 85, respectively, and in KIBA are 1000 and 100, respectively. The specific statistics of the data are shown in Table 1 below.

[0044] Davis: The dataset includes 442 kinase proteins, 68 inhibitor drug small molecules, and 30,056 binding complexes; the affinity strength between target and drug molecules is calculated based on the kinase dissociation constant K d , K d Logarithmic transformation of pK values d , and its calculation formula is shown in the following formula (1).

[0045]

[0046] KIBA: The data in this dataset come from a KIBA method. To solve the problem of data heterogeneity, this method processes and screens 246,088 binding pairs consisting of 467 proteins and 52,498 ligands; ultimately 229 unique proteins and 2,111 unique drug small molecules, as well as 118,254 complex macromolecules were identified.

[0047] In order to improve the learning generalization of the method of the present invention, a nested cross-validation is performed on the benchmark data to construct a training set, a validation set, and a test set of the small molecule-target binding affinity prediction network model;

[0048] Objective: To train a neural network model to process input sequences and predict the corresponding binding affinity scores based on the input protein sequence and drug small molecule SMILES string.

[0049] Specifically, the prediction steps of this embodiment include: Step 1, obtaining the sequence of the target protein and the SMILES string of the drug small molecule,

[0050] The drug molecules are: the drug small molecule Smiles (Simplified Molecular Input Line Entry System) sequence collected from Pubchem, a chemical substance database of the National Center for Biotechnology Information of the United States, using 64 different characters to represent different chemical molecular structure features. The Smiles sequence "CC1C=CC=CC1NC(=O)OCC2=C(OC(=C2)[O-])C" corresponds to Figure 2 The protein is: protein sequences are collected from the protein database UniProt (Universal Protein), using 25 different characters to represent different protein atomic structure features, "AGCTYXKLMN" is an example of a partial protein sequence.

[0051] The present invention regards protein sequence as a string containing semantic information, so word vectors are used for encoding. The lengths of drug molecules and protein sequences are not uniform. In order to improve the learning efficiency of the network model, the usual practice is to truncate or fill the sequence, which will lose feature information or increase noise to a certain extent. The present invention fixes the encoding length of the protein sequence. In the Davis data set, the lengths of the drug small molecule string and protein sequence are 85 and 1200 respectively, and in the KIBA data set, they are 100 and 1000 respectively. Such a length can cover 80% of the protein sequence length in the data set. If the length of the drug small molecule and protein sequence is greater than the fixed value, truncation is performed; otherwise, the sequence is padded with 0.

[0052] Step 2: Encode the SMILES string of the drug small molecule based on word vector encoding to obtain a SMILES word vector matrix. The SMILES word vector matrix obtained by encoding the SMILES string of the drug small molecule based on word vector encoding includes randomly initializing a V*d matrix as the word vector matrix according to the size V of the vocabulary and the embedding dimension d of each word. For each character in the SMILES string, convert it into a corresponding index, and the index is the position of the character in the vocabulary. According to the index of the character, find the corresponding embedding vector from the word vector matrix, obtain the embedding vector of each character in the SMILES string, and construct the SMILES word vector matrix according to the embedding vector of each character.

[0053] This example gives a drug small molecule fragment ["CC 1C=CC=C"], and we can represent each character as a 128-dimensional vector, for example:

[0054] -C:[0.1000,-1.7178,-2.0577,...,0.4555,0.3423]

[0055] -C:[0.2400,-2.4567,-1.4536,...,0.5564,0.2589]

[0056] -1:[0.1600,-1.415,-2.7789,...,0.1236,0.1232]

[0057] -C:[0.1700,-1.5198,-3.7869,...,01367,0.2333]

[0058] -=:[0.0500,-0.3452,-2.3344,...,0.1895,0.2544]

[0059] -C:[0.1900,-1.2356,-1.0088,...,0.5666,0.2677]

[0060] -C:[0.4000,-1.7789,-1.0090,...,0.7789,0.1998]

[0061] -C:[0.3000,-1.0987,-2.2341,...,0.4456,0.1877]

[0062] -=:[0.2600,-1.5620,-2.2346,...,0.7768,0.1980]

[0063] -C:[0.4500,-1.1267,-2.5673,...,0.8999,0.1222]

[0064] These vector representations can be used for tasks such as computing similarities between words, clustering, and classification.

[0065] Specifically, converting input data into a form that can be calculated by the model is called encoding or embedding. In order to better understand this method, we will introduce the encoding method in detail as follows: Encoding is a technology that converts variables into digital form. To adapt to the method model in the present invention, we use word vector encoding for the sequence string of drug small molecule (SMILES) and protein target to facilitate calculation by neural network.

[0066] Word embedding is a technique for mapping words into a low-dimensional vector space. It can represent words as real number vectors by learning the semantic and grammatical information of words in a language, thus facilitating computer processing and understanding of text data. The principle of word embedding is based on the distribution hypothesis, that is, in natural language, similar words usually appear in similar contexts. Therefore, the representation of words in the vector space can be learned by observing the occurrence of words in context.

[0067] A multi-scale convolutional neural network is constructed, and the multi-scale convolutional neural network is used to learn the SMILES word vector matrix and obtain the chemical and structural characteristics of drug small molecules, wherein the multi-scale convolutional neural network is composed of N convolutional layers with different convolution kernel sizes, and the N convolutional layers with different convolution kernel sizes are 9 convolutional layers with different convolution kernel sizes, wherein the 9 different convolution kernel sizes include 1×em_dim, 2×em_dim, 4×em_dim, 8×em_dim, 10×em_dim, 12×em_dim, 16×em_dim, 32×em_dim, and 64×em_dim, wherein em_dim represents the dimension of the SMILES word vector matrix.

[0068] Specifically, in the drug small molecule feature learning part of the present invention, the SMILES word vector matrix is ​​learned through a multi-scale two-dimensional convolution kernel, which can not only extract the local chemical structure of the drug small molecule, but also learn the overall topological structure information of the small molecule, so as to fully understand the chemical and structural characteristics of the drug molecule; the network structure of this part uses a multi-scale two-dimensional convolution model such as Figure 3As shown. We converted the input string into a two-dimensional word vector matrix of size length×128 through word vectors, where length is the fixed length of the SMILES string under different data sets, and 128 is the word vector encoding dimension. In order to enhance the local sensitivity and global difference judgment ability of the model, we introduced multiple convolutional layers with different convolution kernel sizes, so that drug features can be captured at multiple scales. Here, different convolution kernels are used for different convolutional layers. Tensors of different sizes and numbers will be obtained according to the size of the convolution kernel, and then these feature vectors of different scales are fused. A recent study showed that the number of characters of the maximum common substructure (MCS) in most drugs is between 8 and 12. They speculated that the pattern used by chemists to distinguish a group of molecules, such as the maximum common substructure (MCS), is the "chemical word". That is, substructures of this size can best represent the characteristics of drug small molecules. Based on this understanding, the present invention uses large convolution kernels to learn the characteristics of drug small molecules. The specific convolution kernels are shown in Table 2.

[0069] Table 2 Multi-scale large convolution kernels

[0070]

[0071] Note: Length is the length of the drug small molecule string, em_dim is the word vector encoding dimension

[0072] Step 3: After encoding the sequence of the target protein based on the word vector encoding, a protein word vector matrix is ​​obtained, and a one-dimensional convolutional neural network is constructed. The one-dimensional convolutional neural network is used to learn the protein word vector matrix and obtain the structural characteristics of the target protein.

[0073] The number of proteins in the public datasets Davis and Kiba used in this embodiment is 442 and 229 respectively. When the training data is limited, the convolutional neural network often shows good performance. The method of the present invention treats the protein sequence as a string containing semantics. For the semantic information extraction of the string, we use a one-dimensional convolutional neural network, which can capture the structural information of the protein sequence.

[0074] The one-dimensional convolutional neural network model in the present invention is as follows: Figure 4As shown in the figure: In order to obtain more semantic information of protein sequences, we convert the input sequence into a two-dimensional word vector matrix of size length×128 (the specific method is the same as SMILIES string encoding, where V=25, d=128), where length is the fixed length of protein sequences on different datasets, and 128 is the word vector encoding dimension. In order to enhance the local sensitivity and global difference judgment ability of the model, we introduced three layers of one-dimensional convolutional layers and one maximum pooling layer, which can more effectively capture the characteristics of protein sequences.

[0075] Step 4: The chemical and structural features of the drug small molecule and the structural features of the target protein are fused through a splicing operation, and the fused features are expressed as drug-target features to construct an affinity prediction model, which is used to learn the drug-target features and obtain the affinity prediction value between the drug and the protein. The affinity prediction model includes three fully connected layers and two Dropout layers, the three fully connected layers are the first fully connected layer, the second fully connected layer and the third fully connected layer, the two Dropout layers include the first Dropout layer and the second Dropout layer, the first fully connected layer is used to accept the drug-target features and pass them to the hidden layer as input data, wherein the first Dropout layer, the second fully connected layer and the second Dropout layer are used as hidden layers, and the hidden layer performs nonlinear transformation and feature extraction on the input data by applying activation functions; the third fully connected layer uses the ReLU activation function to predict the affinity prediction value between the drug and the protein and outputs it.

[0076] Specifically, in the present invention, the drug-target binding affinity prediction problem is regarded as a regression task. (1) The drug feature learning part extracts the local and topological structure information of the drug small molecule, and (2) the target protein feature learning part extracts the global and local structure information of the protein. The feature vectors of the drug and the target are fused through the concatenation operation to predict the binding affinity. The fused features of the drug-target are input into three fully connected layers; in order to prevent the model from overfitting the training data and reducing the generalization, the regularization method Dropout of the neural network model is used, that is, two Dropout layers are set in the middle of the three fully connected layers; the network structure of this part is as follows Figure 5 shown.

[0077] Other information about the neural network model used in this invention:

[0078] (1) Parameter settings

[0079] The specific parameter settings in the method of the present invention are shown in 3, and the parameter settings of the protein part are shown in 4.

[0080] Table 3 List of model training related parameters and their values

[0081]

[0082]

[0083] Table 4 Protein module parameter list and values

[0084]

[0085] (2) Model training environment settings

[0086] The small molecule-target binding affinity prediction model environment designed in the present invention is coded in Python3.9.13, CUDA 12.0, and the binding affinity prediction model is trained, verified and performance tested on Geforce RTX30706G GPU.

[0087] In order to better understand the method of the present invention, the performance of the proposed affinity prediction method was verified:

[0088] In order to verify the effectiveness of the method proposed in the present invention, an unbiased estimation (Concordance Index, CI) indicator is defined to verify the unbiasedness and consistency of model learning. The function definition of CI is shown in formula (2). The method of obtaining the affinity prediction value between the drug and the protein includes evaluating the prediction result according to formula (2).

[0089]

[0090] Among them, y i ,y j It represents the predicted value of drug-protein affinity obtained by the affinity prediction model. represents the actual affinity value of the drug to the protein, z represents a constant, and when y i >y j ,h is equal to 1, when y i =y j , h is equal to 0.5, otherwise the h function value is 0.

[0091] The performance of the trained affinity prediction method was compared with KronRLS, SimBoost, DeepDTA, GraphDTA, DeepGS and FingerDTA. The average CI and MSE values ​​obtained on the independent test set are shown in 5. It is not difficult to see that the performance of the affinity prediction method proposed in the present invention is very good; in addition, the proposed method is better than the comparison method in CI and MSE performance, which proves the advancement of deep learning methods in drug-target affinity prediction; further, the performance of the proposed method is better than DeepDTA, GraphDTA, and DeepGS. Although the performance is slightly different from FingerDTA, our network structure is simpler, and the model is established based on the study of drug molecules, so it is better than FingerDTA in efficiency and interpretability. This also proves the effectiveness and superiority of multi-scale two-dimensional convolution in global representation learning of drug small molecules.

[0092] Table 5 Average CI and MSE values ​​of the compared methods

[0093]

[0094] Among them, the Chinese meanings of the English words used in each model in the present invention are shown in Table 6.

[0095] Table 6 English vocabulary explanation

[0096]

[0097] In order to predict the binding affinity of drug-target pairs, the present invention establishes a supervised learning model based on the characteristics of protein and drug molecules. The model structure overview is as follows Figure 6 As shown. The method of the present invention uses the sequences of proteins and small drug molecules as input. The methods mentioned in previous studies all use one-dimensional convolution layers to extract structural features from drug strings, but one-dimensional convolution is difficult to obtain global feature information at different scales. First, we use the SMILES string of small drug molecules as a string containing semantic information, so we perform word vector encoding. Because the SMILES string after word vector encoding is itself a two-dimensional matrix, we use a set of multi-scale two-dimensional convolutional networks to extract the semantic information of the sequence. We use large convolution kernels to replace the pooling layer in the traditional convolutional neural network. On the basis of retaining the feature information, the model can obtain richer local and global information from the SMILES sequence. Large two-dimensional convolution kernels with different scales are used to learn the local and global structural features of small drug molecules at different scales, and output the binding affinity prediction value of the drug-target pair.

[0098] The purpose of the present invention is to describe the drug-target binding affinity prediction problem in detail and establish a neural network model; a method of a multi-scale large convolution kernel two-dimensional convolutional neural network is proposed. The proposed deep learning model is trained and tested using the benchmark data sets Davis and KIBA; in the model of the present invention, the mean square error MSE (Mean-Square Error) is used as the loss function, and the optimizer is Adam; the activation function of the fully connected layer is ReLU (Recfied Linear Unit). ReLU and MSE are detailed in formulas (2) and (3), where x represents any real number input to the ReLU function, y i represents the model prediction output value, Represents the actual output value of the sample data, and n represents the number of training data samples.

[0099]

[0100]

[0101] The present invention brings the following beneficial effects:

[0102] The drug small molecule SMILES string word vector matrix is ​​processed as an image, giving full play to the advantages of convolutional neural networks and filling a research gap in deep learning for drug-target binding affinity prediction;

[0103] 2) The present invention replaces the pooling layer in the traditional convolutional neural network with a large convolution kernel, thereby improving the ability of deep learning to understand the characteristics of drug molecules while maximizing the preservation of feature information without loss;

[0104] 3) A multi-scale two-dimensional convolutional neural network is used to extract the characteristics of drug small molecules, which can extract local and overall information of drug small molecules at multiple scales;

[0105] 4) Effectively improve the ability of deep learning in predicting small molecule-target binding affinity, which has important theoretical significance and practical value for drug discovery and optimization, side effect prediction, and disease research and treatment;

[0106] 5) Effectively improves the comprehensibility of the model and provides drug researchers with a more reliable algorithm model.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A drug-target affinity prediction method based on a multi-scale convolutional neural network, characterized in that: include, Step 1: Get the sequence of the target protein and the SMILES string of the drug small molecule. Step 2: Encode the SMILES string of the drug small molecule based on the word vector encoding to obtain the SMILES word vector matrix, and construct a multi-scale convolutional neural network. The multi-scale convolutional neural network is used to learn the SMILES word vector matrix and obtain the chemical and structural characteristics of the drug small molecule. The multi-scale convolutional neural network consists of N convolutional layers with different convolution kernel sizes. Step 3: After encoding the sequence of the target protein based on the word vector encoding, a protein word vector matrix is ​​obtained, and a one-dimensional convolutional neural network is constructed. The one-dimensional convolutional neural network is used to learn the protein word vector matrix and obtain the structural characteristics of the target protein. Step 4: The chemical and structural features of the drug small molecule and the structural features of the target protein are fused through a splicing operation, the fused features are represented as drug-target features, and an affinity prediction model is constructed. The affinity prediction model is used to learn the drug-target features and obtain the affinity prediction value between the drug and the protein.

2. The drug-target affinity prediction method based on a multi-scale convolutional neural network according to claim 1, characterized in that: The SMILES word vector matrix is ​​obtained by encoding the SMILES string of the drug small molecule based on the word vector encoding, including randomly initializing a V*d matrix as the word vector matrix according to the size V of the vocabulary and the embedding dimension d of each word, for each character in the SMILES string, converting it into a corresponding index, the index is the position of the character in the vocabulary, according to the index of the character, looking up the corresponding embedding vector from the word vector matrix, obtaining the embedding vector of each character in the SMILES string, and constructing the SMILES word vector matrix according to the embedding vector of each character.

3. The drug-target affinity prediction method based on a multi-scale convolutional neural network according to claim 1, characterized in that: The affinity prediction model includes three fully connected layers and two Dropout layers, the three fully connected layers are a first fully connected layer, a second fully connected layer and a third fully connected layer, the two Dropout layers include a first Dropout layer and a second Dropout layer, the first fully connected layer is used to accept drug-target features and pass them as input data to a hidden layer, wherein the first Dropout layer, the second fully connected layer and the second Dropout layer are used as hidden layers, and the hidden layer performs nonlinear transformation and feature extraction on the input data by applying an activation function; the third fully connected layer uses a ReLU activation function to predict the affinity prediction value of the drug and the protein and outputs it.

4. The drug-target affinity prediction method based on a multi-scale convolutional neural network according to claim 1, characterized in that: The N convolutional layers with different convolution kernel sizes are 9 convolutional layers with different convolution kernel sizes, wherein the 9 different convolution kernel sizes include 1×em_dim, 2×em_dim, 4×em_dim, 8×em_dim, 10×em_dim, 12×em_dim, 16×em_dim, 32×em_dim, and 64×em_dim, wherein em_dim represents the dimension of the SMILES word vector matrix.

5. The drug-target affinity prediction method based on a multi-scale convolutional neural network according to claim 1, characterized in that: The method of obtaining the affinity prediction value between the drug and the protein includes evaluating the prediction result according to formula (1), Among them, y i ,y j It represents the predicted value of drug-protein affinity obtained by the affinity prediction model. represents the actual affinity value of the drug to the protein, z represents a constant, and when y i >y j ,h is equal to 1, when y i =y j , h is equal to 0.5, otherwise the h function value is 0.

Images