Lead compound screening method based on graph convolutional neural network

Abstract

The present invention discloses a lead compound screening method based on a graph convolutional neural network, comprising: obtaining simplified molecular linear input specifications and molecular descriptor features of multiple compounds; constructing a molecular structure graph model of the compound, and using a graph convolutional neural network to obtain the structural features of the compound; merging the structural features and molecular descriptor features to obtain a merged feature; sequentially randomly modifying the numerical value of one dimension of the molecular descriptor feature in the merged feature to obtain a corresponding plurality of new merged features; inputting the new merged features into a multi-layer perceptron neural network to obtain predicted pharmacodynamic properties; calculating the error between the predicted pharmacodynamic properties and the actual pharmacodynamic properties, and ranking the importance of different dimensions of the molecular descriptor features according to the error value; and screening the lead compound to be screened based on the ranking results. The present invention can quickly screen out compounds with better specific pharmacodynamic properties from a large number of compounds, thereby reducing the cost of drug research and development.

Description

Technical Field

[0001] The present invention relates to the technical field of computer-aided drug design, and in particular to a lead compound screening method based on graph convolutional neural networks. Background Art

[0002] With the continuous development of artificial intelligence (AI) technology, its application areas have gradually expanded beyond computer science to other fields. Drug development, a cutting-edge field in pharmacy, is also experiencing increasing informatization, leading to the growing application of AI in this area. Currently, AI technology is being applied across drug discovery, preclinical research, and clinical research to enhance the precision, purpose, and efficiency of drug development. Drug feature screening, as part of drug discovery, is an early stage of drug development. Its purpose is to identify lead compounds with specific pharmacological activities from a large pool of compounds. Currently, for certain diseases, the lack of information about their therapeutic targets makes it impossible to directly identify active drug molecules that effectively treat the disease from numerous compound libraries. Traditional methods, such as random screening or high-throughput screening, are commonly used for drug analysis and screening. However, these methods often rely on experimental analysis of compound properties and then screen compounds based on these results, resulting in a relatively blind screening process that often consumes significant human, material, and time resources. Summary of the Invention

[0003] In order to solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a lead compound screening method based on graph convolutional neural network.

[0004] The technical solutions of the present invention are as follows:

[0005] A lead compound screening method based on graph convolutional neural network is provided, including:

[0006] Obtain multiple compounds and determine a simplified molecular linear input specification and molecular descriptor characteristics for each compound;

[0007] Based on the simplified molecular linear input specification of the compound, a molecular structure graph model of the compound is constructed, and a pre-trained graph convolutional neural network is used to extract features from the molecular structure graph model to obtain the structural characteristics of the compound;

[0008] Merging the structural features and molecular descriptor features of the compound to obtain a merged feature;

[0009] Starting from the first dimension to the last dimension of the molecular descriptor feature, the values ​​of one dimension of the molecular descriptor feature in the merged feature of the compound are randomly modified in sequence to obtain multiple new merged features corresponding to the compound, wherein the number of the new merged features is the same as the dimension of the molecular descriptor feature;

[0010] The newly combined features of the compounds are input into the pre-trained multi-layer perceptron neural network to obtain the predicted pharmacodynamic properties corresponding to each new combined feature;

[0011] Calculate the error between the predicted pharmacodynamic properties corresponding to different newly combined features and the actual pharmacodynamic properties of the compound, and rank the importance of different dimensions of the molecular descriptor features according to the size of the error value;

[0012] The lead compounds to be screened are screened based on the importance ranking results of different dimensions of the molecular descriptor features.

[0013] In some possible implementations, constructing a molecular structure graph model of a compound based on a simplified molecular linear input specification of the compound includes the following steps:

[0014] According to the simplified molecular linear input specification of the compound, the same type of nodes are used to represent the atoms in the compound, and different node eigenvalues ​​are used to represent different atom types;

[0015] According to the simplified molecular linear input specification of the compound, the same type of edge is used to represent the chemical bonds between atoms in the compound, and different edge weights are used to represent different chemical bond types;

[0016] According to the determined nodes, edges, node eigenvalues ​​and edge weights, a molecular structure graph model corresponding to the compound is constructed.

[0017] In some possible implementations, the structural features and molecular descriptor features of a compound are combined in the following manner:

[0018] The structural features of the compound are converted into an n-dimensional column vector through one-dimensional convolution, and the molecular descriptor features of the compound are converted into an m-dimensional column vector. The n-dimensional column vector corresponding to the structural features and the m-dimensional column vector corresponding to the molecular descriptor features are merged in an end-to-end manner to obtain an n+m-dimensional merged feature.

[0019] In some possible implementations, the graph convolutional neural network and the multilayer perceptron neural network are trained by the following steps:

[0020] Step S510, obtaining a training sample set including a plurality of training samples, wherein the training samples include a molecular structure graph model of a compound, a molecular descriptor feature, and a pharmacodynamic property corresponding to the compound;

[0021] Step S520: inputting the molecular structure graph model of each compound in the training sample into the initialized graph convolutional neural network in sequence to obtain the structural features corresponding to each compound, merging the structural features and molecular descriptor features corresponding to each compound to obtain a merged feature, and inputting each merged feature into the initialized multi-layer perceptron neural network in sequence to obtain the predicted pharmacodynamic properties corresponding to each compound;

[0022] Step S530, comparing the predicted pharmacodynamic properties corresponding to the compound in each training sample with the pharmacodynamic properties to determine the prediction accuracy of the graph convolutional neural network and the multilayer perceptron neural network;

[0023] Step S540, determine whether the prediction accuracy obtained at least twice in a row is greater than the preset accuracy threshold. If so, use the current graph convolutional neural network and multilayer perceptron neural network as the trained neural networks. If not, calculate the loss function, use the loss function to update the parameters of the graph convolutional neural network and the multilayer perceptron neural network, and return to step S520.

[0024] In some possible implementations, the graph convolutional neural network is any one of a residual graph convolutional neural network, a dense graph convolutional neural network, or a U-type graph neural network.

[0025] In some possible implementations, the loss function is:

[0026]

[0027] Where N represents the number of training samples, Indicates the predicted efficacy properties corresponding to the i-th training sample, y i Represents the pharmacodynamic properties corresponding to the i-th training sample.

[0028] In some possible implementations, the parameters of the neural network are updated using the following formula:

[0029]

[0030] Among them, φ represents the parameter set of the graph convolutional neural network and the multilayer perceptron neural network, and η represents the preset learning rate.

[0031] In some possible implementations, the error value is calculated using the following formula:

[0032]

[0033] Among them, Err(j) represents the error value corresponding to the j-th dimension feature of the molecular descriptor feature, M represents the number of compounds used for prediction, Y represents the predicted pharmacodynamic properties corresponding to the new combined features of the i-th compound obtained by modifying the j-th dimension feature of the molecular descriptor feature, i It represents the pharmacodynamic properties corresponding to the i-th compound, and m represents the characteristic dimension of the molecular descriptor.

[0034] In some possible implementations, ranking the importance of different dimensions of molecular descriptor features according to the error values ​​includes:

[0035] Determine the error values ​​corresponding to different dimensions of the molecular descriptor features;

[0036] The importance of different dimensions of molecular descriptor features is ranked in order of error value from small to large.

[0037] In some possible implementations, the method of screening the lead compounds to be screened according to the importance ranking results of different dimensions of the molecular descriptor features includes the following steps:

[0038] Select the dimensions of the molecular descriptor features that are in the top n importance, and calculate the distribution range of the corresponding top n dimensional features in all compounds used for prediction, and obtain the distribution mean, distribution minimum and distribution maximum of the top n dimensional features;

[0039] Select the lead compound that satisfies the following formula from all the lead compounds to be screened;

[0040]

[0041] Among them, x i Indicates the value of the i-th dimension feature among the first n dimensional features corresponding to the lead compound to be screened, μ i Represents the distribution mean of the i-th dimension feature among the first n dimensional features, L represents the preset threshold, s i,max Indicates the maximum value of the distribution of the i-th dimension feature among the first n dimension features, s i,min Indicates the minimum value of the distribution of the i-th dimension feature among the first n dimensional features.

[0042] The main advantages of the technical solution of the present invention are as follows:

[0043] The lead compound screening method based on graph convolutional neural network of the present invention extracts the molecular structure characteristics of the compound by using graph convolutional neural network, modifies the different dimensional characteristics of the molecular descriptor of the compound, and predicts the pharmacodynamic properties corresponding to the compound after the modified molecular descriptor characteristics through multi-layer perceptron neural network. The degree of influence of the different dimensional characteristics of the molecular descriptor on the pharmacodynamic properties is determined according to the prediction results, and then the lead compound is screened according to the degree of influence of the different dimensional characteristics of the molecular descriptor on the pharmacodynamic properties and the molecular descriptor feature information of the lead compound to be screened. It is possible to screen out compounds with better specific pharmacodynamic properties from a large number of compounds, significantly improve the speed of drug development, and reduce the time cost and labor cost of drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] 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 only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a flowchart of a lead compound screening method based on a graph convolutional neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0047] The technical solutions provided by the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0048] See also Figure 1 One embodiment of the present invention provides a method for screening lead compounds based on a graph convolutional neural network, the method comprising the following steps:

[0049] Step S100 , obtaining a plurality of compounds, and determining a simplified molecular linear input specification and a molecular descriptor feature of each compound;

[0050] Step S200: constructing a molecular structure graph model of the compound based on a simplified molecular linear input specification of the compound, and performing feature extraction on the molecular structure graph model using a pre-trained graph convolutional neural network to obtain structural features of the compound;

[0051] Step S300, merging the structural features and molecular descriptor features of the compound to obtain a merged feature;

[0052] Step S400, starting from the first dimension to the last dimension of the molecular descriptor feature, the values ​​of one dimension of the molecular descriptor feature in the merged feature of the compound are randomly modified in sequence to obtain multiple new merged features corresponding to the compound, wherein the number of the new merged features is the same as the dimension of the molecular descriptor feature;

[0053] Step S500 , inputting the newly combined features of the compound into a pre-trained multi-layer perceptron neural network to obtain the predicted pharmacodynamic properties corresponding to each newly combined feature;

[0054] Step S600, calculating the error between the predicted pharmacodynamic properties corresponding to different newly combined features and the actual pharmacodynamic properties of the compound, and ranking the importance of different dimensions of the molecular descriptor features according to the error values;

[0055] Step S700 , screening the lead compounds to be screened according to the importance ranking results of different dimensions of the molecular descriptor features.

[0056] The following is a detailed description of the steps and principles of the lead compound screening method based on graph convolutional neural network provided by one embodiment of the present invention:

[0057] Step S100 : obtaining a plurality of compounds and determining a simplified molecular linear input specification and a molecular descriptor feature of each compound.

[0058] In one embodiment of the present invention, simplified molecular input line entry systems (SMILES) and molecular descriptor features for multiple compounds can be collected through various methods. The multiple compounds can be various existing compounds with known pharmacodynamic properties, or a certain number of compounds can be selected from a given library of lead compounds to be screened and tested to determine their pharmacodynamic properties.

[0059] In step S200 , a molecular structure graph model of the compound is constructed based on a simplified molecular linear input specification of the compound, and a pre-trained graph convolutional neural network is used to extract features from the molecular structure graph model to obtain structural features of the compound.

[0060] Since the molecular structure of a compound consists of atoms connected by chemical bonds, a graphical model is an unstructured data structure capable of storing and processing irregular data. Therefore, the molecular structure of a compound can be modeled using a graphical model.

[0061] The key to building a graph model is determining the nodes and edges within it. A graph G = (V, E) is defined, where V represents the graph's node set, a collection of physical nodes that can store node characteristics. E represents the graph's edge set, which describes the connections between nodes and can represent different edge connections through edge weights. To better describe the connections between nodes and edges in a graph, an adjacency matrix A and a degree matrix D are typically used. The graph's adjacency matrix represents the adjacent relationships between nodes, while the degree matrix is ​​a matrix composed of the degrees of each node. The degree of a node represents the number of edges associated with that node.

[0062] In one embodiment of the present invention, an undirected isomorphic connected graph is used to model the molecular structure graph model of the compound.

[0063] Specifically, based on the simplified molecular linear input specification of the compound, a molecular structure graph model of the compound is constructed, including the following steps:

[0064] Step S210 , according to the simplified molecular linear input specification of the compound, using the same type of nodes to represent atoms in the compound, and using different node eigenvalues ​​to represent different atom types;

[0065] Step S220 , according to the simplified molecular linear input specification of the compound, using the same type of edge to represent the chemical bonds between atoms in the compound, and using different edge weights to represent different chemical bond types;

[0066] Step S230 : constructing a molecular structure graph model corresponding to the compound based on the determined nodes, edges, node feature values, and edge weights.

[0067] In order to improve the efficiency of building graphical models, open source chemical information software packages, such as Python's RDKit toolkit, can be used to assist in construction.

[0068] Furthermore, a pre-trained graph convolutional neural network is used to extract features from the molecular structure graph model of the compound obtained based on the above method to obtain the structural characteristics of the compound.

[0069] Graph convolutional neural network is a kind of neural network introduced by convolutional neural network into graph processing. It can extract the neighborhood information of nodes and pass it to the next layer through learnable parameters. It has the advantages of learnable parameters of neural network.

[0070] The key processing steps of graph convolutional neural networks can be described as follows:

[0071]

[0072] in, Represents the features of the l+1th layer of the graph convolutional neural network after transmission, It represents the symmetric matrix after adding the self-loop. Represents the adjacency matrix after adding the self-loop, represents the features of the lth layer of the graph convolutional neural network, represents the parameters of the lth layer of the graph convolutional neural network, and σ(·) represents the nonlinear activation function.

[0073] The adjacency matrix with self-loop is specifically expressed as:

[0074]

[0075] in, Represents the adjacency matrix after adding the self-loop, I N Represents the identity matrix.

[0076] The parameters of the graph convolutional neural network are determined through training. In one embodiment of the present invention, the training process of the graph convolutional neural network is carried out together with the subsequent training process of the multi-layer perceptron neural network. The input of the graph convolutional neural network is the molecular structure graph model of the compound, and the output is the structural characteristics of the compound.

[0077] Furthermore, in one embodiment of the present invention, in order to fully extract the structural feature information of the compound, the graph convolutional neural network can be a multi-layer graph convolutional neural network. The multi-layer graph convolutional neural network can be any one of a residual graph convolutional neural network (Residual Graph Convolutional Network, ResGCN), a dense graph convolutional neural network (DenseGCN), and a U-type graph neural network (Graph U-Net).

[0078] Step S300: Merge the structural features and molecular descriptor features of the compound to obtain a merged feature.

[0079] In one embodiment of the present invention, the structural features of the compound are converted into an n-dimensional column vector through one-dimensional convolution, and the molecular descriptor features of the compound are converted into an m-dimensional column vector. The n-dimensional column vector corresponding to the structural features and the m-dimensional column vector corresponding to the molecular descriptor features are merged in an end-to-end manner to obtain an n+m-dimensional merged feature, wherein the first n dimensions of the merged feature are the structural features of the compound, and the last m dimensions are the molecular descriptor features of the compound.

[0080] Step S400 , starting from the first dimension to the last dimension of the molecular descriptor feature, the values ​​of one dimension of the molecular descriptor feature in the merged feature of the compound are randomly modified in sequence to obtain multiple new merged features corresponding to the compound, wherein the number of the new merged features is the same as the dimension of the molecular descriptor feature.

[0081] In practical applications, different dimensions of a compound's molecular descriptor features have different effects on the compound's pharmacodynamic properties, with some features having a greater impact on the pharmacodynamic properties, while others have a smaller or negligible impact. In one embodiment of the present invention, a feature importance ranking method is used to determine the relative impact of different dimensions of a compound's molecular descriptor features on the compound's pharmacodynamic properties, thereby screening compounds based on the determined relative impact of different dimensions of the molecular descriptor features on the pharmacodynamic properties.

[0082] Specifically, in one embodiment of the present invention, based on the merged features obtained in step 300, the values ​​of one dimension of the molecular descriptor features in the merged features of the compound are randomly modified in sequence, starting from the first dimension to the last dimension of the molecular descriptor features, to obtain multiple new merged features corresponding to the compound.

[0083] Compared with the merged feature, the new merged feature obtained based on the above method has a different value of one dimension in the subsequent m-dimensional molecular descriptor feature.

[0084] In step S500 , the newly combined features of the compound are respectively input into a pre-trained multi-layer perceptron neural network to obtain the predicted pharmacodynamic properties corresponding to each newly combined feature.

[0085] In one embodiment of the present invention, a multilayer perceptron neural network and a graph convolutional neural network are trained together. The input of the graph convolutional neural network is the molecular structure graph model of the compound, and the output is the structural characteristics of the compound. The input of the multilayer perceptron neural network is the combined characteristics of the compound, and the output is the pharmacodynamic properties of the compound.

[0086] Compared with graph convolutional neural networks, multilayer perceptron neural networks lack the adjacency matrix and degree matrix in the forward propagation process. They are fully connected neural networks that can obtain a suitable model through parameter learning, thereby predicting the specified pharmacodynamic properties. The forward propagation process of each layer of the multilayer perceptron neural network can be expressed as:

[0087]

[0088] in, Represents the features of the l+1th layer of the multilayer perceptron neural network, Represents the features of the lth layer of the multilayer perceptron neural network, represents the parameters of the lth layer of the multilayer perceptron neural network, and σ represents the nonlinear activation function.

[0089] Specifically, in one embodiment of the present invention, the graph convolutional neural network and the multi-layer perceptron can be trained using the following steps:

[0090] Step S510, obtaining a training sample set including a plurality of training samples;

[0091] In one embodiment of the present invention, training samples include molecular structure graph models of compounds, molecular descriptor features, and corresponding pharmacodynamic properties of the compounds. Training samples can be collected through various currently known methods, or by selecting a certain number of compounds from a given library of lead compounds to be screened and conducting experiments on these selected compounds to determine their pharmacodynamic properties, thereby obtaining a certain number of training samples.

[0092] The pharmacological properties may be pre-specified based on actual conditions, such as utilization rate or therapeutic effect within a certain period of time.

[0093] Step S520: inputting the molecular structure graph model of each compound in the training sample into the initialized graph convolutional neural network in sequence to obtain the structural features corresponding to each compound, merging the structural features and molecular descriptor features corresponding to each compound to obtain a merged feature, and inputting each merged feature into the initialized multi-layer perceptron neural network in sequence to obtain the predicted pharmacodynamic properties corresponding to each compound;

[0094] In one embodiment of the present invention, the molecular structure graph model of the compound in the training sample is input into the input end of the initialized graph convolutional neural network, sequentially processed by the parameters of each layer in the initialized graph convolutional neural network, and output from the output end of the initialized graph convolutional neural network. The information output from the output end is the structural feature corresponding to the compound. The merged feature is input into the input end of the initialized multilayer perceptron neural network, sequentially processed by the parameters of each layer in the initialized multilayer perceptron neural network, and output from the output end of the initialized multilayer perceptron neural network. The information output from the output end is the predicted pharmacodynamic properties corresponding to the compound. The initialized graph convolutional neural network and the initialized multilayer perceptron neural network can be untrained neural networks or untrained neural networks, and each layer thereof is provided with initialization parameters, which can be continuously updated and adjusted during the training process of the neural network.

[0095] Step S530, comparing the predicted pharmacodynamic properties corresponding to the compound in each training sample with the pharmacodynamic properties of the compound in the training sample to determine the prediction accuracy of the graph convolutional neural network and the multilayer perceptron neural network;

[0096] In one embodiment of the present invention, the ratio of the difference between the predicted pharmacodynamic property and the pharmacodynamic property corresponding to each training sample to the pharmacodynamic property can be calculated, and the average value of all ratios can be used as the prediction accuracy.

[0097] Step S540, determine whether the prediction accuracy obtained at least twice in a row is greater than the preset accuracy threshold. If so, use the current graph convolutional neural network and multilayer perceptron neural network as the trained neural networks. If not, calculate the loss function, use the loss function to update the parameters of the graph convolutional neural network and the multilayer perceptron neural network, and return to step S520.

[0098] In one embodiment of the present invention, the mean absolute error can be used as the loss function, which can be specifically expressed as:

[0099]

[0100] Where N represents the number of training samples, Indicates the predicted efficacy properties corresponding to the i-th training sample, y i It represents the pharmacodynamic properties corresponding to the i-th training sample. The number of training samples N can be determined according to actual training requirements.

[0101] In one embodiment of the present invention, the parameters of the neural network are optimized and updated using a gradient descent method. Specifically, the chain rule is first used to derive the parameters using a loss function, and then the parameters are updated using the derivative results and a preset learning rate.

[0102] Specifically, the following formula can be used to update the parameters of the neural network:

[0103]

[0104] Among them, φ represents the parameter set of the graph convolutional neural network and the multilayer perceptron neural network, and η represents the preset learning rate.

[0105] Step S600 , calculating the errors between the predicted pharmacodynamic properties corresponding to different newly combined features and the actual pharmacodynamic properties of the compound, and ranking the importance of different dimensions of the molecular descriptor features according to the error values.

[0106] Since the different dimensional features of the molecular descriptor features of the compound have different effects on the pharmacodynamic properties of the compound, some features have a greater effect on the pharmacodynamic properties, while some features have a smaller or negligible effect on the pharmacodynamic properties. In one embodiment of the present invention, by calculating the error between the predicted pharmacodynamic properties corresponding to different newly merged features and the actual pharmacodynamic properties of the compound, the influence of the different dimensional features of the molecular descriptor features on the pharmacodynamic properties is determined according to the size of the error value, and then the importance order of the different dimensions of the molecular descriptor features is determined. Specifically, the larger the error value, the greater the influence of the corresponding dimensional features of the molecular descriptor on the pharmacodynamic properties, and the higher its importance, that is, the higher the order; the smaller the error value, the smaller the influence of the corresponding dimensional features of the molecular descriptor on the pharmacodynamic properties, and the lower its importance, that is, the lower the order.

[0107] In one embodiment of the present invention, for each dimension of the molecular descriptor feature, the error value can be calculated using the following formula:

[0108]

[0109] Among them, Err(j) represents the error value corresponding to the j-th dimension feature of the molecular descriptor feature, M represents the number of compounds used for prediction, Y represents the predicted pharmacodynamic properties corresponding to the new combined features of the i-th compound obtained by modifying the j-th dimension feature of the molecular descriptor feature, i It represents the pharmacodynamic properties corresponding to the i-th compound, and m represents the characteristic dimension of the molecular descriptor.

[0110] Among them, predicting drug efficacy properties It can be expressed as:

[0111]

[0112] f(·) represents the mapping relationship between graph convolutional neural network and multi-layer perceptron neural network. It represents the molecular descriptor feature of the i-th compound obtained by modifying the j-th dimension feature of the molecular descriptor feature, s i represents the molecular structure graph model of the i-th compound, and θ represents the parameter set of the graph convolutional neural network and the multilayer perceptron neural network.

[0113] Furthermore, the importance of different dimensions of the molecular descriptor features is ranked according to the calculated error values ​​corresponding to the different dimensional features of the molecular descriptor features; when the error value is smaller, the importance order of the dimension corresponding to the error value is higher.

[0114] Step S700 , screening the lead compounds to be screened according to the importance ranking results of different dimensions of the molecular descriptor features.

[0115] In one embodiment of the present invention, the lead compounds to be screened are screened according to the importance ranking results of different dimensions of the molecular descriptor features, including the following steps:

[0116] Select the dimensions of the molecular descriptor features that are in the top n importance, and calculate the distribution ranges of the corresponding top n dimensional features in all compounds used for prediction in step S500 to obtain the distribution mean, distribution minimum, and distribution maximum of the top n dimensional features;

[0117] Select the lead compound that satisfies the following formula from all the lead compounds to be screened;

[0118]

[0119] Among them, x i Indicates the value of the i-th dimension feature among the first n dimensional features corresponding to the lead compound to be screened, μ i Represents the distribution mean of the i-th dimension feature among the first n dimensional features, L represents the preset threshold, s i,max Indicates the maximum value of the distribution of the i-th dimension feature among the first n dimension features, s i,min Indicates the minimum value of the distribution of the i-th dimension feature among the first n dimensional features.

[0120] The above screening method can accurately screen out lead compounds with specific pharmacological properties from a large number of lead compounds to be screened, significantly improving the screening quality and reducing the screening cost.

[0121] In one embodiment of the present invention, the specific value of n can be set according to actual conditions. For example, the dimension of the molecular descriptor feature whose corresponding error value is less than a preset error threshold is used as the dimension of the selected molecular descriptor feature.

[0122] The lead compound screening method based on graph convolutional neural network provided by one embodiment of the present invention extracts the molecular structure characteristics of the compound by using graph convolutional neural network, modifies the different dimensional characteristics of the molecular descriptor of the compound, and predicts the pharmacodynamic properties corresponding to the compound after the modified molecular descriptor characteristics through multi-layer perceptron neural network. The degree of influence of the different dimensional characteristics of the molecular descriptor on the pharmacodynamic properties is determined according to the prediction results, and then the lead compound is screened according to the degree of influence of the different dimensional characteristics of the molecular descriptor on the pharmacodynamic properties and the molecular descriptor feature information of the lead compound to be screened. It is possible to screen out compounds with better specific pharmacodynamic properties from a large number of compounds, significantly improve the speed of drug development, and reduce the time cost and labor cost of drug development.

[0123] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In addition, "front", "back", "left", "right", "upper" and "lower" in this document are all referenced to the placement states shown in the accompanying drawings.

[0124] 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 make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A lead compound screening method based on graph convolutional neural network, characterized in that: include: Obtain multiple compounds and determine a simplified molecular linear input specification and molecular descriptor characteristics for each compound; Based on the simplified molecular linear input specification of the compound, a molecular structure graph model of the compound is constructed, and a pre-trained graph convolutional neural network is used to extract features from the molecular structure graph model to obtain the structural characteristics of the compound; Merging the structural features and molecular descriptor features of the compound to obtain a merged feature; Starting from the first dimension to the last dimension of the molecular descriptor feature, the values ​​of one dimension of the molecular descriptor feature in the merged feature of the compound are randomly modified in sequence to obtain multiple new merged features corresponding to the compound, wherein the number of the new merged features is the same as the dimension of the molecular descriptor feature; The newly combined features of the compounds are input into the pre-trained multi-layer perceptron neural network to obtain the predicted pharmacodynamic properties corresponding to each new combined feature; Calculate the error between the predicted pharmacodynamic properties corresponding to different newly combined features and the actual pharmacodynamic properties of the compound, and rank the importance of different dimensions of the molecular descriptor features according to the size of the error value; Based on the importance ranking results of different dimensions of molecular descriptor features, the lead compounds to be screened are screened; The method of screening the lead compounds to be screened according to the importance ranking results of different dimensions of the molecular descriptor features comprises the following steps: Select the dimensions of the molecular descriptor features that are in the top n importance, and calculate the distribution range of the corresponding top n dimensional features in all compounds used for prediction, and obtain the distribution mean, distribution minimum and distribution maximum of the top n dimensional features; Select the lead compound that satisfies the following formula from all the lead compounds to be screened; Among them, x i Indicates the value of the i-th dimension feature among the first n dimensional features corresponding to the lead compound to be screened, μ i Represents the distribution mean of the i-th dimension feature among the first n dimensional features, L represents the preset threshold, s i,max Indicates the maximum value of the distribution of the i-th dimension feature among the first n dimension features, s i,min Indicates the minimum value of the distribution of the i-th dimension feature among the first n dimensional features.

2. The lead compound screening method based on graph convolutional neural network according to claim 1, characterized in that The method of constructing a molecular structure graph model of a compound based on a simplified molecular linear input specification of the compound includes the following steps: According to the simplified molecular linear input specification of the compound, the same type of nodes are used to represent the atoms in the compound, and different node eigenvalues ​​are used to represent different atom types; According to the simplified molecular linear input specification of the compound, the same type of edge is used to represent the chemical bonds between atoms in the compound, and different edge weights are used to represent different chemical bond types; According to the determined nodes, edges, node eigenvalues ​​and edge weights, a molecular structure graph model corresponding to the compound is constructed.

3. The lead compound screening method based on graph convolutional neural network according to claim 1, characterized in that: The structural features and molecular descriptor features of the compound are combined in the following way: The structural features of the compound are converted into an n-dimensional column vector through one-dimensional convolution, and the molecular descriptor features of the compound are converted into an m-dimensional column vector. The n-dimensional column vector corresponding to the structural features and the m-dimensional column vector corresponding to the molecular descriptor features are merged in an end-to-end manner to obtain an n+m-dimensional merged feature.

4. The lead compound screening method based on graph convolutional neural network according to claim 1, characterized in that: The graph convolutional neural network and the multilayer perceptron neural network are trained by the following steps: Step S510, obtaining a training sample set including a plurality of training samples, wherein the training samples include a molecular structure graph model of a compound, a molecular descriptor feature, and a pharmacodynamic property corresponding to the compound; Step S520: inputting the molecular structure graph model of each compound in the training sample into the initialized graph convolutional neural network in sequence to obtain the structural features corresponding to each compound, merging the structural features and molecular descriptor features corresponding to each compound to obtain a merged feature, and inputting each merged feature into the initialized multi-layer perceptron neural network in sequence to obtain the predicted pharmacodynamic properties corresponding to each compound; Step S530, comparing the predicted pharmacodynamic properties corresponding to the compound in each training sample with the pharmacodynamic properties to determine the prediction accuracy of the graph convolutional neural network and the multilayer perceptron neural network; Step S540, determine whether the prediction accuracy obtained at least twice in a row is greater than the preset accuracy threshold. If so, use the current graph convolutional neural network and multilayer perceptron neural network as the trained neural networks. If not, calculate the loss function, use the loss function to update the parameters of the graph convolutional neural network and the multilayer perceptron neural network, and return to step S520.

5. The lead compound screening method based on graph convolutional neural network according to claim 4, characterized in that: The graph convolutional neural network is any one of a residual graph convolutional neural network, a dense graph convolutional neural network, or a U-type graph neural network.

6. The lead compound screening method based on graph convolutional neural network according to claim 4, characterized in that: The loss function is: Where N represents the number of training samples, Indicates the predicted efficacy properties corresponding to the i-th training sample, y i Represents the pharmacodynamic properties corresponding to the i-th training sample.

7. The lead compound screening method based on graph convolutional neural network according to claim 6, characterized in that: The parameters of the neural network are updated using the following formula: Among them, φ represents the parameter set of the graph convolutional neural network and the multilayer perceptron neural network, and η represents the preset learning rate.

8. The lead compound screening method based on graph convolutional neural network according to claim 1, characterized in that: The error value is calculated using the following formula: Among them, Err(j) represents the error value corresponding to the j-th dimension feature of the molecular descriptor feature, M represents the number of compounds used for prediction, Y represents the predicted pharmacodynamic properties corresponding to the new combined features of the i-th compound obtained by modifying the j-th dimension feature of the molecular descriptor feature, i It represents the pharmacodynamic properties corresponding to the i-th compound, and m represents the characteristic dimension of the molecular descriptor.

9. The lead compound screening method based on graph convolutional neural network according to claim 1, characterized in that: The ordering of the importance of different dimensions of the molecular descriptor features according to the error value includes: Determine the error values ​​corresponding to different dimensions of the molecular descriptor features; The importance of different dimensions of molecular descriptor features is ranked in order of error value from small to large.

Images