A method and device for predicting chromatographic retention index based on graph convolutional network
Through the method based on graph convolution network, the structural matrix of the compound is extracted and encoded, which solves the problem of harsh input parameters and limited coverage of the QSRR model in the prediction of chromatographic retention time, and achieves efficient and accurate prediction of the compound.
Patent Information
- Application Number
- CN202111486185.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-07
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-12-07
AI Technical Summary
The existing QSRR models have problems such as strict input parameters, limited range of coverage compounds and limited feature extraction depth in the prediction of chromatographic retention time, resulting in limited prediction effects.
Using a graph convolutional network-based method, the structural matrix is extracted through the SMILES information of the compound, encoded using a stacked autoencoder, and the encoding matrix is input into the graph convolutional neural network model to predict the chromatographic retention index.
Accurate prediction of compounds of different sizes is achieved, the molecular descriptor calculation process is simplified, the prediction accuracy and coverage are improved, and the computing resource requirements are reduced.
Smart Images

Figure CN114121178B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of chromatography application, and in particular to a method and device for predicting a chromatographic retention index based on a graph convolutional network. Background Art
[0002] The research on convolutional neural networks began in the 1980s and 1990s, with time delay networks and LeNet-5 being the earliest convolutional neural networks. After the 21st century, with the introduction of deep learning theory and the improvement of numerical computing equipment, convolutional neural networks have developed rapidly and have been applied to computer vision, natural language processing and other fields.
[0003] Convolutional neural networks are constructed by imitating the visual perception mechanism of organisms and can perform supervised and unsupervised learning. The sharing of convolution kernel parameters within the hidden layer and the sparsity of inter-layer connections enable convolutional neural networks to learn grid-like topology features such as pixels and audio with a small amount of computation, with stable results and without additional feature engineering requirements for the data.
[0004] The BP (Back Propagation) network was proposed by a team of scientists led by Rumelhart and McClelland in 1986. It is a multi-layer feedforward network trained by the error back propagation algorithm and is one of the most widely used neural network models. The BP network can learn and store a large number of input-output pattern mapping relationships without revealing the mathematical equations describing this mapping relationship in advance. Its learning rule is to use the steepest descent method to continuously adjust the network weights and thresholds through back propagation to minimize the sum of squared errors of the network.
[0005] The quantitative structure-chromatographic retention relationship model (QSRR) uses computer-aided tools to perform correlation analysis between the structural information of a compound and the chromatographic retention index, and is often used to optimize chromatographic separation conditions and predict chromatographic retention indices. The computer-aided tools used for QSRR model research mainly include multiple linear regression (SVM), support vector machine (SVM), artificial neural network (ANN), etc. Some studies have used deep learning models such as GNN, LSTM, GRU, etc.
[0006] Most of the models used have shortcomings such as stringent requirements on input parameters, limited range of compound types covered, and limited feature extraction depth.
[0007] The input parameters are demanding: The input parameters of most QSRR models are molecular descriptors that describe the structural characteristics of compounds, which need to be calculated in advance using molecular descriptor-related software. The number and types of molecular descriptors calculated by different software vary greatly, and the calculation of molecular descriptors takes a long time and the steps are cumbersome.
[0008] Limited coverage of compound types: Different descriptor calculation software is applicable to a limited number of compound types, which can easily lead to missing and erroneous molecular descriptor data of some compounds, limiting the scope of application of predictable compounds.
[0009] Limited feature extraction depth: The molecular descriptors on which the model is based may only reflect certain aspects of the structural properties of the compound and cannot fundamentally and comprehensively predict the structural properties of the compound. Summary of the invention
[0010] A technical problem to be solved by an embodiment of the present invention is to provide a method and device for predicting a chromatographic retention index based on a graph convolutional network, so as to solve the problem of chromatographic retention time prediction existing in various models in the prior art.
[0011] According to one aspect of the present invention, a method for predicting a chromatographic retention index based on a graph convolutional network is provided, the method comprising:
[0012] Extract the structural information of the compound based on the SMLIES information of the compound and convert it into a structural matrix;
[0013] The structure matrix is trained and encoded according to a stacked autoencoder to obtain an encoding matrix;
[0014] The encoding matrix is input into the graph convolutional neural network model trained according to the training group and the test group, and the chromatographic retention index is obtained as an output.
[0015] The step of extracting the structural information of the compound according to the SMLIES information of the compound and converting it into a structural matrix includes:
[0016] Parse the SMLIES input of the compound into a syntax tree to obtain nodes and branches at each level;
[0017] Counting the nodes to achieve atomic extraction;
[0018] Parsing the branched chains to extract bonds, subchains and ring structures;
[0019] Based on the atomic, bond, subchain and ring structure information, a structural matrix is obtained, including but not limited to: domain matrix A, feature matrix X, degree matrix D, and Laplace matrix L.
[0020] The stacked autoencoder comprises:
[0021] The input of the autoencoder is [M,M], where M is the number of atoms in the input compound;
[0022] The output of the autoencoder is [N,N], where N is the number of atoms in the output compound;
[0023] Where, M>>N, and N>50;
[0024] For atoms whose atomic number is less than M, the missing positions are filled with 0.
[0025] The stacked autoencoder is divided into four unrelated sub-encoders with different initial parameters but the same structure, and each sub-encoder is trained and encoded for a different matrix.
[0026] The sub-encoder consists of two parts: encoding and decoding;
[0027] The encoding part consists of three neural networks stacked together, with the input being the structure matrix and the output being the encoding matrix;
[0028] The decoding part consists of three stacked neural networks, with the input being the encoding matrix and the output being the structure matrix.
[0029] The graph convolutional neural network model is trained as follows:
[0030] Extracting information from the compounds in the training group and the test group to form a structure matrix;
[0031] Output the encoding matrix through stacking the autoencoder according to the structure matrix;
[0032] The graph convolutional neural network model is input according to the encoding matrix, and the graph convolutional neural network model is adjusted and trained according to the output chromatographic retention index compared with the actual chromatographic retention index of the training group and the test group.
[0033] The graph convolutional neural network model includes:
[0034] The graph convolutional neural network model consists of three graph convolutional layers, three activation layers and three fully connected layers; specifically, it includes:
[0035] Input layer: Normalize the data in the following way:
[0036]
[0037] Among them, X in For each input value of the feature matrix X, X max For X in Maximum value; X min For X in The minimum value in
[0038] Graph convolution layer: The structure of each layer is as follows:
[0039]
[0040] Among them, D is the degree matrix, L is the Laplace matrix, A is the domain matrix, X is the feature matrix after the input layer, H is the input matrix, and W is the convolution layer parameter matrix;
[0041] Activation layer: Lakerelu function is used, and the structure is as follows:
[0042] H out =Lakerelu(H in )
[0043] Fully connected layer: The first layer has N neurons, the second layer has N / 2 neurons, and the third layer has N / 4 neurons.
[0044] Linear output layer: The structure is as follows:
[0045] H out =W1*H in +W2
[0046] Among them, H in is a [n,1] structure matrix, W1 is a [1,n] structure matrix, and W2 is a scalar.
[0047] According to another aspect of the present invention, a chromatographic retention index prediction device based on a graph convolutional network is provided, comprising:
[0048] The structural information extraction module is used to extract the structural information of the compound according to the SMLIES information of the compound and convert it into a structural matrix;
[0049] A stacked autoencoder is used to train and encode the structure matrix to obtain an encoding matrix;
[0050] A graph convolutional neural network model is used to input the encoding matrix into the graph convolutional neural network model trained according to the training group and the test group, and output a chromatographic retention index.
[0051] The structural information extraction module is specifically used for:
[0052] Parse the SMLIES input of the compound into a syntax tree to obtain nodes and branches at each level;
[0053] Counting the nodes to achieve atomic extraction;
[0054] Parsing the branched chains to extract bonds, subchains and ring structures;
[0055] Based on the atomic, bond, subchain and ring structure information, a structural matrix is obtained, including but not limited to: domain matrix A, feature matrix X, degree matrix D, and Laplace matrix L.
[0056] The stacked autoencoder is divided into four unrelated sub-encoders with different initial parameters but the same structure, and each sub-encoder is trained and encoded for a different matrix;
[0057] The graph convolutional neural network model consists of three graph convolutional layers, three activation layers and three fully connected layers.
[0058] According to another aspect of the present invention, there is provided an electronic device, comprising:
[0059] Memory for storing computer programs;
[0060] The processor is used to execute the computer program stored in the memory, and when the computer program is executed, any of the above methods is implemented.
[0061] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, any of the above-mentioned methods is implemented.
[0062] Based on the chromatographic retention index prediction scheme based on graph convolutional network provided by the above embodiment of the present invention, the structural information of the compound is extracted through the SMLIES information of the compound and converted into a structure matrix; the structure matrix is trained and encoded according to the stacked autoencoder to obtain a coding matrix; the coding matrix is input into the graph convolutional neural network model trained according to the training group and the test group, and the chromatographic retention index is obtained as output.
[0063] This embodiment extracts the topological structure and related information of the compound molecule from SMILES, which is reliable and accurate. For chemical molecules of different sizes, they are uniformly encoded into a fixed-size encoding matrix after entering the stacked autoencoder, which is convenient for subsequent convolution operations. In the unified operation process of chemical molecules of different sizes, the stacked autoencoder implemented by the BP neural network is more accurate, fast, and resource-saving. The newly designed graph convolution layer replaces the ordinary convolution layer, and the designed graph convolution network far exceeds the ordinary convolution layer in the processing of topological information. Compared with deep learning models such as LSTM and GNN, the learning model is small in size, easy to implement, and has more controllable internal parameters. It has application scenarios such as edge computing device deployment and more compound property prediction. By optimizing the model parameters, the same or higher prediction accuracy can be achieved with less computing resources.
[0064] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] The accompanying drawings, which constitute a part of the specification, illustrate embodiments of the present invention and, together with the description, serve to explain the principles of the present invention.
[0066] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.
[0067] The present invention may be more clearly understood from the following detailed description with reference to the accompanying drawings, in which:
[0068] Figure 1 A schematic diagram of the principle of a chromatographic retention index prediction method based on a graph convolutional network provided in one embodiment of the present invention.
[0069] Figure 2 A schematic diagram of the functional structure of a structural information extraction module provided by an embodiment of the present invention.
[0070] Figure 3 A schematic diagram of a domain matrix structure provided for one embodiment of the present invention.
[0071] Figure 4 A schematic diagram of a feature matrix structure provided for one embodiment of the present invention.
[0072] Figure 5 A schematic diagram of a degree matrix structure provided by an embodiment of the present invention.
[0073] Figure 6 A schematic diagram of a structural information extraction workflow provided by an embodiment of the present invention.
[0074] Figure 7 A schematic diagram of the overall structure of a stacked autoencoder provided for one embodiment of the present invention.
[0075] Figure 8 A schematic diagram of a sub-encoder structure provided for one embodiment of the present invention.
[0076] Fig. 9 A schematic diagram of the graph convolution model structure provided for one embodiment of the present invention.
[0077] Fig.10 A schematic diagram of a fully connected layer structure provided for one embodiment of the present invention.
[0078] Fig.11 A schematic diagram of the implementation principle of a chromatographic retention index prediction solution based on a graph convolutional network provided in one embodiment of the present invention.
[0079] Fig.12 A schematic diagram of the structure of a chromatographic retention index prediction device based on a graph convolutional network provided in one embodiment of the present invention. DETAILED DESCRIPTION
[0080] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention unless otherwise specifically stated.
[0081] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses.
[0082] The deep learning model solutions with SMILES string as input mainly include LSTM sequence prediction, GNN adversarial network model, etc. The main purpose is to analyze the SMILES string as natural language, discover the structural features of the SMILES string through context association, and then predict the chemical properties of the compound.
[0083] LSTM sequence prediction is to input SMILES into the LSTM network in sequence form to train and discover deep features. GNN adversarial network model is to input SMILES into the LSTM network in sequence form to train and discover deep features.
[0084] The SMILES string of the same compound has multiple correctly spelled formats. Although they all correspond to the same structural information, they will seriously interfere with the LSTM sequence's correct understanding of the input SMILES string, causing model training failure. The molecular formulas of different compounds vary in length, so the length of the SMILES string varies greatly, far exceeding the length range of ordinary natural language. Therefore, the adaptability of the model is very high, which can easily lead to training failure. The existing model is complex, consumes huge computing resources, and has complex internal parameters, which cannot be met by ordinary computer resources.
[0085] The solution of the embodiment of the present invention is not affected by the spelling of the SMILES string. SMILES strings of the same compound with different spelling formats all correspond to the same structural information. It has better compatibility with compounds of different molecular sizes. The autoencoder in the invention specifically solves the problem of different molecular sizes of compounds. The model is concise and reliable, and the internal parameters are well interpretable. The model uses graph convolution instead of ordinary convolution to more efficiently extract feature information in the topological structure of the compound.
[0086] The solution of the embodiment of the present invention uses an autoencoder and a graph convolutional neural network to perform deep learning on the relationship between the topological structure and chromatographic retention characteristics of a compound, and based on this, predict the chromatographic retention index of the compound. First, a structural information extraction module based on a SMILES string is designed; secondly, a stacked autoencoder is designed based on the acquired structural information; then, a graph convolutional network is designed based on the encoded feature matrix, and finally the chromatographic retention index of the compound is obtained.
[0087] The method for using the model of the embodiment of the present invention is as follows: obtaining the SMILES string of the compound to be predicted; obtaining the structural information matrix of the compound to be predicted through the structural information extraction module of the scheme; then entering the trained stacked autoencoder to obtain the feature matrix; inputting the feature matrix into the trained graph convolutional network model to calculate the chromatographic retention index of the compound to be predicted.
[0088] Specifically, Figure 1 As shown, it is a principle flow chart of the chromatographic retention index prediction method based on graph convolutional network provided in this embodiment, wherein:
[0089] Step 11, extracting the structural information of the compound according to the SMLIES information of the compound and converting it into a structure matrix.
[0090] In one embodiment of the present invention, SMILES (Simplified molecular input line entry system) is a specification that uses ASCII strings to clearly describe molecular structures. Examples are shown in Table 1 below.
[0091] Table 1
[0092] SMILES name SMILES name CC Ethane [OH3+] Hydronium ion O=C=O carbon dioxide [235U] Uranium-235 C#N Hydrogen cyanide F / C=C / F E-Difluoroethylene <![CDATA[ CCN (CC)CC]]> Triethylamine F / C=C\F Z-Difluoroethylene CC(=O)O acetic acid <C(=O)O|]]> L-Alanine
[0093] The function of the structural information extraction module in this solution is to extract part of the structural information contained in SMLIES and convert it into a matrix form. Figure 2 shown.
[0094] Specifically, the following steps are included:
[0095] 1) Define the number of atoms in a compound as n.
[0096] 2) Define a domain matrix A[n, n], where the value of any element A[i, j] represents the connection relationship between the i-th atom and the j-th atom. If the i-th atom and the j-th atom are connected, then A[i, j] = 1, otherwise A[i, j] = 0. Figure 3 shown.
[0097] 3) Define the characteristic matrix X[n, n, d], where X is a three-dimensional matrix, the values on the diagonals of the first two dimensions are eigenvalues, and the third dimension has d layers, where d is 5, d1 is the atomic mass, d2 is the type of atomic peripheral bonds, d3 is the number of rings in which the atom participates, d4 is the minimum angle of the atomic peripheral bonds, and d5 is the maximum angle of the atomic peripheral bonds. Figure 4 shown.
[0098] 4) Define the degree matrix D[n, n], where the values on the diagonal of the D matrix are the number of neighboring atoms of each atom. Figure 5 shown.
[0099] 5) Define the Laplace matrix L[n, n], the calculation method of L matrix is as follows:
[0100] L[n,n]=D[n,n]-A[n,n].
[0101] In order to obtain the above structural information matrix, the workflow of the structural information extraction module is designed as follows: 1) First, the syntax tree of the SMLIES input is parsed to obtain nodes and branches at each level; 2) The nodes are counted to achieve atom extraction; 3) The branches are parsed to achieve bond extraction, subchain extraction and ring structure extraction; 4) Based on the information of atoms, bonds, subchains, ring structures, etc., the domain matrix A, feature matrix X, degree matrix D, and Laplace matrix L are obtained. For the specific process, see Figure 6 .
[0102] Step 12: training and encoding the structure matrix according to the stacked autoencoder to obtain an encoding matrix.
[0103] In one embodiment of the present invention, since the number of atoms contained in the compounds is different, the size of the structure information matrix obtained is different, so the stacked autoencoder in this design is uniformly encoded and compressed. The input of the autoencoder is specified as [M, M], where M is the number of atoms. The output of the autoencoder is specified as [N, N], where N is the number of atoms. Set M>>N, and N>50. For atoms with an atomic number less than M, the missing bits are filled with 0. The overall structure of the stacked autoencoder is as follows: Figure 7 shown.
[0104] The stacked encoder is divided into four independent sub-encoders with different initial parameters but the same structure. Each sub-encoder is trained and encoded for a different matrix. The sub-encoder structure is as follows: Figure 8 shown.
[0105] The sub-encoder consists of two parts: encoding and decoding. The encoding part consists of three neural networks stacked, with the input being the structure matrix and the output being the encoding matrix. The decoding part consists of three neural networks stacked, with the input being the encoding matrix and the output being the structure matrix. Through a large amount of data training, the encoding and decoding accuracy of each channel sub-encoder reaches more than 95%.
[0106] Step 13, inputting the encoding matrix into the graph convolutional neural network model trained according to the training group and the test group, and outputting the chromatographic retention index.
[0107] The graph convolutional neural network model is the core part of the present invention. The convolutional network is a convolutional network designed for the characteristics of the graph structure. The model input is the encoding matrix encoded by the stacked encoder (encoding matrix 1 to 4, represented by A, X, D, L in turn).
[0108] In one embodiment of the present invention, the designed graph convolution model structure is as follows: Fig. 9 shown.
[0109] The entire network consists of three graph convolutional layers, three activation layers, and three fully connected layers. The model structure is no different from that of a normal convolutional network, but the convolutional layers are replaced by graph convolutional layers, so the implementation functions are completely different.
[0110] The layers are described as follows:
[0111] (1) Input layer: Normalize the data. The normalization method is as follows:
[0112]
[0113] Among them, X in For each input value of the feature matrix X, X max For X in Maximum value; X min For X in The minimum value in .
[0114] Workflow: First find the maximum and minimum values in the feature matrix X, and then normalize each input value X in Calculate X out .
[0115] (2) Graph convolution layer: The graph convolution layer in this solution is specifically designed for graph structure data and is one of the core innovations of this solution. The structure of each layer is as follows:
[0116]
[0117] Among them, D is the degree matrix, L is the Laplace matrix, A is the domain matrix, X is the feature matrix after the input layer, H is the input matrix, and W is the convolutional layer parameter matrix.
[0118] Workflow: According to the layer structure formula, for each input matrix H in Calculate H out , and update the convolution layer parameter matrix W according to the gradient training results.
[0119] In this embodiment, the feature matrix X is introduced to add node feature information in addition to the domain matrix A. The degree matrix D and the Laplace matrix L are used to introduce self-transmission information. The inverse of the degree matrix D is multiplied on the left and right to achieve matrix normalization.
[0120] (3) Activation layer: The activation layer uses Lakerelu function, and its structure is as follows:
[0121] H out =Lakerelu(H in )
[0122] (4) Fully connected layer: The fully connected layer is a convolutional network fully connected layer. This design uses three layers of fully connected layers. The first layer has N neurons, the second layer has N / 2 neurons, and the third layer has N / 4 neurons, which is one layer more than the commonly used fully connected layer solution. Fig.10 shown.
[0123] (5) Linear output layer: The structure of the linear output layer is as follows:
[0124] H out =W1*H in +W2
[0125] Among them, H in is a [n,1] structure matrix, W1 is a [1,n] structure matrix, and W2 is a scalar.
[0126] Workflow: After passing through the fully connected layer, H is input into the linear output layer, and the output layer parameter matrix W1 and scalar W2 are updated according to the gradient training results.
[0127] Based on the chromatographic retention index prediction scheme based on graph convolutional network provided by the above embodiment of the present invention, the structural information of the compound is extracted through the SMLIES information of the compound and converted into a structure matrix; the structure matrix is trained and encoded according to the stacked autoencoder to obtain a coding matrix; the coding matrix is input into the graph convolutional neural network model trained according to the training group and the test group, and the chromatographic retention index is obtained as output.
[0128] This embodiment extracts the topological structure and related information of the compound molecule from SMILES, which is reliable and accurate. For chemical molecules of different sizes, they are uniformly encoded into a fixed-size encoding matrix after entering the stacked autoencoder, which is convenient for subsequent convolution operations. In the unified operation process of chemical molecules of different sizes, the stacked autoencoder implemented by the BP neural network is more accurate, fast, and resource-saving. The newly designed graph convolution layer replaces the ordinary convolution layer, and the designed graph convolution network far exceeds the ordinary convolution layer in the processing of topological information. Compared with deep learning models such as LSTM and GNN, the learning model is small in size, easy to implement, and has more controllable internal parameters. It has application scenarios such as edge computing device deployment and more compound property prediction. By optimizing the model parameters, the same or higher prediction accuracy can be achieved with less computing resources.
[0129] like Fig.11 As shown in the figure, the model consists of three parts: 1) structural information extraction module, 2) stacked autoencoder, and 3) graph convolutional network.
[0130] Autoencoders and graph convolutional neural networks are used to deeply learn the relationship between the topological structure and chromatographic retention characteristics of compounds, and based on this, the chromatographic retention index of the compounds is predicted. First, a structural information extraction module based on SMILES strings is designed; secondly, a stacked autoencoder is designed based on the acquired structural information; then, a graph convolutional network is designed based on the encoded feature matrix, and finally the chromatographic retention index of the compound is obtained.
[0131] Firstly, a structural formula extraction module was designed for the text descriptor SMILES representing the chemical compound (part of the structural information contained in SMLIES was extracted and converted into a matrix form), and a stacked autoencoder was designed and trained for the acquired structural formula; secondly, the encoded feature matrix was input into the graph convolutional network (GCN), and the GCN convolutional network was trained with the chromatographic retention index as the response; finally, the entire model was verified using the SMILES chemical formula and retention time of the test group.
[0132] The method of using this model is as follows: obtain the SMILES string of the compound to be predicted; obtain the structural information matrix of the compound to be predicted through the structural information extraction module of this scheme; then enter the trained stacked autoencoder to obtain the feature matrix; input the feature matrix into the trained graph convolutional network model to calculate the chromatographic retention index of the compound to be predicted.
[0133] like Fig.12 As shown, a chromatographic retention index prediction device based on a graph convolutional network is provided in an embodiment of the present invention, wherein:
[0134] A structural information extraction module 21 is used to extract the structural information of the compound according to the SMLIES information of the compound and convert it into a structural matrix;
[0135] A stacked autoencoder 22 is used to train and encode the structure matrix to obtain an encoding matrix;
[0136] The graph convolutional neural network model 23 is used to input the encoding matrix into the graph convolutional neural network model trained according to the training group and the test group, and output the chromatographic retention index.
[0137] The structural information extraction module 21 is specifically used for:
[0138] Parse the SMLIES input of the compound into a syntax tree to obtain nodes and branches at each level;
[0139] Counting the nodes to achieve atomic extraction;
[0140] Parsing the branched chains to extract bonds, subchains and ring structures;
[0141] Based on the atomic, bond, subchain and ring structure information, a structural matrix is obtained, including but not limited to: domain matrix A, feature matrix X, degree matrix D, and Laplace matrix L.
[0142] The stacked autoencoder 22 is divided into four unrelated sub-encoders with different initial parameters but the same structure, and each sub-encoder is trained and encoded for a different matrix;
[0143] The graph convolutional neural network model 23 consists of three graph convolutional layers, three activation layers and three fully connected layers.
[0144] An embodiment of the present invention further provides an electronic device, including:
[0145] Memory for storing computer programs;
[0146] The processor is used to execute the computer program stored in the memory, and when the computer program is executed, the method described in any one of the above embodiments is implemented.
[0147] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method described in any of the above embodiments is implemented.
[0148] Based on the chromatographic retention index prediction scheme based on graph convolutional network provided by the above embodiment of the present invention,
[0149] The structural information of the compound is extracted through the SMLIES information of the compound and converted into a structure matrix; the structure matrix is trained and encoded according to the stacked autoencoder to obtain an encoding matrix; the encoding matrix is input into a graph convolutional neural network model trained according to a training group and a test group, and the chromatographic retention index is obtained as an output.
[0150] This embodiment extracts the topological structure and related information of the compound molecule from SMILES, which is reliable and accurate. For chemical molecules of different sizes, they are uniformly encoded into a fixed-size encoding matrix after entering the stacked autoencoder, which is convenient for subsequent convolution operations. In the unified operation process of chemical molecules of different sizes, the stacked autoencoder implemented by the BP neural network is more accurate, fast, and resource-saving. The newly designed graph convolution layer replaces the ordinary convolution layer, and the designed graph convolution network far exceeds the ordinary convolution layer in the processing of topological information. Compared with deep learning models such as LSTM and GNN, the learning model is small in size, easy to implement, and has more controllable internal parameters. It has application scenarios such as edge computing device deployment and more compound property prediction. By optimizing the model parameters, the same or higher prediction accuracy can be achieved with less computing resources.
[0151] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.
[0152] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.
[0153] A person of ordinary skill in the art can understand that all or part of the steps of implementing the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various media that can store program codes.
[0154] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
[0155] The method and apparatus of the present invention may be implemented in many ways. For example, the method and apparatus of the present invention may be implemented by software, hardware, firmware, or any combination of software, hardware, firmware. The above order of steps for the method is only for illustration, and the steps of the method of the present invention are not limited to the order specifically described above, unless otherwise specifically stated. In addition, in some embodiments, the present invention may also be implemented as a program recorded in a recording medium, which includes machine-readable instructions for implementing the method according to the present invention. Thus, the present invention also covers a recording medium storing a program for executing the method according to the present invention.
[0156] The description of the present invention is given for the purpose of illustration and description, and is not intended to be exhaustive or to limit the invention to the disclosed form. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for particular uses.
Claims
1. A chromatographic retention index prediction method based on graph convolutional network, characterized in that: The method comprises: Extract the structural information of the compound based on the SMLIES information of the compound and convert it into a structural matrix; The step of extracting the structural information of the compound according to the SMLIES information of the compound and converting it into a structural matrix includes: Parse the SMLIES input of the compound into a syntax tree to obtain nodes and branches at each level; Counting the nodes to achieve atomic extraction; Parsing the branched chains to extract bonds, subchains and ring structures; A structure matrix is obtained based on the information of atoms, bonds, subchains and ring structures, wherein the structure matrix is: domain matrix A, feature matrix X, degree matrix D and Laplace matrix L; The structure matrix is trained and encoded according to a stacked autoencoder to obtain an encoding matrix; Inputting the encoding matrix into a graph convolutional neural network model trained according to a training group and a test group, and outputting a chromatographic retention index; The stacked autoencoder comprises: The input of the autoencoder is [M,M], where M is the number of atoms in the input compound; The output of the autoencoder is [N,N], where N is the number of atoms in the output compound; Where, M>>N, and N>50; For atoms with atomic number less than M, fill in the missing bits with 0; The stacked autoencoder is divided into four unrelated sub-encoders with different initial parameters but the same structure, and each sub-encoder is trained and encoded for a different matrix; The graph convolutional neural network model comprises: The graph convolutional neural network model consists of three graph convolutional layers, three activation layers and three fully connected layers; specifically, it includes: Input layer: Normalize the data in the following way: Among them, X in For each input value of the feature matrix X, X max For X in Maximum value; X min For X in The minimum value in Graph convolutional layer: The structure of each layer is as follows: Among them, D is the degree matrix, L is the Laplace matrix, A is the domain matrix, X is the feature matrix after the input layer, H is the input matrix, and W is the convolution layer parameter matrix; Activation layer: Lakerelu function is used, and the structure is as follows: H out =Lakereu(H in ) Fully connected layer: The first layer has N neurons, the second layer has N / 2 neurons, and the third layer has N / 4 neurons. Linear output layer: The structure is as follows: H out =W1*H in +W2 Among them, H in is a [n,1] structure matrix, W1 is a [1,n] structure matrix, and W2 is a scalar.
2. The method according to claim 1, characterized in that The sub-encoder consists of two parts: encoding and decoding; The encoding part consists of three neural networks stacked together, with the input being the structure matrix and the output being the encoding matrix; The decoding part consists of three stacked neural networks, with the input being the encoding matrix and the output being the structure matrix.
3. The method according to claim 1, characterized in that The graph convolutional neural network model is trained as follows: Extracting information from the compounds in the training group and the test group to form a structure matrix; Output the encoding matrix through stacking the autoencoder according to the structure matrix; The graph convolutional neural network model is input according to the encoding matrix, and the graph convolutional neural network model is adjusted and trained according to the output chromatographic retention index compared with the actual chromatographic retention index of the training group and the test group.
4. A device for the chromatographic retention index prediction method based on graph convolutional network as claimed in claim 1, characterized in that: include: The structural information extraction module is used to extract the structural information of the compound according to the SMLIES information of the compound and convert it into a structural matrix; A stacked autoencoder is used to train and encode the structure matrix to obtain an encoding matrix; A graph convolutional neural network model, used for inputting the encoding matrix into the graph convolutional neural network model trained according to the training group and the test group, and outputting a chromatographic retention index; The structural information extraction module is specifically used for: Parse the SMLIES input of the compound into a syntax tree to obtain nodes and branches at each level; Counting the nodes to achieve atomic extraction; Parsing the branched chains to extract bonds, subchains and ring structures; A structure matrix is obtained based on the information of atoms, bonds, subchains and ring structures, wherein the structure matrix is: domain matrix A, feature matrix X, degree matrix D and Laplace matrix L; The stacked autoencoder is divided into four unrelated sub-encoders with different initial parameters but the same structure, and each sub-encoder is trained and encoded for a different matrix; The graph convolutional neural network model consists of three graph convolutional layers, three activation layers and three fully connected layers.
