Compound property prediction method based on molecular structure and physical information fusion

Abstract

The invention provides a molecular structure and physical information fusion-based compound property prediction method, a molecular structure and physical information fusion-based compound property prediction system and a computer readable storage medium. The molecular structure and physical information fusion-based compound property prediction method comprises the following steps: acquiring a data set containing molecular SMILES descriptors; converting the data set into a molecular graph embedding vector and a materialization information descriptor, wherein the molecular graph embedding vector is determined by an unsupervised learning graph convolutional neural network model; fusing the molecular map embedded vector and the materialized information descriptor to determine a molecular mixed descriptor; and inputting the molecular mixed descriptors into a molecular property prediction model to predict compound properties, wherein the molecular property prediction model is constructed based on a multi-layer perceptron.

Description

Technical Field

[0001] The present invention relates to the field of chemical synthesis, and in particular to a compound property prediction method based on the fusion of molecular structure and physical information, a compound property prediction system based on the fusion of molecular structure and physical information, and a computer-readable storage medium. Background Art

[0002] Chemical molecular property technology is an important application in cheminformatics and materials science, and is the cornerstone of computer-aided chemical synthesis and molecular design. Predictive models of molecular properties can help virtual screening processes to separate candidate molecules with desired properties. In traditional chemical research, the prediction of molecular properties usually relies on experience and theoretical calculations. However, traditional molecular property prediction methods are often too time-consuming and have limited accuracy when dealing with complex systems. Existing molecular property prediction methods use machine learning models to achieve this. Machine learning models automatically extract useful features and establish mapping models from molecular structure to properties by learning the complex relationship between molecular structure and its properties from a large amount of chemical data. These mapping models usually capture information at different levels of molecules based on different data representation methods (such as molecular sequence-based representation, graph-based representation, 3D structure-based representation, etc.).

[0003] In the prior art, molecular property prediction methods mainly include computational models based on quantum chemical methods, such as density functional theory (DFT), quantitative structure-activity relationship (QSAR) models based on pharmacophores and molecular descriptors, and methods based on statistical learning, etc. However, DFT has high computational costs, QSAR models have poor generalization, and ensemble learning is highly dependent on data.

[0004] In order to overcome the above-mentioned defects of the prior art, there is an urgent need in this field for a compound property prediction technology based on the fusion of molecular structure and physical information, which can circumvent the high computational cost of DFT, the poor generalization of QSAR models, and the strong dependence of ensemble learning on data, and more accurately predict the properties of organic compounds. Summary of the invention

[0005] A brief summary of one or more aspects is given below to provide a basic understanding of these aspects. This summary is not an exhaustive overview of all conceived aspects, and is neither intended to identify the key or critical elements of all aspects nor to define the scope of any or all aspects. Its only purpose is to give some concepts of one or more aspects in a simplified form as a prelude to a more detailed description that will be given later.

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a compound property prediction method based on the fusion of molecular structure and physical information, a compound property prediction system based on the fusion of molecular structure and physical information, and a computer-readable storage medium, which can circumvent the problems of high computational cost of DFT, poor generalization of QSAR model, and strong dependence of ensemble learning on data, and more accurately realize the prediction of properties of organic compounds.

[0007] Specifically, the above-mentioned compound property prediction method based on the fusion of molecular structure and physical information provided according to the first aspect of the present invention includes the steps of: obtaining a data set containing molecular SMILES descriptors; converting the data set into a molecular graph embedding vector and a physicochemical information descriptor, wherein the molecular graph embedding vector is determined by an unsupervised learning graph convolutional neural network model; fusing the molecular graph embedding vector and the physicochemical information descriptor to determine a molecular mixing descriptor; and inputting the molecular mixing descriptor into a molecular property prediction model to predict the properties of the compound, wherein the molecular property prediction model is constructed based on a multi-layer perceptron.

[0008] Preferably, in one embodiment of the present invention, the step of converting the data set into a molecular graph embedding vector and a materialized information descriptor includes: using the RDKit library to convert the molecular SMILES descriptor in the data set into a molecular object; and determining the molecular graph embedding vector and the materialized information descriptor based on the molecular object.

[0009] Preferably, in one embodiment of the present invention, the step of determining the molecular graph embedding vector and the materialized information descriptor based on the molecular object includes: determining the graph structure data based on the molecular object, and converting the graph structure data into the molecular graph embedding vector according to the pre-trained unsupervised learning graph convolutional neural network model.

[0010] Preferably, in one embodiment of the present invention, the molecular object includes properties of atoms and bonds, and the topological structure of the molecule, the graph structure data includes an adjacency matrix and a feature matrix, and the step of determining the graph structure data based on the molecular object includes: taking the atoms of the molecule as graph nodes, the bonds as edges, the properties of the atoms as feature vectors of the graph nodes, and the properties of the bonds as feature vectors of the edges, so as to construct the adjacency matrix and the feature matrix.

[0011] Preferably, in one embodiment of the present invention, the unsupervised learning graph convolutional neural network model includes a linear change layer, a graph convolutional neural network layer and a fully connected neural network layer.

[0012] Preferably, in an embodiment of the present invention, the linear change layer includes a linear layer and an embedded layer.

[0013] Preferably, in an embodiment of the present invention, the step of determining the molecular graph embedding vector and the materialized information descriptor based on the molecular object comprises: calculating the molecular object using an RDKit library to determine the materialized information descriptor.

[0014] Preferably, in one embodiment of the present invention, the physicochemical information descriptor includes the charge distribution, molecular weight, number of atoms, number of rings, type and distribution of functional groups, electrical properties, configuration, topological structure, and biological activity of the molecule.

[0015] Preferably, in an embodiment of the present invention, the step of fusing the molecular graph embedding vector and the materialized information descriptor to determine the molecular mixture descriptor includes: standardizing and normalizing the molecular mixture descriptor.

[0016] Preferably, in an embodiment of the present invention, the molecular property prediction model includes a plurality of fully connected layers and a plurality of random inactivation layers in alternating combination and an output fully connected layer for output.

[0017] In addition, the compound property prediction system based on the fusion of molecular structure and physical information provided according to the second aspect of the present invention includes a memory and a processor. The memory stores computer instructions. The processor is connected to the memory and is configured to execute the computer instructions stored in the memory to implement the compound property prediction method based on the fusion of molecular structure and physical information provided by any one of the above embodiments.

[0018] In addition, the computer-readable storage medium provided according to the third aspect of the present invention stores computer instructions. When the computer instructions are executed by the processor, the compound property prediction method based on the fusion of molecular structure and physical information provided by any of the above embodiments is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The above features and advantages of the present invention can be better understood after reading the detailed description of the embodiments of the present disclosure in conjunction with the following drawings. In the drawings, the components are not necessarily drawn to scale, and components with similar related properties or features may have the same or similar reference numerals.

[0020] Figure 1 A flow chart of a method for predicting compound properties based on the fusion of molecular structure and physical information provided according to some embodiments of the present invention is shown;

[0021] Figure 2 A schematic diagram of a graph convolutional neural network model for unsupervised learning provided according to an embodiment of the present invention is shown;

[0022] Figure 3 A loss convergence curve diagram of an unsupervised learning graph convolutional neural network model during training according to some embodiments of the present invention is shown;

[0023] Figure 4 A schematic diagram of a molecular property prediction model provided according to an embodiment of the present invention is shown;

[0024] Figure 5 The prediction performance of different molecular property prediction models is shown;

[0025] Figure 6A to Figure 6B A schematic diagram showing the prediction effect of a molecular property prediction model provided according to an embodiment of the present invention;

[0026] Figure 7 A schematic diagram showing the fitting effect of the molecular property prediction model provided according to an embodiment of the present invention; and

[0027] Figure 8 A schematic diagram of a compound property prediction system based on the fusion of molecular structure and physical information provided according to some embodiments of the present invention is shown.

[0028] Reference numerals:

[0029] 100: Methods for predicting compound properties;

[0030] S110~S140: steps;

[0031] 800: Compound property prediction system;

[0032] 810: memory;

[0033] 811: Computer readable storage medium; and

[0034] 820: Processor. DETAILED DESCRIPTION

[0035] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Note that the aspects described below in conjunction with the accompanying drawings and specific embodiments are only exemplary and should not be construed as limiting the scope of protection of the present invention in any way.

[0036] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0037] In addition, the terms "upper", "lower", "left", "right", "top", "bottom", "horizontal" and "vertical" used in the following description should be understood as the directions shown in the paragraph and the related drawings. Such relative terms are only used for the convenience of description and do not mean that the device described therein must be manufactured or operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0038] It is understood that although the terms "first", "second", "third", etc. may be used herein to describe various components, regions, layers and / or parts, these components, regions, layers and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers and / or parts. Therefore, the first component, region, layer and / or part discussed below may be referred to as a second component, region, layer and / or part without departing from some embodiments of the present invention.

[0039] As mentioned above, in the prior art, molecular property prediction methods mainly include computational models based on quantum chemical methods, such as density functional theory (DFT), quantitative structure-activity relationship (QSAR) models based on pharmacophores and molecular descriptors, statistical learning-based methods, etc. However, DFT has high computational cost, QSAR models have poor generalization, and ensemble learning has strong dependence on data.

[0040] In order to overcome the above-mentioned defects of the prior art, the present invention provides a compound property prediction method based on the fusion of molecular structure and physical information, a compound property prediction system based on the fusion of molecular structure and physical information, and a computer-readable storage medium, which can circumvent the problems of high computational cost of DFT, poor generalization of QSAR model, and strong dependence of ensemble learning on data, and more accurately realize the prediction of properties of organic compounds.

[0041] In some non-limiting embodiments, the above-mentioned compound property prediction method based on the fusion of molecular structure and physical information provided in the first aspect of the present invention can be implemented via the above-mentioned compound property prediction system based on the fusion of molecular structure and physical information provided in the second aspect of the present invention.

[0042] The following will first describe the working principle of the above-mentioned compound property prediction system based on the fusion of molecular structure and physical information in conjunction with some embodiments of the compound property prediction method based on the fusion of molecular structure and physical information. Those skilled in the art will understand that these embodiments of the compound property prediction method based on the fusion of molecular structure and physical information are only some non-limiting implementation methods provided by the present invention, which are intended to clearly demonstrate the main concept of the present invention and provide some specific schemes that are convenient for the public to implement, rather than to limit the full functions or all working modes of the compound property prediction system based on the fusion of molecular structure and physical information. Similarly, the compound property prediction system based on the fusion of molecular structure and physical information is also only a non-limiting implementation method provided by the present invention, and does not constitute a limitation on the execution subject and execution order of each step in these compound property prediction methods based on the fusion of molecular structure and physical information.

[0043] Please refer to Figure 1 , Figure 1 A flow chart of a method for predicting compound properties based on the fusion of molecular structure and physical information provided according to some embodiments of the present invention is shown.

[0044] like Figure 1 As shown, the compound property prediction method 100 based on the fusion of molecular structure and physical information may include step S110: acquiring a data set containing molecular SMILES descriptors.

[0045] The SMILES (Simplified Molecular Input Line Entry System) representation of molecules is an advanced sequence-based molecular representation. In the SMILES molecular descriptor, atoms and chemical bonds are represented by letters and symbols respectively, and branches are represented by brackets.

[0046] The compound property prediction system based on the fusion of molecular structure and physical information can obtain a data set containing a large number of molecular SMILES descriptors for compound property prediction, and the data set can be the zinc250k data set from the chemical public database.

[0047] Afterwards, the compound property prediction method 100 based on the fusion of molecular structure and physical information may include step S120: converting the data set into a molecular graph embedding vector and a physical and chemical information descriptor, wherein the molecular graph embedding vector is determined by an unsupervised learning graph convolutional neural network model.

[0048] The compound property prediction system preprocesses the data set and converts it into molecular graph embedding vectors and physicochemical information descriptors.

[0049] Here, the compound property prediction system can first use the function module in the RDKit library (an open source chemical information toolkit) (for example, the MolFromSmiles function module of the RDKit library) to convert about 250,000 molecular SMILES descriptors in the data set into molecular objects. Here, molecular objects include the properties of atoms and bonds, and the topological structure of molecules. Those skilled in the art will appreciate that the above-mentioned scheme of converting molecular SMILES descriptors into molecular objects using the RDKit library is only a non-limiting implementation method provided by the present invention, which is intended to clearly demonstrate the main concept of the present invention and provide a specific scheme that is convenient for the public to implement, rather than to limit the scope of protection of the present invention. Optionally, in other embodiments, those skilled in the art may also use other equivalent methods to convert molecular SMILES descriptors based on the concept of the present invention to achieve the same technical effect.

[0050] Afterwards, the compound property prediction system can use the customized MolDataset class function module to receive the converted molecular objects and obtain information such as the two-dimensional topological structure of the molecule. Here, the MolDataset class can be used to process chemical molecular data. In a non-limiting embodiment, the MolDataset class can convert the received molecular objects into molecular structures and extract the characteristics of each atom. Further, the MolDataset class can also define the generation of the adjacency matrix, the one-hot encoding function of the atomic features, the length of the data set, and the acquisition of index items, thereby providing standardized and formatted topological structure data input for the subsequent training of the graph convolutional neural network model for determining the molecular graph embedding vector.

[0051] Please refer to Figure 2 , Figure 2 A schematic diagram of a graph convolutional neural network model for unsupervised learning provided according to an embodiment of the present invention is shown.

[0052] like Figure 2 As shown, the compound property prediction system converts the topological structure of the molecule into graph structure data, which may include an adjacency matrix and a feature matrix determined by the graph structure of the molecule. The compound property prediction system may use the atoms of the molecule as graph nodes of the graph structure, the bonds of the molecule as edges of the graph structure, the properties of the atoms as feature vectors of the graph nodes, and the properties of the bonds as feature vectors of the edges, thereby constructing the adjacency matrix and feature matrix of the molecule.

[0053] Here, the adjacency matrix can be converted into a self-loop by adding a unit matrix. Further, the compound property prediction system can adjust the size of the adjacency matrix of each molecule to ensure that the size of the adjacency matrix of all molecules is consistent.

[0054] The compound property prediction system can use the properties of each atom of the molecule as node features and generate a feature vector for each atom, and each feature vector forms a feature matrix that matches the size of the adjacency matrix A. Here, the feature vector of each atom (i.e., the feature vector of each graph node) can be determined by a series of one-hot encoding functions. The feature vector of each atom can specifically include information such as the symbol, degree, total number of hydrogen atoms, implicit valence, and whether the atom is an aromatic atom.

[0055] In a preferred embodiment, each molecule can be defined as an undirected graph G = (V, E), where V and E represent the node set and edge set, respectively. j ∈V, edge feature vector e j ∈E, node feature vector v j and the edge eigenvector e j are associated with heavy atoms (e.g., C, N, O, F) and their bonds. For an undirected graph G, the adjacency matrix A is a symmetric matrix, where A in the adjacency matrix A ij It is used to indicate whether there is an edge connecting nodes i and j. The feature matrix X can be a matrix with a dimension of |V|×d, where d is the dimension of the feature vector of each atom, that is, the dimension of the node feature vector. Each row X in the feature matrix X i is the feature vector of node i.

[0056] After determining the graph structure data, the compound property prediction system can learn the graph structure data based on the pre-trained unsupervised learning graph convolutional neural network model (GCN, Graph Convolutional Networks), and map the graph structure data to a low-dimensional vector space to convert the graph structure data into a molecular graph embedding vector. The unsupervised learning graph convolutional neural network model can reduce the label requirements for chemical molecules and improve the generalization ability of molecular graph embedding, so that it can be applied to a variety of different tasks related to chemical molecules.

[0057] Please continue to refer to Figure 2 ,The structure of the graph convolutional neural network model for unsupervised learning can be defined at initialization as: including a linear variation layer, multiple graph convolutional neural network layers and a fully connected neural network layer.

[0058] In a preferred embodiment, the linear change layer includes a linear layer and an embedding layer. Through the linear layer, the feature matrix of the input data can be mapped to a high-dimensional space, and the embedding layer can transform the input data to a low-dimensional space. In the forward propagation process of the compound property prediction system, the input feature matrix is ​​first transformed by the linear layer, and it is mapped to a higher-dimensional space to capture more complex feature relationships. Subsequently, these high-dimensional features can be further encoded by the embedding layer to be converted into a low-dimensional continuous vector representation. As a result, the compound property prediction system can retain important information while reducing the data dimension, which is convenient for subsequent model processing and analysis.

[0059] Afterwards, the compound property prediction system propagates features and aggregates and updates graph information through multiple graph convolutional neural network layers (GCL). The formulas for the convolutional layer and output layer of the graph convolutional neural network layer can be:

[0060]

[0061] Among them, H (l) is the node feature matrix of the (l)th layer, W (l) is the weight matrix of the (l)th layer, σ is the activation function (for example, ReLU function); A is the original adjacency matrix, is the normalized adjacency matrix. The normalization of the adjacency matrix can be obtained by the formula Calculated and determined; D is the degree matrix, and the elements D on the diagonal of the degree matrix D ii represents the degree of node i, that is, D ii =∑ j A ij ;I N is the identity matrix, through the formula Can be used to add self-joins.

[0062] Specifically, the feature propagation process of each graph convolutional neural network layer may include: multiplying the node features of the graph structure data (i.e., the feature vectors of the graph nodes) with the adjacency matrix to achieve information transfer between nodes; using an activation function (e.g., ReLU function) to activate the node features to introduce nonlinear transformations; using the Dropout operation (a regularization method) to randomly discard the node features to prevent overfitting. In this way, until the feature propagation process of all graph convolutional neural network layers is applied, the feature vectors of each node are averaged to obtain the intermediate representation of the molecule transformed from the graph structure data.

[0063] The fully connected neural network layer (Output) in the unsupervised learning graph convolutional neural network model can generate and output a molecular graph embedding vector based on the intermediate representation output by multiple graph convolutional neural network layers (GCL).

[0064] So, like Figure 2 The unsupervised learning graph convolutional neural network model shown can map the graph structure data to a low-dimensional space (e.g., from 128 dimensions to 32 dimensions) to determine a molecular graph embedding vector. The molecular graph embedding vector captures the structure and feature information of the graph structure data.

[0065] The offline training phase of the compound property prediction system may include training of the graph convolutional neural network model for unsupervised learning. In a preferred embodiment, the compound property prediction system may divide the data set into a training set and a test set in proportion (for example, divided into a training set and a test set in a ratio of 7:3), load the data set and initialize the weights of the graph convolutional neural network model for unsupervised learning.

[0066] Please refer to Figure 3 , Figure 3 A loss convergence curve diagram of an unsupervised learning graph convolutional neural network model during training provided according to some embodiments of the present invention is shown.

[0067] like Figure 3 As shown, the corresponding graph structure data is determined according to the training set and test set in the data set, and the graph structure data is converted into a molecular graph embedding vector through an unsupervised learning graph convolutional neural network model until the loss value of the loss function (Loss) of the training set and the test set converges to infinitely close to 0, and then the trained unsupervised learning graph convolutional neural network model is saved.

[0068] In addition, the compound property prediction system can also convert the data set into a physicochemical information descriptor. In one embodiment, the descriptor calculation module (e.g., Descriptors function module) in the RDKit library is used to calculate the molecular object converted from the molecular SMILES descriptor to determine the descriptors (about 200 kinds) with physicochemical information contained in the molecular object. After the compound property prediction system performs correlation screening on the physicochemical information, the physicochemical information descriptor of the high-dimensional vector can be obtained. Preferably, the physicochemical information descriptor can select a descriptor containing the charge distribution, molecular weight, number of atoms, number of rings, type and distribution of functional groups, electrical properties, configuration, topological structure and biological activity of the molecule. The physicochemical information descriptor can be used as a one-dimensional characteristic vector of the molecule for subsequent compound property prediction.

[0069] Please continue to refer to Figure 1The compound property prediction method 100 may include step S130: fusing the molecular graph embedding vector and the physicochemical information descriptor to determine a molecular mixture descriptor.

[0070] The compound property prediction system can integrate the function of converting a data set into a molecular graph embedding vector and the function of converting a data set into a physicochemical information descriptor. On this basis, the molecular graph embedding vector is fused with the physicochemical information descriptor, and the molecular graph embedding vector and the physicochemical information descriptor are concatenated to output a molecular mixture descriptor. In a non-limiting embodiment, the compound property prediction system can use the MolCal method (Molecule Calculator) to concatenate the molecular graph embedding vector and the physicochemical information descriptor.

[0071] The compound property prediction method 100 takes into account the problem that the molecular fingerprint characterization features are not comprehensive. By fusing the molecular graph embedding vector with the physicochemical information descriptor, the compound property prediction method 100 can more comprehensively describe the characteristics of the molecule, thereby improving the accuracy and reliability of subsequent compound property predictions. In addition, by using the physicochemical information descriptor to allow the physicochemical information of the molecule to participate in the subsequent compound property prediction, the interpretability of the compound property prediction model can be increased.

[0072] Preferably, the compound property prediction system can standardize and normalize the determined molecular mixture descriptors. Specifically, the compound property prediction system can use the StandardScaler standardization formula for processing:

[0073]

[0074] Among them, z is the eigenvalue of the molecular mixture descriptor after standardization, x is the original eigenvalue of the molecular mixture descriptor, μ is the mean of the eigenvalue of the molecular mixture descriptor, and s is the standard deviation of the eigenvalue of the molecular mixture descriptor.

[0075] Please continue to refer to Figure 1 The compound property prediction method 100 may include step S140: inputting the molecular mixture descriptor into the molecular property prediction model to predict the compound property, and the molecular property prediction model is constructed based on a multi-layer perceptron.

[0076] Through the molecular property prediction model, the compound property prediction system can predict the properties of the compound, and the properties of the compound may include QED (Quantitative Estimate of Drug-likeness), SAS (Synthetic Accessibility Score) and LogP (logarithm of the compound oil-water partition coefficient P). Those skilled in the art will appreciate that, in addition to predicting the QED, SAS, and LogP of the compound, the compound property prediction method 100 provided in the present invention can also be adjusted according to actual conditions to achieve prediction of other compound properties related to the molecular SMILES descriptor.

[0077] The molecular property prediction model constructed based on the multi-layer perceptron can include multiple fully connected layers and multiple random inactivation layers in alternating combination and an output fully connected layer for output. In a preferred embodiment, the structure of the molecular property prediction model can be defined as: an alternating combination of four fully connected layers and four Dropout layers, and the alternating combination of four fully connected layers and four Dropout layers can prevent the molecular property prediction model from overfitting and make the molecular property prediction model have better generalization ability. In this preferred embodiment, the molecular property prediction model also includes an output fully connected layer, through which the dimension is mapped to a single node to achieve the task of regression prediction of molecular properties.

[0078] Please refer to Figure 4 , Figure 4 A schematic diagram of a molecular property prediction model provided according to an embodiment of the present invention is shown.

[0079] exist Figure 4 In the embodiment shown, the molecular property prediction model is composed of two fully connected layers and two random inactivation layers alternately combined. Figure 4 The random dropout layer shown includes an activation function and a Dropout layer. The molecular property prediction model can also include an output fully connected layer for implementing a regression task.

[0080] The offline training phase of the compound property prediction system can also include the training process of the molecular property prediction model. After the data set is transformed and spliced ​​to form a molecular mixture descriptor, it can be divided into a training set, a validation set, and a test set according to the proportion. After that, the molecular property prediction model is trained using an optimizer and a loss function. During the entire training process, the molecular property prediction model will undergo multiple iterative cycles, each cycle including forward propagation and back propagation of the training data. The output and loss value of the molecular property prediction model are calculated through the forward propagation process, and then the parameters of the molecular property prediction model are updated through the back propagation process. In the interval of each cycle, the compound property prediction system can set the molecular property prediction model to evaluation mode, calculate the loss value using the validation set, and calculate the determination coefficient (R 2 , R-squared Coefficient), mean square error (MAE, Mean Absolute Error) and mean absolute error (MSE, Mean Squared Error) are used to evaluate the performance of the molecular property prediction model.

[0081] Specifically, the coefficient of determination R 2 It can be used to measure the proportion of the change in the dependent variable that can be explained by the independent variable, as follows:

[0082]

[0083] Among them, SS res is the residual sum of squares, that is, the sum of the squared differences between the true value and the predicted value, SS tot is the total sum of squares, that is, the sum of the squared differences between the true value and the mean.

[0084] The mean square error (MAE) is the average of the absolute values ​​of the differences between the predicted values ​​and the true values. The formula is as follows:

[0085]

[0086] Where n is the number of samples, y i is the true value, is the predicted value.

[0087] The mean absolute error (MSE) is the average of the squares of the differences between the predicted values ​​and the true values, and its formula is as follows:

[0088]

[0089] Where n is the number of samples, y i is the true value, is the predicted value.

[0090] The following is a specific non-limiting preferred embodiment, based on which the molecular property prediction model proposed in the present invention is further explained.

[0091] In the offline training stage, the compound property prediction system based on the fusion of molecular structure and physical information can first randomly extract 12,000 data points from the zinc250k data set of the chemical public database as samples to form a data set containing molecular SMILES descriptors.

[0092] The 12,000 samples in the data set were randomly divided into training set, validation set and test set in the ratio of 7:1.5:1.5. The training set was used to train the molecular property prediction model in the offline training phase of the compound property prediction system, the validation set was used to observe the trend of the loss value of the molecular property prediction model to optimize the molecular property prediction model, and the test set was used to verify the fitting effect of the molecular property prediction model.

[0093] After that, the compound property prediction system can select stochastic gradient descent (SGD) as the parameter optimization method of the molecular property prediction model. The parameters of the molecular property prediction model are iteratively updated through the SGD parameter optimizer, so that the parameter value is adjusted in the opposite direction of the gradient at each step to minimize the loss function.

[0094] Please refer to Figure 5 , Figure 5 The prediction performance of different molecular property prediction models is shown.

[0095] The molecular property prediction model 501 of the compound property prediction method based on the fusion of molecular structure and physical information provided by the present invention is compared with other molecular property prediction models. Other molecular property prediction models are existing models with higher accuracy, including a model (Rule-based) 502 based on (Morgan fingerprint) rules, a graph convolutional model 503, a contrastive divergence learning (CD-learning) model 504, and a quantum computing-assisted learning (QC-assisted learning) model 505. Different molecular property prediction models are used to predict the properties of compounds, and the true values ​​and predicted values ​​are evaluated to determine the mean square error of each molecular property prediction model.

[0096] Figure 5 The figure shows the mean square error (MAE) of the molecular property prediction model 501 of the compound property prediction method based on the fusion of molecular structure and physical information provided by the present invention and other molecular property prediction methods after multiple trainings on 10,000 sample points and repeated evaluations on 1,000 sample points, to illustrate the prediction performance of each molecular property prediction model.

[0097] The molecular property prediction model 501 based on the compound property prediction method of the fusion of molecular structure and physical information has a mean square error of 0.05±0.008 when predicting QED; 0.35±0.01 when predicting SAS; and 0.24±0.01 when predicting LogP. The model 502 based on the (Morgan fingerprint) rule has a mean square error of 0.19±0.008 when predicting QED; 0.76±0.01 when predicting SAS; and 1.24±0.02 when predicting LogP. The graph convolution model 503 has a mean square error of 0.75±0.02 when predicting QED; 0.88±0.02 when predicting SAS; and 0.81±0.01 when predicting LogP. The mean square error of contrastive divergence learning model 504 in predicting QED is 0.12±0.001; the mean square error in predicting SAS is 0.75±0.09; and the mean square error in predicting LogP is 1.33±0.08. The mean square error of quantum computing-assisted learning model 505 in predicting QED is 0.10±0.002; the mean square error in predicting SAS is 0.66±0.05; and the mean square error in predicting LogP is 1.27±0.05. Figure 5 The index results shown can illustrate that the molecular property prediction model provided by the present invention is superior and has a significant improvement in prediction accuracy compared to existing molecular property prediction models.

[0098] Please continue to refer to Fig. 6A and Figure 6B , Figure 6A to Figure 6B A schematic diagram showing the prediction effect of a molecular property prediction model provided according to an embodiment of the present invention.

[0099] like Fig. 6A As shown in Figure 2, after the molecular property prediction model is trained, it is evaluated in the test set, and the predicted values ​​are closer to the true values ​​in the later stages of the time series. Figure 6B As shown in the figure, after the molecular property prediction model makes predictions on 100 sample points, the predicted values ​​of the properties are closer to the true values ​​in the later stages of the time series, which is in line with the rules in the actual training and prediction process.

[0100] Please refer to further Figure 7 , Figure 7 A schematic diagram showing the fitting effect of a molecular property prediction model provided according to an embodiment of the present invention.

[0101] like Figure 7 As shown in the figure, the molecular property prediction model has a determination coefficient R when predicting QED. 2is 0.674, the mean absolute error MSE is 0.005, and the mean square error MAE is 0.058; when predicting SAS, the determination coefficient R 2 is 0.894, the mean absolute error MSE is 0.245, and the mean square error MAE is 0.358; when predicting LogP, the determination coefficient R 2 The mean absolute error (MSE) is 0.934, the mean absolute error (MSE) is 0.119, and the mean square error (MAE) is 0.241. Among the evaluation indicators, the mean absolute error (MSE) and the mean square error (MAE) are both small, and the determination coefficient R 2 They are all close to 1, indicating that the molecular property prediction model has a good fitting effect.

[0102] In summary, the compound property prediction method based on the fusion of molecular structure and physical information provided by the present invention can extract useful information from known molecular structures and property data by constructing a neural network architecture, and use this information to predict the properties of unknown compounds. Specifically, the use of an unsupervised graph convolutional neural network model reduces the need for labeled samples and reduces dependence on traditional chemical calculations; by fusing the molecular graph embedding vector with the physicochemical information descriptor of the molecule, the molecular features involved in the molecular property prediction model are made more specific and complete, and the interpretability of the molecular property prediction model is increased. In addition, the molecular property prediction model provided by the present invention has a higher determination coefficient R 2 and lower mean absolute error (MSE) and mean square error (MAE).

[0103] Please refer to Figure 8 , Figure 8 A schematic diagram of a compound property prediction system based on the fusion of molecular structure and physical information provided according to some embodiments of the present invention is shown.

[0104] like Figure 8 As shown, the compound property prediction system 800 based on the fusion of molecular structure and physical information may be configured with a memory 810 and a processor 820. The memory 810 may be configured with a computer-readable storage medium 811, and the computer-readable storage medium 811 stores computer instructions. The processor 820 is connected to the memory 810 and is configured to execute the computer instructions stored in the memory 810 to implement the above-mentioned compound property prediction method based on the fusion of molecular structure and physical information.

[0105] Although the above methods are illustrated and described as a series of actions for simplicity of explanation, it should be understood and appreciated that these methods are not limited by the order of the actions, because according to one or more embodiments, some actions may occur in a different order and / or concurrently with other actions from those illustrated and described herein or not illustrated and described herein but understandable to those skilled in the art.

[0106] Those skilled in the art will appreciate that information, signals, and data may be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips cited throughout the above description may be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, optical fields or optical particles, or any combination thereof.

[0107] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of the two. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps are generally described above in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. The technician may implement the described functionality in different ways for each specific application, but such implementation decisions should not be interpreted as resulting in a departure from the scope of the present invention.

[0108] The various illustrative logic modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or executed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in cooperation with a DSP core, or any other such configuration.

[0109] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. The software module may reside in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read and write information from / to the storage medium. In an alternative, a storage medium may be integrated into a processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and the storage medium may reside in a user terminal as discrete components.

[0110] In one or more exemplary embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented as a computer program product in software, each function may be stored on or transmitted by a computer-readable medium as one or more instructions or codes. Computer-readable media include both computer storage media and communication media, including any medium that facilitates the transfer of a computer program from one place to another. Storage media may be any available medium that can be accessed by a computer. As an example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer. Any connection is also properly referred to as a computer-readable medium. For example, if the software is transmitted from a website, a server, or other remote source using a coaxial cable, a fiber optic cable, a twisted pair, a digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwaves, the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwaves are included in the definition of the medium. Disk and disc as used herein include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and Blu-ray disc, wherein disk often reproduces data magnetically, while disc reproduces data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0111] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein, but should be granted the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting compound properties based on the fusion of molecular structure and physical information, characterized in that: Includes steps: Get a dataset containing SMILES descriptors of molecules; Converting the data set into a molecular graph embedding vector and a materialized information descriptor, wherein the molecular graph embedding vector is determined by an unsupervised learning graph convolutional neural network model; fusing the molecular graph embedding vector and the materialized information descriptor to determine a molecular mixture descriptor; as well as The molecular mixture descriptor is input into a molecular property prediction model to predict the properties of the compound, and the molecular property prediction model is constructed based on a multi-layer perceptron.

2. The method for predicting compound properties according to claim 1, wherein: The step of converting the data set into a molecular graph embedding vector and a materialized information descriptor comprises: Converting the molecular SMILES descriptors in the data set into molecular objects using the RDKit library; and The molecular graph embedding vector and the materialized information descriptor are determined based on the molecular object.

3. The method for predicting compound properties according to claim 2, characterized in that: The step of determining the molecular graph embedding vector and the materialized information descriptor based on the molecular object comprises: Graph structure data is determined based on the molecular object, and the graph structure data is converted into the molecular graph embedding vector according to the pre-trained unsupervised learning graph convolutional neural network model.

4. The method for predicting compound properties according to claim 3, characterized in that: The molecular object includes properties of atoms and bonds, and a topological structure of the molecule, the graph structure data includes an adjacency matrix and a feature matrix, and the step of determining the graph structure data based on the molecular object includes: The atoms of the molecule are used as graph nodes, the bonds are used as edges, the properties of the atoms are used as feature vectors of the graph nodes, and the properties of the bonds are used as feature vectors of the edges, so as to construct the adjacency matrix and the feature matrix.

5. The method for predicting compound properties according to claim 1, wherein: The unsupervised learning graph convolutional neural network model includes a linear change layer, a graph convolutional neural network layer and a fully connected neural network layer.

6. The method for predicting compound properties according to claim 5, characterized in that: The linear variation layer includes a linear layer and an embedding layer.

7. The method for predicting compound properties according to claim 2, wherein: The step of determining the molecular graph embedding vector and the materialized information descriptor based on the molecular object comprises: The molecular object is calculated using the RDKit library to determine the physicochemical information descriptor.

8. The method for predicting compound properties according to claim 1, wherein: The physicochemical information descriptors include charge distribution, molecular weight, number of atoms, number of rings, type and distribution of functional groups, electrical properties, configuration, topological structure, and biological activity of the molecule.

9. The method for predicting compound properties according to claim 1, wherein: The step of fusing the molecular graph embedding vector and the materialized information descriptor to determine the molecular mixture descriptor comprises: The molecular mixture descriptors are standardized and normalized.

10. The method for predicting compound properties according to claim 1, wherein: The molecular property prediction model includes a plurality of fully connected layers and a plurality of random inactivation layers in alternating combination and an output fully connected layer for output.

11. A compound property prediction system based on the fusion of molecular structure and physical information, characterized in that: include: a memory having computer instructions stored thereon; as well as A processor is connected to the memory and is configured to execute computer instructions stored in the memory to implement the compound property prediction method based on the fusion of molecular structure and physical information as described in any one of claims 1 to 10.

12. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the compound property prediction method based on the fusion of molecular structure and physical information as described in any one of claims 1 to 10 is implemented.

Images