CN111916143A - Molecular activity prediction method based on fusion of multiple substructure features - Google Patents

Abstract

The invention discloses a molecular activity prediction method based on the fusion of multiple substructure features. The neural network is trained by extracting the substructure features of molecular graphs, which overcomes the problems of closed loop extraction of substructures, poor network prediction accuracy and difficult calculation in the prior art. question. The steps realized by the present invention are: 1) transforming drug molecule information into a molecular feature matrix; 2) selecting an initial node; 3) obtaining multiple substructures; 4) calculating the similarity of the substructures; 5) fusing the substructure feature matrix; 6) train the neural network; 7) judge whether the training neural network converges; 8) obtain the molecular activity to be predicted. The invention has the advantages of distinguishing differences between different substructures, solving the problem of noise in molecular graphs, and predicting molecular activity with high precision.

Description

Molecular activity prediction method based on fusion of multiple substructure features

technical field

The invention belongs to the field of biotechnology, and further relates to a molecular activity prediction method based on fusion of multiple structural features in the field of bioactivity technology. The present invention can use the substructure information of a class of drug molecules and their corresponding biological activities to predict the influence of an unknown class of drug molecules on the biological activity.

Background technique

Molecular activity prediction technology refers to the use of a class of drug molecule structure information and its corresponding biological activity impact training neural network model, the model can use the unknown similar drug molecule structure information to predict its biological activity impact. In this way, the model can screen out the molecules of drug compounds that are more suitable for biological activity analysis in the biological laboratory from a large range of similar molecules. In order to convert drug molecules into information that can be recognized by computers, the molecular structure information is converted into molecular graphs, and the influence of drug molecules on biological activity is quantified as molecular graph labels. At present, molecular activity prediction technology can not only simplify the drug development process, reduce the potential safety hazards of biological experiments, but also save the cost of biological experiments. Current molecular activity prediction techniques face the challenge of node label noise in molecular graphs.

In his paper "Deep Graph Kernels" (knowledge discovery and data mining conference 2015), Pinar Yanardag proposed a method to predict the molecular activity of a class of molecular graphs by comparing the similarity of their substructures between them. This method divides a class of molecular graphs into a training set and a test set, divides the molecular graph of the training set into multiple substructures, uses the training set substructures and labels to train the neural network model, and finally uses the molecular graph substructures of the test set to train the neural network model. The similarity of the molecular graph substructures of the ensemble yields the molecular graph label of the test ensemble. The disadvantage of this method is that since a molecular map is randomly divided into a variety of different substructures, different substructures may predict different results, resulting in a decrease in the accuracy of predicted molecular map labels.

In his paper "Graph classification using structural attention" (knowledge discovery and data mining conference 2018), J.B.Lee proposed an attentional neural network-based molecular activity prediction method. The method divides a class of molecular graphs into a training set and a test set, uses the attention mechanism to find the substructure of the molecular graph in the training set, and uses its substructure and molecular graph labels to train the LSTM model. Finally, the model uses the attention mechanism to find the substructure of the test set molecular graph and predict its test set molecular graph label. The disadvantage of this method is that the training network cannot distinguish the difference between the substructures due to the closed loop of the found molecular graph substructures.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a molecular activity prediction method based on the fusion of multiple substructure features in view of the shortcomings of the above-mentioned existing molecular activity prediction technologies, which is used to solve the problem of noisy molecules in the process of molecular activity prediction. It is difficult to extract structural features and the prediction accuracy is poor.

The idea of realizing the purpose of the present invention is: according to the unique structural features and node label features in the molecular graph substructure, use a random walk to extract the molecular graph substructure set, improve the molecular graph substructure feature information, and combine some of the substructures. The fusion features are input into the constructed neural network training model to achieve the purpose of more accurate and rapid prediction of molecular activity.

The specific implementation steps of the present invention include the following:

(1) Obtain the feature matrix corresponding to the drug molecule information:

After one-hot encoding the atoms in a drug molecule based on bytes, a one-hot encoding feature matrix is obtained, the key-value pairs between the drug atoms are represented as a neighborhood feature matrix, and the activity of the drug molecule is based on the word. One-hot encoding is performed on the section to obtain the one-hot encoding label feature matrix;

(2) Select the initial node:

(2a) The atoms of the drug molecule are represented as nodes, the chemical bonds between atoms are represented as edges, and the activity of drug molecules is represented as molecular graph labels, and the molecular graph is composed of nodes, edges and molecular graph labels;

(2b) Use the Betweenness method to calculate the centrality value of each node in the molecular graph, and select the node with the highest node centrality value as the initial node;

(3) Extract multiple substructure features of molecular graph:

Starting from the initial node, use the random walk method to select l non-repetitive nodes from the molecular graph that are less than the number of nodes in the molecular graph to form the substructure of the molecular graph, and use the same method to select a set of substructures;

(4) Calculate the similarity of substructures:

(4a) each substructure in the substructure set is based on node coding, and obtains the characteristic matrix of this substructure;

(4b) Using the similarity formula, calculate the similarity of every pair of substructures in the substructure set:

Among them, J _m,n represents the similarity between the mth substructure and the nth substructure in the substructure set, g represents the feature matrix corresponding to the mth substructure in the substructure set, and p represents the nth substructure in the substructure set. Eigen matrix, |·| represents the matrix modulo operation, ∩ represents the intersection operation, and ∪ represents the union operation;

(4c) Store all substructures whose similarity is greater than or equal to a threshold into a similar set, and then store the remaining substructures into a dissimilar set, where the threshold is in the range of (0.5, 1), according to different molecules Selection of the number of nodes in the graph class;

(5) Fusion substructure feature matrix:

Averaging all substructure feature matrices in the similar set to obtain a fused substructure feature matrix;

(6) Training the neural network:

(6a) Input the fused substructure features into a 4-layer multilayer perceptron neural network, output the predicted molecular map label, and use the cross-entropy loss function to calculate the real molecular map label corresponding to the predicted molecular map label loss value between;

(6b) Arbitrarily select two substructure features from the dissimilar set, input the selected two substructure features into a 4-layer multilayer perceptron neural network, output the predicted molecular graph label, and use the cross entropy loss function to calculate The loss value between the real molecular map labels corresponding to the predicted molecular map labels;

(6c) Superimpose the above two loss values to obtain the loss value of the training neural network;

(7) Judging whether the loss value of the training neural network converges, if so, stop training, obtain a trained multilayer perceptron neural network, and execute step (8), otherwise, execute step (3);

(8) Input the same molecular graph to be predicted into the trained multilayer perceptron neural network, output the molecular graph label, and obtain the activity type corresponding to the molecular graph label.

Compared with the prior art, the present invention has the following advantages:

First, since the fusion substructure feature matrix of the present invention averages all substructure feature matrices in the similar set, a fused substructure feature matrix is obtained, and the fused substructure feature matrix is input into the training network to obtain the molecular map label. , overcomes the problem in the prior art that a molecular map is randomly divided into a variety of different substructures, and different substructures may predict different results, resulting in a decrease in the accuracy of the predicted molecular map label, so that the present invention has excellent extraction molecular map. The characteristics of substructure features improve the accuracy of predicting molecular map labels.

Second, the present invention uses the random walk method to select non-closed-loop substructures from the molecular graph, and divide them into similar substructure sets and dissimilar substructure sets to train the network, which overcomes the problem of finding in the prior art. The molecular graph substructures have closed loops, which leads to the problem that the training network cannot distinguish the differences between the substructures, so that the present invention can distinguish the differences between different substructures.

Description of drawings

Fig. 1 is the flow chart of the present invention;

Fig. 2 is the schematic diagram that the present invention utilizes random walk to extract molecular graph substructure process;

FIG. 3 is a schematic diagram of the present invention fusing similar set substructure features.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings.

1, the specific steps of the present invention will be further described.

In step 1, a feature matrix corresponding to the drug molecule information is obtained.

After one-hot encoding the atoms in a drug molecule based on bytes, a one-hot encoding feature matrix is obtained, the key-value pairs between the drug atoms are represented as a neighborhood feature matrix, and the activity of the drug molecule is based on the word. Sections are one-hot encoded to obtain the one-hot encoded label feature matrix.

Step 2, select the initial node.

The atoms of drug molecules are represented as nodes, the atoms of drug molecules are represented as nodes, the atomic properties of drug molecules are represented as node labels, the chemical bonds between atoms are represented as edges, and the activity of drug molecules is represented as molecular graph labels. Links and Molecular Graph labels make up the Molecular Graph.

Using the Betweenness method, the centrality value of each node in the molecular graph is calculated, and the node with the highest node centrality value is selected as the initial node.

Step 3, extracting multiple sub-structure features of the molecular graph.

In the following, the process of extracting the substructure of the molecular graph by using random walk will be described in detail with reference to FIG. 2 .

The nodes with black edges in Figure 2 represent the nodes selected by the random walk method, the nodes without black edges represent the unselected nodes, the numbers in the nodes represent the serial numbers of atoms in the molecule, and c _t-1 represents the random walk The node pointed to at time t-1, c _t represents the node pointed to by the random walk at time t. Figure 2(a) shows a schematic diagram of node 4 being selected by random walk at time t-1. Figure 2(b) shows a schematic diagram of the process of backtracking to node 9. It can be seen from Figure 2(b) that there is no unselected node on the edge of node 9. In order to find the unselected node, c _t-1 There may be unpicked nodes before continuing to backtrack. Figure 2(c) shows a schematic diagram of c _t-1 backtracking to node 2. It can be seen from Figure 2(c) that node 1 is not selected in current node 2. Figure 2(d) shows a schematic diagram of selecting node 1. From 2(d), it can be seen that since node 1 is not selected, node 1 is selected by random walk, and c _t is set to point to node 1 and its edge is set to black.

As can be seen from Figure 2, starting from the initial node, using the random walk method, select l non-repetitive nodes from the molecular graph that are less than the number of nodes in the molecular graph to form the substructure of the molecular graph, and use the same method to select a substructure set.

Step 4: Calculate the similarity of substructures.

In the first step, each substructure in the substructure set is encoded based on the node, and the feature matrix of the substructure is obtained;

The second step is to use the similarity formula to calculate the similarity of every pair of substructures in the substructure set:

Among them, J _m,n represents the similarity between the mth substructure and the nth substructure in the substructure set, pm represents the feature matrix corresponding to the mth substructure in the substructure set, and p _n represents the _nth substructure in the substructure set. The characteristic matrix of the structure, |·| represents the matrix modulo operation, ∩ represents the intersection operation, and ∪ represents the union operation.

The third step, |p _m ∩ p _n | represents the Jaccard similarity and Hamming distance of the node sequence features of the substructures p _m and p _n . The specific process is as follows: First, the node labels of the corresponding substructures are obtained according to the substructures Jump sequences, for example, substructure p _m has node sequence features [1, 1, 2, 2, 3] and p _n has node sequence features [1, 2, 3, 2, 3], where elements are represented as node labels . According to [1, 1, 2, 2, 3], the jump sequence [0, 1, 0, 1] is obtained, and p _n = [1, 2, 3, 2, 3] to obtain the jump sequence [1, 1, 1] , 1], where 1 in the jump sequence means that the labels of adjacent nodes in the substructure sequence have changed, and 0 means that there has been no change. Then, the node label hopping sequences of the two paths are normalized by the Jaccard similarity measure and the Hamming distance measure into the similarity value.

The fourth step is to store all substructures whose similarity is greater than or equal to the threshold into the similar set, and then store the remaining substructures into the dissimilar set. The threshold is in the range of (0.5, 1), according to different The number of nodes in the molecular graph class is selected

Step 5, fuse the substructure feature matrix.

The following describes the process of substructure fusion in similar sets in detail with reference to FIG. 3 .

In Figure 3, the substructure is represented as a row of node sequences, and the node sequences can be represented as feature matrices. A row of node sequences corresponds to a feature matrix. Figure 3(a) shows a schematic diagram of the feature matrices of the three substructures. Figure 3(b) shows the fused substructure feature matrix. The three substructure feature matrices in Fig. 3(a) are averaged to obtain the substructure feature matrix after fusion as shown in Fig. 3(b).

A fused substructure feature matrix is obtained by averaging all substructure feature matrices in the similar set. A substructure can be encoded into a feature matrix according to its node features, and the feature matrices of multiple substructures in a similar set are averagely fused into a substructure feature matrix.

Step 6, train the neural network.

The first step is to input the fused substructure features into a 4-layer multilayer perceptron neural network, output the predicted molecular map label, and use the cross-entropy loss function to calculate the real map label corresponding to the predicted molecular map label. The loss value between λ ₁ .

The second step is to arbitrarily select two substructures from the dissimilar set, input the selected two substructure features into a 4-layer multilayer perceptron neural network, and output the predicted molecular graph labels. Using the cross entropy loss function, calculate The loss value λ ₂ between the true graph labels corresponding to the predicted molecular graph labels.

The third step, according to the following formula, get the neural network loss value L

L=pλ ₁ +(1-p)λ ₂

Among them, p represents the bias value, which is selected according to the number of nodes in different molecular graph classes in the range of (0.8, 1).

Step 7, judge whether the loss value of the neural network converges, if so, stop training, obtain a trained multilayer perceptron neural network, and execute step (8), otherwise, execute step (3).

Step 8: Input the molecular graph of the same type to be predicted into the trained multilayer perceptron neural network, output graph labels, and obtain the activity types corresponding to the labels.

The effect of the present invention is further described below in conjunction with the simulation experiment:

1. Simulation experimental conditions:

The hardware platform of the simulation experiment of the present invention is: the processor is Intel i5 3470 CPU, the main frequency is 3.20GHz, and the memory is 4GB.

The software platform of the simulation experiment of the present invention is: Ubuntu operating system and python 3.6.

The input molecular map data set used in the simulation experiment of the present invention is the following four types of molecular compounds, NCI-1, NCI-33, NCI-1, NCI-33, NCI-83 and NCI-123. NCI-1 is a balanced dataset of compound datasets for screening activity in non-small cell lung cancer, and there are two types of activity markers. NCI-33 is a balanced dataset of compound datasets for screening activity in melanoma, and there are two types of activity markers. , NCI-83 is a balanced dataset of compound datasets for breast cancer screening activity, there are two types of activity markers, and NCI-123 is a balanced dataset dataset of compound datasets for breast cancer screening activity, with two activity markers .

2. Simulation content and result analysis:

The simulation experiment of the present invention adopts the present invention and three existing technologies (Kernel-GK method, Deep-SP method and GAM method) to classify molecular graphs, respectively, to obtain classification results.

In the simulation experiments, the three existing technologies used are:

The existing Kernel-GK method refers to the molecular graph prediction method proposed by Nino Shervashidze et al. in "Efficient graphletkernels for large graph comparison" 2009 Conference on Artificial Intelligence and Statistics, referred to as the similar substructure graph kernel method.

The existing Deep-SP method refers to the molecular graph prediction method proposed by Pinar Yanardag et al.

The existing GAM method refers to the molecular graph prediction method proposed by J.B.Lee et al. in "Graph classification using structural attention", the 2018 knowledge discovery and data mining conference, referred to as the attention neural network graph classification method.

The classification results of the three methods were evaluated by the average precision evaluation index. Use the following formula to calculate the classification accuracy of the datasets NCI1, NCI33, NCI83 and NCI123, and draw all the calculation results into Table 1:

Table 1. Quantitative analysis table of the present invention and each prior art classification result in simulation experiment

It can be seen from Table 1 that the classification accuracy AA index of the present invention is higher than that of the three prior art methods, which proves that the present invention can obtain higher molecular activity prediction accuracy.

The above simulation experiments show that the method of the present invention utilizes the idea of random walk to identify different substructure features of molecular graphs, wherein the open-loop requirement for extracting substructures overcomes the problem of extracting substructures of molecular graphs in the prior art method. There is a closed loop, which leads to the disadvantage that the training network cannot distinguish the differences between substructures. In addition, through the feature fusion of similar substructures, the difference between the features of different substructures is reduced, and the problem that the prediction accuracy of molecular activity is reduced due to the different prediction results of different substructures in the existing technical methods is solved. Compared with other comparison methods, the multi-layer neural network prediction model proposed by the present invention has short training time and high network generalization, and is a very practical molecular activity prediction method.

Claims (3)

1. A molecular activity prediction method based on multi-substructure feature fusion is characterized in that a random walk method is used for extracting a plurality of substructure features of a molecular diagram, and the fused substructure features are input into a trained multilayer neural network to predict molecular activity, and the method specifically comprises the following steps:

(1) obtaining a characteristic matrix corresponding to the drug molecule information:

carrying out one-hot encoding on atoms in a drug molecule based on bytes to obtain one-hot encoding characteristic matrix, expressing bond value pairs between the drug atoms into a neighborhood characteristic matrix, and carrying out one-hot encoding on the activity of the drug molecule based on bytes to obtain one-hot encoding label characteristic matrix;

(2) selecting an initial node:

(2a) representing atoms of drug molecules into nodes, representing chemical bonds among the atoms into connecting edges, representing the activity of the drug molecules into component subgraph labels, and forming component subgraphs by the nodes, the connecting edges and the molecular label labels;

(2b) calculating the centrality value of each node in the molecular graph by using a Betweenness method, and selecting the node with the highest centrality value as an initial node;

(3) extracting a plurality of substructure characteristics of the molecular diagram:

starting from an initial node, selecting l substructures of the component subgraph without repeated node groups, wherein the number of the substructures is less than that of the nodes of the molecular subgraph, from the molecular subgraph by using a random walk method, and selecting one substructure set by using the same method;

(4) calculating the similarity of the substructures:

(4a) coding each substructure in the substructure set based on nodes to obtain a characteristic matrix of the substructure;

(4b) and calculating the similarity of every two substructures in the substructure set by using a similarity formula:

wherein, J_m,nRepresenting the similarity between the mth substructure and the nth substructure in the substructure set, g representing a characteristic matrix corresponding to the mth substructure in the substructure set, p representing the characteristic matrix of the nth substructure in the substructure set, | · | representing matrix modulo operation, | · representing intersection operation, and u representing union operation;

(4c) storing all the substructures with the similarity greater than or equal to a threshold value into a similar set, and storing the rest substructures into a different set, wherein the threshold value is in a range of (0.5,1), and is selected according to the number of nodes in different molecular diagram classes;

(5) fusion substructure feature matrix:

averaging all the substructure feature matrixes in the similar set to obtain a fused substructure feature matrix;

(6) training a neural network:

(6a) randomly selecting two substructure characteristics from different sets, inputting the two substructure characteristics into a multilayer perceptron neural network with 4 layers, outputting predicted molecular icon labels, and calculating loss values between the real molecular icon labels corresponding to the predicted molecular icon labels by using a cross entropy loss function;

(6b) inputting the fused substructure characteristics into a 4-layer multi-layer perceptron neural network, outputting predicted molecular icon labels, and calculating loss values between the predicted molecular icon labels and the real molecular icon labels corresponding to the predicted molecular icon labels by using a cross entropy loss function;

(6c) superposing the two loss values to obtain a loss value of the training neural network;

(7) judging whether the loss value of the trained neural network is converged, if so, stopping training to obtain the trained multilayer perceptron neural network, and executing the step (8), otherwise, executing the step (3);

(8) inputting the same kind of molecular graph to be predicted into the trained multilayer perceptron neural network, and outputting the molecular graph label to obtain the activity type corresponding to the molecular graph label.

2. The method for predicting the activity of a molecule based on the fusion of various substructure features according to claim 1, wherein the step of the random walk method in step (3) is: and selecting unselected nodes in the node neighborhood of the molecular graph by using a random walk method, and backtracking to the previously selected nodes if the unselected nodes do not exist in the current node neighborhood in the selection process, wherein the node neighborhood represents all other node sets connected with the node in the molecular graph.

3. The method for predicting the activity of a molecule based on the fusion of various substructure features according to claim 1, wherein the step of adding the two loss values in step (6c) is:

firstly, inputting the fused feature matrix into a neural network by using a cross entropy loss function to obtain a loss value lambda₁Inputting the selected different set of substructures into a neural network to obtain a loss value lambda₂。

Secondly, obtaining a neural network loss value L according to the following formula:

L＝pλ₁+(1-p)λ₂

wherein p represents a bias value, which is selected in the range of (0.8,1) according to the number of nodes in different molecular graphs.

Images