A method for predicting structure of small molecule substances based on a multi-layer perceptron network model

Abstract

The application discloses a small molecule structure prediction method based on a multilayer perceptron network model. The input of the small molecule structure prediction model based on the multilayer perceptron network is mass spectrum data, and the output is a molecular fingerprint of a substance to which the input mass spectrum data belongs, namely a fixed-length binary code; the small molecule structure prediction model comprises a feature extraction network and a prediction network, the feature extraction network comprises n series of feedforward neural networks with the same structure, a residual network and an output layer, and n is greater than or equal to 1. Each feedforward neural network comprises a full connection layer and a nonlinear activation layer. The small molecule structure prediction model based on the multilayer perceptron network is provided, and a modified residual network is introduced to improve the performance of the feature network in calculating the relationship between mass spectrum data and reduce the space-time complexity.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of small molecule substance structure prediction, and particularly relates to a small molecule substance structure prediction method based on a multilayer perception network model. TECHNICAL BACKGROUND

[0002] Most drugs are composed of small molecules, which are the universal language of cell communication, metabolism and signal transduction, and form extremely diverse chemical substances. When the small molecule substance structure is obtained, certain spectrum prediction or accurate classification of the small molecule substance can be performed according to the existing related spectrum, and therefore it is necessary to construct the small molecule substance structure.

[0003] Generally, researchers obtain mass spectrum data through a mass spectrometer, and then perform spectrum matching through fragmentation mechanism or a spectrum library to perform substance structure speculation, but factors such as mass spectrometry technology, cost problems, human factors and external interference factors make the obtained spectrum not an ideal spectrum or the calculation time is longer, so it is easy to cause substance classification error or spectrum matching error, and thus errors are easily caused in substance structure speculation.

[0004] Therefore, it is necessary to establish a new small molecule structure prediction model to avoid as many interference factors as possible and more accurately predict the substance structure, thereby helping researchers to quickly perform substance classification and spectrum matching. SUMMARY

[0005] A first object of the application is that the current mainstream prediction model has high time complexity and space complexity, and only explains the molecular structure from the data-driven perspective, and the application provides a small molecule substance structure prediction method based on a multilayer perception network model, effectively solves the shortcomings of the existing method, and achieves good prediction results on actual data.

[0006] The method of the application comprises the following steps:

[0007] Step (1), constructing a data set

[0008] All secondary small molecule mass spectrum data and SMILES molecular codes are obtained from a global natural product social molecular network, and then all the obtained mass spectrum data are set as one-dimensional array information of the same length after preprocessing, the index is the mass-to-charge ratio, and the information in the index is intensity information normalized to 0-1. The SMILES molecular code is converted into a fixed-length binary expression molecular fingerprint.

[0009] Step (2), establishing a small molecule structure prediction model based on a multilayer perception network, and training the same by using a training set

[0010] The input of the small molecule structure prediction model based on the multi-layer perception network is mass spectrum data, and the output is the molecular fingerprint of the substance to which the input mass spectrum data belongs, that is, a fixed-length binary code; the small molecule structure prediction model based on the multi-layer perception network comprises a feature extraction network and a prediction network, and specifically comprises the following:

[0011] 2-1 Constructing the feature extraction network to extract the features of the input mass spectrum data

[0012] The feature extraction network comprises n serially connected feedforward neural networks with the same structure and a residual network, and n≥1. Each feedforward neural network comprises a fully connected layer and a nonlinear activation layer.

[0013] Considering that simply using the fully connected layer and the nonlinear activation layer can only calculate the weight of the mass spectrum intensity information under different mass-to-charge ratios instead of calculating the intrinsic relationship between the mass spectrum data, and in order to reduce the time complexity and the space complexity as much as possible and closely relate the relationship between the mass spectrum data, a modified residual network is added to improve the performance of the feature network in calculating the relationship between the mass spectrum data and reduce the time and space complexity, that is,

[0014] A t =sigmod(WA t-1 +b)+A t-1

[0015] Wherein A t is the output feature of the residual network at the current time, A t-1 is the output feature of the residual network at the last time, W is a weight matrix, b is a bias, and sigmod is a nonlinear activation function.

[0016] 2-2 Constructing the prediction network to output the molecular fingerprint of the current identified substance according to the extracted low-dimensional mass spectrum features.

[0017] The prediction network is composed of a fully connected layer, and the low-dimensional mass spectrum features output by the feature extraction network are taken as the input, and the binary code of the molecular fingerprint of the current unknown substance is output.

[0018] 2-3 Loss function during training

[0019] l(x,y)=L={l1,...,l N} T ,l n =(x n -y n ) 2

[0020] Loss=mean(L)

[0021] Wherein, Loss represents the mean square error loss, l is the mean loss corresponding to the mass spectrum data, x n , yn respectively represent the binary encoding of the predicted and actual molecular fingerprints.

[0022] 2-4 Network training using Adam algorithm.

[0023] Step (3), applying a small molecule structure prediction model based on a multi-layer perceptron network

[0024] 3-1 After pre-processing the electrospray ion source mass spectrometry data of the substance to be detected, input the data into the trained feature extraction network for feature extraction, and output the corresponding low-dimensional mass spectrometry features.

[0025] 3-2 Input the obtained low-dimensional mass spectrometry features into the trained prediction network, and output the molecular fingerprint of the substance to be detected.

[0026] A second object of the present application is to provide a computer-readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method.

[0027] A third object of the present application is to provide a computing device comprising a memory and a processor, the memory having stored therein executable code, the processor executing the executable code to implement the method.

[0028] The present application has the following beneficial effects:

[0029] 1. Transfer learning - using small molecule data in the global natural product social molecular network as input for the multi-layer perceptron network, greatly increasing the sample data for model training, effectively solving the small sample problem of supervised learning methods in other small molecule structure prediction tasks.

[0030] 2. Small molecule library matching - effectively solving the problem of matching errors caused by external factors causing spectrum disturbances, making the library matching problem more accurate.

[0031] 3. Small molecule classification task - effectively solving the problem of classification errors caused by external factors causing spectrum disturbances, directly using substance structure information for classification, making the classification effect better.

[0032] 4. The present application proposes a small molecule structure prediction model based on a multi-layer perceptron network, which introduces a modified residual network to improve the performance of the feature network in calculating the relationship between mass spectrometry data and reduce the space-time complexity.

[0033] 5. The application proposes to maximize the inter-class difference and minimize the intra-class difference in the design of the loss function of the small molecule structure prediction model based on the multi-layer perception network, so that the extracted features are more accurate, and the performance of the detection model is improved.

[0034] 6. The application proposes that the small molecule structure prediction model based on the multi-layer perception network is effective on the commonly used mass spectrum data (electrospray ion source), and is verified on public data sets and actual data sets, and high prediction accuracy is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is the architecture diagram of the small molecule structure prediction model based on the multi-layer perception network of the application;

[0036] Figure 2 is the structure diagram of the feature extraction network in the small molecule structure prediction model based on the multi-layer perception network of the application;

[0037] Figure 3 is the main structure of the prediction network in the small molecule structure prediction model based on the multi-layer perception network of the application. DETAILED DESCRIPTION

[0038] The application will be further described below in conjunction with the drawings.

[0039] A small molecule substance structure prediction model based on a multi-layer perception network, the specific steps are as follows:

[0040] Step 1, pre-processing the mass spectrum data used for model input

[0041] The mass spectrum can be represented as L=(m x ,a x ), where the mass-to-charge ratio m x ={m1,m2,m3...,m n}, and the relative intensity a x ={a1,a2,a3...,a n}.

[0042] Convert all mass spectrum data {L1,L2,L3...,L n} to standard format

[0043] Where the mass-to-charge ratio is defined as follows:

[0044]

[0045] Make its step size 1, and update the corresponding relative intensity at each mass-to-charge ratio:

[0046]

[0047] After that, the relative intensity is normalized by dividing the base peak:

[0048]

[0049] Step 2, establishing a prediction model based on a multi-layer perception network

[0050] As shown in Figure 1 , the input of the prediction model is composed of small molecule mass spectrometry data processed by the global natural product social molecule network, the multi-layer perception network is used to train the mass spectrometry feature extraction network, and then the output mass spectrometry features are sent to the prediction network to output the molecular fingerprint code of the substance, i.e. the fixed-length binary code.

[0051] 2-1 Constructing a feature extraction network to extract features of input mass spectrometry data

[0052] The specific structure of the feature extraction network is shown in Figure 2 , which includes three serially connected feedforward neural networks with the same structure and a residual network.

[0053] Each feedforward neural network includes a fully connected layer, a nonlinear activation layer, and a fully connected layer connected in sequence; wherein the fully connected layer network size is 1024, and the activation function is a nonlinear activation function.

[0054] The first feedforward neural network input is the preprocessed mass spectrometry data, the second and third feedforward neural network inputs are the sum of the output of the previous feedforward neural network and the output of the residual network, and the output is a low-dimensional mass spectrometry feature.

[0055] The residual network is used to improve the performance of the feature network in calculating the relationship between mass spectrometry data and reduce the space-time complexity, i.e.

[0056] A t = sigmod(WA t-1 +b) + A t-1

[0057] Where A t is the output feature of the current time residual network, A t-1 is the output feature of the previous time residual network, W is the weight matrix, b is the bias, and sigmod is the nonlinear activation function.

[0058] 2-2 Constructing a prediction network to output the molecular fingerprint of the current substance to be measured according to the extracted mass spectrometry features Figure 3 The extracted mass spectrometry features are obtained through a fully connected layer.

[0059] Step 3, training the prediction model based on the multi-layer perception network

[0060] 3-1 Constructing training data set

[0061] Preprocessing operation is performed on all mass spectrum data to convert it into standard format. The label of the training sample is the encoding of the molecular fingerprint of the current substance, i.e. a fixed-length binary string.

[0062] 3-2 Designing loss function

[0063] The loss function used in training is shown in Figure 1 The total loss function design is as follows

[0064] l(x,y)=L={l1,...,l N} T ,l n =(x n -y n ) 2

[0065] Loss=mean(L)

[0066] Wherein, Loss represents the mean square error loss, l is the average loss corresponding to the mass spectrum data x n , y n respectively represent the encoding of the predicted and actual molecular fingerprint.

[0067] 3-3 Network training using Adam algorithm, in the training process, when the loss obtained converges and no longer changes significantly within 5 consecutive iteration times, the training of the model is stopped.

[0068] Step 4, applying the prediction model based on the multilayer perception network

[0069] 4-1 Input the mass spectrum data of the substance to be predicted into the feature extraction network of the multilayer perception network module to obtain the corresponding low-dimensional mass spectrum features.

[0070] 4-2 Input the mass spectrum features into the prediction network to output the molecular fingerprint of the substance to be detected.

[0071] The effects of the present application are demonstrated as follows

[0072] 1. Selection of data set

[0073] The small molecule structure prediction model based on the multilayer perception network proposed in the present application is verified on the commonly used electrospray ion source data set, wherein the model is trained on the electrospray ion source public data set, and the proposed prediction model is applied on the electrospray ion source actual data set, and the detection accuracy of the detection model is counted.

[0074] The detection accuracy Acc of the model is defined as follows:

[0075]

[0076] 2. Model verification results

[0077] 2-1 Electrospray ion source actual data set

[0078] In this data set, a total of 4154 small molecules were included. The experimental results obtained by verifying on the data set are as follows

[0079] Table 1. Prediction results of the model of the application on the electrospray ion source actual data set

[0080] MSE cosine_similarity Invention model 0.012775 84.3459％

[0081] The above table verifies that the proposed prediction model achieves a high prediction accuracy on the electrospray ion source actual mass spectrum data set, and therefore the model further proves that it is reliable in the small molecule structure prediction task based on the electrospray ion source mass spectrum.

Claims

1. A method for predicting the structure of small molecules based on a multilayer perceptron network model, characterized in that... The method includes the following steps: Step (1) Constructing the dataset All secondary small molecule mass spectrometry data and their SMILES molecular codes were obtained from the global natural product social molecular network. Then, all the obtained mass spectrometry data were preprocessed and set as one-dimensional array information of the same length. The index is the mass-to-charge ratio, and the information in the index is the intensity information normalized to 0~1. The SMILES molecular codes were transformed into molecular fingerprints of fixed-length binary expression. Step (2): Establish a small molecule structure prediction model based on a multilayer perceptron network and train it using a training set; The small molecule structure prediction model based on a multilayer perceptron network takes mass spectrometry data as input and outputs the molecular fingerprint of the substance to which the input mass spectrometry data belongs, i.e., a fixed-length binary code; the specific steps are as follows: 2-1 Construct a feature extraction network to extract features from the input mass spectrometry data. The feature extraction network comprises n cascaded feedforward neural networks with identical structures, and a residual network, where n≥1; Each feedforward neural network includes a fully connected layer and a non-linear activation layer; The first feedforward neural network input is the preprocessed mass spectrometry data, and the input of each of the other feedforward neural networks is the sum of the output of the previous feedforward neural network and the output of the residual network. The output is a low-dimensional mass spectrometry feature. The residual network is used to improve the performance of the feature network in calculating the relationships between mass spectrometry data and reduce the time and space complexity. ； in The output characteristics of the residual network at the current time. The output characteristics of the residual network at the previous time step. This is the weight matrix. For bias, It is a non-linear activation function; The preprocessing of mass spectrometry data specifically involves: Mass spectrometry This indicates that the mass-to-charge ratio is... relative strength ; All mass spectrometry data Convert to standard format ; The mass-to-charge ratio is defined as follows: ； Set its step size to 1, and simultaneously update the relative intensity corresponding to each mass-to-charge ratio: ； Then normalize the relative intensity by dividing it by the base peak: ； 2-2 Construct a prediction network to output the molecular fingerprint of the currently identified substance based on the extracted low-dimensional mass spectrometry features; The prediction network consists of a fully connected layer, which takes the low-dimensional mass spectrometry features output by the feature extraction network as input and outputs the binary code of the molecular fingerprint of the unknown substance. Step (3): Apply a small molecule structure prediction model based on a multilayer perceptron network; 3-1 The electrospray ionization mass spectrometry data of the substance to be detected is preprocessed and then input into a trained feature extraction network for feature extraction, and the corresponding low-dimensional mass spectrometry features are output. 3-2 Input the acquired low-dimensional mass spectrometry features into the trained prediction network to output the molecular fingerprint of the substance to be detected.

2. The method according to claim 1, characterized in that... The loss function used during the training of the small molecule structure prediction model based on a multilayer perceptron network is as follows: ； ； Where Loss represents the mean squared error loss. This corresponds to the average loss of the mass spectrometry data. These represent the binary codes for the predicted and actual molecular fingerprints, respectively.

3. The method according to claim 2, characterized in that... The small molecule structure prediction model based on a multilayer perceptron network is trained using the Adam algorithm.

4. A computer-readable storage medium, characterized in that... It stores a computer program that, when executed in a computer, causes the computer to perform the method described in any one of claims 1-3.

5. A computing device, comprising a memory and a processor, characterized in that... The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-3.

Images