Molecular mass spectrum prediction method based on deep learning feature fusion
Through the deep learning feature fusion method, small molecule features are extracted using MLP and GCN models and combined with LSTM models, the problem of low prediction performance of small molecule mass spectrometry in the existing technology is solved, and more accurate spectrogram prediction and rich mass spectrometry are achieved. Library generation, suitable for compound identification and other applications.
Patent Information
- Application Number
- CN202410920533.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-10
- Publication Date
- 2024-11-15
AI Technical Summary
The prior art has problems with data quality and quantity sensitivity in small molecule mass spectrometry prediction. Traditional methods are difficult to effectively process high noise mass spectrometry data, and the database scale and diversity are insufficient, resulting in low prediction performance.
Using a feature fusion method based on deep learning, chemical molecular coding features and molecular three-dimensional structural features are extracted through MLP and GCN models, and feature fusion is performed by combining LSTM models. Mass spectrometry spectrogram is further predicted through multi-layer MLP models to construct multi-layer perceptrons and Deep learning model of graph neural network.
More accurate prediction of small-molecule mass spectrometry spectrometry was achieved, a rich mass spectrometer library was generated, and the accuracy and efficiency of compound identification were improved. The experimental results performed better than other models on multiple data sets.
Smart Images

Figure CN118965252A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of small molecule mass spectrometry prediction, and in particular to a molecular mass spectrometry prediction method based on deep learning feature fusion. Background Art
[0002] Mass spectrometry plays an important role in the fields of chemistry, biomedicine and environmental science, and can reveal the quality and structure of molecules. Molecular structure prediction provides key support for new drug development, biomarker identification and environmental pollutant tracking by predicting the mass spectra of molecules. However, mass spectrometry data is affected by experimental conditions and noise, the database size and diversity are insufficient, and traditional molecular structure characterization methods are limited. Deep learning technology shows its potential in big data processing and complex pattern recognition, and has unique advantages in the field of spectrum prediction. However, the current deep learning algorithm is sensitive to the quality and quantity of data, and the performance of spectrum prediction needs to be further improved. Therefore, it is necessary to design a new set of deep learning models to adapt to high-noise mass spectrometry data, and make full use of existing mass spectrometry database resources to improve the accuracy and efficiency of small molecule mass spectrometry prediction. The present invention predicts spectra based on small molecule structures, constructs a deep learning fusion model, can accurately predict mass spectra based on small molecule compound structures, and can further generate a rich mass spectrometry library based on this technology, which can be further applied to applications such as compound identification. Summary of the invention
[0003] The purpose of the present invention is to provide a molecular mass spectrometry prediction method based on deep learning feature fusion in view of the shortcomings of the prior art.
[0004] The object of the present invention is achieved through the following technical solution: a molecular mass spectrometry prediction method based on deep learning feature fusion, comprising:
[0005] Based on the first MLP model and the GCN model, chemical molecular coding features and molecular three-dimensional structural features are extracted from the molecular structure respectively;
[0006] Splicing chemical molecular coding features and molecular three-dimensional structural features;
[0007] Input the concatenated feature vectors into the LSTM model to obtain the fused feature vector;
[0008] Based on the second MLP model, mining the key feature vectors required for predicting the molecular mass spectrum from the fused feature vectors;
[0009] The key feature vector is input into the third MLP model, and the predicted spectrogram data is output.
[0010] Furthermore, the molecular structure includes a small molecule structure.
[0011] Furthermore, before extracting the chemical molecular coding features and the molecular three-dimensional structural features from the molecular structure, feature characterization is performed, including:
[0012] First MLP model feature characterization: converting the molecular structure into SMILES code, and then converting the SMILES code into a molecular fingerprint in binary format;
[0013] GCN model feature characterization: converting the molecular structure into a molecular graph format, wherein the molecular structure includes atoms in the molecular structure and chemical bonds between atoms.
[0014] Furthermore, a plurality of residual blocks are connected in series to construct an MLP model, each residual block includes a fully connected layer, a nonlinear activation layer, a normalization layer, and another fully connected layer as a closed loop; the MLP model includes a first MLP model, a second MLP model and a third MLP model.
[0015] Furthermore, the multiple residual blocks are 6 residual blocks.
[0016] Furthermore, the three-dimensional structural features of the molecule are extracted from the molecular structure based on the GCN model, including:
[0017] Input the adjacency matrix A and the feature matrix X; X contains the initial feature information of each node, that is, the type of atoms in the molecular structure; A represents the connection relationship between nodes in the graph, that is, the chemical bonds between atoms;
[0018] Process the adjacency matrix by Add a self-join and normalize, where is the degree matrix of A+I, I is the identity matrix;
[0019] use To aggregate the features of adjacent nodes to update the feature representation of each node, and finally obtain the three-dimensional structure features of the molecule;
[0020] in, is the processed adjacency matrix, is the degree matrix of A+I, is the normalization of the degree matrix, H(l) is the node feature representation of the lth layer, and for the first layer, H (0) =X,W (l) is a learnable weight matrix and σ is a nonlinear activation function.
[0021] Furthermore, the LSTM model includes several layers of LSTM.
[0022] Furthermore, the several layers of LSTM are three layers of LSTM.
[0023] The beneficial effect of the present invention is that the small molecule mass spectrometry prediction method based on the deep learning feature fusion of multi-layer perceptron (MLP) and graph neural network (GCN) constructed by the present invention realizes more accurate prediction of small molecule mass spectrometry spectra, and can further generate a rich mass spectrometry library based on this technology, which can be further applied to applications such as compound identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.
[0025] Figure 1 It is a flow chart of a molecular mass spectrum prediction method using deep learning feature fusion provided by an embodiment of the present invention;
[0026] Figure 2 is a diagram of an MLP network structure provided by an embodiment of the present invention;
[0027] Figure 3 is a Res block structure diagram provided by an embodiment of the present invention;
[0028] Figure 4 This is a GCN network structure diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0029] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.
[0030] The present invention provides a molecular mass spectrometry prediction method based on deep learning feature fusion, such as Figure 1 As shown, the following steps are included:
[0031] Step 1: Extract chemical molecular coding features and molecular three-dimensional structural features from the molecular structure based on the first MLP model and the GCN model respectively;
[0032] Before extracting chemical molecular coding features and molecular three-dimensional structural features from the molecular structure, feature characterization is performed, including:
[0033] First MLP model feature representation: The first MLP model first receives molecular structures as input. These molecular structures are then converted into SMILES codes, and the SMILES codes are further converted into molecular fingerprints (FingerPrint) in binary format through a conversion process. The purpose of this process is to encode complex molecular structure information into a form suitable for deep learning model processing, thereby providing a basis for subsequent feature learning and prediction.
[0034] GCN model feature representation: In the input preprocessing stage, the model receives the structural data of a molecule as input. This data contains the atoms in the molecule (node features) and the chemical bonds between them (edge indices). This information is then converted into a molecular graph format that can be processed by the graph neural network.
[0035] The present invention adopts a dual-path feature encoding strategy to obtain a comprehensive feature representation of the molecular structure. Each path uses a different advanced feature extraction model, that is, feature fusion encoding based on the features obtained by the two deep learning models of MLP and GCN. Specifically:
[0036] The MLP model is constructed by connecting multiple residual blocks in series. Each residual block contains a fully connected layer, a nonlinear activation layer, a normalization layer, and another fully connected layer as a closed loop, such as Figure 3 As shown in the figure, the residual block is designed to solve the gradient vanishing or gradient exploding problems that may occur during deep network training. By introducing short-circuit connections, the gradient is allowed to flow directly through some layers. The input of the residual block is the output from the residual block of the previous level. The MLP model includes a first MLP model, a second MLP model and a third MLP model. The first MLP model is used to extract chemical molecule coding features from molecular structures.
[0037] Based on the GCN model, the three-dimensional structural features of the molecule are extracted from the molecular structure, such as Figure 4 As shown, A, B, C, and D represent four nodes, including:
[0038] Input adjacency matrix A and feature matrix X; X contains the initial feature information of each node. These features can be the attributes of the node, which are the types of atoms in the chemical structure of small molecules; A represents the connection relationship between nodes in the graph, that is, the chemical bonds between atoms. If node i is connected to node j, then A ij is 1 (or the weight of the edge), otherwise it is 0;
[0039] Process the adjacency matrix by Add a self-join and normalize, where is the degree matrix of A+I, I is the identity matrix;
[0040] use To aggregate the features of adjacent nodes to update the feature representation of each node. After multiple such updates, the features of the nodes will be able to capture topological information in a wider range. Specifically, the first layer of GCN is responsible for receiving the initial node features and performing the first feature transformation and aggregation of neighbor information. The middle layer of GCN further refines these features to enhance the model's perception of higher-level chemical environments; the last layer of GCN integrates features at a more abstract level and outputs a feature representation from a global perspective, which ultimately results in the three-dimensional structural features of the molecule.
[0041] in, is the processed adjacency matrix, is the degree matrix of A+I, is the normalization of the degree matrix, H (l) is the node feature representation of the lth layer. For the first layer, H (0) =X,W (l) is a learnable weight matrix and σ is a nonlinear activation function.
[0042] In one embodiment, six residual blocks are connected in series to construct an MLP model, such as Figure 2 shown.
[0043] Step 2: Concatenate and normalize the chemical molecule coding features and the molecular three-dimensional structure features, and then input them into the LSTM model to obtain a fused feature vector;
[0044] Specifically, the two sets of feature vectors of the feature encoding part are merged into an LSTM model including several layers (in one embodiment, the number of layers is 3) of LSTM. The ability of LSTM lies in its mechanism for processing sequence data, which can capture the long-distance dependencies between features. In this structure, LSTM not only fuses the information from two different encoders, but also maintains and strengthens the complex interactions between these features, thereby generating a high-dimensional fused feature vector containing rich contextual information.
[0045] Step 3: Based on the second MLP model, key feature vectors required for predicting molecular mass spectra are mined from the fused feature vectors;
[0046] Specifically, the high-dimensional fusion features output by feature fusion are used as the input of the second MLP model. The second MLP uses residual blocks containing multiple fully connected layers, nonlinear activation layers, and normalization layers to learn, analyze, and further abstract the fused features. The hierarchical structure design of MLP provides the model with sufficient capabilities to mine the features required for predicting molecular spectra from high-dimensional fusion features.
[0047] Step 4: Input the key feature vector into the third MLP model and output the predicted spectrogram data.
[0048] Specifically, spectrogram prediction is implemented through the third MLP model, which receives the output feature vector of the feature fusion stage as input. The predicted spectrogram data is output by the MLP deep learning model, which has a fixed dimension and is convenient for comparison and analysis with the real spectrogram data. The goal of this stage is to map the learned deep features to the predicted representation of the spectrogram for accuracy analysis and further research.
[0049] Figure 1 This is the framework diagram of the entire model. The molecular structure and its corresponding mass spectrometry spectrum are used as the training set to train the entire model.
[0050] In one embodiment, the effect of the present invention is experimentally verified using three public mass spectrometry databases, GNPs, MassBank and NIST05. Table 1, Table 2 and Table 3 are the correlation experimental results of different classification models on the three data sets of GNPs, MassBank and NIST05 respectively. The experimental results show that the correlation experimental results of the feature fusion model (MLP+GCN) on all three data sets are the best. Specifically, on the GNPs data set, the Jaccard similarity of the feature fusion model reaches 0.57, on the MassBank data set, it reaches 0.56, and on the NIST05 data set, it reaches 0.58. The correlation scores of other models are generally lower than those of the feature fusion model, and the performance is poor. The feature fusion model makes full use of the advantages of different models by combining the features extracted by the GCN and MLP networks, so that the prediction results are more accurate. Therefore, from the experimental results, the small molecule mass spectrometry method based on the feature fusion of MLP+GCN of the present invention performs best on these three data sets.
[0051] Table 1 Correlation experimental results of different classification models on GNPs dataset
[0052]
[0053] Table 2 Correlation experimental results of different classification models on the MassBank dataset
[0054]
[0055] Table 3 Correlation experimental results of different classification models on the NIST05 dataset
[0056]
[0057]
[0058] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.
Claims
1. A molecular mass spectrometry prediction method based on deep learning feature fusion, characterized in that: include: Based on the first MLP model and the GCN model, chemical molecular coding features and molecular three-dimensional structural features are extracted from the molecular structure respectively; Splicing chemical molecular coding features and molecular three-dimensional structural features; Input the concatenated feature vectors into the LSTM model to obtain the fused feature vector; Based on the second MLP model, mining the key feature vectors required for predicting the molecular mass spectrum from the fused feature vectors; The key feature vector is input into the third MLP model, and the predicted spectrogram data is output.
2. The method according to claim 1, characterized in that The molecular structure includes a small molecule structure.
3. The method according to claim 1, characterized in that: Before extracting the chemical molecular coding features and the molecular three-dimensional structural features from the molecular structure, feature characterization is performed, including: First MLP model feature characterization: converting the molecular structure into SMILES code, and then converting the SMILES code into a molecular fingerprint in binary format; GCN model feature characterization: converting the molecular structure into a molecular graph format, wherein the molecular structure includes atoms in the molecular structure and chemical bonds between atoms.
4. The method according to claim 1, characterized in that An MLP model is constructed by connecting multiple residual blocks in series, each residual block includes a fully connected layer, a nonlinear activation layer, a normalization layer, and another fully connected layer as a closed loop; the MLP model includes a first MLP model, a second MLP model and a third MLP model.
5. The method according to claim 4, characterized in that The multiple residual blocks are 6 residual blocks.
6. The method according to claim 1, characterized in that Based on the GCN model, the three-dimensional structural features of the molecule are extracted from the molecular structure, including: Input the adjacency matrix A and the feature matrix X; X contains the initial feature information of each node, that is, the type of atoms in the molecular structure; A represents the connection relationship between nodes in the graph, that is, the chemical bonds between atoms; Process the adjacency matrix by Add a self-join and normalize, where is the degree matrix of A+I, I is the identity matrix; use To aggregate the features of adjacent nodes to update the feature representation of each node, and finally obtain the three-dimensional structure features of the molecule; in, is the processed adjacency matrix, is the degree matrix of A+I, is the normalization of the degree matrix, H (l) is the node feature representation of the lth layer. For the first layer, H (0) =X,W (l) is a learnable weight matrix and σ is a nonlinear activation function.
7. The method according to claim 1, characterized in that The LSTM model includes several LSTM layers.
8. The method according to claim 7, characterized in that The several layers of LSTM are three layers of LSTM.
Citation Information
Patent Citations
Molecular feature extraction and performance prediction method based on image convolution
CN113409893A
Molecular attribute prediction method based on mass spectrum and graph structure
CN114724646A
Transform and graph neural network-combined drug target interaction prediction method
CN116417093A
Cited By
Molecular structure estimation method, electronic device and computer readable storage medium
CN120473022A
Molecular structure inference method, electronic device, and computer-readable storage medium
CN120473022B
