A method and system for molecular characterization of biomedical materials based on multi-view fusion

Abstract

The present invention discloses a method and system for molecular characterization of biomedical materials based on multi-view fusion, belonging to the technical fields of molecular characterization learning of biomedical materials, material gene engineering, and deep learning, and solving the problem that the characterization of molecules learned by the prior art is incomplete. The present invention defines strong and weak associations for different views of material molecules, models different views of material molecules, and learns the embedded characterization of material molecules in a single view; fuses the embedded characterizations of the strongly associated views of material molecules; maps the embedded characterizations of material molecules distributed in different semantic spaces to the same semantic space; calculates the total consistency loss between the embedded characterizations of material molecules in different views after mapping; fuses the embedded characterizations of material molecules in the weakly associated views, and maps the fused result to obtain the final embedded characterization of the molecule; calculates the supervised loss according to the true labels of the molecular physical and chemical properties, and optimizes the molecular characterization learning model in combination with the total consistency loss, and finally obtains the optimal multi-view molecular fusion embedded characterization. The present invention is used for molecular characterization learning of biomedical materials.

Description

Technical Field

[0001] A method and system for molecular characterization of biomedical materials based on multi-view fusion, which are used for molecular characterization learning of biomedical materials, and belong to the technical fields of molecular characterization learning of biomedical materials, material gene engineering, and deep learning. Background Technique

[0002] Molecular characterization learning is a popular research issue in materials genomics and has very important significance in work such as virtual screening of biomedical materials, prediction of material properties, and synthesis of biomedical materials. In-depth study of the characterization method of molecules can accelerate the research and development process of new biomedical materials and increase the probability of successfully developing new biomedical materials. However, due to the complexity of the structure of biomedical material molecules themselves, it is difficult to construct a suitable molecular characterization of biomedical materials. How to accurately predict the physical and chemical properties of biomedical material molecules based on the molecular characterization of biomedical materials has become the focus of current research. Currently, there is literature (such as Zhenqin Wu, Bharath Ramsundar, Evan N. Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S. Pappu, Karl Leswing, Vijay Pande, MoleculeNet: A Benchmark for Molecular Machine Learning, arXiv preprint, arXiv: 1703.00564, 2017) that designs a benchmark platform for measuring the performance of molecular characterization, enabling researchers to judge the performance strength of the current molecular characterization learning model through several material molecular datasets.

[0003] The existing methods for molecular characterization learning of biomedical materials mainly fall into two categories, namely text-based characterization learning and graph-based characterization learning.

[0004] Text-based characterization learning:

[0005] (1) Obtain the Simplified Molecular Input Line Entry Specification (SMILES) of biomedical material molecules with the help of databases such as ChEMBL. This is a chemical language used to describe the molecular structure and is often used to store the structural information of molecules.

[0006] (2) Based on the grammar rules of SMILES, perform a tokenization operation on the SMILES text representing biomedical material molecules, so that each group of characters after segmentation represents only a single atom, chemical bond, or structural information.

[0007] (3) The segmented characters are deduplicated and a dictionary is constructed. According to the dictionary, the original simplified molecular linear input specification (SMILES) string described in text vocabulary is converted into a tensor form represented by numerical values.

[0008] (4) With the help of recurrent neural networks (RNNs), such as the LSTM neural network proposed in the paper (Hochreiter S, Schmidhuber J. Long short-term memory [J]. Neural computation, 1997, 9 (8): 1735-1780.), we can learn the representation of biomedical material molecules after tensorization, and optimize the constructed recurrent neural network model based on the physical and chemical properties of some biomedical material molecules as labels.

[0009] Graph-based representation learning:

[0010] (1) The SMILES form of biomedical material molecules was converted into molecular graph form with the help of RDKit (a Python software package for cheminformatics).

[0011] (2) Statistically collect some physical and chemical properties of atoms in chemical molecules, as well as some physical and chemical information of chemical bonds, and process this information into tensors as the initial characteristics of atoms and chemical bonds.

[0012] (3) With the help of graph neural network models, such as the GCN model proposed in the paper (Kipf TN, Welling M. Semi-supervised classification with graph convolutional networks [J]. arXiv preprint arXiv: 1609.02907, 2016), the graph-level representation of biomedical material molecules can be obtained through this model. At the same time, the constructed network model can be optimized based on some physical and chemical properties of the material molecules as labels.

[0013] Therefore, the prior art has the following technical problems:

[0014] (1) Text-based representation learning ignores the spatial information of molecules. The embedded representation of molecules learned by this type of model only conforms to the grammatical rules specified by a specific chemical language (such as SMILES). The embedded representation of material molecules does not contain their structural features, making it impossible for the embedded representation of material molecules to fully describe the encoded material molecules, resulting in a significant reduction in the accuracy of subsequent models in predicting the properties of material molecules.

[0015] (2) Graph - based embedding representation learning usually only focuses on the planar topological structure of material molecules. This type of embedding representation method is more suitable for dealing with material molecules such as cyclic compounds, but has poor ability in the embedding representation learning of long - chain molecules. As a result, the obtained embedding representation of the material molecule cannot comprehensively describe the encoded material molecule, leading to a significant reduction in the accuracy of the subsequent model's prediction of the properties of the material molecule;

[0016] (3) Existing models almost only learn the embedding representation of material molecules from a single perspective, ignoring the interactivity of material molecule information under different views. The learned molecular representation may be incomplete, resulting in the obtained embedding representation of the material molecule being unable to comprehensively describe the encoded material molecule, and causing a significant reduction in the accuracy of the subsequent model's prediction of the properties of the material molecule;

[0017] (4) Most of the existing models currently require manual parameter adjustment. When using manual parameter adjustment, each time a hyperparameter is adjusted, the model needs to be re - run and trained from scratch, which consumes a large amount of time and GPU computing power. Summary of the Invention

[0018] Aiming at the problems in the above research, the purpose of the present invention is to provide a method and system for biomedical material molecule representation based on multi - view fusion, so as to solve the problem that the representation of molecules learned by the existing technology is incomplete, resulting in a significant reduction in the accuracy of subsequent prediction of the properties of material molecules.

[0019] To achieve the above - mentioned purpose, the present invention adopts the following technical solutions:

[0020] A method for biomedical material molecule representation based on multi - view fusion, comprising the following steps:

[0021] Step 1: Learn the embedding representation of each view of the material molecule:

[0022] Define the strong and weak associations of different views of each material molecule, model different views of each material molecule, and learn the embedding representation of the material molecule under each view;

[0023] Step 2: Fusion of strongly associated embedding representations:

[0024] Fuse the embedding representations of the strongly associated views of each material molecule;

[0025] Step 3: Mapping of multi - view embedding representations:

[0026] Based on the fused embedding representation obtained in Step 2, map the embedding representations of each material molecule distributed in different semantic spaces to the same semantic space;

[0027] Step 4: Calculate the total consistency loss:

[0028] Calculate the total consistency loss between the embedded representations of each material molecule in different views after mapping based on the result obtained in step 3;

[0029] Step 5: Fusion and mapping of multi-view embedded representations:

[0030] Fuse the embedded representations of each material molecule in the weakly associated views, and map the fused result to obtain the final embedded representation of the material molecule;

[0031] Step 6: Representation loss calculation:

[0032] Calculate the supervised loss according to the true labels of the molecular physical and chemical properties, and optimize the embedded representation learning model of each material molecule by combining the total consistency loss, and finally obtain the optimal multi-view fusion embedded representation of each material molecule.

[0033] Furthermore, the specific steps of step 1 are as follows:

[0034] Step 1.1: Each material molecule has the original features in the text view, the original features in the topological structure view, and the original features in the spatial view. The topological structure view and the spatial view are both stored and represented in the form of graphs. Then, define the topological structure view and the spatial structure view of the material molecule as strongly associated views, and define the text view of the material molecule and the topological structure view and the spatial view as weakly associated views;

[0035] Step 1.2: Based on the text view, the topological structure view, and the spatial view, model the original features of each material molecule in the text view, the original features in the topological structure view, and the original features in the spatial view respectively, and obtain the text view model, the topological structure view model, and the spatial view model. Among them, the text view model is a BiGRU model or an LSTM model, and the topological structure view model and the spatial view model are GINE models, graph attention models, or HetGNN models;

[0036] Step 1.3: Based on the text view model, the topological structure view model, and the spatial view model respectively, obtain the embedded representations h s 、h tp and h sp of each material molecule in the text view, the topological structure view, and the spatial view, and represent the real number field of dimension 1×d.

[0037] Furthermore, the specific steps of step 2 are as follows:

[0038] Step 2.1: Extract the chemical bonds of the embedded representation h sp of each material molecule in the spatial view to obtain the chemical bond embedded representation;

[0039] Step 2.2: Replace the edge feature information of each material molecule in the topological structure view with the chemical bond embedding representation in the spatial view;

[0040] Step 2.3: Based on the graph isomorphism convolutional neural network GINE, re - perform the embedding representation learning on the modified topological structure view obtained in Step 2.2 to obtain the fused embedding representation of each material molecule's strong - correlation view That is, obtain the embedding representation of the graph view

[0041] Furthermore, the specific steps of Step 3 are as follows:

[0042] Step 3.1: Construct a multi - layer perceptron A required for the multi - view embedding representation mapping. Among them, the multi - layer perceptron A is used to map the embedding representation from dimension d to dimension e after passing through a linear mapping and a non - linear transformation in sequence for different - dimensional embedding representations;

[0043] Step 3.2: Based on the multi - layer perceptron A, map the embedding representation h of the text view of each material molecule s to obtain z s ;

[0044] Step 3.3: Based on the multi - layer perceptron A, map the embedding representation h of the graph view of each material molecule g to obtain z g .

[0045] Furthermore, the specific steps of Step 4 are as follows:

[0046] Step 4.1: According to the embedding representation z of the text view of each material molecule s , and the embedding representation z of the graph view of each material molecule g , calculate the consistency loss between the graph view and the text view That is, the ratio of the positive - sample cosine similarity to the sum of all negative - sample cosine similarities. Among them, the positive - sample cosine similarity is The sum of all negative - sample cosine similarities is represents the embedding representation of the i - th material molecule in the text view, represents the embedding representation of the i - th material molecule in the graph view, represents the embedding representation of the j - th material molecule in the text view;

[0047] Step 4.2: According to the embedding representation z of the text view of each material molecule s , and the graph - view embedding representation z of each material molecule g , calculate the consistency loss between the text view and the graph view

[0048] Step 4.3. Calculate the total consistency loss under two views based on and to obtain the total consistency loss of the weakly associated multi-views That is, the total consistency loss of the weakly associated multi-views is obtained

[0049] Furthermore, the specific steps of Step 5 are as follows:

[0050] Step 5.1. Perform a row splicing operation on the matrix of the embedded representation z s of the text view of each material molecule and the embedded representation z g of the graph view to obtain the overall embedded representation z tmp of each material molecule under all weakly associated views;

[0051] Step 5.2. Construct a multi-layer perceptron B required for fusing the overall multi-view representation. The multi-layer perceptron B is used to linearly map and non-linearly transform the overall embedded representation of different dimensions in sequence, and map the overall embedded representation z tmp from 2e dimensions to d dimensions;

[0052] Step 5.3. Map the overall embedded representation z tmp of each material molecule based on the multi-representation layer perceptron B to obtain the final fused representation z f of each material molecule.

[0053] Furthermore, the specific steps of Step 6 are as follows:

[0054] Step 6.1. Construct a multi-layer perceptron C for predicting the properties of material molecules. Among them, the multi-layer perceptron C linearly maps and non-linearly transforms the fused representation of different dimensions in sequence, and maps the fused representation z f from d dimensions to 1 dimension; f Step 6.2. Predict the properties based on the fused representation

[0055] of the i-th material molecule by the multi-layer perceptron C to obtain the probability that the i-th material molecule is predicted to be classified as a positive sample

[0056] Step 6.3. Calculate the cross-entropy loss of the i-th material molecule in supervised classification learning according to the true label y i of the i-th material molecule. The true label y is obtained from the given dataset, and the true label y i refers to the true property of the i-th material molecule; i Step 6.4. According to the total consistency loss of the material molecules

[0057] ​ and cross-entropy loss obtain the total loss of the embedded representation learning model α is a given balancing parameter used to balance the ratio between the two errors. The embedded representation learning model includes a text view model, a topological structure view model, a spatial view model, a multi-layer perceptron A, a multi-layer perceptron B, and a multi-layer perceptron C;

[0058] Step 6.5: For the i-th material molecule, if the number of learning rounds of the embedded representation learning model is less than the given N times, according to the total loss perform backpropagation on the embedded representation learning model, update the parameters of the embedded representation learning model, and continue to predict the properties of the fused representation of the i-th material molecule otherwise, obtain the optimal multi-view fusion embedded representation of the i-th material molecule

[0059] A biomedical material molecule characterization system based on multi-view fusion, including:

[0060] Learning embedded representation module:

[0061] Define the strong and weak associations of different views of each material molecule, model different views of each material molecule, and learn the embedded representations of material molecules under each view;

[0062] Step 2: Strongly associated embedded representation fusion module:

[0063] Fuse the embedded representations of the strongly associated views of each material molecule;

[0064] Step 3: Multi-view embedded representation mapping module:

[0065] Based on the fused embedded representation obtained in Step 2, map the embedded representations of each material molecule distributed in different semantic spaces to the same semantic space;

[0066] Step 4: Calculate the total consistency loss module:

[0067] Calculate the total consistency loss between the embedded representations of each material molecule under different views after mapping based on the result obtained in Step 3;

[0068] Step 5: Multi-view embedded representation fusion mapping module:

[0069] Fuse the embedded representations of each material molecule under the weakly associated views, and map the fused result to obtain the final embedded representation of the material molecule;

[0070] Step 6: Characterization loss calculation module:

[0071] Calculate the supervised loss according to the true labels of the molecular physicochemical properties, and optimize the embedding representation learning model of each material molecule by combining the total consistency loss, and finally obtain the optimal multi-view fusion embedding representation of each material molecule.

[0072] Furthermore, the specific implementation steps of the learning embedding representation module are as follows:

[0073] Step 1.1: Each material molecule has the original features under the text view, the original features under the topological structure view, and the original features under the spatial view. Both the topological structure view and the spatial view are stored and represented in the form of graphs. Then, define the topological structure view and the spatial structure view of the material molecule as strongly associated views, and define the text view of the material molecule and the topological structure view and the spatial view as weakly associated views;

[0074] Step 1.2: Based on the text view, the topological structure view, and the spatial view, respectively model the original features under the text view, the original features under the topological structure view, and the original features under the spatial view of each material molecule, and obtain the text view model, the topological structure view model, and the spatial view model respectively. Among them, the text view model is a BiGRU model or an LSTM model, and the topological structure view model and the spatial view model are GINE models, graph attention models, or HetGNN models;

[0075] Step 1.3: Based on the text view model, the topological structure view model, and the spatial view model respectively, obtain the embedding representations h s 、h tp and h sp of each material molecule under the text view, the topological structure view, and the spatial view, and represent the real number field of dimension 1×d;

[0076] The specific implementation steps of the strongly associated embedding representation fusion module are as follows:

[0077] Step 2.1: Extract the chemical bonds of the embedding representation h sp of each material molecule under the spatial view to obtain the chemical bond embedding representation;

[0078] Step 2.2: Replace the feature information of the edges in the topological structure view of each material molecule with the chemical bond embedding representation of the spatial view;

[0079] Step 2.3: Based on the graph isomorphism convolutional neural network GINE, re-perform embedding representation learning on the modified topological structure view obtained in Step 2.2 to obtain the embedding representation after fusion of the strongly associated views of each material molecule,

[0080] Furthermore, the specific steps of the multi-view embedding representation mapping module are as follows:

[0081] Step 3.1: Construct a multi-layer perceptron A required for multi-view embedding representation mapping. Among them, the multi-layer perceptron A is used to sequentially perform a linear mapping and a non-linear transformation on the embedding representations of different dimensions, and map the embedding representations from d dimensions to e dimensions.

[0082] Step 3.2: Based on the multi-layer perceptron A, map the embedding representation h of the text view of each material molecule s to obtain z s ;

[0083] Step 3.3: Based on the multi-layer perceptron A, map the embedding representation h of the graph view of each material molecule g to obtain z g ;

[0084] The specific implementation steps of the total consistency loss calculation module are as follows:

[0085] Step 4.1: According to the embedding representation z of the text view of each material molecule s , and the embedding representation z of the graph view of each material molecule g , calculate the consistency loss between the graph view and the text view That is, the ratio of the positive sample cosine similarity to the sum of all negative sample cosine similarities. Among them, the positive sample cosine similarity is The sum of all negative sample cosine similarities is represents the embedding representation of the i-th material molecule in the text view, represents the embedding representation of the i-th material molecule in the graph view, represents the embedding representation of the j-th material molecule in the text view;

[0086] Step 4.2: According to the embedding representation z of the text view of each material molecule s , and the embedding representation z of the graph view of each material molecule g , calculate the consistency loss between the text view and the graph view

[0087] Step 4.3: Based on and calculate the total consistency loss under the two views That is, the total consistency loss of the weakly associated multi-view is obtained

[0088] The specific implementation steps of the multi-view embedding representation fusion mapping module are as follows:

[0089] Step 5.1: The embedding representation z of the text view of each material molecules With the embedded representation z of the graph view g Perform row concatenation operation on the matrix to obtain the overall embedded representation z of each material molecule under all weakly associated views tmp ;

[0090] Step 5.2, Construct a multi-layer perceptron B required for fusing the overall multi-view representation. The multi-layer perceptron B is used to linearly map and non-linearly transform the overall embedded representation of different dimensions in sequence, and map the overall embedded representation z tmp From 2e dimensions to d dimensions;

[0091] Step 5.3, Based on the multi-layer perceptron B, map the overall embedded representation z of each material molecule tmp To obtain the final fused representation z of each material molecule f ;

[0092] The specific steps of the representation loss calculation module are as follows:

[0093] Step 6.1, Construct a multi-layer perceptron C for predicting the properties of material molecules. Among them, the multi-layer perceptron C linearly maps and non-linearly transforms the fused representations of different dimensions in sequence, and maps the fused representation z f From d dimensions to 1 dimension; f In sequence, through a linear mapping and a non-linear transformation, map the fused representation z

[0094] Step 6.2, Based on the multi-layer perceptron C, predict the properties of the fused representation of the i-th material molecule To obtain the probability that the i-th material molecule is predicted to be classified as a positive sample

[0095] Step 6.3, According to the true label y of the i-th material molecule i , Calculate the cross-entropy loss of the i-th material molecule in supervised classification learning True label y i Is obtained from the given dataset, and the true label y i Refers to the true property of the i-th material molecule;

[0096] Step 6.4, According to the total consistency loss of the material molecules And the cross-entropy loss To obtain the total loss of the embedded representation learning model α is a given balance parameter used to balance the ratio between the two errors. The embedded representation learning model includes a text view model, a topological structure view model, a spatial view model, a multi-layer perceptron A, a multi-layer perceptron B, and a multi-layer perceptron C;

[0097] Step 6.5. For the i-th material molecule, if the number of learning rounds of the embedding representation learning model is less than the given N times, according to the total loss perform backpropagation on the embedding representation learning model, update the parameters of the embedding representation learning model, and continue to predict the properties of the fused representation of the i-th material molecule Otherwise, obtain the optimal multi-view fusion embedding representation of the i-th material molecule

[0098] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0099] First, the present invention fully considers the characteristics of biomedical molecules (i.e., material molecules) under multiple views. The embedding representation of the material molecules learned by the embedding representation learning model fully takes into account the characteristics of each view. Through the interaction between the features under different views, the finally learned material molecule representation is more comprehensive. That is, the embedding representation of the material molecules integrates the effective information of multiple views of the biomedical material molecules. The finally learned embedding representation of the material molecules not only has the grammatical constraints of the specific material molecule language in the text view, but also can contain the spatial three-dimensional features and planar topological features of the molecule (that is, the embedding representation of the material molecule in the text view contains its structural features, and the topological structure view and the spatial view have higher embedding representation learning ability for long-chain molecules), so that the prediction of the subsequent model is more accurate (chemical molecules naturally have two-dimensional planar structures and three-dimensional stereoscopic information, which is not possessed by general data. Therefore, the present invention fully considers the characteristics of the physical and chemical properties of chemical material molecules themselves, and uses the chemical bonds that coexist in the two-dimensional planar view and the three-dimensional space view as a bridge to realize the fusion of the two-dimensional features and three-dimensional features of the material molecules);

[0100] Second, the present invention adopts an adaptive fusion strategy based on consistency loss in the fusion stage of the embedding representations of material molecules in multiple views. The fusion process does not require manual intervention, reducing the inconvenience brought by manual parameter tuning to model optimization. That is, in the weak-correlation view fusion stage, the fusion strategy first uses the row splicing method to obtain the overall embedding representation under the two views, and then uses a multi-layer perceptron to learn the fused embedding representation. The fusion strategy introduces a consistency loss learning strategy between the two views (the graph view and the text view). By this strategy, the consistency loss under the two views is calculated, and the consistency loss is used as an auxiliary fusion strategy to assist in achieving good fusion under the two views. This fusion strategy does not involve the learning of the latent space and the feature distribution of the subspace, does not require learning the self-expressive subspace, and does not need to calculate the similarity matrix. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1is a schematic diagram of the framework of the present invention, where GRU represents the GRU model, GNN represents the GNN model, MLP represents the multi-layer perceptron, and g seq (.) represents the text view encoder of the material molecule using the GRU model, and g tp (.) represents the topological view (topological structure view) encoder of the material molecule using the GNN model, and g sp (.) represents the spatial view encoder of the material molecule using the GNN model, and f prooj (.) represents the mapping function that maps the embedded representation h of the text view using the multi-layer perceptron A s and the embedded representation h after fusing with the strong correlation view of the material molecule g ; f fuse (.) represents the mapping function that maps the overall embedded representation z of the material molecule using the multi-layer perceptron B tmp ; f pred (·) represents the mapping function that maps the fused representation zf of the material molecule by the multi-layer perceptron C to the specific label value;

[0102] Figure 2 is a schematic diagram showing the influence of each variant of the embedded representation learning model on the experimental results in Experiment 2 of the present invention; Specific Embodiments

[0103] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0104] The present invention simultaneously models the molecular information of material molecules under different views; by constructing the consistency loss of the embedded representations under different views, maps the molecular embedded representations in different semantic spaces to the same semantic space; and adopts an adaptive embedded representation fusion method to fuse the embedded representations of material molecules under different views, without the need to manually adjust the hyperparameters of the model. The specific implementation steps are as follows:

[0105] Step 1. Learn the embedded representations of each view of the material molecule:

[0106] Define the strong and weak correlations of different views of each material molecule, model different views of each material molecule, and learn the embedded representations of the material molecule under each view; the specific steps are as follows:

[0107] Step 1.1. Each material molecule has the original features under the text view, the original features under the topological structure view, and the original features under the spatial view. The topological structure view and the spatial view are both stored and represented in the form of a graph. Then, define the topological structure view and the spatial structure view of the material molecule as strong correlation views, and define the text view of the material molecule and the topological structure view and the spatial view as weak correlation views;

[0108] Step 1.2: Based on the text view, topological structure view, and spatial view, model the original features of each material molecule under the text view, topological structure view, and spatial view respectively, to obtain a text view model, a topological structure view model, and a spatial view model. Among them, the text view model is a BiGRU model or an LSTM model, and the topological structure view model and the spatial view model are GINE models, graph attention models, or HetGNN models;

[0109] Step 1.3: Based on the text view model, topological structure view model, and spatial view model respectively, obtain the embedding representations h s 、h tp and h sp of each material molecule under the text view, topological structure view, and spatial view. and represent the real number field of dimension 1×d. The embedding representations of the i-th material molecule under the text view, topological structure view, and spatial view are known

[0110] Step 1.3: Based on the text view model, topological structure view model, and spatial view model respectively, obtain the embedding representations h s 、h tp and h sp of the material molecule under the text view, topological structure view, and spatial view. and represent the real number field of dimension 1×d.

[0111] Step 2: Strongly correlated embedding representation fusion:

[0112] Fuse the embedding representations of the strongly correlated views of the material molecule; the specific steps are as follows:

[0113] Step 2.1: Extract the chemical bonds of the embedding representation h sp of each material molecule under the spatial view to obtain a chemical bond embedding representation;

[0114] Step 2.2: Replace the feature information of the edges in the topological structure view of each material molecule with the chemical bond embedding representation of the spatial view;

[0115] Step 2.3: Based on the graph isomorphism convolutional neural network GINE, re-learn the embedding representation of the modified topological structure view obtained in Step 2.2 to obtain the embedding representation after fusion of the strongly correlated views of each material molecule That is, obtain the embedding representation of the graph view The embedding representation of the graph view of the i-th material molecule is

[0116] Step 3. Mapping of multi-view embedded representations:

[0117] Based on the fused embedded representations obtained in Step 2, map the embedded representations of material molecules distributed in different semantic spaces to the same semantic space. The specific steps are as follows:

[0118] Step 3.1. Construct a multi-layer perceptron A required for mapping multi-view embedded representations. Among them, the multi-layer perceptron A is used to map the embedded representations from dimension d to dimension e after passing through a linear mapping and a non-linear transformation in sequence for embedded representations of different dimensions.

[0119] Step 3.2. Based on the multi-layer perceptron A, map the embedded representation h of the text view of each material molecule s to obtain z s . The embedded representation of the text view of the i-th material molecule after mapping is

[0120] Step 3.3. Based on the multi-layer perceptron A, map the embedded representation h of the graph view of each material molecule g to obtain z g . The embedded representation of the graph view of the i-th material molecule after mapping is

[0121] Step 4. Calculate the total consistency loss:

[0122] Based on the results obtained in Step 3, calculate the total consistency loss between the embedded representations of material molecules in different views after mapping. The specific steps are as follows:

[0123] Step 4.1. According to the embedded representation z of the text view of each material molecule s , and the embedded representation z of the graph view of the material molecule g , calculate the consistency loss between the graph view and the text view The consistency loss between the graph view and the text view of the i-th material molecule is That is, the ratio of the cosine similarity of the positive sample to the sum of the cosine similarities of all negative samples. Among them, the cosine similarity of the positive sample is The sum of the cosine similarities of all negative samples is represents the embedded representation of the i-th material molecule in the text view, represents the embedded representation of the i-th material molecule in the graph view, represents the embedded representation of the j-th material molecule in the text view;

[0124] Step 4.2. According to the embedded representation z of the text view of each material molecule s, and the graphical view embedding representation z of each material molecule g , calculate the consistency loss between the text view and the graphical view Consistency loss between the text view and the graphical view The consistency loss between the text view and the graphical view can be calculated by The calculation process is obtained, but the corresponding parameters need to be replaced according to the requirements of the text view for the graphical view. The consistency loss between the text view and the graphical view of the i-th material molecule is

[0125] Step 4.3, based on and Calculate the total consistency loss under the two views That is, the total consistency loss of the weakly associated multi-view is obtained The total consistency loss of the i-th material molecule is

[0126] Step 5, Fusion and mapping of multi-view embedding representations:

[0127] Fuse the embedding representations of the material molecules under the weakly associated views, and map them after fusion to obtain the final embedding representation of the material molecules; the specific steps are:

[0128] Step 5.1, the embedding representation z of the text view of each material molecule s and the embedding representation z of the graphical view g Perform a row concatenation operation of the matrices to obtain the overall embedding representation z of each material molecule under all weakly associated views tmp , and the overall embedding representation of the i-th material molecule is

[0129] Step 5.2, construct a multi-layer perceptron B required for fusing the overall multi-view representation. The multi-layer perceptron B is used to linearly map and non-linearly transform the overall embedding representations of different dimensions in sequence, and map the overall embedding representation from 2e dimensions to d dimensions;

[0130] Step 5.3, based on the multi-layer perceptron B, map the overall embedding representation z of each material molecule tmp to obtain the final fused representation z of each material molecule f , where represents the fused representation of the i-th material molecule.

[0131] Step 6, Representation loss calculation:

[0132] Calculate the supervised loss according to the true labels of the molecular physicochemical properties, and optimize the embedding representation learning model of the material molecules in combination with the total consistency loss, and finally obtain the optimal multi-view fusion embedding representation of the material molecules. The specific steps are:

[0133] Step 6.1: Construct a multilayer perceptron C for predicting the molecular properties of materials, where the multilayer perceptron C represents the fusion of different dimensions z f After a linear mapping and a nonlinear transformation, the fusion representation z f Map from d dimensions to 1 dimension;

[0134] Step 6.2: Fusion representation of the i-th material molecule based on the multilayer perceptron C Perform property prediction and obtain the probability that the i-th material molecule is predicted to be classified as a positive sample

[0135] Step 6.3: According to the true label y of the i-th material molecule i , calculate the cross entropy loss of the i-th material molecule in supervised classification learning The true label y i Obtained from the given dataset, the true label y i refers to the true property of the i-th material molecule:

[0136] Step 6.4: Total consistency loss based on material molecules and cross entropy loss Get the total loss of the embedding representation learning model α is a given balance parameter used to balance the ratio between the two errors. The embedding representation learning model includes a text view model, a topological structure view model, a spatial view model, a multilayer perceptron A, a multilayer perceptron B, and a multilayer perceptron C.

[0137] Step 6.5: Based on the total loss Perform back propagation on the embedding representation learning model, update the embedding representation learning model parameters, and go to step 6.2 to perform fusion representation on the i-th material molecule. Perform property prediction, otherwise, obtain the optimal multi-view fusion embedding representation of the i-th material molecule

[0138] Example

[0139] Given a biomedical material molecule, the material molecule has original features under the text view, original features under the topological structure view, and original features under the spatial view. Since the topological structure view and the spatial structure view are both stored in the form of graphs, the topological structure view and the spatial structure view (spatial view) of the material molecule are defined as strongly associated views, while the text view of the material molecule is significantly different from the graph-based storage form, so the text view of the material molecule and the other two views are defined as weakly associated views. We use different neural network models to model the original features under different views.

[0140] Among them, the original features of the text view are characterized by a bidirectional gated recurrent unit (BiGRU) in a recurrent neural network (RNN) model (the reference is Cho K, Van B, Bahdanau D, et al. On the properties of neural machine translation: Encoder-decoder approaches[J]. arXiv preprint arXiv:1409.1259, 2014.), and the embedded representation of the original features of the material molecules obtained by learning this model in the text view is The original features in the topological structure view and the original features in the spatial structure view are characterized by a graph isomorphism convolutional neural network GINE with edge attribute information in a graph convolutional network (GCN) model (the reference is Hu, Weihua, et al. "Strategies for Pre-training Graph Neural Networks." International Conferenceon Learning Representations. 2019.), and the embedded representations obtained by learning these models are respectively known Among them, the symbol represents the real number field of dimension 1×d, and this symbol represents the same meaning in the following text.

[0141] Given a trainable graph neural network fusion function, a graph isomorphism convolutional neural network GINE with edge attribute information is used as the neural network function for fusing the graph features of material molecules. This function can fuse the embedded representation h tp of the topological structure view of the material molecules and the embedded representation h sp of the spatial structure view of the material molecules, and the embedded representation of the fused material molecules represents the embedded representation of the material molecules based on the graph view.

[0142] The multi - layer perceptron A for the embedded representation of material molecules under multiple views can map the molecular embedded representations distributed in different semantic spaces to the same semantic space, enabling subsequent fusion operations on the material molecular embedded representations under different views. That is, given a multi - layer perceptron A, this multi - layer perceptron A can map the embedded representation h of the text view of the material molecule s , and the embedded representation h of the fused graph view g to the same semantic space, and the mapped embedded representations of the material molecule are respectively written as

[0143] The main role of the consistency loss calculation is to obtain the similarity degree of material molecules under different views, enabling interaction between the embedded representations of material molecules under different views. Given a batch, which contains N material molecules. Taking the i - th material molecule in this batch of material molecules as an example, after mapping, the embedded representation of the material molecule in the text view is The embedded representation of the material molecule in the graph view is The consistency loss between the graph view and the text view and the consistency loss between the text view and the graph view The above two losses are both calculated by the cosine similarity metric function. According to and the total consistency loss of the i - th material molecule under multiple views is obtained represents the sum of the consistency loss between the graph view and the text view and the consistency loss between the text view and the graph view.

[0144] There are significant differences in the descriptions of molecules between weakly associated views. Fusing weakly associated views can enhance the performance of molecular representations. Given a multi - layer perceptron B, the multi - layer perceptron B fuses the embedded representation z of the text view of the material molecule s with the embedded representation z of the graph view g and then maps the resulting overall embedded representation z of the material molecule tmp to obtain the fused embedded representation of the material molecule

[0145] Calculate the supervised loss according to the true labels of the molecular physicochemical properties, and combine the consistency loss to jointly optimize the embedded representation learning model of the material molecule. Given a multi - layer perceptron C, the multi - layer perceptron C can use the fused embedded representation z f after fusion to predict the molecular properties on a specific downstream dataset. The probability that the material molecule i is predicted to be a positive sample It can be calculated by this multi-layer perceptron C. The true property label (true label) yi of the material molecule can be obtained from the original dataset (referring to the dataset used to train the entire model. Before training, we first select a dataset from the candidate datasets (such as BACE, BBBP, etc.) that contains the SMILES information of the molecules and the true label values of the properties to be predicted). Cross-entropy loss is used as the loss for supervised learning in this task. Combined with the consistency loss of the material molecule under the weakly associated multi-view The total loss of the embedding representation learning model is obtained as: where α is a balancing parameter used to balance the ratio between the two errors. The entire model (embedding representation learning model) learns the optimal multi-view fusion embedding representation by minimizing the total loss Thus, the optimal multi-view fusion embedding representation is learned.

[0146] Experiment 1: Proof and analysis of the effectiveness of the present invention

[0147] To verify that the present invention can effectively predict the physicochemical properties of molecules, the present invention conducted comparative experiments and analysis with 9 benchmark algorithms on 7 publicly available real material molecule datasets. The present invention conducted experiments on the following 7 datasets, and the statistical information of these 7 datasets is shown in Table 1. Among them, the number of predicted properties represents the actual number of physicochemical properties of molecules that need to be predicted in the current dataset.

[0148] The present invention selected a total of 9 commonly used benchmark algorithms for comparison to verify the effectiveness of the method proposed by the present invention. These methods are respectively: the molecular embedding representation learning algorithms GCN, GIN, GAT, GraphConv, D-MPNN, and AttentiveFP based on traditional graph models, GROVER, MolCLR, and PretrainedGNN based on pre-training. The above 9 algorithms are widely used in the representation learning of molecules in biomedical and materials science.

[0149] Table 1 shows the dataset statistical information:

[0150]

[0151]

[0152] (1) Experiment setting: In the present invention, all steps use Xavier initialization, and the optimizer uses Adam. Among them, the embedding dimension d = 256, e = 512, the number of training times is 1000, the batch size is 128, and the learning rate is 0.1.

[0153] (2) Experiment result evaluation and analysis:

[0154] Since all the benchmark datasets used in this invention are for classification tasks, the evaluation criterion for the experimental results uses the area under the curve (AUC) of the Receiver Operating Characteristic (ROC) curve to evaluate the model. For datasets with multiple labels, we use the average ROC-AUC value as its final performance result.

[0155] This invention patent was compared with benchmark algorithms on the above 7 datasets. Table 2 shows the average values and standard deviations of the performance evaluation index ROC-AUC (%) of this invention and the benchmark methods in the task of predicting molecular physicochemical properties, and the best experimental results are highlighted in bold.

[0156] The results show that in the four datasets of BBBP, Tox21, ClinTox, and HIV, this invention performs the best among all the benchmark algorithms. Compared with the current best experimental results, the AUC of this invention has increased by 0.9%, 1.6%, 1.0%, and 1.4% respectively. This indicates that MORN can better capture the commonalities and differences between the graphical and text views and can utilize the multi-view information of molecules.

[0157] Compared with supervised methods, this invention performs the best in almost all benchmark methods, except that SIDER is 0.4% lower than GIN. The reason for this situation may be that the labels of SIDER are more related to the backbone structure of molecules, while GIN, as a convolutional neural network that can effectively capture graph structures, can learn the structural features of material molecules more comprehensively from the topological space. And this invention also takes into account some semantic information in the SMILES format of material molecules, which has a slightly weaker relationship with the labels compared to molecular structure information. Therefore, the effect on this dataset is relatively weaker than that of the neural network model completely based on graph structures. From the experimental results of the above 7 datasets, since this invention considers both the text view and the topological space view simultaneously, the learned embedding representation contains more information than single-view models, which means it may be closer to the true representation space of material molecules.

[0158] Table 2 Results of predicting molecular physicochemical properties:

[0159]

[0160]

[0161] Meanwhile, the experimental performances of the present invention and the pre-trained model were compared. It can be seen that the present invention is inferior to MolCLR on BACE and ToxCast, being 2.2% and 3.5% lower respectively. The reason is that MolCLR obtained important molecular features on millions of data through contrastive learning in the pre-training stage, which greatly improved its classification ability on the BACE and ToxOast datasets. However, benefiting from the two-stage learning method, the present invention can also capture the prominent features of molecules with label limitations. The present invention outperforms the pre-trained model on five of the seven datasets.

[0162] In addition, it was also observed that, compared with the pre-trained learning model, the present invention has good performance and a lower standard deviation. Therefore, the present invention has better repeatability and superior stability.

[0163] Experiment 2: Analysis of the effectiveness of each step in the present invention

[0164] To verify the influence of each step in the embedding representation learning of the material molecules proposed in the present invention patent under multi-views on the experimental results, we designed three model variants, namely w / o graph, w / o sequence, w / o contrast, which respectively represent only including the text view, only including the topological structure view or the spatial view, and not calculating the consistency loss. The performances of the present invention and its variants on the above datasets are as Figure 2 shown. Observing this figure, it can be found that the AUC of the present invention is higher than that of its variants on all seven datasets, proving that the multi-view fusion strategy proposed by the present invention is effective.

[0165] Observing the experimental results, on the four datasets of Tox21, Clinton, SIDER and HIV, the w / o sequence variant is more effective than w / o graph. On the datasets BBBP and BACE, the experimental results of w / o sequence are lower than those of w / o graph. The experimental results of the two variants on the Toxcast dataset are very close. The reason for the obvious difference in performance is related to the physicochemical properties of the dataset itself.

[0166] Analyze the results in terms of molecular length. On the datasets BACE and BBBP, the average lengths of the molecules in SMILES sequence format are 61.6 and 41.3 respectively. However, the values for the dataset Tox21 MIV are 31.5 and 32.4, which are lower than those of the first two datasets. For longer molecules, the sequence-based model performs better than the graph-based model because it can capture more context information for each chemical symbol, thus obtaining better feature descriptions. When the molecular length is small, the GNN-based model performs slightly better than the sequence-based model, indicating that through adjacency information aggregation, the representation obtained by GNN is closer to the true representation space of the molecule and can achieve better classification results.

[0167] In addition, consistency contrastive learning across multiple perspectives also plays an important role. The contrastive learning consistency loss promotes the experimental performance of the present invention in all seven datasets, and the improvement is most obvious in the three datasets BACE, BBBP, and Tox21. The contrastive learning module makes good use of the consistency and complementarity of features under different views, enabling the finally obtained fused embedding representation to cover more similar features under the two views (text view and graph view). Therefore, the consistency contrastive loss proposed in this experiment has a non-negligible effect on the present invention. In summary, the three variants of the present invention play an important role in the characterization of material molecules, and the fusion strategy has significantly improved the experimental results of the model we finally proposed.

[0168] The above are only representative embodiments among the numerous specific application scopes of the present invention, and do not constitute any limitation to the protection scope of the present invention. Any technical solutions formed by transformation or equivalent replacement fall within the scope of the rights protection of the present invention.

Claims

1. A method for molecular characterization of biomedical materials based on multi-view fusion, characterized in that, It includes the following steps: Step 1. Learn the embedded representations of each view of the material molecules: Define the strong and weak correlations of different views of each material molecule, model different views of each material molecule, and learn the embedded representations of the material molecules under each view; Each material molecule has the original features under the text view, the original features under the topological structure view, and the original features under the spatial view. The topological structure view and the spatial view are both stored and represented in the form of graphs. Then, define the topological structure view and the spatial structure view of the material molecule as strongly correlated views, and define the text view of the material molecule as weakly correlated with the topological structure view and the spatial view; Step 2. Fusion of strongly correlated embedded representations: Fuse the embedded representations of the strongly correlated views of each material molecule; Step 3. Mapping of multi-view embedded representations: Based on the fused embedded representations obtained in Step 2, map the embedded representations of each material molecule distributed in different semantic spaces to the same semantic space; Step 4. Calculate the total consistency loss: Calculate the total consistency loss between the embedded representations of each material molecule under different views after mapping based on the results obtained in Step 3; Step 5. Fusion and mapping of multi-view embedded representations: Fuse the embedded representations of each material molecule under the weakly correlated views, and map the fused result to obtain the final embedded representation of the material molecule; Step 6. Calculate the representation loss: Calculate the supervised loss according to the true labels of the molecular physicochemical properties, and optimize the embedded representation learning model of each material molecule in combination with the total consistency loss, and finally obtain the optimal multi-view fusion embedded representation of each material molecule.

2. The method for molecular characterization of biomedical materials based on multi-view fusion according to claim 1, wherein Step 1 further includes: Step 1.

1. Based on the text view, the topological structure view, and the spatial view, model the original features of each material molecule under the text view, the original features under the topological structure view, and the original features under the spatial view respectively, and obtain the text view model, the topological structure view model, and the spatial view model respectively. Among them, the text view model is a BiGRU model or an LSTM model, and the topological structure view model and the spatial view model are GINE models, graph attention models, or HetGNN models; Step 1.2: Obtain the embedding representations h s , h tp , and h sp of each material molecule in the text view, topological structure view, and spatial view respectively, and represent the real number field of dimension 1×d.

3. A method for molecular characterization of biomedical materials based on multi-view fusion according to claim 2, wherein The specific steps of Step 2 are: Step 2.1: Extract the embedded representation h of each material molecule in the spatial view sp of the chemical bonds to obtain the chemical bond embedded representation; Step 2.

2. Replace the feature information of the edges in the topological structure view of each material molecule with the chemical bond embedded representation of the spatial view; Step 2.3: Based on the graph isomorphism convolutional neural network GINE, re - perform embedding representation learning on the modified topological structure view obtained in Step 2.2 to obtain the fused embedding representation of the strong - association views of each material molecule That is, obtain the embedding representation of the graph view 4. A method for molecular characterization of biomedical materials based on multi-view fusion according to claim 3, characterized in that, The specific steps of Step 3 are: Step 3.

1. Construct a multi-layer perceptron A required for mapping multi-view embedded representations. The multi-layer perceptron A is used to linearly map and non-linearly transform the embedded representations of different dimensions in sequence, and map the embedded representations from d dimensions to e dimensions; Step 3.

2. Map the embedded representation h of the text view of each material molecule based on the multi-layer perceptron A s to obtain z s ; Step 3.

3. Embed the graph view representation h of each material molecule in Multi-Layer Perceptron A g Perform mapping to obtain z g .

5. A method for molecular characterization of biomedical materials based on multi-view fusion according to claim 4, characterized in that, The specific steps of Step 4 are: Step 4.

1. Calculate the consistency loss between the graph view and the text view based on the embedding representation z of the text view of each material molecule s and the embedding representation z of the graph view of each material molecule g That is, the ratio of the positive sample cosine similarity to the sum of all negative sample cosine similarities, where the positive sample cosine similarity is The sum of all negative sample cosine similarities is represents the embedding representation of the i-th material molecule in the text view, represents the embedding representation of the i-th material molecule in the graph view, represents the embedding representation of the j-th material molecule in the text view;​ Step 4.

2. Calculate the consistency loss between the text view and the graph view based on the embedded representation z of the text view of each material molecule s and the embedded representation z of the graph view of each material molecule g ​ Step 4.3, based on and calculate the total consistency loss under the two views i.e., obtain the total consistency loss of the weakly correlated multi-views 6. A method for molecular characterization of biomedical materials based on multi-view fusion according to claim 5, characterized in that The specific steps of Step 5 are: Step 5.

1. Concatenate the embedding representations z of the text views of each material molecule s with the embedding representations z of the graph views g by row-wise matrix concatenation operation to obtain the overall embedding representation z of each material molecule under all weakly associated views tmp ; Step 5.2, construct the multi-layer perceptron B required for multi-view global representation fusion. The multi-layer perceptron B is used to linearly map and non-linearly transform the global embeddings of different dimensions in sequence, mapping the global embedding z tmp from 2e dimensions to d dimensions; Step 5.3: Based on the multi-representation layer perceptron B, map the overall embedded representation z of each material molecule tmp to obtain the final fused characterization z of each material molecule f .

7. A method for molecular characterization of biomedical materials based on multi-view fusion according to claim 6, characterized in that, The specific steps of Step 6 are: Step 6.

1. Construct a multi-layer perceptron C for predicting the properties of material molecules. Among them, the multi-layer perceptron C sequentially performs a linear mapping and a non-linear transformation on the fused representation z of different dimensions, and maps the fused representation z from d dimensions to 1 dimension. f through a linear mapping and a non-linear transformation in sequence, and map the fused representation z f from d dimensions to 1 dimension; Step 6.2: Based on the multi-layer perceptron C, perform property prediction on the fused representation of the i-th material molecule to obtain the probability that the i-th material molecule is predicted to be classified as a positive sample Step 6.

3. Calculate the cross-entropy loss of the \(i\)-th material molecule in supervised classification learning according to the true label \(y\) of the \(i\)-th material molecule i ; the true label \(y\) is obtained from the given dataset, and the true label \(y\) refers to the true property of the \(i\)-th material molecule The true label \(y\) i is obtained from the given dataset, and the true label \(y\) i refers to the true property of the \(i\)-th material molecule; Step 6.

4. Obtain the total loss of the embedding representation learning model according to the total consistency loss of the material molecules and the cross-entropy loss to obtain the total loss of the embedding representation learning model α is a given balancing parameter used to balance the ratio between the two errors. The embedding representation learning model includes a text view model, a topological structure view model, a spatial view model, a multi-layer perceptron A, a multi-layer perceptron B, and a multi-layer perceptron C; Step 6.

5. For the i-th material molecule, if the number of learning rounds of the embedding representation learning model is less than the given N times, according to the total loss perform backpropagation on the embedding representation learning model, update the parameters of the embedding representation learning model, and continue to predict the properties of the fusion representation of the i-th material molecule otherwise, obtain the optimal multi-view fusion embedding representation of the i-th material molecule 8. A biomedical material molecular characterization system based on multi-view fusion, characterized in that, It includes: Embedded representation learning module: Define the strong and weak correlations of different views of each material molecule, model different views of each material molecule, and learn the embedded representations of the material molecules under each view; Each material molecule has original features in the text view, original features in the topological structure view, and original features in the spatial view. The topological structure view and the spatial view are both stored and represented in the form of graphs. Then, the topological structure view and the spatial structure view of the material molecule are defined as strongly correlated views, and the text view of the material molecule is defined as a weakly correlated view with the topological structure view and the spatial view; Strongly correlated embedding representation fusion module: Fuse the embedding representations of the strongly correlated views of each material molecule; Multi-view embedding representation mapping module: Based on the obtained fused embedding representation, map the embedding representations of each material molecule distributed in different semantic spaces to the same semantic space; Total consistency loss calculation module: Calculate the total consistency loss between the embedding representations of each material molecule under different views after mapping based on the obtained results; Multi-view embedding representation fusion and mapping module: Fuse the embedding representations of each material molecule under the weakly correlated views, and map the fused result to obtain the final embedding representation of the material molecule; Representation loss calculation module: Calculate the supervised loss according to the true labels of the molecular physicochemical properties, and optimize the embedding representation learning model of each material molecule by combining the total consistency loss. Finally, obtain the optimal multi-view fusion embedding representation of each material molecule.

9. The biomedical material molecular characterization system based on multi-view fusion according to claim 8, characterized in that The specific implementation steps of the learning embedding representation module further include: Step 1.1: Model the original features of each material molecule in the text view, the original features in the topological structure view, and the original features in the spatial view based on the text view, the topological structure view, and the spatial view respectively, and obtain a text view model, a topological structure view model, and a spatial view model. Among them, the text view model is a BiGRU model or an LSTM model, and the topological structure view model and the spatial view model are GINE models, graph attention models, or HetGNN models; Step 1.2: Obtain the embedding representation h of each material molecule in the text view, topological structure view and spatial view based on the text view model, topological structure view model and spatial view model respectively. s 、h tp and h sp , and represents the real number field of dimension 1×d; The specific implementation steps of the strongly correlated embedding representation fusion module are: Step 2.1: Extract the embedded representation h of each material molecule in the spatial view sp of the chemical bond to obtain the chemical bond embedded representation; Step 2.2: Replace the feature information of the edges in the topological structure view of each material molecule with the chemical bond embedding representation of the spatial view; Step 2.3: Based on the graph isomorphism convolutional neural network GINE, re - perform embedded representation learning on the modified topological structure view obtained in Step 2.2 to obtain the embedded representation after the fusion of the strongly correlated views of each material molecule That is, obtain the embedded representation of the graph view 10. A molecular characterization system for biomedical materials based on multi-view fusion according to claim 9, characterized in that, The specific steps of the multi-view embedding representation mapping module are: Step 3.1: Construct a multi-layer perceptron A required for multi-view embedding representation mapping. Among them, the multi-layer perceptron A is used to map the embedding representation from d dimensions to e dimensions after passing through a linear mapping and a non-linear transformation in sequence for embedding representations of different dimensions; Step 3.2: Map the embedded representation h of the text view of each material molecule based on the multi-layer perceptron A s to obtain z s ; Step 3.

3. Embedding representation h of the graph views of each material molecule based on the multi-layer perceptron A g Perform mapping to obtain z g ; The specific implementation steps of the total consistency loss calculation module are: Step 4.

1. Calculate the consistency loss between the graph view and the text view based on the embedding representation z of the text view of each material molecule s and the embedding representation z of the graph view of each material molecule g That is, the ratio of the positive sample cosine similarity to the sum of all negative sample cosine similarities, where the positive sample cosine similarity is The sum of all negative sample cosine similarities is represents the embedding representation of the i-th material molecule in the text view represents the embedding representation of the i-th material molecule in the graph view represents the embedding representation of the j-th material molecule in the text view Step 4.

2. Calculate the consistency loss between the text view and the graph view based on the embedded representation z of the text view of each material molecule s and the embedded representation z of the graph view of each material molecule g ​ Step 4.

3. Based on and calculate the total consistency loss under the two views That is, the total consistency loss of the weakly associated multi-view is obtained The specific implementation steps of the multi-view embedding representation fusion and mapping module are: Step 5.

1. Concatenate the embedding representations z of the text views of each material molecule s with the embedding representations z of the graph views g by row-wise matrix concatenation operation to obtain the overall embedding representation z of each material molecule under all weakly associated views tmp ; Step 5.

2. Construct a multi-layer perceptron B required for multi-view global representation fusion. The multi-layer perceptron B is used to linearly map and non-linearly transform the global embeddings of different dimensions in sequence, and map the global embedding z tmp from 2e dimensions to d dimensions; Step 5.

3. Based on the multi-representation layer perceptron B, map the overall embedded representation z of each material molecule tmp to obtain the final fused representation z of each material molecule f ; The specific steps of the representation loss calculation module are: Step 6.1: Construct a multi-layer perceptron C for predicting the properties of material molecules. Among them, the multi-layer perceptron C sequentially performs a linear mapping and a non-linear transformation on the fused representation z of different dimensions, and maps the fused representation z from d dimensions to 1 dimension. f through a linear mapping and a non-linear transformation in sequence, and map the fused representation z f from d dimensions to 1 dimension; Step 6.2: Based on the multi-layer perceptron C, perform property prediction on the fusion representation of the i-th material molecule to obtain the probability that the i-th material molecule is predicted and classified as a positive sample Step 6.

3. Calculate the cross-entropy loss of the \(i\)-th material molecule in supervised classification learning according to the true label \(y\) of the \(i\)-th material molecule i of the \(i\)-th material molecule The true label \(y\) i is obtained from the given dataset, and the true label \(y\) i refers to the true property of the \(i\)-th material molecule; Step 6.

4. Obtain the total loss of the embedding representation learning model according to the total consistency loss of the material molecules and the cross-entropy loss to obtain the total loss of the embedding representation learning model ɑ is a given balance parameter used to balance the ratio between the two errors. The embedding representation learning model includes a text view model, a topological structure view model, a spatial view model, a multi-layer perceptron A, a multi-layer perceptron B, and a multi-layer perceptron C; Step 6.

5. For the i-th material molecule, if the number of learning rounds of the embedding representation learning model is less than the given N times, according to the total loss perform backpropagation on the embedding representation learning model, update the parameters of the embedding representation learning model, and continue to predict the properties of the fusion representation of the i-th material molecule otherwise, obtain the optimal multi-view fusion embedding representation of the i-th material molecule

Images