Graph neural network and chemical fingerprint-based carbohydrate biomacromolecule property prediction method and system
By combining undirected graphs and chemical fingerprints, the problem of capturing glycan structural features in existing technologies has been solved, achieving higher-precision prediction of glycan properties and supporting related scientific research and applications.
Patent Information
- Application Number
- CN202511660220.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-13
- Publication Date
- 2026-02-13
AI Technical Summary
Existing technologies struggle to effectively capture the complex branching and tree-like structural features of sugar chains, limiting the performance of sugar chain property prediction models and failing to meet high-precision requirements.
By constructing an undirected graph, using monosaccharides and glycosidic bonds as real nodes, introducing virtual nodes, and combining chemical fingerprint information, a multi-layer graph convolutional layer is used to learn the topological structure and molecular fingerprint features, thereby achieving the fusion of graph structure and chemical information.
It significantly improves the accuracy of predicting properties such as glycan classification and immunogenicity, provides a more comprehensive and reliable basis for judging glycan properties, and supports glycan-related scientific research and applications, such as host-pathogen interactions and glycan vaccine development.
Smart Images
Figure CN121528348A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of glycoinformatics technology, specifically to a method and system for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting. Background Technology
[0002] Polysaccharides are complex, highly branched polymers formed by monosaccharides linked by various glycosidic bonds, and are among the most diverse categories of biological macromolecules. They play an irreplaceable role in key biological processes such as cell signaling, protein folding, immune recognition, and host-microbe interactions.
[0003] Unlike biomolecules such as DNA, RNA, and proteins, which are synthesized directly under the guidance of gene templates, glycan synthesis is not directly regulated by gene templates and exhibits nonlinear, multi-branched structural characteristics, posing a significant challenge to computational modeling of glycans. In existing techniques, Bojar et al. decomposed glycan sequences into substructural fragments as "glycan words" input into a language model for glycan classification and pathogenicity contribution prediction. While this approach can capture some nonlinear features, it struggles to fully learn complex branching and tree-like structures.
[0004] Subsequent methods have encoded glycans as graph structures, using monosaccharides and glycosidic bonds as graph nodes, and aggregating neighbor features through multi-layer graph convolution to implicitly learn branching information. While these methods have made some progress in capturing branching connection patterns, they generally neglect the chemical information of glycans as macromolecules, resulting in limited performance of the prediction models and failing to meet the demands for high-precision property prediction. Summary of the Invention
[0005] The purpose of this invention is to propose a method and system for predicting macromolecular properties. By effectively fusing the graphical structure information of glycans with chemical fingerprint information, a more biologically meaningful representation of glycans is obtained, which significantly improves the accuracy of tasks such as glycan classification and immunogenicity prediction.
[0006] According to a first aspect of the embodiments of this disclosure, a method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting is provided, comprising the following steps: An undirected graph was constructed based on the IUPAC sequence of glycans, with monosaccharides and glycosidic bonds as real nodes of the graph, while virtual nodes were also introduced; and molecular fingerprint information was extracted from the glycan sequence. Construct an original adjacency matrix based on the connection relationships of real nodes, and then construct a fused adjacency matrix containing the connection relationships between virtual nodes and all real nodes based on the original adjacency matrix. The node features and molecular fingerprints are processed to achieve dimensional unification, and the molecular fingerprint features are assigned to the virtual nodes. A multi-layer graph convolutional layer is used. The first few layers use the original adjacency matrix to learn the topological structure information of the graph, and the last layer uses a fused adjacency matrix to achieve the fusion of graph structure features and molecular fingerprint features. The node features output by each graph convolution layer are pooled to obtain graph-level features. All graph-level features are then concatenated to form the final glycan representation, which is used for the prediction of the properties of carbohydrate biomolecules.
[0007] In one embodiment, the specific process of extracting molecular fingerprint information from the glycan sequence is as follows: the IUPAC sequence of the glycan is converted into a SMILES string using the GlyLEs tool, and then the molecular fingerprint is extracted from the SMILES string using the RDKit tool.
[0008] In one embodiment, the molecular fingerprint includes a MACCS 167-bit fingerprint, a PubChem 881-bit fingerprint, and an ECFP 1024-bit fingerprint, with the three fingerprints combined to form a 2072-dimensional molecular fingerprint vector.
[0009] In one embodiment, the specific process of unifying the dimensions of node features and molecular fingerprints is as follows: all nodes, including virtual nodes, are mapped to a 256-dimensional vector space, and the molecular fingerprint vector is mapped to 256 dimensions through a fully connected layer, so as to achieve the consistency of the dimensions of node features and molecular fingerprint features.
[0010] In one embodiment, the fully connected layer used for molecular fingerprint vector dimensionality reduction adopts a “2072→128→256” network structure, which transforms the 2072-dimensional molecular fingerprint vector into 256-dimensional vector through two linear transformations.
[0011] In one embodiment, the graph convolutional layer is a TransformerConv layer, with a total of four layers, each configured with eight attention heads; the first three layers use the original adjacency matrix for message passing to learn topological structure information, and the fourth layer uses a fused adjacency matrix to achieve feature fusion.
[0012] In one embodiment, the node features output by each graph convolution layer are pooled using a global summation pooling method, resulting in 256-dimensional graph-level features after each pooling layer. All graph-level features are concatenated to form a 1024-dimensional final glycan representation. The final glycan representation is then input into a classification head composed of a multilayer perceptron to complete property prediction. The classification head adopts a network structure of "1024→512→64→number of categories".
[0013] According to a second aspect of the present disclosure, a system for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprints is provided, comprising: The graph construction and fingerprint extraction module constructs an undirected graph based on the IUPAC sequence of glycans, using monosaccharides and glycosidic bonds as real nodes of the graph, while also introducing virtual nodes; and extracts molecular fingerprint information from the glycan sequence. The adjacency matrix construction module constructs an original adjacency matrix based on the connection relationships of real nodes, and then constructs a fused adjacency matrix containing the connection relationships between virtual nodes and all real nodes based on the original adjacency matrix. The feature dimension unification and assignment module performs dimension unification processing on node features and molecular fingerprints, and assigns molecular fingerprint features to virtual nodes. The multi-layer convolutional feature learning and fusion module uses multi-layer graph convolutional layers. The first few layers use the original adjacency matrix to learn the topological structure information of the graph, and the last layer uses the fusion adjacency matrix to achieve the fusion of graph structure features and molecular fingerprint features. The graph-level feature aggregation and prediction module pools the node features output from each graph convolution layer to obtain graph-level features, and concatenates all graph-level features to form the final glycan representation, which is used for the prediction of the properties of carbohydrate biomacromolecules.
[0014] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprints.
[0015] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprints.
[0016] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. By innovatively combining the glycan diagram structure (the topological relationship between monosaccharides and glycosidic bonds) with chemical fingerprint information (molecular structural features), we break through the limitations of traditional methods that only focus on a single feature. This allows the glycan representation to simultaneously cover spatial connectivity rules and chemical essential properties. Compared with models that rely solely on topological or chemical information, this approach can more comprehensively and deeply characterize glycan properties, providing more valuable basic data for subsequent predictions.
[0017] 2. Unlike traditional models that tend to ignore topological or chemical information, this invention allows the model to simultaneously learn the topological structure (such as branching and connection patterns) and chemical structure (such as molecular-level feature details) of sugar chains. This avoids prediction bias caused by missing key information, ensures that the model's judgment of sugar chain properties is based on complete information, and improves the comprehensiveness and reliability of learning.
[0018] 3. Based on comprehensive glycan characterization and complete information learning, it can more accurately capture key factors affecting glycan properties (such as the impact of chemical modifications on glycan immunogenicity). In tasks such as glycan classification and immunogenicity prediction, it can output more accurate results compared with existing mainstream models, providing more reliable technical support for glycan-related scientific research (such as host-pathogen interactions) and applications (such as glycan-based vaccine development). Attached Figure Description
[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.
[0020] Figure 1 This is a flowchart of a method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting. Detailed Implementation
[0021] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0024] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.
[0025] Example 1: This embodiment provides a method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting, including the following steps: Step 1: Construct an undirected graph based on the IUPAC sequence of glycans, using monosaccharides and glycosidic bonds as real nodes of the graph, while also introducing virtual nodes; and extract molecular fingerprint information from the glycan sequence. Specifically, taking one sugar chain as an example, the sugar chain of the sequence Man(a1-3)[Man(a1-6)]Man(b1-4)GlcNAc, where the node list is [Man,a1-3,Man,b1-4,GlcNAc,Man,a1-6,Virtual], and the SMILES string is O1C(O)[C@H](NC(C)=O)[C@@H](O[C@H Extract a 2072-bit fingerprint vector from the given data.
[0026] Step 2: Construct the original adjacency matrix based on the actual node connections. A ori In the original adjacency matrix A ori Based on this, virtual nodes are connected to other nodes to construct a fused adjacency matrix that includes the connection relationships between virtual nodes and all real nodes. A full ; In this embodiment, for the sugar chain Man(a1-3)[Man(a1-6)]Man(b1-4)GlcNAc, an adjacency matrix is constructed. A ori and A full : A ori A full Step 3: Perform dimensional unification processing on node features and molecular fingerprints, and assign the molecular fingerprint features to virtual nodes; Specifically, nodes are mapped to 256-dimensional vectors using torch.nn.Embedding, and 2072-dimensional fingerprint features are mapped to 256-dimensional feature vectors using a fully connected layer (2072→128→256).
[0027] The graph data is processed in batches, with virtual nodes uniformly set as the last node in each graph. To accurately locate virtual nodes, a Boolean mask tensor is generated for each graph, where the position corresponding to the virtual node is marked as True, and the rest of the nodes are marked as False. This mask tensor can be used to quickly locate the virtual nodes in each graph and replace their features with molecular fingerprint features.
[0028] In this embodiment, for the sugar chain with the sequence Man(a1-3)[Man(a1-6)]Man(b1-4)GlcNAc, its node list is [Man,a1-3,Man,b1-4,GlcNAc,Man,a1-6,Virtual], the node embedding matrix h has a shape of [8,256], the fingerprint embedding is fp, and the virtual_mask obtained after masking is [False, False, False, False, False, False, False, True]. h[virtual_mask] will return the row corresponding to the True position. The virtual node features are replaced with fingerprint features by h[virtual_mask] = fp.
[0029] Step 4: Employ multi-layer graph convolutional layers. The first few layers use the original adjacency matrix to learn the topological structure information of the graph, and the last layer uses a fused adjacency matrix to achieve the fusion of graph structure features and molecular fingerprint features. Specifically, a 4-layer TransformerConv architecture is used, with 8 attention heads per layer. The output of each layer is concatenated to form a 256-dimensional array. Layers 1-3 use... A oriThe topology of the message passing learning graph is constructed, and node features x1, x2, and x3 are obtained respectively. The fourth layer uses... A full The fingerprint and graph structure are fused to obtain x4.
[0030] Step 5: Pool the node features output by each graph convolution layer to obtain graph-level features, and concatenate all graph-level features to form the final glycan representation, which is used for the task of predicting the properties of carbohydrate biomolecules.
[0031] Specifically, the node features obtained from each layer are globally summed and pooled to obtain 256-dimensional graph representations g1, g2, g3, and g4. The graph representations of the four layers are concatenated into a 1024-dimensional vector to obtain the final representation of the glycan. The final representation is then input into a multilayer perceptron. The classification head structure is 1024→512→64→number of categories, as shown in Table 1. The average accuracy of each model on five independent training runs (N=5) on a single test set is given. The TFP-Net of this invention has achieved better results in the glycan property prediction task.
[0032] Table 1 Comparison of prediction results for glycan properties using different deep learning models Example 2: This embodiment provides a system for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting, including: The graph construction and fingerprint extraction module constructs an undirected graph based on the IUPAC sequence of glycans, using monosaccharides and glycosidic bonds as real nodes of the graph, while also introducing virtual nodes; and extracts molecular fingerprint information from the glycan sequence. The adjacency matrix construction module constructs an original adjacency matrix based on the connection relationships of real nodes, and then constructs a fused adjacency matrix containing the connection relationships between virtual nodes and all real nodes based on the original adjacency matrix. The feature dimension unification and assignment module performs dimension unification processing on node features and molecular fingerprints, and assigns molecular fingerprint features to virtual nodes. The multi-layer convolutional feature learning and fusion module uses multi-layer graph convolutional layers. The first few layers use the original adjacency matrix to learn the topological structure information of the graph, and the last layer uses the fusion adjacency matrix to achieve the fusion of graph structure features and molecular fingerprint features. The graph-level feature aggregation and prediction module pools the node features output from each graph convolution layer to obtain graph-level features, and concatenates all graph-level features to form the final glycan representation, which is used for the prediction of the properties of carbohydrate biomacromolecules.
[0033] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.
[0034] Example 3: An electronic device is provided for running the aforementioned "Prediction Method for Carbohydrate Biomacromolecule Properties Based on Graph Neural Networks and Chemical Fingerprints". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps one through five of the method described in Embodiment 1, specifically including but not limited to: Step 1: Construct an undirected graph based on the IUPAC sequence of glycans, using monosaccharides and glycosidic bonds as real nodes of the graph, while also introducing virtual nodes; and extract molecular fingerprint information from the glycan sequence. Step 2: Construct the original adjacency matrix based on the connection relationships of the real nodes, and then construct a fused adjacency matrix containing the connection relationships between the virtual nodes and all real nodes based on the original adjacency matrix; Step 3: Perform dimensional unification processing on node features and molecular fingerprints, and assign the molecular fingerprint features to virtual nodes; Step 4: Employ multi-layer graph convolutional layers. The first few layers use the original adjacency matrix to learn the topological structure information of the graph, and the last layer uses a fused adjacency matrix to achieve the fusion of graph structure features and molecular fingerprint features. Step 5: Pool the node features output by each graph convolution layer to obtain graph-level features, and concatenate all graph-level features to form the final glycan representation, which is used for the task of predicting the properties of carbohydrate biomolecules.
[0035] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.
[0036] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to perform steps one through five of the method described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.
[0037] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.
[0038] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
[0039] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.
Claims
1. A method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting, characterized in that, Includes the following steps: An undirected graph was constructed based on the IUPAC sequence of glycans, with monosaccharides and glycosidic bonds as real nodes of the graph, while virtual nodes were also introduced. Molecular fingerprint information was extracted from the glycan sequence; Construct an original adjacency matrix based on the connection relationships of real nodes, and then construct a fused adjacency matrix containing the connection relationships between virtual nodes and all real nodes based on the original adjacency matrix. The node features and molecular fingerprints are processed to achieve dimensional unification, and the molecular fingerprint features are assigned to the virtual nodes. A multi-layer graph convolutional layer is used. The first few layers use the original adjacency matrix to learn the topological structure information of the graph, and the last layer uses a fused adjacency matrix to achieve the fusion of graph structure features and molecular fingerprint features. The node features output by each graph convolution layer are pooled to obtain graph-level features. All graph-level features are then concatenated to form the final glycan representation, which is used for the prediction of the properties of carbohydrate biomolecules.
2. The method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting according to claim 1, characterized in that, The specific process for extracting molecular fingerprint information from glycan sequences is as follows: the IUPAC sequence of the glycan is converted into a SMILES string using the GlyLEs tool, and then the molecular fingerprint is extracted from the SMILES string using the RDKit tool.
3. The method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting according to claim 1 or 2, characterized in that, The molecular fingerprint includes a MACCS 167-bit fingerprint, a PubChem 881-bit fingerprint, and an ECFP 1024-bit fingerprint. The combination of these three fingerprints forms a 2072-dimensional molecular fingerprint vector.
4. The method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting according to claim 1, characterized in that, The specific process of unifying the dimensions of node features and molecular fingerprints is as follows: all nodes, including virtual nodes, are mapped to a 256-dimensional vector space, and the 2072-dimensional molecular fingerprint vector is mapped to 256 dimensions through a fully connected layer, so as to achieve the consistency of the dimensions of node features and molecular fingerprint features.
5. The method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting according to claim 4, characterized in that, The fully connected layer used for molecular fingerprint vector dimensionality reduction adopts a network structure of "2072→128→256", which transforms the 2072-dimensional molecular fingerprint vector into 256-dimensional vector through two linear transformations.
6. The method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting according to claim 1, characterized in that, The graph convolutional layer is a TransformerConv layer, with a total of four layers, each configured with eight attention heads; the first three layers use the original adjacency matrix for message passing to learn topological information, and the fourth layer uses a fused adjacency matrix to achieve feature fusion.
7. The method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting according to claim 1, characterized in that, The node features output by each graph convolution layer are pooled using a global summation pooling method, resulting in 256-dimensional graph-level features after each pooling layer. All graph-level features are concatenated to form a 1024-dimensional final glycan representation. The final glycan representation is then input into a classification head composed of a multilayer perceptron to complete property prediction. The classification head adopts a network structure of "1024→512→64→number of categories".
8. A system for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprinting, characterized in that, include: The graph construction and fingerprint extraction module constructs an undirected graph based on the IUPAC sequence of glycans, using monosaccharides and glycosidic bonds as real nodes of the graph, while also introducing virtual nodes. Molecular fingerprint information was extracted from the glycan sequence; The adjacency matrix construction module constructs an original adjacency matrix based on the connection relationships of real nodes, and then constructs a fused adjacency matrix containing the connection relationships between virtual nodes and all real nodes based on the original adjacency matrix. The feature dimension unification and assignment module performs dimension unification processing on node features and molecular fingerprints, and assigns molecular fingerprint features to virtual nodes. The multi-layer convolutional feature learning and fusion module uses multi-layer graph convolutional layers. The first few layers use the original adjacency matrix to learn the topological structure information of the graph, and the last layer uses the fusion adjacency matrix to achieve the fusion of graph structure features and molecular fingerprint features. The graph-level feature aggregation and prediction module pools the node features output from each graph convolution layer to obtain graph-level features, and concatenates all graph-level features to form the final glycan representation, which is used for the prediction of the properties of carbohydrate biomacromolecules.
9. An electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprints as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method for predicting the properties of carbohydrate biomacromolecules based on graph neural networks and chemical fingerprints as described in any one of claims 1-7.