Sparse knowledge graph entity alignment method and system

Abstract

The invention provides a sparse knowledge graph entity alignment method and system. The method comprises the following steps: acquiring structure data of chemical molecules, and identifying a minimum ring structure based on topological connection and bond energy distribution; clustering molecules with the same ring structure and the bond energy deviation smaller than a threshold value into a molecular group, extracting a shared minimum ring structure as a substructure node, creating a molecular node, and constructing a bidirectional correlation map of the molecules and the substructure node; extracting absorption peak data of the compound sample by utilizing spectral imaging, and generating an energy level spectrum; searching a substructure node combination corresponding to the homonymous molecule based on a map, and obtaining an embedded vector; and segmenting the energy level atlas into energy fragments according to a substructure path, extracting a distribution vector, performing multi-scale fusion on the distribution vector, a path feature and an embedded vector, and generating a composite feature vector to identify an equivalent entity and establish aligned connection. According to the method, high-precision alignment of the functional equivalent entities in the chemical sparse knowledge graph is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cheminformatics and artificial intelligence, and particularly relates to a sparse knowledge graph entity alignment method and system. BACKGROUND

[0002] In the field of chemical database integration, cross-database molecular entity alignment is a core task for building a unified knowledge graph. Typical scenarios include the existence of isomeric naming differences in different databases for molecular entities such as solvents and catalysts. The structural description of such entities often presents sparsity, only containing simplified topological connections or bond type information, lacking complete atomic coordinates or three-dimensional conformation data, and it is difficult to directly compare key structural features such as ring configurations or functional groups.

[0003] At present, the mainstream scheme adopts an embedding representation driven entity alignment method. Through a word embedding model or a graph neural network, a molecular entity and its text attributes are mapped into a low-dimensional vector. Then, the name string similarity and attribute vector similarity are comprehensively calculated, and the equivalence of entities is determined based on a preset threshold. If the comprehensive similarity of two entities exceeds the threshold, an alignment connection is established.

[0004] However, this scheme, although realizing automatic alignment, exposes significant defects in the chemical molecular scenario. Embedding generation relies on explicit relationship triples, but molecular entities in actual databases often exist in isolation, lacking associated reaction data, resulting in low embedding quality. Similarity calculation only focuses on surface text or simple molecular formulas, and cannot capture functional equivalence, mistakenly judging functionally similar but structurally different molecules as unrelated entities. SUMMARY

[0005] The present application provides a sparse knowledge graph entity alignment method and system to solve the problem of low entity alignment accuracy in the prior art.

[0006] In a first aspect, the present application provides a sparse knowledge graph entity alignment method, comprising:

[0007] Obtaining structural data of a chemical molecule, based on the topological connection relationship and bond energy distribution in the structural data, traversing the connection path between atoms in the structural data to identify the smallest ring structure;

[0008] Clustering chemical molecules with the same smallest ring structure and a bond energy distribution deviation less than a threshold into a molecular group, extracting a shared smallest ring structure in the molecular group as a substructure node, and creating a molecular node for each molecule in the molecular group, establishing a bidirectional connection relationship between the molecular node and the corresponding substructure node, constructing a bidirectional association graph, and recording the atomic connection path features of the shared smallest ring structure through the substructure node;

[0009] extracting absorption peak data of a compound sample in a high-frequency recognition area by spectral imaging, converting the absorption peak data into a bond vibration gradient, and generating a vibration energy level atlas;

[0010] based on the atomic connection path features in the bidirectional association atlas, retrieving a substructure node combination corresponding to a molecular node with the same chemical name as the compound sample, and obtaining an embedding vector of the molecular node corresponding to the substructure node combination;

[0011] segmenting the vibration energy level atlas according to the atomic connection path corresponding to the substructure node to obtain a distribution vector of the energy segment, fusing the distribution vector with the connection path features of the substructure node in multiple scales, and simultaneously introducing the embedding vector of the molecular node to generate a composite feature vector fusing bond energy and connection features to identify equivalent entities and establish alignment connection.

[0012] Optionally, after obtaining the embedding vector of the molecular node corresponding to the substructure node combination, the method further comprises:

[0013] extracting an atomic connection path feature vector of the molecular node corresponding to the substructure node combination, and combining the topological fingerprint feature sequence of the same name molecule in the path order;

[0014] performing a data enhancement operation on the topological fingerprint feature sequence to generate a perturbation feature sequence, combining the original topological fingerprint feature sequence with the perturbation feature sequence as a positive sample pair, and randomly selecting a topological fingerprint feature sequence of a different molecular node to combine with the original topological fingerprint feature sequence of the current molecular node as a negative sample pair;

[0015] inputting the positive sample pair and the negative sample pair into a similarity measure function to calculate the similarity scores of the two sequences in the positive sample pair and the similarity scores of the two sequences in the negative sample pair;

[0016] adjusting the parameters of the embedding vector generation model through backpropagation to maximize the similarity scores of the positive sample pairs and minimize the similarity scores of the negative sample pairs, iteratively updating until convergence, and outputting the optimized embedding vector.

[0017] clustering chemical molecules with the same minimum cyclic structure and a bond energy distribution deviation less than a threshold value into a molecular group, extracting a shared minimum cyclic structure in the molecular group as a substructure node, creating a molecular node for each molecule in the molecular group, establishing a bidirectional connection relationship between the molecular node and the corresponding substructure node, constructing a bidirectional association atlas, and recording the atomic connection path features of the shared minimum cyclic structure through the substructure node, comprising:

[0018] Calculate the bond energy distribution difference value of the same minimal ring structure in different chemical molecules. If the bond energy distribution difference value is less than a preset deviation threshold, the corresponding chemical molecules are classified into the same molecular group.

[0019] The smallest ring structure common to all chemical molecules in each molecular group is extracted and defined as the shared smallest ring structure. The atomic connection sequence and bond energy strength sequence of the shared smallest ring structure are combined into a topological feature group.

[0020] A hash function is performed on the topological feature group to generate a fixed-length unique identifier. A unique substructure node is created based on the unique identifier, and an independent molecular node is created for each chemical molecule in the molecular group.

[0021] A bidirectional connection relationship is established between the molecular node and the corresponding substructure node. The bidirectional connection relationship includes the inclusion relationship of the molecular node pointing to the substructure node, and the belonging relationship of the substructure node pointing to the molecular node.

[0022] All the molecular nodes, substructure nodes, and bidirectional connection relationships are integrated to construct a bidirectional correlation map, and the atomic connection path characteristics of the shared minimum ring structure are recorded through the substructure nodes.

[0023] Optionally, the bond energy distribution difference value of the same minimal ring structure in different chemical molecules is calculated. If the bond energy distribution difference value is less than a preset deviation threshold, the corresponding chemical molecules are classified into the same molecular group, including:

[0024] Extract the bond length data and electronegativity difference values ​​of all chemical bonds contained in the same minimal ring structure, and calculate the mean bond length and mean electronegativity difference of the same minimal ring structure in different chemical molecules respectively.

[0025] Based on the mean bond length and the mean electronegativity difference, the bond energy distribution difference value of the same minimal ring structure among different chemical molecules is calculated. The bond energy distribution difference value is the weighted sum of the variance of the mean bond length and the variance of the mean electronegativity difference.

[0026] If the bond energy distribution difference is less than a preset deviation threshold, the corresponding chemical molecules are classified into the same molecular group.

[0027] Optionally, the vibrational energy level spectrum is segmented into energy fragments according to the atomic connection paths corresponding to the substructure nodes, and the distribution vectors of the energy fragments are extracted. The distribution vectors are then fused with the connection path features of the substructure nodes using multi-scale features. Simultaneously, the embedding vectors of the molecular nodes are introduced to generate a composite feature vector that integrates bond energy and connection features to identify equivalent entities and establish aligned connections. This includes:

[0028] Based on the atomic connection path characteristics of the substructure nodes, the vibrational energy segments corresponding to the chemical bonds are located in the vibrational energy level spectrum, and the vibrational direction distribution and energy intensity distribution of each energy segment are extracted to generate an energy distribution vector.

[0029] The energy distribution vector and the atomic connection path features of the substructure nodes are fused hierarchically according to a preset ratio. The first level fuses the chemical bond type and vibration direction features, and the second level fuses the bond energy sequence and energy intensity distribution.

[0030] The embedding vector of the molecular node is spliced ​​with the multi-scale features after hierarchical fusion to generate an initial composite feature. The initial composite feature is then enhanced to retain the local features of bond vibration gradients and the global features of molecular structure, thus generating a composite feature vector.

[0031] Based on the composite feature vector, the similarity value of different knowledge graphs is calculated. When the similarity value is greater than the preset alignment threshold, they are identified as equivalent entities, and alignment connections are established between the equivalent entities.

[0032] Optionally, absorption peak data of the compound sample in the high-frequency recognition region is extracted using spectral imaging, and the absorption peak data is converted into bond vibration gradients to generate a vibrational energy level spectrum, including:

[0033] The absorption spectrum of the compound sample in a preset frequency band is acquired by a high-frequency spectral imaging device, and all peak positions and corresponding absorption peak intensities are extracted from the absorption spectrum to generate absorption peak data.

[0034] Based on the position of each peak in the absorption peak data, a preset bond vibration frequency mapping table is matched to determine the chemical bond type and vibration mode corresponding to each absorption peak.

[0035] The vibrational energy gradient of the chemical bond is calculated based on the absorption peak intensity value. The vibrational energy gradient includes the bond stretching vibration gradient and the bond bending vibration gradient.

[0036] By integrating the vibrational modes and vibrational energy gradients of all chemical bond types and sorting them according to the spatial position of the chemical bonds in the molecule, a hierarchical vibrational energy level map is generated.

[0037] Optionally, based on the atomic connection path features in the bidirectional correlation map, substructure node combinations corresponding to molecular nodes with the same chemical name as the compound sample are retrieved, and the embedding vectors of the molecular nodes corresponding to the substructure node combinations are obtained, including:

[0038] Based on the atomic connection path features in the bidirectional correlation map, molecular nodes with the same chemical name as the compound sample are matched;

[0039] Starting from the matched molecular node, all substructure nodes associated with the molecular node are located through bidirectional connection relationships, a combination of substructure nodes is generated, and based on the atomic connection path features, the set of molecular nodes associated with the combination of substructure nodes in the bidirectional association map is retrieved.

[0040] The embedding vector of each molecular node in the molecular node set is obtained from the preset embedding vector library. The embedding vector is generated by learning the global features of the molecular structure.

[0041] Secondly, this application provides a sparse knowledge graph entity alignment system, including:

[0042] The identification module is used to acquire the structural data of chemical molecules, and based on the topological connection relationships and bond energy distribution in the structural data, traverse the interatomic connection paths in the structural data to identify the smallest ring structure.

[0043] A construction module is used to cluster chemical molecules with the same minimum ring structure and a bond energy distribution deviation of less than a threshold into a molecular group, extract the shared minimum ring structure in the molecular group as a sub-structure node, create a molecular node for each molecule in the molecular group, establish a bidirectional connection relationship between the molecular node and the corresponding sub-structure node, construct a bidirectional association map, and record the atomic connection path features of the shared minimum ring structure through the sub-structure node.

[0044] The conversion module is used to extract absorption peak data of compound samples in the high-frequency recognition region using spectral imaging, convert the absorption peak data into bond vibration gradients, and generate vibrational energy level spectra.

[0045] The retrieval module is used to retrieve the substructure node combination corresponding to the molecular node with the same chemical name as the compound sample based on the atomic connection path features in the bidirectional correlation map, and to obtain the embedding vector of the molecular node corresponding to the substructure node combination.

[0046] The alignment module is used to segment the vibrational energy level spectrum into energy segments according to the atomic connection paths corresponding to the substructure nodes, extract the distribution vector of the energy segments, perform multi-scale feature fusion of the distribution vector and the connection path features of the substructure nodes, and simultaneously introduce the embedding vector of the molecular nodes to generate a composite feature vector that fuses bond energy and connection features to identify equivalent entities and establish alignment connections.

[0047] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement a sparse knowledge graph entity alignment method as described in the first aspect above.

[0048] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a sparse knowledge graph entity alignment method as described in the first aspect.

[0049] This application identifies the smallest ring structure by acquiring structural data of chemical molecules and traversing interatomic connection paths based on topological connections and bond energy distribution. This overcomes the limitations of sparsity in traditional structure descriptions, accurately capturing the configuration of core functional groups and providing structural consistency criteria for aligning heterogeneous molecules. By clustering molecules with the same smallest ring structure and bond energy deviations less than a threshold into molecular groups, extracting shared substructure nodes, and constructing bidirectional correlation maps, it achieves functional equivalence aggregation of structurally similar molecules. Simultaneously, by recording atomic connection path characteristics through substructure nodes, it establishes a structured index foundation for cross-database entity alignment. Finally, it extracts the number of absorption peaks from compound samples through spectral imaging. The data is converted into bond vibration gradients to generate vibrational energy level maps, which can supplement the missing bond energy dynamic features in structural data. The uniqueness of absorption peaks in high-frequency identification regions can solve the semantic ambiguity caused by differences in heterogeneous naming. Based on bidirectional correlation maps, the substructure node combinations of molecules with the same name are retrieved and the embedding vectors are obtained, which can integrate structural topology and semantic association, and eliminate the heterogeneity of identifiers caused by abbreviations or common names. By segmenting the vibrational energy level map according to substructure nodes and extracting the distribution vector, and fusing it with connection path features and embedding vectors at multiple scales to generate composite feature vectors, it can coordinate static structure and dynamic bond energy features to accurately identify functionally equivalent entities and avoid misjudgments caused by structural differences.

[0050] Furthermore, by extracting the topological fingerprint feature sequences of substructure node combinations corresponding to molecular nodes, perturbation feature sequences are generated through data augmentation, and positive and negative sample pairs are constructed. After inputting the similarity metric function to calculate the score, backpropagation is used to optimize the embedding vector generation model, maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. The technical effect lies in enhancing the model's robustness to subtle differences in the structure of homologous molecules, such as optical isomerism in cyclic configurations, through the serialization representation and contrastive learning mechanism of topological fingerprints. Simultaneously, by using negative samples to force the differentiation of functionally non-equivalent molecules, the generalization ability and alignment accuracy of the embedding vectors under sparse data are significantly improved.

[0051] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart of a sparse knowledge graph entity alignment method provided in this application is shown;

[0054] Figure 2 A scene diagram illustrating a sparse knowledge graph entity alignment method provided in this application is shown.

[0055] Figure 3 This paper presents a schematic diagram of the structure of a sparse knowledge graph entity alignment system provided in this application.

[0056] Figure 4 A schematic diagram of the structure of a computing device provided in this application is shown. Detailed Implementation

[0057] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0058] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0059] Research has found that in the field of chemical database integration, existing entity alignment techniques struggle to effectively address the sparsity and functional equivalence of molecular structures. Specifically, structural descriptions are incomplete; for example, solvents, catalysts, and other molecular entities are often represented in simplified topological connections or bond types across different databases, lacking three-dimensional conformational data, making direct comparison of key structural features impossible. Furthermore, traditional embedding methods rely on explicit relational triples to generate vector representations, but actual molecular entities often exist in isolation, lacking associated reaction data and failing to capture deep structural relationships, resulting in low embedding quality. In addition, similarity calculations focus only on textual or molecular formula surface features, neglecting molecular functional equivalence. For instance, molecules with different structures may possess similar catalytic activities due to the same cyclic configuration, leading to the misclassification of functionally equivalent molecules as irrelevant entities. Therefore, a new entity alignment method specifically for chemical molecular scenarios is urgently needed to accurately capture structural-functional equivalence under sparse data conditions.

[0060] To address the aforementioned issues, this invention proposes a sparse knowledge graph entity alignment method based on substructure decoupling and vibrational energy level fusion. Its core lies in constructing a bidirectional association graph through ring-shaped topological clustering and fusing bond vibrational gradient features to achieve functional equivalence perception. This method, by fusing vibrational energy levels and topological structures, overcomes the limitations of traditional text similarity, achieving precise alignment of molecules with "different structures but the same function." The bidirectional association graph maps isolated molecules to shared substructures, improving embedding quality even in the absence of reaction triples through substructure nodes. Minimal ring-shaped structure clustering reduces redundant comparisons, and energy fragment segmentation lowers the complexity of high-dimensional vibrational data processing, significantly optimizing computational efficiency. This method significantly improves the robustness of chemical molecule entity alignment: while preserving key structural features such as ring configurations and bond vibrational modes, it addresses data sparsity through substructure decoupling and resolves functional equivalence misjudgment through multimodal fusion, providing high-precision knowledge graph support for scenarios such as drug design and catalyst screening.

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] Figure 1 A flowchart of a sparse knowledge graph entity alignment method is provided in this application embodiment, such as... Figure 1 As shown, the method includes:

[0063] 101. Obtain the structural data of chemical molecules, and based on the topological connections and bond energy distribution in the structural data, traverse the interatomic connection paths in the structural data to identify the smallest ring structure;

[0064] In the above steps, structural data refers to a two-dimensional connection table or graphical model describing the structure of a chemical molecule, where atoms are abstracted as nodes and chemical bonds as edges. The data includes information such as the type of atom (carbon, oxygen) and the type of bond (single bond, double bond, etc.). Topological connections refer to the connection paths formed between atoms through chemical bonds, such as adjacency matrices or adjacency lists. Bond energy distribution refers to the energy characteristic values ​​of different types of chemical bonds. For example, the bond energy of a carbon-carbon single bond is approximately 347 kJ / mol, and the bond energy of a carbon-carbon double bond is approximately 614 kJ / mol. These energy values ​​are used to help evaluate the stability or reactivity of ring structures. The smallest ring structure refers to the smallest indivisible ring structure in the molecular diagram, such as a 6-atom ring of benzene or a 3-atom ring of cyclopropane, which does not contain smaller sub-rings.

[0065] In this embodiment, a depth-first search algorithm is first used to traverse the atomic connection paths in the molecular structure data, and the smallest ring structure is identified by combining bond energy features. The chemical molecular structure is converted into an undirected graph model, where each atom corresponds to a node in the graph, and each chemical bond corresponds to an edge. The bond type, such as single bond, double bond, and corresponding bond energy value in kJ / mol is recorded. For example, the bond energy of a carbon-carbon single bond is approximately 347 kJ / mol, and the bond energy of a carbon-carbon double bond is approximately 614 kJ / mol. A depth-first search traversal is initiated from any atomic node. During the traversal, the node visit order is recorded and a tree structure is generated. Edges not included in this tree structure are called closed-loop bonds. The number of closed-loop bonds is mathematically related to the number of molecular rings, and the calculation formula is: R = E - V + 1, where R represents the number of rings, E represents the total number of edges, and V represents the total number of nodes. For example, the decane molecule (C 10 H 22 If a ring contains 10 atomic nodes (V=10) and 9 chemical bonds (E=9), then the ring number R = 9 - 10 + 1 = 0. Cyclohexane (C6H...) 12The given equation contains 6 atomic nodes and 6 chemical bonds, so R = 6 - 6 + 1 = 1. For each detected closed loop bond, the continuity of the atoms at both ends in the depth-first search path needs to be determined. When the paths of the atoms at both ends are continuous, all atoms on the path form a primitive ring. If the path is interrupted, such as by branching structures, the shortest path between the atoms at both ends is found through the spanning tree backtracking function. The closed loop bond is then merged with the backtracked path to form a primitive ring. For example, when a closed loop bond connecting atoms 1 and 11 is detected, if the access path from atom 1 to atom 11 is continuous, a primitive ring containing the atoms of the path is directly formed. When a closed loop bond connecting atoms 1 and 16 is detected with an interrupted path, the spanning tree is backtracked through the parent node pointer to find the shortest path from atom 1 to 2 to 3 to ... to 16, and this path is merged with the closed loop bond to form a new primitive ring. After collecting all primitive rings, the minimum weighted loop basic algorithm is used to filter the minimum ring set. First, the number of atoms N in each primitive ring is calculated, and then the average bond energy is calculated. The calculation formula is as follows:

[0066]

[0067] Where K represents the number of chemical bonds within the ring, E i This represents the bond energy value of the i-th bond. During filtering, the primitive ring with the smallest N value is prioritized; if N values ​​are the same, then... Rings with higher bond values, such as cyclohexyl rings containing 6 atoms (N=6) and 6 bonds, would have a total bond energy of 3600 kJ / mol. If a cyclobutane structural unit ring exists, N = 4, assuming... =650 kJ / mol, then the cyclobutane ring is preferentially selected as the smallest ring. Finally, the output results will be used for subsequent processing such as molecular skeleton simplification or reactivity analysis.

[0068] In a practical application, in a case study of optimizing a chemical intermediate synthesis process, researchers analyzed a candidate molecule containing 7 carbon atoms and 1 oxygen atom. The atoms were numbered 1 to 8, and their topological connections were as follows: atoms 1 and 2 are connected by a single bond, 2 and 3 by a double bond, 3 and 4 by a single bond, 4 and 5 by a single bond, 5 and 6 by a single bond, and 6 and 1 by a single bond, forming a basic cyclic pathway. Simultaneously, atoms 4 and 7, 7 and 8, and 8 and 5 form a branched structure with single bonds. Furthermore, atoms 3 and 8 are directly connected by a single bond. To improve the stability of this molecule in the catalytic hydrogenation process at Plant A, it is necessary to identify its smallest ring structure to assess ring strain. Researchers transformed the molecule into a graph model with 8 nodes and 9 edges, employing a depth-first search to traverse the path starting from atom 1: Paths 1, 2, 3, 4, 5, 6, 1 were marked as closed loop bond 6-1; when extending the path to atoms 4, 7, 8, 5, since atom 5 was already in the path, closed loop bond 8-5 was marked; when exploring the path to atoms 3 and 8, since atom 8 already existed, closed loop bond 3-8 was marked. The number of rings was calculated using the elementary ring formula: the total number of bonds 9 minus the number of atoms 8 plus 1 equals 2 elementary rings. Further analysis of the closed loop paths revealed that closed loop bond 6-1 corresponds to a six-membered ring composed of atoms 1, 2, 3, 4, 5, and 6; closed loop bond 3-8, combined with the sub-paths of atoms 3 and 5, forms a five-membered ring composed of atoms 3, 4, 5, and 8; while closed loop bond 8-5, because atom 5 was already included in the aforementioned rings, did not generate a new ring. By comparing the number of atoms within the ring, the five-membered ring, containing only 5 atoms, was identified as the smallest ring structure. This identification result guides process engineers to optimize the hydrogenation reaction temperature parameters and avoid side reactions caused by the high strain of the small ring.

[0069] In the overall scheme of step 101 above, by integrating molecular topological connections and bond energy distribution characteristics, a systematic traversal of interatomic connection paths in chemical structures is achieved, accurately identifying the smallest ring structures. During the traversal process, path priority is optimized based on bond energy data to ensure that the algorithm prioritizes ring structures with stable bond angles and minimal bond tension. The combination of depth-first search and the smallest ring set algorithm significantly improves ring identification efficiency and effectively avoids the generation of redundant rings. The hierarchical decoding mechanism and primitive ring extraction technology can accurately decompose complex structures such as fused rings and bridged rings. The algorithm's comprehensive consideration of atomic connections and ring stability provides a reliable molecular ring skeleton analysis basis for drug design, while the enhanced noise resistance ensures high-precision identification performance in hand-drawn molecular diagrams and large-scale compound screening scenarios.

[0070] 102. Cluster chemical molecules with the same minimum ring structure and a bond energy distribution deviation less than a threshold into a molecular group, extract the shared minimum ring structure in the molecular group as a substructure node, create a molecular node for each molecule in the molecular group, establish a bidirectional connection relationship between the molecular node and the corresponding substructure node, construct a bidirectional correlation map, and record the atomic connection path features of the shared minimum ring structure through the substructure node;

[0071] Optionally, step 102 may specifically include the following steps:

[0072] 1021. Calculate the bond energy distribution difference value of the same minimal ring structure in different chemical molecules. If the bond energy distribution difference value is less than a preset deviation threshold, the corresponding chemical molecules are classified into the same molecular group.

[0073] Specifically, step 1021 may include the following processes: extracting the bond length data and electronegativity difference values ​​of all chemical bonds contained in the same minimal ring structure; calculating the mean bond length and mean electronegativity difference values ​​of the same minimal ring structure in different chemical molecules; calculating the bond energy distribution difference value of the same minimal ring structure among different chemical molecules based on the mean bond length and mean electronegativity difference values, wherein the bond energy distribution difference value is the weighted sum of the variance of the mean bond length and the variance of the mean electronegativity difference value; if the bond energy distribution difference value is less than a preset deviation threshold, then the corresponding chemical molecules are classified into the same molecular group.

[0074] 1022. Extract the smallest ring structure common to all chemical molecules in each molecular group, define it as the shared smallest ring structure, and combine the atomic connection sequence and bond energy strength sequence of the shared smallest ring structure into a topological feature group;

[0075] 1023. Execute a hash function on the topological feature group to generate a fixed-length unique identifier, create a unique substructure node based on the unique identifier, and create an independent molecular node for each chemical molecule in the molecular group;

[0076] 1024. Establish a bidirectional connection relationship between the molecular node and the corresponding substructure node, wherein the bidirectional connection relationship includes the inclusion relationship of the molecular node pointing to the substructure node, and the belonging relationship of the substructure node pointing to the molecular node.

[0077] 1025. Integrate all the molecular nodes, substructure nodes and bidirectional connection relationships to construct a bidirectional correlation map, and record the atomic connection path characteristics of the shared minimum ring structure through the substructure nodes.

[0078] In the above steps, a molecular group refers to a collection of chemical molecules with the same minimum ring structure and minimal differences in bond energy distribution. The difference is calculated by weighting the variance of the mean bond length and the variance of the mean electronegativity difference, and must be less than a preset threshold. A shared minimum ring structure refers to the minimum ring common to all molecules in the molecular group. Its atomic connection sequence and bond energy strength sequence are combined to form a topological feature set, which is used to describe the geometric and energy properties of the ring. A substructure node is a unique node generated based on the topological feature set. It is created by converting the topological feature set into a fixed-length unique identifier using a hash function. A molecular node represents the entity of a single chemical molecule. Bidirectional connections include the "inclusion relationship" of molecular nodes pointing to substructure nodes, indicating that the molecule contains the ring structure, and the "attribution relationship" of substructure nodes pointing to molecular nodes, indicating which molecules the ring belongs to. A bidirectional correlation map is a map formed by integrating all nodes and connections, used to record the association between ring structures and molecules.

[0079] In this embodiment, step 1021 first analyzes the bond energy distribution of the same minimal ring structure in different chemical molecules, extracting the bond length data of all chemical bonds within the ring and the electronegativity difference values ​​of the bonding atom pairs. Then, the average value of these bond lengths and the average value of the electronegativity difference are calculated for each molecule. Next, the variances of the average bond lengths and the average electronegativity difference between different molecules are calculated. These two types of variances are weighted according to a preset weight ratio, such as 0.6 for the bond length variance and 0.4 for the electronegativity variance, and then weighted and summed to obtain the bond energy distribution difference value. If the bond energy distribution difference value is less than a preset deviation threshold, the energy characteristics of these molecules are considered sufficiently similar, and they are grouped into the same molecular group. For example, in the analysis of the benzene ring molecular group, the average benzene ring bond length of molecule A is 0.140 nm, and that of molecule B is 0.139 nm; the average electronegativity difference of molecule A is 0.34, and that of molecule B is 0.36. The calculated bond length variance is 0.0000005, and the calculated electronegativity variance is 0.0002. The weighted sum is 0.6 × 0.0000005 + 0.4 × 0.0002 = 0.0000803. This value is less than the preset threshold of 0.0001. Therefore, molecules A and B are classified into the same benzene molecular group.

[0080] Secondly, in step 1022, the smallest ring structure common to all chemical molecules in the established molecular group is extracted and defined as the shared smallest ring structure. The atomic connection sequence and bond energy strength sequence of the shared structure are combined into a topological feature set. The atomic connection sequence describes the closed-loop connection order of atoms within the ring, and the bond energy strength sequence records the bond energy value corresponding to each chemical bond within the ring. For example, in the benzene molecular group, the shared structure is a six-atom benzene ring, with an atomic connection sequence of a closed-loop path of C1-C2-C3-C4-C5-C6-C1, and a bond energy strength sequence containing a carbon-carbon bond energy of 614 kJ / mol and a carbon-hydrogen bond energy of 413 kJ / mol. This information is combined to form a complete topological feature set.

[0081] Next, a hash function is performed on the topological feature group in step 1023 to generate a fixed-length unique identifier. The atomic connection sequence and bond energy sequence are concatenated into a complete string, which is then input into a hash algorithm such as SHA-256 for calculation. The resulting hash value serves as the unique identifier for the substructure node. Based on this identifier, substructure nodes representing shared ring structures are created, and independent molecular nodes are created for each chemical molecule within the molecular group. For example, the atomic sequence "C1C2C3C4C5C6" of the benzene ring is concatenated with the bond energy sequence "614,413" to form the complete string "C1C2C3C4C5C6,614,413". After SHA-256 calculation, a 64-bit hash value "e3b0c44298fc1c14..." is generated, and the substructure node "Benzene Ring-1" is created accordingly. Molecular nodes are also created for benzene molecule A and benzene molecule B.

[0082] Then, in step 1024, two pairs of directed connections are established between molecular nodes and substructure nodes. The first pair is an inclusion relationship where a molecular node points to a substructure node, indicating that the molecule contains this ring structure. The second pair is a membership relationship where a substructure node points to a molecular node, indicating which molecules the ring structure belongs to. For example, molecular node A establishes an inclusion relationship with the substructure node "benzene ring-1", indicating that molecule A contains this benzene ring structure; at the same time, the substructure node "benzene ring-1" establishes membership relationships with both molecular nodes A and B, indicating that the benzene ring structure belongs to both molecules A and B. These two pairs of connections form a complete bidirectional relationship chain.

[0083] Finally, step 1025 integrates all molecular nodes, substructure nodes, and bidirectional connections to construct a bidirectional correlation graph. Substructure nodes store the atomic connection path characteristics of shared minimum ring structures. The entire graph supports bidirectional query functionality: a molecular node can be used to query the ring structures it contains, and a substructure node can be used to query all molecules containing that ring. For example, in the final constructed bidirectional correlation graph, molecular nodes A and B are connected to the substructure node "benzene ring-1" through inclusion relationships. This node stores the closed-loop path information of C1-C2-C3-C4-C5-C6-C1; simultaneously, the "benzene ring-1" node is connected back to molecular nodes A and B through membership relationships. Using this graph, inputting "molecular node A" will query that it contains a benzene ring structure, and inputting "benzene ring-1" will query molecules A and B containing that ring.

[0084] In practical applications, in a catalyst development case, when researchers analyzed six candidate molecules M1 to M6 containing six-membered carbon rings, they first extracted the bond length data of the six C-C bonds within the six-membered ring and the difference in electronegativity of the bonding atoms. Then, they calculated the differences in bond energy distribution of the same ring in different molecules. Specifically, the bond length sequence of the ring in molecule M1 was taken as 1.54 × 10⁻⁶. ―10 m, 1.53×10 ―10 m, 1.55×10 ―10 m, 1.52×10 ―10 m, 1.54×10 ―10 m, 1.53×10 ―10 For m, the mean bond length μ1 = (1.54 + 1.53 + 1.55 + 1.52 + 1.54 + 1.53) ÷ 6 = 1.535. Since the difference in electronegativity among all bonds is 0, the mean δ1 = 0. When comparing M1 and M2, the variance of the mean bond length Δμ = (1.535 - 1.538). 2=0.000009, electronegativity variance Δδ=0, combined with weighting coefficients α=0.7 and β=0.3, the synthesized difference value D=0.7×0.000009+0.3×0=0.0000063. If D is less than the preset threshold of 0.00001, they are grouped together. At this time, M1 to M4 cluster into the first molecular group, and M5 to M6 cluster into the second molecular group. Next, the atomic connection path C1-C2-C3-C4-C5-C6-C1 and the bond energy sequence [347kJ / mol×6] of the shared six-membered ring of the first molecular group are extracted and the SHA-256 hash function is input to generate a 128-bit identifier "9F86D081" to create a unique substructure node S1. At the same time, independent molecular nodes N1 to N4 are constructed for M1 to M4, and two bidirectional connections are established: the inclusion relationship between molecular nodes and substructure nodes and the affiliation relationship between substructure nodes and molecular nodes. Finally, a bidirectional correlation map was constructed by integrating 4 molecular nodes, 1 substructure node, and 8 bidirectional connection edges. This map stores the characteristics of the atomic connection paths within the ring through the S1 node, such as the C1-C2 bond length of 1.54 × 10⁻⁶. ―10 m and bond angle 109.5°. This technical solution successfully guided Plant A to screen out highly stable molecular groups for high-temperature catalytic reactions, and utilized hash identifiers to ensure compatibility with molecular naming differences in Company B's heterogeneous database.

[0085] In the overall scheme of step 102 above, molecular intelligent clustering is achieved by comparing the bond energy distribution differences of the same minimum ring structure. The weighted sum of the mean variance of bond length and the mean of electronegativity difference is used as the bond energy distribution difference value for precise screening, and only molecules with deviations less than a preset threshold are retained to establish a homogeneous molecular group. The minimum ring structure common to all chemical molecules in the molecular group is extracted to form a shared substructure, and its atomic connection sequence and bond energy intensity sequence are integrated into a topological feature group. A hash function is used to generate a unique identifier to create a substructure node. At the same time, molecular nodes are generated independently for each molecule in the molecular group, and bidirectional connection relationships including affiliation and inclusion relationships are established. Finally, a bidirectional association map is constructed. This map dynamically records the atomic connection path characteristics and bond energy intensity sequences of the shared minimum ring structure through substructure nodes, realizing a bidirectional mapping between molecular topological features and cluster structure, and providing a systematic framework for molecular structure tracing and cluster relationship analysis.

[0086] 103. Extract absorption peak data of compound samples in the high-frequency recognition region using spectral imaging, convert the absorption peak data into bond vibration gradients, and generate vibrational energy level spectra;

[0087] Optionally, step 103 may specifically include the following steps:

[0088] 1031. Acquire the absorption spectrum of the compound sample in a preset frequency band using a high-frequency spectral imaging device, extract all peak positions and corresponding absorption peak intensities in the absorption spectrum, and generate absorption peak data;

[0089] 1032. Based on the position of each peak in the absorption peak data, match a preset bond vibration frequency mapping table to determine the chemical bond type and vibration mode corresponding to each absorption peak;

[0090] 1033. Calculate the vibrational energy gradient of the chemical bond based on the absorption peak intensity value, wherein the vibrational energy gradient includes the bond stretching vibration gradient and the bond bending vibration gradient;

[0091] 1034. Integrate the vibrational modes and vibrational energy gradients of all chemical bond types, and sort them according to the spatial position of the chemical bonds in the molecule to generate a hierarchical vibrational energy level map.

[0092] In the above steps, the high-frequency identification region refers to a preset frequency band in spectral imaging where chemical bond vibration characteristics are significant, such as 4000–400 cm⁻¹ in the infrared spectrum. -1 Absorption peak data includes the wavenumber of all peak positions within the frequency band and the corresponding absorption peak intensity values ​​and absorbance; bond vibration gradient is a physical quantity calculated from the absorption peak intensity, used to quantify the stretching vibration energy along the bond axis and the bending vibration energy of the bond angle change, and its value is obtained by multiplying the intensity by the gradient coefficient corresponding to the bond type; stretching vibration gradient and bending vibration gradient correspond to the energy quantification results of the above two vibration modes, respectively; vibrational energy level spectrum is a structural diagram arranged hierarchically according to the spatial position of chemical bonds in the molecule, integrating bond type, vibration mode and gradient value, used to describe the dynamic state of each bond in the molecule.

[0093] In this embodiment of the application, the compound sample is first acquired in a preset frequency band, such as the mid-infrared region 4000–400 cm⁻¹, using a high-frequency spectral imaging device, such as a Fourier transform infrared spectrometer, in step 1031. -1 The absorption spectrum was analyzed, and a filter was used to eliminate noise interference. Instrument drift error was corrected using a baseline correction algorithm such as polynomial fitting. Subsequently, all local maxima in the spectrum were identified as peak positions, and their wavenumbers and corresponding absorbance values ​​were recorded to generate a structured absorption peak data table. For example, the infrared spectrum of ethyl acetate (CH3COOC2H5) was collected at 1720 cm⁻¹. -1 The peak intensity of the carbonyl (C=O) stretching vibration is 0.85 at 1240 cm⁻¹. -1 The peak intensity of CO stretching vibration was identified at 0.62 at 2980 cm⁻¹. -1 The peak intensity of the methyl (CH) stretching vibration was identified at 0.78, forming the data table: [(1720,0.85),(1240,0.62),(2980,0.78)].

[0094] Secondly, based on the absorption peak data in step 1032, the wavenumber of each peak is matched with a preset bond vibration frequency mapping table, such as the standard infrared characteristic peak database. If the peak position is within the typical vibration frequency threshold of a specific chemical bond, such as 1650–1750 cm⁻¹ for the C=O bond... -1 If the peaks match, the chemical bond type and vibrational mode are determined to be stretching or bending; unmatched peaks are marked as unknown bond types to be analyzed. For example, ethyl acetate at 1720 cm⁻¹... -1 The peak matching database confirmed it as a carbonyl (C=O) stretching vibration, 1370 cm⁻¹. -1 The peaks match the methyl (CH) bending vibration, while the peak at 1150 cm⁻¹ corresponds to the methyl (CH) bending vibration. -1 The peak was marked as unknown because it did not fall into any key band.

[0095] Next, the bond vibration gradient is calculated based on the absorption peak intensity value in step 1033: the stretching vibration gradient formula is G. s =k s ·I, where k s The stretching vibration coefficient is determined by the bond type, such as C=O bond k s =1200J / mol), where I is the absorption peak intensity; the bending vibration gradient formula is G b =k b ·I, where k b For example, k is the bending vibration coefficient, such as for C-H bonds. b =850 J / mol, the two types of gradient values ​​reflect the energy intensity of axial deformation and angular deformation of the bond, respectively. For example, the C=O bond of ethyl acetate (I = 0.85) has a stretching vibration gradient of G. s =1200×0.85=1020J / mol, its C-H bond bending vibration gradient, I=0.42, is G b =850×0.42=357J / mol.

[0096] Finally, step 1034 integrates the vibrational modes and gradient data of all chemical bonds, sorting them according to the molecular spatial structure hierarchy: first the main chain bonds such as C=C and C=O, then the side chain bonds such as CH, generating a vibrational energy level map. Each node in the map contains the bond type, vibrational mode, gradient value, and three-dimensional coordinates, supporting the retrieval of bond dynamic parameters by spatial location. For example, the vibrational energy level map hierarchy of ethyl acetate is as follows: carbonyl node C=O, molecular center coordinates: stretching vibration gradient 1020 J / mol; ester node CO, adjacent C=O: stretching vibration gradient 744 J / mol; ethyl chain node CH2-CH3, terminal: CH bending vibration gradient 357 J / mol.

[0097] In practical applications, in a project analyzing the active sites on the surface of a chemical catalyst, researchers used a Fourier transform infrared spectroscopy imaging system with a spectral range of 4000 cm⁻¹. -1 Up to 400cm -1 Scanning of novel catalyst samples. First, samples were acquired using a high-frequency spectral imaging device at a preset frequency band of 2800 cm⁻¹. -1 Up to 3200cm -1 The infrared absorption spectrum detected three significant absorption peaks at positions of 2945 cm⁻¹. -1 3020cm -1 3180cm -1 The corresponding absorption peak intensity values ​​are A = 0.85, A = 0.92, and A = 0.78. Based on a preset bond vibration frequency mapping table, such as the CH stretching vibration range of 2800 cm⁻¹. -1 Up to 3000cm -1 OH stretching vibration range 3200cm -1 Up to 3600cm -1 2945cm -1 The peak is attributed to the stretching vibration of the CH bond in the alkyl chain, 3020 cm⁻¹ -1 The peak is attributed to the stretching vibration of the CH bond in the olefin, at 3180 cm⁻¹. -1 The peak is attributed to the stretching vibration of the OH bond in carboxylic acid. The energy gradient of the chemical bond vibrations is then calculated: Bond stretching vibration gradient: Based on the positive correlation between absorption peak intensity A and vibrational energy E, the energy gradient of the alkyl CH bond vibration is calculated as ΔE1 = 0.85 × 2945 = 2503.25 cm⁻¹. -1 ; Bond bending vibration gradient: combined with 1420cm -1 The weak absorption peak detected at A = 0.35 is attributed to the CH bending vibration. The bending vibration gradient ΔE2 is calculated to be 0.35 × 1420 = 497 cm⁻¹. -1 Integrating all bond vibration modes and energy gradients, vibrational energy level spectra are generated by sorting them according to molecular spatial hierarchy: Surface adsorbed molecules: carboxylic acid OH bond, 3180 cm⁻¹ -1 , ΔE=0.78×3180=2476.4cm -1 With olefins, CH bond, 3020cm -1 ΔE = 2788.4cm -1 Constitutes a high-energy vibrational layer; underlying framework structure: alkyl CH bonds, 2945 cm⁻¹ -1 ΔE = 2503.25cm -1 With CC key, 1100cm -1 ΔE = 0.28 × 1100 = 308 cm -1This forms a low-energy vibrational layer. The spectrum shows that the vibrational energy gradient of the carboxylic acid groups on the catalyst surface is higher than that of the underlying alkyl chains, guiding Plant A to optimize the carboxylic acid modification process on the catalyst surface to increase the density of reactive sites.

[0098] In the overall scheme of step 103 above, the absorption spectral characteristics of the compound sample within a preset frequency band are accurately acquired using high-frequency spectral imaging technology, and the position and intensity values ​​of each absorption peak are extracted. Combined with a bond vibration frequency mapping table, the peak positions are converted into key information about the chemical bond type and vibrational mode. The energy gradients of bond stretching and bending vibrations are quantified using the absorption peak intensity values. Finally, the vibrational modes and energy gradient parameters of all chemical bonds in the molecule are integrated, and a hierarchical vibrational energy level map is generated based on the spatial topological relationship of the chemical bonds, realizing a physical mapping from spectral signals to the dynamic characteristics of molecular vibrations. This map systematically presents the vibrational energy level distribution and spatial correlation of chemical bonds, providing a quantifiable data foundation for molecular conformation analysis and the study of the dynamic properties of chemical bonds.

[0099] 104. Based on the atomic connection path features in the bidirectional correlation graph, retrieve the substructure node combinations corresponding to the molecular nodes with the same chemical name as the compound sample, and obtain the embedding vector of the molecular node corresponding to the substructure node combination.

[0100] Optionally, step 104 may specifically include the following steps:

[0101] 1041. Based on the atomic connection path features in the bidirectional correlation map, match molecular nodes with the same chemical name as the compound sample;

[0102] 1042. Starting from the matched molecular node, locate all substructure nodes associated with the molecular node through bidirectional connection relationship, generate substructure node combination, and retrieve the set of molecular nodes associated with the substructure node combination in the bidirectional association map based on the atomic connection path features.

[0103] 1043. Obtain the embedding vector of each molecular node in the molecular node set from the preset embedding vector library. The embedding vector is generated through global feature learning of molecular structure.

[0104] In the above steps, the bidirectional association graph is a graph data structure that stores the relationship between chemical molecules and substructures. Molecular nodes represent complete chemical molecules, such as benzene molecules, while substructure nodes represent shared minimal ring structures, such as benzene rings. Atomic connection path features describe the connection order between atoms in a substructure, such as the closed-loop path C1→C2→C3. Embedding vectors are fixed-dimensional numerical vectors learned through graph neural networks or algorithms such as Node2Vec, such as 64-dimensional vectors [0.3, -0.5, ...], used to characterize the global structural features of molecules or substructures. Substructure node combinations refer to the set of all substructure nodes associated with the same molecular node through bidirectional connection relationships, such as the benzene ring and methyl ring nodes associated with the benzene molecule. The embedding vector library is a pre-generated database storing the embedding vectors of all molecular nodes, generated through global molecular structure feature learning, such as GNN aggregating atomic features.

[0105] In this embodiment, step 1041 first searches the bidirectional correlation spectrum for all molecular nodes with the same name based on the chemical name of the compound sample, such as "benzene". The matching is based on the consistency of atomic connection path characteristics: if the ring structure atomic sequence of a molecular node in the spectrum, such as the C1-C2-C3-C4-C5-C6 closed ring of benzene, completely matches the sample, it is determined to be a node with the same name. For example, when searching for a compound sample named "ethyl acetate", "molecular node A" is found in the bidirectional correlation spectrum. Its substructure contains the carbonyl (C=O) atomic sequence "C1=O" and the ester (CO) sequence "C2-O3", which are consistent with the atomic path of the sample, so the match is successful.

[0106] Secondly, starting from the matched molecular node in step 1042, all associated substructure nodes are located through bidirectional connection relationships: First, along the inclusion relationship molecular node to the substructure node, the substructure contained in the molecule is obtained, such as the benzene molecule pointing to the benzene ring node; second, along the affiliation relationship substructure node to the molecular node, other molecular nodes sharing the same substructure are obtained, forming a set of molecular nodes. For example, starting from the benzene molecule "molecular node A", its associated benzene ring node is located through the inclusion relationship; then, through the affiliation relationship of the benzene ring node, the set of molecular nodes sharing the ring {"molecular node A", "molecular node B"} is located, forming the molecular set corresponding to the substructure node combination "benzene ring-1".

[0107] Finally, in step 1043, the embedding vector of each node in the molecular node set is extracted from the preset embedding vector library. This vector is generated by learning the global molecular structure through a graph neural network (such as GCN): the molecular graph is input into the GNN, and features are aggregated through multiple layers of neighbors, as shown in the formula:

[0108]

[0109] Where i is the target node and l is the network layer number. This represents the set of neighbors of node i. The formula aggregates neighbor features through weighted aggregation and linear transformation, and finally outputs a new node embedding through an activation function. The output is a low-dimensional vector representing atomic connections and bond strength. For example, for the molecular node set {"molecular node A", "molecular node B"}, its pre-trained vector is read from the embedding vector library (e.g., "molecular node A" corresponds to [0.2, -0.3, 0.7]). This vector is generated by the GNN averaging the features of the 6 carbon atoms of the benzene ring and is used for downstream tasks such as drug similarity calculation.

[0110] In practical applications, during the screening of compound structural variants for a pharmaceutical company A, researchers need to retrieve all compounds labeled "benzodiazepine" from a graph database. The study first identified the molecular structure and its derived structures based on the atomic connection path characteristics in the bidirectional correlation map, such as the bond length between nitrogen atoms in the core seven-membered ring being 0.142 nm and the average bond angle being 112°. This led to the matching of three molecular nodes with the same name: N1, N2, and N3. Subsequently, through the bidirectional connection relationships in the map, the associated substructure node combination {benzene ring S1 hash identifier 9F86D081, seven-membered nitrogen heterocycle S2 hash identifier 5A03D0B2, chlorine substituent S3 hash identifier C7B2E4A1} was located from node N1. Based on the atomic path characteristics... The algorithm retrieves a set of four molecular nodes {M} = {N1, N3, N5, N7} associated with the combination. Finally, it obtains a 128-dimensional embedding vector for each node in the set from a pre-defined embedding vector library. For example, N1 corresponds to vector [0.72, -1.3, ..., 0.15], and N3 corresponds to [-0.21, 0.88, ..., -0.42]. These vectors incorporate molecular topological connectivity, such as the number of neighboring atoms of N5 = 6, and global structural features such as the number of chiral centers in the three-dimensional conformation = 2, which are used for subsequent training of the molecular activity prediction model.

[0111] In the overall scheme of step 104 above, the topological correlation characteristics of the bidirectional correlation map are used to quickly match the corresponding molecular nodes based on the chemical name of the compound sample, and locate the combination of associated substructure nodes along the bidirectional connection relationship; through the retrieval mechanism of atomic connection path features, the set of all molecular nodes associated with the target substructure combination in the map is dynamically obtained, and then the deep representation vectors of molecular nodes are extracted in batches from the pre-trained embedding vector library; this vector is generated by global feature learning of molecular structure, which can effectively encode spatial conformation and chemical bond properties, realize efficient information mapping across data sources, and provide a quantitative analysis hub for the correlation between spectral features and molecular structure.

[0112] 105. The vibrational energy level spectrum is segmented into energy segments according to the atomic connection paths corresponding to the substructure nodes. The distribution vector of the energy segments is extracted. The distribution vector is fused with the connection path features of the substructure nodes at multiple scales. At the same time, the embedding vector of the molecular nodes is introduced to generate a composite feature vector that fuses bond energy and connection features to identify equivalent entities and establish alignment connections.

[0113] Optionally, step 105 may specifically include the following steps:

[0114] 1051. Based on the atomic connection path characteristics of the substructure nodes, locate the vibrational energy segments corresponding to the chemical bonds in the vibrational energy level spectrum, extract the vibrational direction distribution and energy intensity distribution of each energy segment, and generate an energy distribution vector;

[0115] 1052. The energy distribution vector and the atomic connection path features of the substructure nodes are fused hierarchically according to a preset ratio. The first level fuses the chemical bond type and vibration direction features, and the second level fuses the bond energy sequence and energy intensity distribution.

[0116] 1053. The embedding vector of the molecular node is spliced ​​with the multi-scale features after hierarchical fusion to generate an initial composite feature. The initial composite feature is then enhanced to retain the local features of bond vibration gradient and the global features of molecular structure to generate a composite feature vector.

[0117] 1054. Based on the composite feature vector, calculate the similarity value of different knowledge graphs. When the similarity value is greater than the preset alignment threshold, identify them as equivalent entities and establish alignment connections between the equivalent entities.

[0118] In addition, step 105 can also involve introducing an optimized embedding vector to generate a composite feature vector that fuses bond energy and connectivity features to identify equivalent entities and establish aligned connections.

[0119] In the above steps, atomic connection path features refer to the connection order between atoms in the ring structure stored in the substructure node, such as the C1→C2→C3 closed ring path of the benzene ring; the energy distribution vector is the energy fragment quantification data extracted from the vibrational energy level spectrum, including vibrational direction, intensity distribution of stretching or bending, and energy value sequence; multi-scale feature fusion refers to the hierarchical integration of features of different granularities according to a preset ratio, such as the first layer fusing bond type and vibrational direction, and the second layer fusing bond energy and energy intensity; the composite feature vector is a high-dimensional vector that fuses molecular embedding vector to represent the global structure, energy distribution vector to represent local vibration, and connection path features to represent the topology; the similarity value is calculated by cosine similarity or Euclidean distance and is used to determine the equivalence of entities in different knowledge graphs; equivalent entities refer to molecules or substructure nodes with the same chemical function or structure; alignment connection is the bidirectional association relationship established between the identified equivalent entities.

[0120] In this embodiment, firstly, in step 1051, based on the atomic connection path of the substructure node, such as the C1-C6 closed ring of the benzene ring, the energy segment corresponding to the chemical bond, such as the C-C bond, is located in the vibrational energy level spectrum. The vibrational direction distribution, the intensity ratio of stretching vibration and bending vibration, the energy intensity distribution, and the gradient value sequence of each bond are extracted for each energy segment. The directional components are decomposed by wavelet transform and normalized to generate a fixed-dimensional energy distribution vector. For example, in the energy segment of the benzene ring substructure, the vibrational direction distribution of the C-C bond is 70% stretching vibration and 30% bending vibration, and the energy intensity distribution is [614, 598, 602] kJ / mol. After wavelet decomposition and normalization, the directional distribution vector [0.70, 0.30] and the intensity distribution vector [0.35, 0.25, 0.40] are generated, which are combined to form the final energy distribution vector [0.70, 0.30, 0.35, 0.25, 0.40].

[0121] Secondly, in step 1052, the energy distribution vector and atomic connection path features are fused in layers: First-level fusion: chemical bond type encoding (e.g., C bond type encoding 101) and vibration direction distribution vector (e.g., [0.70, 0.30]) are spliced ​​together to output a 128-dimensional feature vector; Second-level fusion: the bond energy sequence (e.g., [614, 598, 602]) and energy intensity distribution vector (e.g., [0.35, 0.25, 0.40]) are input into the fully connected layer, and after weighted summation, a 64-dimensional feature vector is output. For example, the connection path feature of the benzene ring "C1C2C3C4C5C6" is encoded as [1, 0, 1, ...], and spliced ​​with the direction vector [0.70, 0.30] to generate the first-level vector; the bond energy sequence [614, 598, 602] and the intensity vector [0.35, 0.25, 0.40] are fused through the fully connected layer to generate the second-level vector [0.42, -0.15, ...].

[0122] Next, in step 1053, the embedding vector of the molecular node, such as the GNN output vector [0.2, -0.3, 0.7] of benzene molecule A, is concatenated with the layered fused features to generate an initial composite feature. Through feature enhancement operations, such as 1D convolution to extract local bond vibration gradient features and self-attention mechanism to capture global molecular structure dependence, local vibration details and global topological association are preserved, and a 256-dimensional composite feature vector is output. For example, the molecular embedding vector [0.2, -0.3, 0.7] is concatenated with the first layer vector [1, 0, 1, ..., 0.70, 0.30] and the second layer vector [0.42, -0.15] to form a 290-dimensional initial vector. Local features such as C-C bond vibration peaks are extracted by the convolutional layer and global features such as benzene ring symmetry are weighted by the attention layer to generate a composite vector [0.18, -0.22, ..., 0.31].

[0123] Finally, based on the composite feature vector, step 1054 is used to calculate the cosine similarity of entities in different knowledge graphs:

[0124]

[0125] in, and This represents two feature vectors to be compared, typically numerical arrays. If the similarity is greater than a preset threshold, such as 0.85, they are considered equivalent entities, and a bidirectional alignment connection is established between them, representing inclusion and attribution relationships. For example, if the cosine similarity between the composite vector of molecule A and the vector of molecule B is 0.92 > 0.85, they are identified as equivalent benzene ring entities; a bidirectional connection is established between them, supporting cross-spectral queries, allowing the location of vibrational energy level data of molecule B through molecule A.

[0126] In practical applications, in a drug molecule similarity analysis project, researchers need to compare the vibrational properties of specific chemical bonds in two candidate molecules X and Y to determine their structural equivalence. First, based on the atomic connection path characteristics of substructure node N1, a ring structure containing six carbon atoms with bond length sequences of 0.138 nm, 0.140 nm, 0.139 nm, 0.141 nm, 0.137 nm, and 0.142 nm, the energy segment corresponding to the carbon-carbon double bond (C=C) is located in the vibrational energy level spectrum: at a wavenumber of 1580 cm⁻¹. -1 The extracted vibrational direction distribution is perpendicular to the molecular plane, with a direction vector [0, 0.92, 0.39]. The energy intensity distribution is Gaussian, with a peak intensity of 0.87 and a half-width / half-height of 15 cm. -1This generates an energy distribution vector V1 = [1580, 0.87, 15, 0.92, 0.39]. Subsequently, multi-scale feature fusion is performed: the first level combines the chemical bond type (C=C) of the substructure nodes with vibrational direction features to generate feature F1 = [C=C, perpendicular, 0.92]; the second level weights and fuses the mean of the bond energy sequence with the peak energy intensity according to a preset weight of 6:4, i.e., F2 = 0.6 × 0.1395 + 0.4 × 0.87 = 0.0837 + 0.348 = 0.4317. Next, the embedding vector of molecular node M1, such as [0.72, -1.3, 0.15], is introduced and concatenated with the hierarchical features F1 and F2 to form the initial composite feature C. init =[0.72,-1.3,0.15,C=C,Vertical,0.92,0.4317], and then the bond vibration gradient, such as 1580cm, is preserved through feature enhancement operations. -1 Up to 1600cm -1 The slope of the interval is 0.85, which, along with the global topological features of the molecule, ultimately generates a 128-dimensional composite feature vector C. final The similarity between molecules X and Y is calculated based on this vector using the cosine similarity formula. and The ratio of the dot product to the modulus is used to determine an equivalent entity if the result exceeds a preset threshold. An alignment connection is then established in the knowledge graph, linking it to the same pharmacophore database ID. This process, through vibrational energy segmentation and cross-level feature fusion, achieves collaborative analysis of dynamic bond vibration characteristics and static structural features, providing a new approach for identifying isomer active groups.

[0127] In the overall scheme of step 105 above, the energy distribution vector corresponding to the chemical bond is extracted and the vibration direction and energy intensity information are integrated; the vibration direction features in the distribution vector are initially fused with the chemical bond type of the substructure node through a preset hierarchical fusion mechanism, and the energy intensity distribution and bond energy sequence are further fused to form multi-scale correlation features; then the embedding vector of the molecular node is spliced ​​and feature enhancement operation is performed to retain the local gradient characteristics of bond vibration and the global characterization of molecular conformation, generating a composite feature vector that contains both static topological intensity and dynamic vibration mode; based on this vector, cross-source entity similarity is calculated, and when it exceeds the preset alignment threshold, it is automatically identified as an equivalent entity, and a bidirectional alignment connection is established between the substructure and the vibration energy fragment, realizing cross-modal collaborative mapping of molecular topological path features and spectral vibration energy level data.

[0128] The following is a complete embodiment for steps 101 to 105:

[0129] like Figure 2As shown, in a drug development project, researchers needed to determine whether two candidate molecules shared the same anti-inflammatory active group. First, by analyzing the molecular topological connections and bond energy distribution, the smallest common ring structure—the carbon-benzene ring—was identified. The average bond length of the benzene ring in molecule X was 0.1402 nm, while that in molecule Y was 0.1405 nm. Their standard deviation difference of 0.0003 nm was below the threshold of 0.002 nm, therefore, they were clustered into the same molecular group, and the benzene ring was extracted as the shared substructure node N1.

[0130] Subsequently, a bidirectional correlation graph was constructed: molecular nodes M were created for X and Y. X M Y and establish a bidirectional connection. N1 records the bond length sequence and mean characteristics of the benzene ring. Next, infrared spectroscopy is used to extract the 1580–1620 cm⁻¹ region. ―1 Absorption peak data in the high-frequency region. Molecular X at 1600 cm⁻¹ ―1 An absorption peak intensity of 0.85 was detected at 1595 cm⁻¹. ―1 An absorption peak intensity of 0.82 was detected. After converting it into a bond vibration gradient, a vibrational energy level spectrum was generated, yielding the vibrational direction vectors for X as [0.89, 0.31] and Y as [0.92, 0.28]. Based on the bidirectional spectrum retrieval of N1 and the acquisition of molecular node embedding vectors, the energy fragment distribution vector was further extracted: V X =[1600,0.85,0.89]、V Y =[1595,0.82,0.92].

[0131] In the feature fusion stage, a multi-scale strategy is adopted: atomic paths and vibrations are fused, and the average bond length of N1 (0.1402 nm) is combined with the vibration direction to generate feature F. X =[0.1402,0.89]、F Y = [0.1402, 0.92]; Weighted energy fusion, calculating F by weighting the standard deviation of bond energy (0.0018) and absorption peak intensity at a ratio of 6:4. 2X =0.34108, F 2Y =0.32908. Embedded vector concatenation generates a composite feature vector C. X =[0.72,―1.3,0.15,0.1402,0.89,0.34108] and C Y = [0.68,―1.4,0.18,0.1402,0.92,0.32908].

[0132] Finally, C is calculated using cosine similarity. X With C Y The similarity is 0.992 (dot product 3.2873, modulus || C).X ||≈1.82、||C Y If the value is approximately 1.79, exceeding the threshold of 0.95, X and Y are determined to be equivalent entities, and an alignment connection is established in the knowledge graph. This method solves the misalignment problem caused by neglecting energy distribution in traditional methods by fusing static structural features and dynamic bond vibration data.

[0133] Optionally, after obtaining the embedding vectors of the molecular nodes corresponding to the substructure node combination, the method further includes: extracting the atomic connection path feature vectors of the molecular nodes corresponding to the substructure node combination, and combining them in the path order to generate a topological fingerprint feature sequence of molecules with the same name; performing data augmentation on the topological fingerprint feature sequence to generate a perturbation feature sequence, combining the original topological fingerprint feature sequence with the perturbation feature sequence to form a positive sample pair, randomly selecting the topological fingerprint feature sequences of different molecular nodes, and combining them with the original topological fingerprint feature sequence of the current molecular node to form a negative sample pair; inputting the positive sample pair and the negative sample pair into a similarity metric function, calculating the similarity score between the two sequences in the positive sample pair and the similarity score between the two sequences in the negative sample pair; adjusting the parameters of the embedding vector generation model through backpropagation to maximize the similarity score of the positive sample pair and minimize the similarity score of the negative sample pair, iteratively updating until convergence, and outputting the optimized embedding vector.

[0134] In the above steps, the atomic connection path feature vector refers to a numerical vector generated based on the atomic connection order of substructure nodes. A hash function maps chemical bond types (e.g., single / double bonds), atom types (e.g., C / N / O), and path topology to a fixed-dimensional feature representation, such as a 1024-dimensional vector, to quantify local molecular connection patterns. The topological fingerprint feature sequence is a sequence formed by arranging the feature vectors of all substructure nodes of the molecule according to the atomic connection path order, representing the overall molecular topology. The perturbation feature sequence is a new sequence generated after applying data augmentation operations such as subgraph permutation and bond breaking to the original topological fingerprint feature sequence, enhancing the model's robustness by introducing controllable noise. Positive sample pairs consist of the original topological fingerprint sequence of the same molecule and its perturbation sequence, used to train the model to identify structural equivalence. Negative sample pairs are randomly paired from the original sequences of different molecules, forcing the model to distinguish non-equivalent structures. The similarity metric function calculates the similarity score between the vector spaces of two sequences, using the formula: Where A·B is the vector dot product, and ||A|||B|| is the vector magnitude.

[0135] In this embodiment, the atomic connection path feature vectors corresponding to the substructure node combinations are first extracted: Based on the molecular graph structure, the atomic connection paths of each substructure node, such as the benzene ring, are identified as closed loops such as C1→C2→C3. Using the topological fingerprint algorithm of RDKit, the path feature vector is calculated to traverse all subgraphs in the path with a preset minimum path length (e.g., minPath=1) and a maximum path length (e.g., MaxPath=7). For each subgraph, the atomic type (e.g., carbon atom encoding 12), bond type (e.g., double bond encoding 2), and aromaticity are hashed to generate a 32-bit hash value. After modulo operation, the hash value is mapped to a 2048-dimensional binary vector and arranged according to the atomic connection path order to generate the topological fingerprint feature sequence 6. Next, data augmentation operations are performed to generate a perturbation feature sequence, including randomly permuting the substructure order in the sequence with a probability p=0.2, such as swapping the positions of the benzene ring and methyl group, or breaking non-critical chemical bonds according to a Bernoulli distribution p=0.15, such as breaking the O-H bond of the hydroxyl group, and regenerating the subgraph hash vector to form a perturbation sequence. Then, training sample pairs are constructed, and the original sequence S is... orig Its perturbation sequence S aug Combining into positive sample pairs (S) orig ,S aug Simultaneously, topological fingerprint sequences S of different molecules were randomly selected. orig With, S neg Combined into negative sample pairs (S) orig ,S neg Finally, the similarity score between positive and negative sample pairs is calculated using a similarity metric function, employing a contrastive loss function:

[0136]

[0137] Where τ is the temperature hyperparameter with a default value of 0.05, backpropagation is performed to iteratively optimize the parameters of the embedding vector generation model, so that the scores of positive sample pairs approach 1 and the scores of negative sample pairs approach 0, until the loss converges and the optimized embedding vector is output.

[0138] In practical applications, in a drug development project, researchers used quantum-inspired optimization frameworks such as Q-Drug to optimize the structure of candidate molecules. First, the target molecule, such as a benzimidazole derivative, was encoded into a 128-dimensional binary embedding vector using a discrete variational autoencoder. For example, molecule M... refThe vector representation is [1,0,1,0,0,1,...]. This process is based on training on the ZINC-250k drug dataset, ensuring that the vector preserves the molecular topology and physicochemical properties. Subsequently, a target function similar to the Ising energy is constructed using binary embedding. For example, drug bioactivity (pIC50) and drug similarity (QED) are proportionally fused into a function E = -10 × QED - pIC50. A simulated bifurcation algorithm is then used to search for the minimum energy in the discrete domain; this step only requires 1 / 10 of the time of traditional methods to locate the global optimum. Finally, the optimized binary vector is decoded into a new molecular structure; for example, the optimal vector [0,1,1,0,1,...] is converted into a structurally modified molecule M. opt The method achieved a QED value of 0.948 and significantly enhanced bioactivity. Furthermore, its synthetic feasibility was verified through synthetic pathway analysis, ensuring that the new molecule reduced toxicity while retaining core pharmacologically active groups such as the benzene ring and nitrogen heterocycles. This method, combining quantum-inspired optimization with deep learning, overcomes the limitations of continuous-domain optimization, providing an efficient solution to complex problems such as the activity cliff.

[0139] In the overall scheme described above, the atomic connection path feature vectors of the molecular nodes associated with the substructure nodes are extracted, and molecular fingerprint features with topological sequence uniqueness are generated according to the path order. The adversarial enhancement of the feature space is achieved by constructing positive and negative sample pairs by perturbing the original sequence. In the contrastive learning framework, the self-consistency of positive sample pairs and the heterogeneity difference of negative sample pairs are calculated simultaneously using a similarity metric function, driving the dynamic optimization of the embedding vector model parameters. This forces the model to map the original topological sequence and its perturbed sequence to the adjacent embedding space, while moving away from the topological feature representation of random negative samples. After iterative training, the embedding vector output by the model has strong robustness to molecular topological noise and significantly improves the measurement accuracy of chemical structure similarity relationships, providing a highly generalizable topological semantic representation base for molecular clustering and cross-source retrieval tasks.

[0140] Figure 3 This application provides a schematic diagram of the structure of a sparse knowledge graph entity alignment system, as shown in the embodiments below. Figure 3 As shown, the system includes:

[0141] The identification module 31 is used to acquire the structural data of chemical molecules, and based on the topological connection relationship and bond energy distribution in the structural data, traverse the interatomic connection paths in the structural data to identify the smallest ring structure.

[0142] The construction module 32 is used to cluster chemical molecules with the same minimum ring structure and a bond energy distribution deviation of less than a threshold into a molecular group, extract the shared minimum ring structure in the molecular group as a sub-structure node, create a molecular node for each molecule in the molecular group, establish a bidirectional connection relationship between the molecular node and the corresponding sub-structure node, construct a bidirectional correlation map, and record the atomic connection path features of the shared minimum ring structure through the sub-structure node.

[0143] The conversion module 33 is used to extract the absorption peak data of the compound sample in the high-frequency recognition region using spectral imaging, convert the absorption peak data into bond vibration gradients, and generate a vibrational energy level spectrum.

[0144] The retrieval module 34 is used to retrieve the substructure node combination corresponding to the molecular node with the same chemical name as the compound sample based on the atomic connection path features in the bidirectional correlation map, and to obtain the embedding vector of the molecular node corresponding to the substructure node combination.

[0145] Alignment module 35 is used to segment the vibrational energy level spectrum into energy segments according to the atomic connection paths corresponding to the substructure nodes, extract the distribution vector of the energy segments, perform multi-scale feature fusion of the distribution vector and the connection path features of the substructure nodes, and simultaneously introduce the embedding vector of the molecular nodes to generate a composite feature vector that fuses bond energy and connection features to identify equivalent entities and establish alignment connections.

[0146] Figure 3 The sparse knowledge graph entity alignment system described above can perform... Figure 1 The implementation principle and technical effects of the sparse knowledge graph entity alignment method described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the sparse knowledge graph entity alignment system in the above embodiments perform operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0147] In one possible design, Figure 3 The sparse knowledge graph entity alignment system of the embodiment shown can be implemented as a computing device, such as... Figure 4 As shown, the computing device may include a storage component 41 and a processing component 42;

[0148] The storage component 41 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 42.

[0149] The processing component 42 is used for the above Figure 1 The embodiment describes a sparse knowledge graph entity alignment method.

[0150] The processing component 42 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0151] Storage component 41 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0152] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0153] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0154] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0155] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0156] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown is a sparse knowledge graph entity alignment method.

[0157] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0158] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0159] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for aligning entities in a sparse knowledge graph, characterized in that, include: Obtain structural data of chemical molecules, and based on the topological connections and bond energy distribution in the structural data, traverse the interatomic connection paths in the structural data to identify the smallest ring structure; Chemical molecules with the same minimum ring structure and a bond energy distribution deviation less than a threshold are clustered into molecular groups. The shared minimum ring structure in the molecular groups is extracted as a substructure node, and a molecular node is created for each molecule in the molecular group. A bidirectional connection relationship is established between the molecular node and the corresponding substructure node to construct a bidirectional correlation map. The atomic connection path features of the shared minimum ring structure are recorded through the substructure node. Absorption peak data of compound samples in the high-frequency recognition region are extracted using spectral imaging, and the absorption peak data are converted into bond vibration gradients to generate vibrational energy level spectra. Based on the atomic connection path features in the bidirectional correlation map, substructure node combinations corresponding to molecular nodes with the same chemical name as the compound sample are retrieved, and the embedding vectors of the molecular nodes corresponding to the substructure node combinations are obtained. The vibrational energy level spectrum is segmented into energy fragments according to the atomic connection paths corresponding to the substructure nodes. The distribution vectors of the energy fragments are extracted, and the distribution vectors are fused with the connection path features of the substructure nodes at multiple scales. At the same time, the embedding vectors of the molecular nodes are introduced to generate a composite feature vector that fuses bond energy and connection features to identify equivalent entities and establish aligned connections.

2. The method according to claim 1, characterized in that, After obtaining the embedding vectors of the molecular nodes corresponding to the substructure node combination, the method further includes: Extract the atomic connection path feature vectors of the molecular nodes corresponding to the substructure node combination, and combine them in the path order to generate the topological fingerprint feature sequence of molecules with the same name; Perform data augmentation on the topological fingerprint feature sequence to generate a perturbation feature sequence. Combine the original topological fingerprint feature sequence with the perturbation feature sequence to form a positive sample pair. Randomly select the topological fingerprint feature sequences of different molecular nodes and combine them with the original topological fingerprint feature sequence of the current molecular node to form a negative sample pair. Input the positive sample pairs and negative sample pairs into the similarity measurement function to calculate the similarity score between the two sequences in the positive sample pair and the similarity score between the two sequences in the negative sample pair; The parameters of the embedding vector generation model are adjusted by backpropagation to maximize the similarity score of positive sample pairs and minimize the similarity score of negative sample pairs. The process is iteratively updated until convergence, and the optimized embedding vector is output.

3. The method according to claim 1, characterized in that, Chemical molecules with the same minimum ring structure and bond energy distribution deviation less than a threshold are clustered into molecular groups. The shared minimum ring structure within each molecular group is extracted as a sub-structure node. Molecular nodes are created for each molecule in the molecular group. A bidirectional connection relationship is established between the molecular nodes and their corresponding sub-structure nodes to construct a bidirectional correlation graph. The atomic connection path characteristics of the shared minimum ring structure are recorded through the sub-structure nodes, including: Calculate the bond energy distribution difference value of the same minimal ring structure in different chemical molecules. If the bond energy distribution difference value is less than a preset deviation threshold, the corresponding chemical molecules are classified into the same molecular group. The smallest ring structure common to all chemical molecules in each molecular group is extracted and defined as the shared smallest ring structure. The atomic connection sequence and bond energy strength sequence of the shared smallest ring structure are combined into a topological feature group. A hash function is performed on the topological feature group to generate a fixed-length unique identifier. A unique substructure node is created based on the unique identifier, and an independent molecular node is created for each chemical molecule in the molecular group. A bidirectional connection relationship is established between the molecular node and the corresponding substructure node. The bidirectional connection relationship includes the inclusion relationship of the molecular node pointing to the substructure node, and the belonging relationship of the substructure node pointing to the molecular node. All the molecular nodes, substructure nodes, and bidirectional connection relationships are integrated to construct a bidirectional correlation map, and the atomic connection path characteristics of the shared minimum ring structure are recorded through the substructure nodes.

4. The method according to claim 3, characterized in that, Calculate the bond energy distribution difference of the same minimal ring structure in different chemical molecules. If the bond energy distribution difference is less than a preset deviation threshold, the corresponding chemical molecules are classified into the same molecular group, including: Extract the bond length data and electronegativity difference values ​​of all chemical bonds contained in the same minimal ring structure, and calculate the mean bond length and mean electronegativity difference of the same minimal ring structure in different chemical molecules respectively. Based on the mean bond length and the mean electronegativity difference, the bond energy distribution difference value of the same minimum ring structure among different chemical molecules is calculated. The bond energy distribution difference value is the weighted sum of the variance of the mean bond length and the variance of the mean electronegativity difference. If the bond energy distribution difference is less than a preset deviation threshold, the corresponding chemical molecules are classified into the same molecular group.

5. The method according to claim 1, characterized in that, The vibrational energy level spectrum is segmented into energy fragments according to the atomic connection paths corresponding to the substructure nodes. Distribution vectors of these energy fragments are extracted, and multi-scale feature fusion is performed between these distribution vectors and the connection path features of the substructure nodes. Simultaneously, the embedding vectors of the molecular nodes are introduced to generate a composite feature vector that integrates bond energy and connection features to identify equivalent entities and establish aligned connections. This includes: Based on the atomic connection path characteristics of the substructure nodes, the vibrational energy segments corresponding to the chemical bonds are located in the vibrational energy level spectrum, and the vibrational direction distribution and energy intensity distribution of each energy segment are extracted to generate an energy distribution vector. The energy distribution vector and the atomic connection path features of the substructure nodes are fused hierarchically according to a preset ratio. The first level fuses the chemical bond type and vibration direction features, and the second level fuses the bond energy sequence and energy intensity distribution. The embedding vector of the molecular node is spliced ​​with the multi-scale features after hierarchical fusion to generate an initial composite feature. The initial composite feature is then enhanced to retain the local features of bond vibration gradients and the global features of molecular structure to generate a composite feature vector. Based on the composite feature vector, the similarity value of different knowledge graphs is calculated. When the similarity value is greater than the preset alignment threshold, they are identified as equivalent entities, and alignment connections are established between the equivalent entities.

6. The method according to claim 1, characterized in that, Absorption peak data of a compound sample in the high-frequency recognition region is extracted using spectral imaging. This absorption peak data is then converted into bond vibrational gradients to generate a vibrational energy level spectrum, including: The absorption spectrum of the compound sample in a preset frequency band is acquired by a high-frequency spectral imaging device, and all peak positions and corresponding absorption peak intensities are extracted from the absorption spectrum to generate absorption peak data. Based on the position of each peak in the absorption peak data, a preset bond vibration frequency mapping table is matched to determine the chemical bond type and vibration mode corresponding to each absorption peak. The vibrational energy gradient of the chemical bond is calculated based on the absorption peak intensity value. The vibrational energy gradient includes the bond stretching vibration gradient and the bond bending vibration gradient. By integrating the vibrational modes and vibrational energy gradients of all chemical bond types and sorting them according to the spatial position of the chemical bonds in the molecule, a hierarchical vibrational energy level map is generated.

7. The method according to claim 1, characterized in that, Based on the atomic connection path features in the bidirectional correlation graph, substructure node combinations corresponding to molecular nodes with the same chemical name as the compound sample are retrieved, and the embedding vectors of the molecular nodes corresponding to the substructure node combinations are obtained, including: Based on the atomic connection path features in the bidirectional correlation map, molecular nodes with the same chemical name as the compound sample are matched; Starting from the matched molecular node, all substructure nodes associated with the molecular node are located through bidirectional connection relationships, a combination of substructure nodes is generated, and based on the atomic connection path features, the set of molecular nodes associated with the combination of substructure nodes in the bidirectional association map is retrieved. The embedding vector of each molecular node in the molecular node set is obtained from the preset embedding vector library. The embedding vector is generated by learning the global features of the molecular structure.

8. A sparse knowledge graph entity alignment system, characterized in that, include: The identification module is used to acquire the structural data of chemical molecules, and based on the topological connections and bond energy distribution in the structural data, traverse the interatomic connection paths in the structural data to identify the smallest ring structure. A construction module is used to cluster chemical molecules with the same minimum ring structure and a bond energy distribution deviation of less than a threshold into a molecular group, extract the shared minimum ring structure in the molecular group as a substructure node, create a molecular node for each molecule in the molecular group, establish a bidirectional connection relationship between the molecular node and the corresponding substructure node, construct a bidirectional association map, and record the atomic connection path features of the shared minimum ring structure through the substructure node. The conversion module is used to extract absorption peak data of compound samples in the high-frequency recognition region using spectral imaging, convert the absorption peak data into bond vibration gradients, and generate vibrational energy level spectra. The retrieval module is used to retrieve the substructure node combination corresponding to the molecular node with the same chemical name as the compound sample based on the atomic connection path features in the bidirectional correlation map, and to obtain the embedding vector of the molecular node corresponding to the substructure node combination. The alignment module is used to segment the vibrational energy level spectrum into energy segments according to the atomic connection paths corresponding to the substructure nodes, extract the distribution vector of the energy segments, perform multi-scale feature fusion of the distribution vector and the connection path features of the substructure nodes, and simultaneously introduce the embedding vector of the molecular nodes to generate a composite feature vector that fuses bond energy and connection features to identify equivalent entities and establish alignment connections.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a sparse knowledge graph entity alignment method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The system contains a computer program that, when executed by a computer, implements a sparse knowledge graph entity alignment method as described in any one of claims 1 to 7.

Images