Cancer molecular subtype recognition method based on self-adaptive pellet multi-view image clustering

Abstract

The invention belongs to the technical field of biological information, and particularly relates to a cancer molecular subtype recognition method based on self-adaptive pellet multi-view image clustering. The method comprises the following steps: acquiring a multi-omics data set of cancer molecules, and constructing a pellet set for each kind of omics data in the multi-omics data set; constructing a biological network structure chart according to the particle ball set; fusing the biological network structure diagrams of the omics data to obtain a unified graph; inputting the unified graph into a pre-trained heterogeneous graph neural network for processing to obtain a cancer molecular subtype recognition result; according to the method, multi-scale biological structure features in multiple omics data can be captured at the same time, collaborative optimization of molecular network topology and patient characterization is achieved, and therefore the accuracy of cancer molecular subtype recognition results is improved.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a method for identifying cancer molecular subtypes based on adaptive granulocyte multi-view graph clustering. Background Technology

[0002] Multi-view data refers to data that describes the same object or instance from multiple different perspectives or modalities, and it is widely present in various application scenarios in the real world. For example, in the Internet field, images on social media platforms are often accompanied by text tags and user descriptions, forming a visual-text multi-view representation; in the biomedical field, cancer samples from the same group of patients often contain multi-omics data such as genomic variations, transcriptomic expression, and epigenetic regulation, forming a complementary molecular characterization system.

[0003] Multi-view clustering has become a key research frontier, achieving unsupervised classification by leveraging the consistency and complementarity between views. Existing multi-view clustering methods can be divided into two major paradigms: traditional methods and deep learning methods. Traditional methods are mainly based on subspace learning, graph clustering, and multi-kernel learning, but these methods generally suffer from limited representational capabilities and high computational complexity, making them difficult to adapt to high-dimensional nonlinear data distributions in real-world scenarios.

[0004] In recent years, deep multi-view clustering has rapidly developed due to the powerful nonlinear representation capabilities of deep neural networks. Despite significant progress, deep multi-view clustering still faces a key bottleneck in topological modeling—its inability to effectively utilize the intrinsic spatial relationships between instances. To address this issue, researchers have proposed deep multi-view graph clustering methods, aiming to improve clustering performance through graph structure modeling. Although several important advances have been made in this field, existing deep multi-view graph clustering methods still have two major limitations: first, most methods ignore the inherent topological relationships in the data, making it difficult to capture local geometric details and global semantic consistency; second, reliance on fixed k-nearest neighbor graphs cannot adapt to dynamic changes in data density.

[0005] The goal of multi-omics integrative analysis in cancer is to learn a unified consensus molecular subtyping result by effectively integrating complementary biological information from heterogeneous omics data. Traditional methods typically construct a fixed molecular interaction network (such as a k-NN-based gene co-expression network) for each type of omics data, and then perform network fusion or feature learning. However, such methods face three major challenges:

[0006] 1. Limited Structural Modeling: A single network structure built on original molecular feature points cannot simultaneously capture local fine-grained molecular interactions (such as protein complex formation) and global coarse-grained pathway regulatory relationships (such as metabolic pathway synergy) in the data, resulting in insufficient utilization of biological structural information.

[0007] 2. Rigid neighborhood definition: Using a fixed neighborhood size (k value) cannot adapt to the uneven distribution of molecular feature density in different genomic regions. In high-density regions (such as mutation hotspots), important connections may be lost, while in low-density regions (such as rare variant regions), noisy connections are easily introduced, affecting the discovery of real biological associations.

[0008] 3. Separation of representation learning and structural perception: The construction of molecular networks and the learning of patient representations are often two independent stages, failing to form a closed-loop system that reinforces and optimizes each other, thus limiting the model's overall understanding of complex biological mechanisms.

[0009] In summary, there is an urgent need for a multi-view clustering method for cancer molecular subtype identification that can simultaneously capture multi-scale biological structural features from multi-omics data and achieve synergistic optimization of molecular network topology and patient representation, thereby improving the accuracy of cancer molecular subtype identification results. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention proposes a cancer molecular subtype identification method based on adaptive granular multi-view graph clustering. This method includes:

[0011] S1: Obtain a multi-omics dataset of cancer molecules and construct a set of particles for each omics data in the multi-omics dataset;

[0012] S2: Construct a biological network structure diagram based on the granule set;

[0013] S3: A unified graph is obtained by fusing biological network structure diagrams from each omics dataset;

[0014] S4: Input the unified graph into a pre-trained heterogeneous graph neural network for processing to obtain the results of cancer molecular subtype identification.

[0015] Preferably, the process of constructing a set of particles for each omics data in a multi-omics dataset includes:

[0016] S11: Initialize the molecular feature data of all patient samples under each omics dataset into a single sphere, and initialize the feature weight vector;

[0017] S12: Calculate the weighted distribution metric of the particle based on the feature weight vector; calculate the distribution metric of the sub-particle based on the weighted distribution metric of the sub-particle;

[0018] S13: Determine whether the parent particle satisfies the splitting criterion. If it does, split the parent particle to obtain multiple child particles. The splitting criterion is that the weighted distribution metric of the parent particle is greater than the distribution metric of the child particles.

[0019] S14: Repeat steps S22~S23 until all spheres no longer satisfy the splitting criterion, thus obtaining a set of spheres.

[0020] Furthermore, the formula for calculating the weighted distribution metric of the granules is as follows:

[0021]

[0022] in, This represents a weighted distribution measure of the particles. Represents the feature weight vector. This represents the characteristics of the j-th patient sample in the granulocytes under the v-th omics data. Represents the characteristic center of particle i. This represents the number of patient samples with granulocyte i under the v-th omics data;

[0023] The formula for calculating the particle distribution metric is:

[0024]

[0025] in, This represents a measure of the distribution of sub-particles. This represents a weighted distribution measure of the daughter sphere i after splitting. This represents the number of patient samples with parent granulocytes under the v-th omics data. This indicates the number of daughter cells after splitting.

[0026] Preferably, the process of constructing a biological network structure diagram includes:

[0027] S21: Calculate the local molecular density and global molecular density of each sphere;

[0028] S22: Calculate the granulocyte size factor based on the number of patient samples; calculate the number of internal connections of granulocytes based on the ratio of local molecular density to global molecular density and the granulocyte size factor;

[0029] S23: Calculate the global distance and global distance reference value of the particle based on the center distance between the particle and other particles; use the ratio of the global distance of the particle to the global distance reference value as the adaptive coefficient of the particle;

[0030] S24: Calculate the number of external connections of the spheres based on the adaptive coefficient and the sphere size factor;

[0031] S25: Construct a first-layer local graph within the granules based on the number of internal connections within the granules; construct a second-layer global graph between the granules based on the number of external connections outside the granules;

[0032] S26: Add the local map within the first layer of granules to the global map between the second layer of granules to obtain the biological network structure map.

[0033] Furthermore, the formulas for calculating local molecular density and global molecular density are as follows:

[0034]

[0035]

[0036] in, Represents local molecular density. Represents global molecular density. This represents the characteristics of the j-th patient sample in the granulocytes under the v-th omics data. Represents the characteristic center of particle i. This represents the set of patient sample features for granulocytes i under the v-th omics data. This represents the set of particles under the v-th omics data. This represents the median. This represents the Euclidean norm.

[0037] Furthermore, the formula for calculating the number of internal connections within a granule is:

[0038]

[0039]

[0040] in, This represents the number of internal connections of particle i. This represents the ratio of the local molecular density to the global molecular density of particle i. This represents the size factor of particle i. This represents the minimum number of connections within a granule. This represents the maximum number of connections within a granule. This indicates rounding down. Let represent the number of patient samples with granulocyte i under the v-th omics data.

[0041] Furthermore, the formulas for calculating the global distance of the particle and the global distance reference value are as follows:

[0042]

[0043]

[0044] in, This represents the global distance of particle i. This represents the global distance reference value for the pellet. This represents the set of particles under the v-th omics data. This represents the center-to-center distance between particle i and particle j.

[0045] Furthermore, the formula for calculating the number of external connections of a granule is:

[0046]

[0047] in, This represents the number of external connections of particle i. Represents the adaptive coefficient of particle i. This represents the size factor of particle i. This represents the minimum number of external connections of a particle. This represents the maximum number of external connections of a particle.

[0048] Furthermore, the process of constructing the local map within the first layer of granules includes:

[0049] When the number of patient samples contained in a granule is no greater than the number of connections within the granule, the patient samples are connected using a fully connected topology pattern to construct the first layer of local graphs within the granules.

[0050] When the number of patient samples contained in a granulocyte is greater than the number of internal connections within the granulocyte, each patient sample is only associated with its nearest neighbor. Connecting the samples constructs a local map within the first layer of granules. This represents the number of connections within the granule.

[0051] Furthermore, the process of constructing the second-layer global graph between particles includes:

[0052] When the number of patient samples contained in a granule is not greater than the number of external connections of the granule, a second-layer global graph between granules is established using a fully connected topology pattern among the central samples of the granules.

[0053] When the number of patient samples contained in a granulocyte is greater than the number of external connections of the granulocyte, the central sample of each granulocyte is only connected to the nearest neighbor of the organ. Connecting the samples constructs a local map within the first layer of granules. This represents the number of external connections of the granule.

[0054] The beneficial effects of this invention are as follows:

[0055] This invention clusters multi-view data into adaptive spheres with semantic-geometric consistency, and then constructs a two-layer graph structure—the first layer captures local data structure through adaptive connections within the spheres, and the second layer models the global topology through centroid connectivity between spheres, thereby efficiently encoding multi-scale structural patterns. This invention abandons the traditional fixed k-nearest neighbor rule and introduces an adaptive k-selection mechanism. This mechanism dynamically adjusts the neighborhood size based on local data density, eliminating the need for manual parameter tuning and significantly enhancing robustness to variable data distributions. This invention can simultaneously capture multi-scale biological structural features in multi-omics data and achieve synergistic optimization of molecular network topology and patient representation, thereby improving the accuracy of cancer molecular subtype identification results. Attached Figure Description

[0056] Figure 1 This is a flowchart of the cancer molecular subtype classification method based on adaptive hierarchical granular clustering in this invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] This invention proposes a method for identifying cancer molecular subtypes based on adaptive granulocyte multi-view graph clustering, such as... Figure 1 As shown, the method includes the following:

[0059] S1: Obtain a multi-omics dataset of cancer molecules and construct a set of particles for each omics data in the multi-omics dataset.

[0060] The process of constructing a set of particles for each omics dataset in a multi-omics dataset specifically includes:

[0061] S11: Initialize the molecular feature data of all patient samples under each omics dataset into a single sphere and initialize the feature weight vector.

[0062] In the v-th omics data type, we initialize the molecular feature data of all patient samples as a weighted sphere. and initialize the feature weight vector. =[ ,…, ] ⊤ Each component is initialized to (d is the feature dimension). This initialization satisfies the normalization condition. This ensures that all feature dimensions are considered in a balanced manner during the initial stage.

[0063] Particles in omics data type v Defined as:

[0064]

[0065] Its molecular feature center and distribution radius It is given by the following formula:

[0066]

[0067]

[0068] The set of spheres for omics data type v is defined as follows:

[0069]

[0070] in, This represents the number of pellets after splitting.

[0071] S12: Calculate the weighted distribution metric of the particle based on the feature weight vector; calculate the distribution metric of the sub-particle based on the weighted distribution metric of the sub-particle.

[0072] Weighted distribution measure ( By calculating the spheres The sum of the weighted distances from each patient sample to its feature center, divided by the number of samples. The distribution of biological characteristics inside the granules was obtained:

[0073]

[0074] in, This represents a weighted distribution measure of the particles. Represents the feature weight vector. This represents the characteristics of the j-th patient sample in the granulocytes under the v-th omics data. Represents the characteristic center of particle i. This represents the number of patient samples with granulocyte i under the v-th omics data.

[0075] Assuming the parent particle can split into potential daughter particles, the distribution metric for daughter particles is defined as follows:

[0076]

[0077] in, This represents a measure of the distribution of sub-particles. This represents a weighted distribution measure of the daughter sphere i after splitting. This represents the number of patient samples with parent granulocytes under the v-th omics data. This indicates the number of daughter cells after splitting.

[0078] S13: Determine whether the parent particle satisfies the splitting criterion. If it does, split the parent particle to obtain multiple child particles. The splitting criterion is that the weighted distribution metric of the parent particle is greater than the distribution metric of the child particles.

[0079] Based on a weighted distribution metric and splitting criterion, a binary strategy is used to recursively divide the parent particle sphere, generating two child particles spheres. Prototype sphere The initial partitioning uses the LW-K-means algorithm to generate k patient subgroups to establish a primary structure. Subsequent optimization employs a binary split (k′=2) to achieve adaptive local biological feature density. Only when the following conditions are met... Effective splitting is triggered at certain times; this hybrid strategy balances global structure preservation with local density adaptation.

[0080] S14: Repeat steps S22~S23 until all spheres no longer satisfy the splitting criterion, thus obtaining a set of spheres.

[0081] The top-down partitioning algorithm is executed iteratively until all particles are identified. ∈ None of them satisfy the splitting criterion. Finally, the optimized set of particles for this omics data type is obtained. = Each granule contains patient samples that are highly similar at the molecular level.

[0082] This invention adaptively partitions the input multi-omics data, generating a series of spheres with biologically consistent granularity, forming a multi-granular hierarchical biomedical representation ranging from fine to coarse. The original data points are abstracted into multiple spheres. Each sphere is represented by its centroid and radius, and is recursively binary-split using a weighted distribution metric. This significantly reduces the data size and suppresses noise interference while preserving the main data structure.

[0083] S2: Construct a biological network structure diagram based on the granule set.

[0084] To address the challenge of selecting the optimal neighborhood size k in biological network construction, this invention proposes a novel molecular density-aware adaptive mechanism. This mechanism optimizes network connections both within (local) and between (global) molecular feature units, abandoning the traditional fixed k-value strategy. Specifically:

[0085] S21: Calculate the local molecular density and global molecular density of each sphere.

[0086] First, the local molecular density characteristics are calculated for each particle. The formulas for calculating the local molecular density and global molecular density of each particle are as follows:

[0087]

[0088]

[0089] in, Represents local molecular density. Represents global molecular density. This represents the characteristics of the j-th patient sample in the granulocytes under the v-th omics data. Represents the characteristic center of particle i. This represents the set of patient sample features for granulocytes i under the v-th omics data. This represents the set of particles under the v-th omics data. This represents the median. This represents the Euclidean norm.

[0090] S22: Calculate the granulocyte size factor based on the number of patient samples; calculate the number of internal connections of granulocytes based on the ratio of local molecular density to global molecular density and the granulocyte size factor.

[0091] Calculate the ratio of local molecular density to global molecular density. High-density areas ) typically contain more complex local structures, requiring a larger number of internal connections, while low-density regions ( More sparse connections are needed to prevent noise propagation. The number of connections within each sphere is dynamically determined by incorporating the sphere size factor, using the following formula:

[0092]

[0093]

[0094] in, This represents the number of internal connections of particle i. This represents the ratio of the local molecular density to the global molecular density of particle i. This represents the size factor of particle i. This represents the minimum number of connections within a granule. This represents the maximum number of connections within a granule. This indicates rounding down to the nearest integer.

[0095] S23: Calculate the global distance and global distance reference value of the particle based on the center distance between the particle and other particles; use the ratio of the global distance of the particle to the global distance reference value as the adaptive coefficient of the particle.

[0096] Intergranular distances are calculated based on relative biological distances and spatial distribution characteristics. The number of connections for each particle at the global level is dynamically determined by calculating the relative distances between particle centers.

[0097]

[0098] Calculate global distance reference value :

[0099]

[0100] Calculate global distance :

[0101]

[0102] Based on this, define the distance adaptive coefficient. :

[0103]

[0104] S24: Calculate the number of external connections of the sphere based on the adaptive coefficient and the sphere size factor.

[0105] when When the region is relatively sparse, it indicates that the connections between particles need to be strengthened to maintain the connectivity of the global structure; when This indicates that the region is relatively dense, therefore it is necessary to reduce the connections between particles to avoid redundant connections. The final parameters incorporate the scale factor. (and (Keep the calculation method consistent)

[0106]

[0107] in, This represents the number of external connections of particle i. Represents the adaptive coefficient of particle i. This represents the minimum number of external connections of a particle. This represents the maximum number of external connections of a particle.

[0108] This adaptive mechanism dynamically assigns an appropriate neighborhood size to each granule by separately quantifying local molecular density and global relative biological distance, thereby enabling the constructed biological network structure to more realistically and robustly reflect the biological topological relationships of molecular data.

[0109] S25: Construct a first-layer local graph within the granules based on the number of internal connections within the granules; construct a second-layer global graph between the granules based on the number of external connections within the granules.

[0110] This invention proposes a highly efficient two-layer biological network construction module. By integrating the local interaction network within mitochondria and the global regulatory network between mitochondria, it comprehensively captures structural information from local molecular interactions to global pathway regulation in multi-omics data, laying a solid foundation for subsequent molecular typing tasks. Its core idea is that larger neighborhoods are needed in high-molecular-density regions (such as mutation hotspots) to characterize complex biological networks, while sparser connections are needed in low-molecular-density regions (such as rare mutation regions) to prevent noise interference. Specifically:

[0111] Partial view of the first layer of granules Construction: For each sphere Based on the number of patient samples it contains With adaptive parameters The relationship is analyzed using a differentiated connection strategy: when the sphere contains fewer samples ( When the granulocyte contains a large number of samples, a fully connected topology mode is adopted to preserve the molecular relationships between all patient sample pairs within the granulocyte, ensuring the local structural integrity of the sparse genomic region; when the granulocyte contains a large number of samples ( When using -Nearest neighbor connection mode, each patient sample is only associated with its nearest neighbor. The samples are connected, and the computational complexity is controlled while maintaining the main local molecular interaction patterns.

[0112] Second layer global map of intergranular space Construction: Based on the set of sphere centers, according to the total number of spheres of the current omics data type and adaptive parameters The relationship is to implement a hierarchical connection strategy: when the total number of particles is small ( When the total number of granulocytes is large, a fully connected network at the granulocyte center is established to ensure the complete preservation of global biological pathway relationships; when the total number of granulocytes is large ( When using -Nearest neighbor connectivity pattern: each granule center connects only with its closest neighbor in its biological space. The centers are interconnected to construct a sparse but representative global control network framework.

[0113] S26: Add the local map within the first layer of granules to the global map between the second layer of granules to obtain the biological network structure map.

[0114] Summing the local graph within the first-layer granulocytes with the global graph matrix between the second-layer granulocytes yields the final biological network structure representation of this omics data type:

[0115]

[0116] This fusion network simultaneously encodes local molecular interaction features and global pathway regulatory relationships, forming a multi-scale network representation from microscopic molecular mechanisms to macroscopic biological processes, providing a solid biological foundation for subsequent graph neural network learning.

[0117] S3: The biological network structure diagram of each omics data is integrated to obtain a unified diagram.

[0118] The biological network structure diagram integrates data from each omics group. Specifically, it establishes connections between each sample point in each view, links identical sample points to each other, and connects all identical sample points to form a unified diagram.

[0119] S4: Input the unified graph into a pre-trained heterogeneous graph neural network for processing to obtain the results of cancer molecular subtype identification.

[0120] A unified graph is input into a pre-trained heterogeneous graph neural network (HBR), which performs message propagation and aggregation across omics data types. The HBR is based on learned consensus biological representations. A soft allocation matrix is ​​generated through clustering layers. First, calculate the soft-assignment distribution. and target distribution :

[0121] ,

[0122] in, Indicates patient sample Belongs to molecular subtype The probability of is given by the elements in the allocation matrix P. Indicates the number of molecular subtypes. Based on This sets a "high benchmark" for the optimization process. It tells the model that it should be even more confident in patients for whom it already has some knowledge, and that it can ensure a reasonable proportion of patients in each subtype.

[0123] During the training of heterogeneous graph neural networks, through Divergence loss function drives soft assignment results To an optimized target distribution Alignment is used to achieve joint optimization of model parameters and molecular subtype centers:

[0124]

[0125] After optimization, based on the soft allocation matrix Each patient sample is assigned a molecular subtype label, and the final clustering result, i.e., the cancer molecular subtype identification result, is directly output.

[0126] In summary, the structural modeling paradigm of this invention shifts from the existing "single-scale" to a "multi-granularity collaborative" approach: abandoning the traditional method of directly constructing a single biological network on raw molecular feature data, it pioneers a modeling path of "multi-omics data → granules → two-layer biological network". By introducing the intermediate abstraction of "granules", the biological consistency of molecular data (like a functional module) and the spatial proximity characteristics of molecular features are encapsulated in different granularity levels, thereby naturally and explicitly capturing multi-scale biological structural features from microscopic molecular interactions to macroscopic pathway regulation. Each granule represents a highly consistent patient subgroup at the molecular level, providing a reliable biological basis for subsequent accurate subtyping. This invention innovates the neighborhood definition method: from "fixed and rigid" to "density adaptive": freeing itself from dependence on a preset k value, it invents a neighborhood size adaptive selection mechanism based on molecular density. This mechanism dynamically senses the density variations of molecular feature distribution across different genomic regions, allocating appropriate neighborhood ranges to regions with varying molecular densities (such as mutation hotspots and rare variant regions). This constructs a biological network structure that better reflects the true biological topology, fundamentally improving the model's robustness and generalization ability in complex multi-omics data. This adaptive mechanism ensures the full capture of complex molecular interaction networks in high-density regions while effectively avoiding noise interference in low-density regions. Through these improvements, this invention can simultaneously capture multi-scale biological structural features in multi-omics data and achieve synergistic optimization of molecular network topology and patient characterization, thereby improving the accuracy of cancer molecular subtype identification results.

[0127] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for cancer molecular subtype identification based on adaptive granular ball multi-view graph clustering, characterized in that, The method comprises the following steps: S1: obtaining a multi-omics dataset of cancer molecules, and constructing a granule set for each omics data in the multi-omics dataset; S2: constructing a biological network structure diagram according to the granule set; S3: fusing the biological network structure diagrams of each omics data to obtain a unified diagram; S4: inputting the unified diagram into a pre-trained heterogeneous graph neural network for processing to obtain a cancer molecule subtype identification result.

2. The method according to claim 1, wherein, The process of constructing a granule set for each omics data in the multi-omics dataset comprises the following steps: S11: initializing the molecular feature data of all patient samples under each omics data as a granule, and initializing a feature weight vector; S12: calculating a weighted distribution metric of the granule according to the feature weight vector; and calculating a sub-granule distribution metric according to the weighted distribution metric of the sub-granule; S13: determining whether the parent granule meets a splitting criterion, if yes, splitting the parent granule to obtain a plurality of sub-granules; wherein the splitting criterion is that the weighted distribution metric of the parent granule is greater than the sub-granule distribution metric; S14: repeating steps S22-S23 until all granules do not meet the splitting criterion, to obtain the granule set.

3. The method of claim 2, wherein the method is based on adaptive granular ball multi-view graph clustering for cancer molecular subtype identification. The formula for calculating the weighted distribution metric of the granule is: ； wherein, denotes a weighted distribution metric of granulocytes, denotes a feature weight vector, denotes a feature of the jth patient sample in the granulocytes under the vth omics data, denotes a feature center of the granulocytes i, denotes the number of patient samples of the granulocytes i under the vth omics data; The formula for calculating the sub-granule distribution metric is: ； wherein, represents a sub-particle distribution metric, represents a weighted distribution metric of the sub-particles i after splitting, represents the number of patient samples for the parent-particle under the vth omics data, represents the number of sub-particles after splitting.

4. The method of claim 1, wherein, The process of constructing a biological network structure diagram comprises the following steps: S21: calculating a local molecular density and a global molecular density of each granule; S22: calculating a granule size factor according to the number of patient samples of the granule; and calculating a granule internal connection number according to the ratio of the local molecular density to the global molecular density and the granule size factor; S23: calculating a global distance of the granule and a global distance reference value according to the center distance between the granule and other granules; and taking the ratio of the global distance of the granule to the global distance reference value as an adaptive coefficient of the granule; S24: calculating a granule external connection number according to the adaptive coefficient and the granule size factor; S25: constructing a first layer granule intra-local graph according to the granule internal connection number; and constructing a second layer granule inter-global graph according to the granule external connection number; S26: adding the first layer granule intra-local graph and the second layer granule inter-global graph to obtain the biological network structure diagram.

5. The method of claim 4, wherein, The formulas for calculating the local molecular density and the global molecular density are respectively: ； ； wherein, denotes the local molecular density, denotes the global molecular density, denotes the feature of the jth patient sample in the granulocyte under the vth omic data, denotes the feature center of the granulocyte i, denotes the set of patient sample features of the granulocyte i under the vth omic data, denotes the set of granulocytes under the vth omic data, denotes the median, denotes the Euclidean norm.

6. The method of claim 4, wherein the method is based on adaptive granular ball multi-view graph clustering for cancer molecular subtype identification. The formula for calculating the granule internal connection number is: ； ； wherein, represents the internal connectivity number of granulocyte i, represents the ratio of local molecular density to global molecular density of granulocyte i, represents the size factor of granulocyte i, represents the minimum value of the internal connectivity number of granulocyte, represents the maximum value of the internal connectivity number of granulocyte, represents the floor function, represents the number of patient samples of granulocyte i under the vth omics data.

7. The method of claim 4, wherein the method is based on adaptive granular ball multi-view graph clustering for cancer molecular subtype identification. The formulas for calculating the global distance of the granule and the global distance reference value are respectively: ； ； wherein, denotes the global distance of granule i, denotes the global distance reference value of granules, denotes the set of granules under the vth omic data, denotes the center distance between granule i and granule j.

8. The method of claim 4, wherein the method is based on adaptive granular ball multi-view graph clustering for cancer molecular subtype identification. The formula for calculating the granule external connection number is: ； wherein, represents the number of external connections of the granule i, represents the adaptation coefficient of the granule i, represents the scale factor of the granule i, represents the minimum value of the number of external connections of the granule, represents the maximum value of the number of external connections of the granule.

9. The method of claim 4, wherein, The process of constructing the first layer granule intra-local graph comprises the following steps: When the number of patient samples contained in the granule is not greater than the granule internal connection number, a full connection topology mode is adopted to construct the first layer granule intra-local graph between the patient samples; When the number of granules contained in the patient sample is greater than the number of internal connections of the granules, each patient sample is only connected to its nearest sample to construct the first layer of the local graph inside the granules, is the number of internal connections of the granules.

10. The method of claim 4, wherein, The process of constructing the second layer granule inter-global graph comprises the following steps: When the number of patient samples contained in the granule is not greater than the granule external connection number, a full connection topology mode is adopted to construct the second layer granule inter-global graph between the center samples of the granule; When the number of samples contained by the granule is greater than the number of connections outside the granule, the center sample of each granule is only connected to the nearest sample of the first layer of the granule to construct the first layer of the local graph inside the granule, is the number of connections outside the granule.

Images