Intelligent classification and automatic interpretation method based on gene set enrichment entries

Abstract

This invention provides an intelligent classification and automatic interpretation method based on gene set enrichment entries, comprising: inputting gene set enrichment results and constructing a gene-entry binary matrix accordingly; calculating IDF-weighted cosine similarity to obtain a similarity matrix S0; generating a similarity matrix S1 through Siamese network, TF-IDF, or a large language embedding model; fusing S0 and S1 with certain weights to obtain the final similarity matrix; converting the similarity matrix into a distance matrix, constructing an affinity matrix through KNN sparsification and Gaussian kernel weighting; performing spectral clustering on the affinity matrix, and automatically determining the number of clusters based on silhouette coefficient and Bootstrap stability; using a large language model to perform structured interpretation of the clusters and outputting functional grouping results; finally, calculating data separation degree and semantic separation degree to check trend consistency. This invention achieves automated classification and standardized interpretation of gene enrichment entries, overcoming the problems of information loss, strong subjectivity, and difficulty in batch processing of traditional methods.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics analysis technology, and in particular to an intelligent classification and automatic interpretation method based on gene set enrichment entries. Background Technology

[0002] With the widespread application of high-throughput omics technologies (such as transcriptomics and proteomics), gene function enrichment analysis has become a core tool for interpreting the biological significance of gene lists. This analysis identifies significantly enriched biological functional entries (such as GO terms, KEGG pathways, etc.) by comparing the intersection of the target gene set with the known functional gene set, thereby helping researchers understand the biological processes behind the data. Traditional enrichment analyses often output a large number of entries, with significant functional redundancy and semantic overlap, posing a significant challenge to the interpretation of the results. To address this, gene enrichment entry clustering technology has emerged. This technology aims to group functionally similar entries into several meaningful categories, thereby simplifying the result structure, improving interpretability, and assisting researchers in identifying core biological themes.

[0003] Currently, gene enrichment item clustering mainly falls into four technical categories: First, semantic similarity clustering, with representative methods including rrvgo (using GO semantic similarity for clustering), simplifiedEnrichment (using binary cut algorithm and semantic similarity clustering), and REVIGO (removing redundant enrichment items based on semantic similarity); second, gene overlap clustering, with representative methods including GeneSetCluster (based on gene overlap), GOMCL (using Markov clustering (MCL) algorithm to handle gene overlap), and EnrichmentMap (a network visualization method combining gene overlap and semantic similarity); third, hybrid methods, with representative methods including ClueGO (iterative clustering based on kappa statistics) and GSCluster (network weighted gene set clustering analysis combining gene set overlap and protein-protein interaction networks); and fourth, simplified methods, with representative methods including manually selected GO subsets (GO slims) and topGO (selecting representative enrichment items based on saliency). However, existing technical approaches have the following drawbacks: (1) Redundancy removal leads to information loss: REVIGO filters terms through semantic similarity thresholds, resulting in the loss of specific functional information; GO slims uses predefined simplified sets, which mask hierarchical details, leading to information loss. Oversimplification can distort biological interpretations, especially in disease mechanism research. (2) Clustering algorithms have limitations: hierarchical clustering cannot capture nonlinear relationships and is sensitive to initial distance; Markov clustering has complex parameter adjustment and unstable results; (3) Limitations of similarity measurement methods: the kappa statistic is sensitive to the size of the gene set; the calculation is complicated; the Jaccard coefficient cannot be used to measure the non-matching similarity of most gene set enrichment entries, and cannot be effectively classified. (4) Insufficient automation: Most tools require manual specification of the number of clusters; (5) There are platform dependencies, the operating environment is complex and it is not easy to deploy: for example, rrvgo depends on Bioconductor, GOMCL requires Python+MCL environment, and EnrichmentMap depends on Cytoscape; (6) Lack of efficient automatic interpretation of biological meaning: Existing methods mostly rely on manual summarization of the biological meaning of clustering results, which is highly subjective and difficult to process in batches, affecting the consistency and traceability of results. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent classification and automatic interpretation method based on gene set enrichment entries, to overcome the information loss problem caused by oversimplification in existing technologies; to solve the technical defects of insufficient capture of nonlinear similarity relationships; to achieve intelligent determination of cluster numbers and reduce subjective intervention; to provide a lightweight cross-platform solution and reduce deployment complexity; and to achieve automated and standardized biological meaning induction and interpretation based on large language models, thereby improving analysis efficiency and result consistency.

[0005] To achieve the above-mentioned objectives, the technical solution provided by this invention is as follows: A method for intelligent classification and automatic interpretation of gene set enrichment entries, the method comprising: S101. Input the gene set enrichment results, which include several gene enrichment entries; S102. Construct a gene-entry binary matrix using gene enrichment entries; S103. Calculate the similarity matrix based on the gene-entry binary matrix; S104. Construct an affinity matrix based on the similarity matrix; S105. Perform spectral clustering on the affinity matrix and output the cluster label to which each gene enrichment entry belongs; S106. Use a large language model to perform a structured interpretation of each cluster, and output structured gene enrichment item functional grouping results, AI interpretation results, and semantic-data trend consistency check results.

[0006] Furthermore, the gene set enrichment entries include an entry identifier, a list of genes associated with the entry, and a description of the entry's function.

[0007] Furthermore, a gene-entry binary matrix is ​​constructed using gene enrichment entries, specifically including the following operations: S201. Represent each gene enrichment entry as a binary vector with a length equal to the union of all genes. For genes that exist in the gene enrichment entry, set their corresponding position in the vector to 1; otherwise, set it to 0. S202. Form a gene-entry binary matrix from the binary vectors corresponding to all gene enrichment entries. Each row of the gene-entry binary matrix corresponds to a binary vector, and each column corresponds to a gene.

[0008] Furthermore, step S103 specifically includes the following operations: S301. Calculate the inverse document frequency for each column of the gene-entry binary matrix X to obtain the IDF weight of each column; S302. Multiply the gene-entry binary matrix X column by the corresponding IDF weights to obtain X. w ; S303, Regarding X w Calculate the cosine similarity to obtain the similarity matrix S0.

[0009] Furthermore, in step S103, the similarity matrix is ​​calculated by enabling the Siamese network, specifically including the following operations: S401. Determine whether a similarity matrix S1 is provided externally. If yes, proceed to step S403; otherwise, proceed to step S402. S402. Determine whether the Siamese network weights and metadata are provided. If yes, input the Siamese network weights and metadata into the Siamese network for inference and generate the similarity matrix S1. If no, train the Siamese network first and then generate the similarity matrix S1 through the trained Siamese network. S403. Perform a weighted summation of the similarity matrices S0 and S1 to obtain the final similarity matrix S.

[0010] Furthermore, in step S103, if the twin network is not enabled, perform the following operations: S501. Based on the item function description text, construct a semantic similarity matrix S using TF-IDF or a large language embedding model. sem ; S502, regarding the similarity matrices S0 and S sem Perform a weighted summation to obtain the final similarity matrix S.

[0011] Furthermore, step S104 specifically includes the following operations: S601. Convert the similarity matrix into a distance matrix D; S602. For each gene enrichment entry in the distance matrix D, find the n closest samples, keep only the edges in the neighborhood of the n samples, and set the rest of the edges to 0 to achieve KNN sparsity. S603. For each retained neighborhood edge, calculate the Gaussian kernel weight using the local scale parameter to obtain the original affinity matrix; S604. Symmetricize the original affinity matrix by setting the diagonal to 1 and output the final affinity matrix.

[0012] Furthermore, in step S105, when performing spectral clustering, the number of clusters can be manually specified by the user or automatically determined. Automatic determination specifically includes the following operations: S701. Set the search range for the cluster number k to obtain multiple candidate k values; S702. Calculate the silhouette coefficient and Bootstrap stability for each candidate k value; S703. Perform a fusion score on each candidate k value based on the contour coefficient and Bootstrap stability; S704. Based on the fusion score results of each candidate k value, determine the optimal K value and output it.

[0013] Furthermore, step S106 specifically includes the following operations: S801. Calculate the data separation degree and semantic separation degree for each cluster. S802. Determine whether the data separation degree and semantic separation degree are greater than 0 respectively. If they are, the check is considered to pass; otherwise, the check is considered to fail. S803, Output semantic-data trend consistency check results.

[0014] Furthermore, in step S106, when using a large language model to perform structured interpretation of each cluster, it supports multi-round optimization and dialogue memory management based on user feedback.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Compared with existing technologies that use Jaccard coefficients or kappa statistics to measure the similarity between enriched items, this invention uses cosine similarity to construct a similarity matrix between items, which has low computational complexity and can avoid bias in gene set size. 2. This invention can determine the optimal number of clusters by combining the silhouette coefficient and Bootstrap stability without pre-setting the number of clusters, thus achieving automated setting; 3. This invention uses a spectral clustering algorithm to group gene set enrichment items based on the similarity matrix. The spectral clustering algorithm is more suitable for processing complex data such as non-convex data and non-linear structures, and is often superior to traditional clustering algorithms. 4. This invention automatically summarizes clustering results through a large language model, which greatly improves analysis efficiency, result consistency and traceability, and significantly reduces human subjectivity and batch processing difficulty; 5. This invention introduces a semantic-data trend consistency check mechanism, which objectively evaluates the clustering quality by comparing the separation degree of clustering results in the gene similarity space and semantic embedding space, providing users with quantifiable clustering reliability indicators and making up for the lack of internal quality assessment in existing methods. 6. This invention supports multiple rounds of optimization of AI interpretation results based on user feedback, and utilizes a dialogue memory management mechanism to retain historical interaction context, thereby improving the flexibility of human-machine collaboration and the efficiency of result iteration optimization. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the overall process of an intelligent classification and automatic interpretation method based on gene set enrichment entries provided in an embodiment of the present invention. Detailed Implementation

[0018] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0019] Reference Figure 1 This embodiment provides an intelligent classification and automatic interpretation method based on gene set enrichment entries. The method includes the following steps: S101. Input the gene set enrichment results, which include several gene enrichment entries.

[0020] S102. Construct a gene-entry binary matrix using gene enrichment entries.

[0021] S103. Calculate the similarity matrix based on the gene-entry binary matrix.

[0022] S104. Construct an affinity matrix based on the similarity matrix.

[0023] S105. Perform spectral clustering on the affinity matrix and output the cluster label to which each gene enrichment entry belongs.

[0024] S106. Use a large language model to perform a structured interpretation of each cluster, and output structured gene enrichment item functional grouping results, AI interpretation results, and semantic-data trend consistency check results.

[0025] In step S101, the input gene set enrichment results can be in .csv, .txt, or .xlsx format. The input gene set enrichment result file must simultaneously contain three columns: the enrichment entry identifier (Term), the list of genes associated with the entry (Symbols), and the entry functional description (Description), to ensure consistency in subsequent semantic fusion and AI interpretation. The enrichment entry identifier can be a GO ID, pathway ID, etc.

[0026] In a preferred embodiment, step S102, which involves constructing a gene-entry binary matrix using gene enrichment entries, specifically includes the following operations: S201. Represent each gene enrichment entry as a binary vector with a length equal to the union of all genes. For genes present in the gene enrichment entry, set their corresponding position in the vector to 1; otherwise, set it to 0.

[0027] S202. Form a gene-entry binary matrix from the binary vectors corresponding to all gene enrichment entries. Each row of the gene-entry binary matrix corresponds to a binary vector, and each column corresponds to a gene.

[0028] In this implementation, for each gene enrichment entry, a vector of length N can be constructed, where N is the total number of unique genes appearing in all entries. If a gene belongs to an entry, the corresponding position in the binary vector is set to 1; otherwise, it is set to 0. The binary vectors of all entries are stacked row-wise to form an m×n gene-entry binary matrix, where m is the number of entries, n is the total number of unique genes, and the elements of the gene-entry binary matrix are... This indicates whether the i-th gene enrichment entry contains the j-th gene.

[0029] For example, assuming there are 3 gene enrichment entries and 5 unique genes, the corresponding gene-entry binary matrix can be represented as:

[0030] Based on this, step 103 specifically includes the following operations: S301. Calculate the inverse document frequency for each column of the gene-entry binary matrix X to obtain the IDF weight for each column. The IDF weight can reduce the interference of frequently occurring high-frequency genes on similarity. The expression for the IDF weight is: IDF(j) = log((m+1) / (df(j)+1)) + 1 Where m is the total number of entries, and df is a statistical matrix showing how many entries each gene appears in. df(j) represents the number of entries containing gene j.

[0031] S302. Multiply the gene-entry binary matrix X column by the corresponding IDF weights to obtain X. w .

[0032] S303, Regarding X w Calculate the cosine similarity to obtain the similarity matrix S0.

[0033] Cosine similarity measures the proximity of two vectors in a direction, independent of their length, making it particularly suitable for gene sets with significant differences in size. Compared to Jaccard coefficients and kappa statistics, cosine similarity can be directly calculated using vector magnitudes, making it suitable for large-scale matrix operations with low computational complexity and high efficiency. It is also more suitable for high-dimensional sparse vectors and can be directly used for subsequent clustering analysis without additional transformations. This method improves the efficiency and mathematical applicability of similarity calculation while preserving gene information.

[0034] As a further optional implementation, step S103 may also enable the Siamese network to calculate the similarity matrix, specifically including the following operations: S401. Determine whether a similarity matrix S1 is provided externally. If yes, proceed to step S403; otherwise, proceed to step S402.

[0035] S402. Determine whether the Siamese network weights and metadata are provided. If yes, input the Siamese network weights and metadata into the Siamese network for inference and generate the similarity matrix S1. If no, train the Siamese network first, and then generate the similarity matrix S1 through the trained Siamese network.

[0036] In this implementation, the metadata includes a gene dictionary, input dimensions, embedding dimensions, etc. If the user does not provide a similarity matrix S1 from an external source, nor does the user provide pre-trained Siamese network weights and metadata, the system will automatically execute the following training process: Automatic labeling: Based on the functional description text of each gene enrichment entry in the input file, the cosine similarity between each pair is calculated after TF-IDF vectorization, and a similarity label matrix Y is generated as a supervision signal for Siamese network training.

[0037] Sample pair sampling: From the label matrix Y, select item pairs with a similarity of not less than the first similarity threshold as positive samples, and item pairs with a similarity of not more than the second similarity threshold as negative samples, thus forming the training dataset.

[0038] Network training: The Siamese network is trained until convergence by taking the binary gene vectors of any two entries in the training dataset as input, the similarity value in Y as the target, and the mean squared error (MSE) as the loss function.

[0039] Model saving: After training is complete, save the network weight parameters and metadata to a local file.

[0040] In this implementation, if the Siamese network weights and metadata are not provided, the Siamese network needs to be trained. During training, the same sample gene set enrichment result file as in step S101 is first input, and the functional description text of the entries is vectorized using TF-IDF. The semantic similarity (cosine) between entry pairs is calculated, and the calculated semantic similarity is used as the supervision signal for the Siamese network. The input to the Siamese network is a binary gene vector, which is mapped to the embedding space using MLP+Dropout. The optimizer for the Siamese network is Adam.

[0041] During the training phase, to enable the Siamese network to learn embedding representations consistent with text semantics from binary genes, positive and negative sample pairs need to be constructed. Positive sample pairs consist of two entries that are highly similar in text semantics; negative sample pairs consist of two entries that are very dissimilar in text semantics. For example, the first similarity threshold for two entries in a positive sample pair can be set to no less than 0.65; the second similarity threshold for two entries in a negative sample pair can be set to no more than 0.20. Understandably, both the first and second similarity thresholds can be adjusted according to data characteristics. The output of the Siamese network during the training phase consists of Siamese network weights and metadata.

[0042] During the inference phase, the Siamese network uses the gene dictionary in the metadata to align the input gene enrichment entry vectors, then computes the embedding vectors forward, calculates the embedding cosine similarity to construct the similarity matrix S1, and outputs it. The advantage of the above automatic training mode is that it does not require users to prepare additional labeled data or pre-trained models. Users only need to provide the same gene enrichment result file as in step S101 to automatically complete model training and generate the S1 matrix, greatly reducing the user's learning curve.

[0043] S403. Perform a weighted summation of the similarity matrices S0 and S1 to obtain the final similarity matrix S.

[0044] For example, the final similarity matrix S can be calculated as follows: S = α × S0 + (1 - α) × S1 Where α = 0.7, the default value of α can be adjusted by the parameter.

[0045] As another possible implementation, if twin networks are not enabled, the following operations are performed: S501. Based on the item function description text, construct a semantic similarity matrix S using TF-IDF or a large language embedding model. sem .

[0046] In this implementation, users can choose to use TF-IDF (construct TF-IDF vectors based on the item function description text and calculate cosine similarity) or embedding (batch obtain the embedding vectors of the item function description text through the language large model embedding API and then calculate cosine similarity).

[0047] S502, regarding the similarity matrices S0 and S sem A weighted summation is performed to obtain the final similarity matrix S. In this implementation, the final similarity matrix S is expressed as: S = (1-β)×S0 + β×S sem Where β = 0.3. The default value of β can be adjusted via parameters.

[0048] In this implementation, if the user does not enable the twin network and does not specify the semantic fusion method, the final similarity matrix S is equal to S0.

[0049] Based on the foregoing embodiments, as a further preferred implementation, step S104 specifically includes the following operations: S601. Convert the similarity matrix into a distance matrix D.

[0050] In this embodiment, the elements D of the distance matrix D ij =1-S ij S ij Let be the j-th element in the i-th row of the similarity matrix S.

[0051] S602. For each gene enrichment entry in the distance matrix D, find the n closest samples, keep only the edges within the neighborhood of the n samples, and set the rest of the edges to 0, thus achieving KNN sparsity. n = max(2, min(10, N - 1)), where N is the total number of entries.

[0052] S603. For each retained neighborhood edge, calculate the Gaussian kernel weight using the local scale parameter to obtain the original affinity matrix.

[0053] S604. Symmetricize the original affinity matrix by setting the diagonal to 1, and output the final affinity matrix. Symmetricization ensures the symmetry required for subsequent spectral clustering. Setting the diagonal to 1 maximizes self-affinity and avoids zero rows in the matrix during subsequent calculations.

[0054] In another preferred embodiment, in step S105, when performing spectral clustering, the number of clusters is either manually specified by the user or automatically determined. Automatic determination specifically includes the following operations: S701. Set the search range for the cluster number k to obtain multiple candidate k values.

[0055] For example, the search range of k can be set to [2, N-1], where N is the total number of gene enrichment entries.

[0056] S702. Calculate the silhouette coefficient and Bootstrap stability for each candidate k value.

[0057] The silhouette coefficient measures whether samples within the same cluster are sufficiently compact, and whether samples between different clusters are sufficiently separated. It is an internal evaluation metric, meaning it does not require real class labels and evaluates the quality of the clustering result solely based on its structure. Its value ranges from [-1, 1]. A value close to 1 indicates reasonable clustering with high cohesion and good separation; a value close to 0 indicates that samples are located on the boundary between two clusters; and a value close to -1 indicates that samples may have been assigned to incorrect clusters. Bootstrap stability measures the robustness of clustering results to data perturbations.

[0058] S703. Perform a fusion score on each candidate k value based on the contour coefficient and Bootstrap stability.

[0059] S704. Based on the fusion score results of each candidate k value, determine the optimal K value and output it.

[0060] In this embodiment, after spectral clustering is completed, Fisher's exact test is performed for verification: for each cluster, the significance of overlap between genes within the cluster and genes in individual entries is calculated, and cluster-level and entry-level statistical results are output. Cluster-level statistical results include the number of genes, minimum FDR, and number of significant entries; entry-level statistical results include the number of overlapping genes, p-value, and FDR. Fisher's exact test can determine whether the gene set within a cluster significantly enriches a specific biological function. If entries within a cluster are functionally related, their gene overlap should be significantly higher than expected randomly. The smaller the p-value of the Fisher test, the more likely the gene overlap within the cluster is not accidental, and the clustering is biologically reasonable. The minimum FDR and number of significant entries in the statistical results can help users determine which clusters are truly meaningful, filtering out noisy clusters with insignificant p-values ​​or too few genes, and prioritizing the interpretation of highly significant clusters.

[0061] In addition, this stage allows visualization of the preprocessed results of the input and output data, including cosine similarity matrix heatmap, affinity matrix heatmap, and PCA clustering map.

[0062] In yet another preferred embodiment, step S106 specifically includes the following operations: S801. For each cluster c, calculate the data separation degree SepS(c) and the semantic separation degree SepE(c).

[0063] In this implementation, the formula for calculating the data separation degree SepS(c) is as follows: SepS(c) = mean_{i,j c, i≠j}(S_ij) - mean_{i c, j c}(S_ij) Where, mean_{i,j c, i≠j}(S_ij) represents the average value of all distinct pairs (i,j) within cluster c in the similarity matrix S_ij; mean_{i c, j c}(S_ij) represents the average value of the corresponding values ​​of the entries within cluster c and all entries outside cluster c in the similarity matrix S_ij.

[0064] The semantic separation degree SepE(c) is calculated as follows: SepE(c) = mean_{i,j c, i≠j}(E_ij) - mean_{i c, j c}(E_ij) Where, mean_{i,j c, i≠j}(E_ij) represents the average value of the corresponding values ​​of all distinct entry pairs (i,j) within cluster c in the semantic similarity matrix E_ij; mean_{i c, j c}(E_ij) represents the average value of the corresponding values ​​of the entries within cluster c and all entries outside the cluster in the semantic similarity matrix E_ij.

[0065] S802. Determine whether the data separation degree and semantic separation degree are greater than 0 respectively. If they are, the check is considered to have passed; otherwise, the check is considered to have failed.

[0066] S803, Output semantic-data trend consistency check results.

[0067] As another preferred implementation, in step S106, when using a large language model to perform structured interpretation of each cluster, it supports multi-round optimization and dialogue memory management based on user feedback.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made 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 intelligent classification and automatic interpretation of gene set enrichment entries, characterized in that, The method includes: S101. Input the gene set enrichment results, which include several gene enrichment entries; S102. Construct a gene-entry binary matrix using gene enrichment entries; S103. Calculate the similarity matrix based on the gene-entry binary matrix; S104. Construct an affinity matrix based on the similarity matrix; S105. Perform spectral clustering on the affinity matrix and output the cluster label to which each gene enrichment entry belongs; S106. Use a large language model to perform a structured interpretation of each cluster, and output structured gene enrichment item functional grouping results, AI interpretation results, and semantic-data trend consistency check results.

2. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 1, characterized in that, The gene set enrichment entries include an entry identifier, a list of genes associated with the entry, and a description of the entry's function.

3. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 2, characterized in that, Constructing a gene-entry binary matrix from gene enrichment entries involves the following operations: S201. Represent each gene enrichment entry as a binary vector with a length equal to the union of all genes. For genes that exist in the gene enrichment entry, set their corresponding position in the vector to 1; otherwise, set it to 0. S202. Form a gene-entry binary matrix from the binary vectors corresponding to all gene enrichment entries. Each row of the gene-entry binary matrix corresponds to a binary vector, and each column corresponds to a gene.

4. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 3, characterized in that, Step S103 specifically includes the following operations: S301. Calculate the inverse document frequency for each column of the gene-entry binary matrix X to obtain the IDF weight of each column; S302. Multiply the gene-entry binary matrix X column by the corresponding IDF weights to obtain X. w ; S303, Regarding X w Calculate the cosine similarity to obtain the similarity matrix S0.

5. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 4, characterized in that, In step S103, the similarity matrix is ​​also calculated by enabling the Siamese network, specifically including the following operations: S401. Determine whether a similarity matrix S1 is provided externally. If yes, proceed to step S403; otherwise, proceed to step S402. S402. Determine whether the Siamese network weights and metadata are provided. If yes, input the Siamese network weights and metadata into the Siamese network for inference and generate the similarity matrix S1. If no, train the Siamese network first and then generate the similarity matrix S1 through the trained Siamese network. S403. Perform a weighted summation of the similarity matrices S0 and S1 to obtain the final similarity matrix S.

6. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 4, characterized in that, In step S103, if the twin network is not enabled, perform the following operations: S501. Based on the item function description text, construct a semantic similarity matrix S using TF-IDF or a large language embedding model. sem ; S502, regarding the similarity matrices S0 and S sem Perform a weighted summation to obtain the final similarity matrix S.

7. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 4, characterized in that, Step S104 specifically includes the following operations: S601. Convert the similarity matrix into a distance matrix D; S602. For each gene enrichment entry in the distance matrix D, find the n closest samples, keep only the edges in the neighborhood of the n samples, and set the rest of the edges to 0 to achieve KNN sparsity. S603. For each retained neighborhood edge, calculate the Gaussian kernel weight using the local scale parameter to obtain the original affinity matrix; S604. Symmetricize the original affinity matrix by setting the diagonal to 1 and output the final affinity matrix.

8. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 1, characterized in that, In step S105, when performing spectral clustering, the number of clusters can be manually specified by the user or automatically determined. Automatic determination specifically includes the following operations: S701. Set the search range for the cluster number k to obtain multiple candidate k values; S702. Calculate the silhouette coefficient and Bootstrap stability for each candidate k value; S703. Perform a fusion score on each candidate k value based on the contour coefficient and Bootstrap stability; S704. Based on the fusion score results of each candidate k value, determine the optimal K value and output it.

9. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 1, characterized in that, Step S106 specifically includes the following operations: S801. Calculate the data separation degree and semantic separation degree for each cluster. S802. Determine whether the data separation degree and semantic separation degree are greater than 0 respectively. If they are, the check is considered to pass; otherwise, the check is considered to fail. S803, Output semantic-data trend consistency check results.

10. The intelligent classification and automatic interpretation method based on gene set enrichment entries according to claim 1, characterized in that, In step S106, when using a large language model to perform structured interpretation of each cluster, it supports multi-round optimization and dialogue memory management based on user feedback.

Images