A machine learning-based insulin resistance improving drug target prioritization system

Abstract

The application discloses an insulin resistance improving drug target point priority sequencing system based on machine learning, acquires a function tendency value according to the similarity between the expression level of a candidate gene in a tissue type and an expression mode having a bidirectional regulation effect on insulin resistance in the type, determines a cross-tissue polarity divergence degree based on the difference degree of the function tendency value of the candidate gene in each tissue type after upstream transcription regulation correction, acquires a network function consistency value according to the consistency degree of the function tendency of the candidate gene and the genes interacting with the candidate gene in the tissue type, acquires a tissue pathway conflict value according to the regulation difference of the function regulation direction of the signal pathway in which the candidate gene is involved in the tissue type, and comprehensively acquires a cross-tissue antagonistic potential value of the candidate gene and determines a target point priority sequence. The application realizes specific target point accurate sequencing by multi-dimensional quantification and cross-verification of cross-tissue function antagonistic characteristics.

Description

Technical Field

[0001] This invention belongs to the field of biomedicine, specifically relating to a machine learning-based drug target priority ranking system for improving insulin resistance. Background Technology

[0002] In the development of drugs to improve insulin resistance, there are a vast number of potential targets, but biological validation is time-consuming and costly, making it impossible to conduct experimental evaluations of all candidate molecules individually. Therefore, it is essential to prioritize targets early in the research and development process, focusing limited resources on the molecules most likely to succeed. Scientific and precise prioritization can significantly reduce the failure rate of clinical trials in later stages of new drug development.

[0003] Currently, commonly used target prioritization methods mainly utilize genome-wide association analysis (GWAS) to identify risk sites related to insulin resistance, or rely on differentially expressed genes obtained through transcriptome sequencing, using statistical thresholds such as p-values ​​and fold changes to screen candidate targets. However, these existing technologies all have significant limitations: insulin resistance involves multi-tissue synergy, and the functional roles of the same gene may be drastically different in the liver, skeletal muscle, and adipose tissue. For example, some genes promote gluconeogenesis in the liver but promote glucose uptake in skeletal muscle. In this context, if prioritization relies solely on sequencing evidence from a single tissue source, the aforementioned tissue-specific effects will be lost; and if multi-tissue data are simply averaged, the functional antagonistic relationships between tissues will be masked, leading to the erroneous downgrading or omission of key targets with tissue-targeted regulatory potential. Summary of the Invention

[0004] To address the technical problem that existing ranking methods cannot effectively quantify cross-tissue functional antagonistic features, leading to misjudgment of key tissue-specific targets, the present invention aims to provide a machine learning-based priority ranking system for insulin resistance improvement drug targets.

[0005] The specific technical solution for achieving the objective of this invention is as follows:

[0006] A machine learning-based drug target prioritization system for improving insulin resistance includes the following modules:

[0007] The data acquisition module is used to obtain the expression levels of each candidate gene in samples from different tissue types;

[0008] The tissue divergence quantification module is used to obtain the functional tendency value of the candidate gene in each tissue type based on the similarity between the expression level of the candidate gene in each tissue type and the expression pattern that has a bidirectional regulatory effect on insulin resistance in that tissue type.

[0009] Based on the degree of difference in the functional tendency values ​​of candidate genes after upstream transcriptional regulation characteristics in different tissue types, the cross-tissue polarity divergence of candidate genes is determined.

[0010] The network consistency assessment module is used to obtain the network functional consistency value of a candidate gene based on the degree of consistency between the candidate gene and the functional tendency distribution of its interacting genes in different tissue types.

[0011] The pathway conflict analysis module is used to obtain the tissue pathway conflict value of candidate genes based on the degree of regulatory differences between the functional regulatory directions of the signaling pathways in which candidate genes participate in different tissue types.

[0012] The antagonistic potential analysis module is used to obtain the cross-tissue antagonistic potential value of candidate genes based on the relative expression level of candidate genes in various tissue types, cross-tissue polarity divergence, network functional consistency value, and tissue pathway conflict value.

[0013] The priority ranking module is used to arrange all candidate genes according to the cross-tissue antagonistic potential value to obtain the target priority sequence.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0015] In the process of prioritizing drug targets for improving insulin resistance, this approach obtains functional tendency values ​​that accurately reflect the tissue-specific microenvironment by introducing expression pattern similarity analysis of bidirectional regulatory effects and correction of upstream transcriptional regulatory features. Based on the degree of difference of the correction values ​​among different tissues, the functional conflict of candidate genes among different tissue types is quantified to obtain the degree of cross-tissue polarity divergence, thus capturing the inter-tissue functional antagonistic features lost by traditional methods due to information averaging.

[0016] Further integrating network functional consistency values ​​determined based on candidate genes and their interacting genes, and tissue pathway conflict values ​​determined based on the functional regulation direction of the signaling pathways involved, multiple cross-validations of the antagonistic characteristics of the target are performed from two dimensions: protein-level interaction distribution and signal regulation. This makes up for the shortcomings of single expression level data being easily interfered with and lacking support from physiological antagonistic mechanisms.

[0017] This approach synthesizes a cross-tissue antagonistic potential value by comprehensively considering relative expression levels, cross-tissue polarity divergence, network functional consistency values, and tissue pathway conflict values, and prioritizes all candidate genes. This completely overcomes the problem that existing ranking methods cannot effectively quantify cross-tissue functional antagonistic characteristics, effectively preventing key antagonistic targets with important tissue-specific therapeutic value from being misjudged or masked, thereby significantly improving the accuracy of insulin resistance drug target screening and the success rate of clinical translation.

[0018] The present invention will be further described below with reference to specific embodiments. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the machine learning-based drug target priority ranking system architecture for improving insulin resistance according to the present invention.

[0020] Figure 2 This is a schematic diagram of the organizational divergence quantification module architecture of the present invention.

[0021] Figure 3 This is a schematic diagram of the computer device architecture in an embodiment of the present invention. Detailed Implementation

[0022] Example

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0024] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0025] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0026] Combination Figure 1 A machine learning-based drug target prioritization system for improving insulin resistance includes the following modules:

[0027] Data acquisition module 110 is used to acquire the expression levels of each candidate gene in samples of different tissue types;

[0028] In this approach: during the development of drugs to improve insulin resistance, genes with potential drug target value identified across the entire genome or through preliminary screening will be used as candidate genes.

[0029] Tissue type refers to key physiological tissues that participate in the body's metabolic regulation and are closely related to the process of insulin resistance. In this embodiment, tissue types include liver, skeletal muscle, and adipose tissue.

[0030] Each tissue type contains multiple biological samples. High-throughput transcriptome sequencing data of each sample are obtained. The TPM value (transcriptional per million bases) of each candidate gene in all samples of each tissue type is extracted as the expression level to characterize the transcriptional activity level of the candidate gene in the specific biological sample microenvironment.

[0031] The tissue divergence quantification module 120 is used to obtain the functional tendency value of the candidate gene in each tissue type based on the similarity between the expression level of the candidate gene in each tissue type and the expression pattern that has a bidirectional regulatory effect on insulin resistance in that tissue type. Then, based on the degree of difference between the functional tendency values ​​of the candidate gene in different tissue types after being corrected by upstream transcriptional regulation features, the cross-tissue polarity divergence degree of the candidate gene is determined.

[0032] In order to assess whether candidate genes tend to improve or worsen metabolism in a specific tissue microenvironment, the inherent functional tendency of candidate genes in the corresponding tissue type is quantified by analyzing the similarity between the expression level of candidate genes in each tissue type and the expression pattern of those tissue types that have a bidirectional regulatory effect on insulin resistance, thus obtaining functional tendency values.

[0033] Considering that gene function in a specific tissue depends not only on its own expression pattern but also on the driving regulation of upstream transcription factors, and that the same gene is usually regulated by different transcription factors in different tissues, the functional tendency value after correction for upstream transcriptional regulation characteristics can accurately reflect the functional role of candidate genes in the regulatory logic of the real tissue microenvironment. Based on this, the degree of functional conflict of candidate genes across tissues is quantified according to the degree of difference of this correction value across different tissues, resulting in the cross-tissue polarity divergence degree. This index can effectively capture the inter-tissue functional antagonism characteristics lost by traditional methods due to the simple averaging of multi-tissue information.

[0034] More specifically, in this embodiment, combined with Figure 2The organizational divergence quantification module 120 specifically includes a functional tendency analysis unit 121, a regulatory factor screening unit 122, a functional tendency correction unit 123, and a polarity divergence analysis unit 124.

[0035] The process of obtaining functional aptitude values ​​includes:

[0036] Functional propensity analysis unit 121: The process of obtaining the tissue regulatory transcription factor set of candidate genes in each tissue type and calculating the functional propensity value of each transcription factor in the tissue regulatory transcription factor set includes:

[0037] Obtain the reference gene set for improvement and the reference gene set for deterioration for each tissue type, and denote them as the reference gene set;

[0038] The expression levels of candidate genes in all samples of each tissue type are ranked to obtain expression vectors;

[0039] For each reference gene set, the correlation coefficient between the candidate gene and the expression vector of all genes in the reference gene set is calculated, and the arithmetic mean of the normalized results of all correlation coefficients is taken as the functional similarity between the candidate gene and the reference gene set.

[0040] Calculate the difference and sum of the functional similarity between the candidate gene and the improvement reference gene set and the deterioration reference gene set for each tissue type, and use the ratio of the difference to the sum as the functional tendency value of the candidate gene in that tissue type;

[0041] Specifically, in each tissue type, the candidate gene is compared with the improvement reference gene set and the deterioration reference gene set of that tissue to obtain a difference value, a sum value, and then a ratio (i.e., functional tendency value d).

[0042] If the candidate gene involves n tissue types, in this scheme n=3, namely liver, skeletal muscle, and fat, then one candidate gene corresponds to three differences and three sums, thus obtaining n ratios.

[0043] It should be noted that the development of insulin resistance exhibits significant tissue specificity, necessitating the acquisition of reference gene sets for improvement and deterioration in each tissue type. Based on publicly available biomolecular pathway databases such as Reactome and the Kyoto Encyclopedia of Genes and Genomes (KEGG), as well as literature data mining, known functional genes associated with the pathological process of insulin resistance were retrieved. For each tissue type, these genes were categorized according to their physiological regulatory effects on insulin resistance when overexpressed or activated in that tissue type. An improvement reference gene set was constructed from genes proven to improve insulin resistance, and a deterioration reference gene set was constructed from genes proven to worsen insulin resistance. To ensure the reliability of expression pattern alignment, the total number of genes in each reference gene set was no less than 30.

[0044] The correlation coefficient between the expression vectors of candidate genes and genes in the reference gene set reflects the degree of synergistic co-expression between the expression level change trend of candidate genes and genes that have a bidirectional regulatory effect on insulin resistance (improvement or deterioration). Functional similarity characterizes the degree of fit between the expression pattern of candidate genes and the improvement or deterioration of insulin resistance; the larger the value, the more synergistic the expression pattern of candidate genes is with the corresponding functional expression pattern of the reference gene set. The Pearson correlation coefficient ranges from [-1, 1]. To avoid the cancellation of positive and negative correlation coefficients during summation, half of the sum of the correlation coefficient and the constant 1 is taken to normalize the correlation coefficient. The influence of differences in gene background expression abundance is eliminated by the ratio of the difference in functional similarity between the improved and deteriorated reference gene sets to the sum of the similarities, resulting in a functional tendency value ranging from [-1, 1]. The absolute value of the functional tendency value reflects the strength of the regulatory effect of the candidate gene; a positive value indicates that the candidate gene tends to improve insulin resistance in the tissue type, while a negative value indicates that the candidate gene tends to worsen insulin resistance in the tissue type. Since both the improving reference gene set and the worsening reference gene set originate from the same tissue type and are related to insulin resistance, the sum of the above values ​​cannot be zero.

[0045] In this embodiment of the invention, the correlation coefficient is the Pearson correlation coefficient. In other embodiments, it can also be the Spearman correlation coefficient or the Kendall rank correlation coefficient, etc.

[0046] The tissue-regulated transcription factor set for candidate genes in each tissue type was obtained, and the propensity function (PFC) of each transcription factor in this set was calculated. The transcription factors themselves were treated as targets, and the same calculation method as for the PFC of candidate genes in each tissue type was used. It should be noted that publicly available transcriptional regulation knowledge databases such as TRRUST or hTFtarget, as well as public high-throughput omics datasets such as chromatin immunoprecipitation sequencing, were accessed. Transcription factors with physical binding evidence or known driver relationships to candidate genes were retrieved. Combined with the transcriptional maps of each tissue type, transcription factors with clear regulatory effects on candidate genes in that tissue type were screened to constitute the tissue-regulated transcription factor set for candidate genes in that tissue type.

[0047] The regulatory factor screening unit 122 is used to calculate the correlation coefficient between the expression vectors of all transcription factors in the tissue regulatory transcription factor set of each tissue type and the expression vector of the corresponding candidate gene, which is used as the regulatory association strength between the corresponding transcription factor and the candidate gene; from the regulatory association strength between the candidate gene and all transcription factors in the tissue regulatory transcription factor set of each tissue type, the transcription factors corresponding to the regulatory association strength greater than the preset regulatory threshold are selected as the positive regulatory factors of the candidate gene in each tissue type.

[0048] It should be noted that the expression levels of each transcription factor in the tissue-regulating transcription factor set of candidate genes in each tissue type are ranked across all samples of that tissue type, resulting in the expression vector of the corresponding transcription factor in that tissue type. The expression vectors of candidate genes and transcription factors, and elements with the same subscript in all expression vectors, correspond to the same sample. The regulatory association strength characterizes the degree of synchronization between the co-expression of candidate genes and transcription factors at the transcriptional level; the higher the value, the stronger the driving effect of the transcription factor on the expression of the candidate gene, and the tighter the regulatory relationship. To eliminate biological background noise and weak ineffective regulation, positive regulatory factors that exert a driving effect are screened through a preset regulatory threshold.

[0049] In this embodiment of the invention, the preset regulatory threshold can be set to 0.05 based on biostatistical experience. Alternatively, the average regulatory association strength between the candidate gene and all transcription factors in the tissue regulatory transcription factor set of each tissue type, or a preset quantile, such as the top 50% quantile, can be used as the preset regulatory threshold.

[0050] Functional propensity correction unit 123 is used to perform normalized weighted calculation of the functional propensity values ​​of each positive regulatory transcription factor, using the strength of the regulatory association between the candidate gene and its positive regulatory factor in each tissue type as weights, to obtain the transcriptional regulation propensity value of the candidate gene in each tissue type; and to perform weighted summation of the functional propensity value and the transcriptional regulation propensity value to obtain the corrected propensity value of the candidate gene in each tissue type.

[0051] In this embodiment, the correction tendency value is expressed by the formula:

[0052] ;

[0053] ;

[0054] In the formula, represents the corrected propensity value for the candidate gene in the t-th tissue type; represents the functional predisposition value of the candidate gene in the t-th tissue type; The transcriptional regulation tendency value of the candidate gene in the t-th tissue type; This represents the total number of positive regulatory factors for the candidate gene in the t-th tissue type; The regulatory association strength between a candidate gene and its nth positive regulatory factor in the tth tissue type; This represents the functional tendency value of the candidate gene as the nth positive regulatory factor in the tth tissue type; Preset fusion weights.

[0055] It should be noted that the transcriptional regulation propensity value characterizes the functional tendency of candidate genes driven by the upstream microenvironment of each tissue type. The larger the absolute value, the higher the synergy between the candidate gene and the corresponding upstream regulatory network in promoting the improvement or deterioration of insulin resistance. By weighted summing the functional propensity value and the transcriptional regulation propensity value to obtain the corrected propensity value, it can comprehensively characterize the superimposed effect of the candidate gene's own expression characteristics and the tissue-specific regulatory environment. The larger the absolute value, the more reliable the reflection of the candidate gene's functional tendency in each tissue type.

[0056] The preset fusion weight is used to balance the contribution of the candidate gene's own functional tendency value and transcriptional regulatory tendency value in the correction process. Considering that the gene's own expression pattern is a direct reflection of its functional role in the tissue, while upstream transcriptional regulatory information is the corroboration and correction of this functional role, the gene's own expression should be dominant, and the preset fusion weight should be greater than 0.5. In this embodiment, the preset fusion weight is set to 0.7. The preset fusion weight can effectively introduce the correction effect of upstream regulation while preserving the original expression information.

[0057] It is important to note that if a candidate gene does not have a positive regulatory factor in a certain tissue type, the initial functional tendency value of the candidate gene should be assigned to the transcriptional regulatory tendency value; if... Then the functional tendency value will be assigned to the correction tendency value.

[0058] The polarity divergence analysis unit 124 is used to determine whether the signs of the corrected tendency values ​​of the candidate gene are the same in all tissue types. If so, the cross-tissue polarity divergence of the candidate gene is assigned to zero; otherwise, the range of the corrected tendency values ​​of the candidate gene in all tissue types is calculated as the cross-tissue polarity divergence of the candidate gene.

[0059] If the signs of the corrected tendency values ​​for a candidate gene are the same across all tissue types, it indicates that the candidate gene acts in the same direction across all tissue types and does not possess cross-tissue antagonistic properties; in this case, the cross-tissue polarity divergence is assigned a value of zero. Conversely, if the signs are different, it indicates that the candidate gene exhibits functional antagonism across different tissue types. Using the range as the cross-tissue polarity divergence can effectively eliminate the influence of absolute numerical scales to accurately quantify relative fluctuations. The greater the cross-tissue polarity divergence, the more severe the functional conflict of the candidate gene across different tissue types and the more significant its cross-tissue antagonistic properties.

[0060] Furthermore, when the mean of the corrected tendency values ​​of candidate genes across all tissue types is zero, or the range is greater than or equal to a preset upper limit of divergence, it indicates that the inter-tissue antagonistic effect has reached an extreme state. To prevent numerical explosion from disrupting the global feature distribution, the cross-tissue polarity divergence of candidate genes is assigned the preset upper limit of divergence. Specifically, the corrected tendency values ​​of all candidate genes across all tissue types are calculated, and their corresponding ranges are calculated. After eliminating invalid extreme values ​​caused by a mean of zero, a specific percentile of all remaining ranges, such as the 95th percentile, is used as the preset upper limit of divergence.

[0061] The network consistency assessment module 130 is used to obtain the network functional consistency value of the candidate gene based on the degree of consistency between the candidate gene and the functional tendency distribution of its interacting genes in different tissue types, including:

[0062] The interaction gene set is obtained by identifying the genes that interact with the candidate gene in all tissue types. This can be achieved, for example, by using publicly available protein interaction databases such as BioGRID (Biological General Repository for Interaction Datasets) to extract genes that have a clear physical binding or functional association with the candidate gene and label them as interaction genes. Then, by combining the gene expression profiles of each tissue type, the interaction genes that are effectively expressed in all tissue types can be screened to form the interaction gene set.

[0063] The interaction gene set of candidate genes is arranged according to the corrected propensity values ​​of all interacting genes in each tissue type to obtain the interaction network propensity vector of that tissue type.

[0064] The network functional consistency value of the candidate gene is obtained by averaging the normalized cosine similarity of the interaction network tendency vectors of every two tissue types across all tissue types. Cosine similarity reflects the degree of consistency of local regulatory network features in the direction of action between corresponding tissue types. The network functional consistency value obtained by averaging all cosine similarities comprehensively quantifies the synergistic level of the interaction protein population around the candidate gene in the cross-tissue direction of action. The larger the network functional consistency value, the more consistent the functional tendency of the candidate gene's interaction network is across tissues; the smaller the value, the more significant functional conflict exists between different tissues. Since the range of cosine similarity is [-1, 1], this embodiment normalizes the cosine similarity by taking half of the sum of the cosine similarity and the constant 1, so that the range of the network functional consistency value is [0, 1].

[0065] Furthermore, the correction tendency values ​​here are obtained using the same method as those for candidate genes in each tissue type, calculating the correction tendency values ​​of each interacting gene in the interacting gene set across all tissue types. If an interacting gene is not expressed in a specific tissue type or does not belong to the interacting gene set of that tissue type, the correction tendency value of that gene in that tissue type is set to zero. In the interaction network tendency vectors of all tissue types, the correction tendency values ​​at the same index correspond to the same interacting gene. When the interaction network tendency vector of a certain tissue type is zero, the cosine similarity between the interaction network tendency vectors of that tissue type and those of other tissue types is set to zero.

[0066] The pathway conflict analysis module 140 is used to obtain the tissue pathway conflict value of candidate genes based on the degree of regulatory differences between the functional regulatory directions of the signaling pathways in which candidate genes participate in different tissue types, including:

[0067] Obtain a set of insulin-related pathways consisting of several signaling pathways. Each signaling pathway contains several genes and their interconnections, and the regulatory direction of the genes and the type of tissue they act on are pre-labeled. The regulatory direction labels include an improvement direction value of 1 and a deterioration direction value of -1.

[0068] Here, when obtaining a set of insulin-related pathways consisting of several signaling pathways, one can use publicly available pathway databases such as Reactome and KEGG to search for signaling pathways related to insulin resistance and construct an insulin-related pathway set. For each signaling pathway, all genes contained in the pathway and the topological connections between genes (including interaction types such as activation, inhibition, and binding) are extracted from the signaling pathway database file (such as KGML files from KEGG or Systems Biology Markup Language (SBML) files from Reactome). Simultaneously, combining literature mining and pathway function annotation, the signaling pathways are pre-labeled with regulatory direction indicators and tissue types of action: regulatory direction indicators include an improvement direction value of 1 and a deterioration direction value of -1, where an improvement direction value indicates that overall activation of the signaling pathway can improve insulin resistance, and a deterioration direction value indicates that overall activation of the signaling pathway can worsen insulin resistance; the tissue type of action indicates the tissue type in which the signaling pathway mainly functions, such as liver, skeletal muscle, or adipose tissue.

[0069] For each signaling pathway in the insulin-related pathway set, the number of genes in the signaling pathway that are directly related to the candidate gene is counted as the node connectivity. The ratio of the node connectivity to the total number of genes in the signaling pathway is used as the importance weight of the candidate gene in the signaling pathway.

[0070] Two tissue types are randomly selected and denoted as the first tissue type and the second tissue type, respectively; the candidate participation pathway set is composed of all signaling pathways in the insulin-related pathway set that contain any candidate gene;

[0071] In the candidate pathway set, the signaling pathway with the first tissue type and the regulation direction identifier being the improvement direction value is selected as the first tissue-matched functional pathway, and the signaling pathway with the second tissue type and the regulation direction identifier being the deterioration direction value is selected as the second tissue-matched functional pathway.

[0072] Calculate the product of the importance weight of the candidate gene in each tissue-matching functional pathway and the regulatory direction identifier. The sum of the corresponding products of the candidate gene in the same tissue-matching functional pathway is used as the functional pathway weighted regulation value of each tissue-matching functional pathway. The absolute difference between the functional pathway weighted regulation values ​​of the first tissue-matching functional pathway and the second tissue-matching functional pathway is used as the local pathway conflict value between the first tissue type and the second tissue type.

[0073] The sum of the local pathway conflict values ​​between all pairs of tissue types is calculated as the tissue pathway conflict value of the candidate gene.

[0074] To overcome the assumption that different genes contribute equally in the same pathway, this study aims to obtain importance weights that characterize the core hub status and regulatory influence of candidate genes in specific signal transduction cascade networks. To uncover functional antagonism between candidate genes in different tissues, we extracted first-tissue matched functional pathways that improve function in the first tissue and second-tissue matched functional pathways that worsen function in the second tissue. Based on importance weights and regulatory direction identifiers, we calculated the weighted regulatory values ​​of functional pathways to quantify the regulatory effect of candidate genes in specific regulatory directions within a single tissue type microenvironment. The absolute difference between the weighted regulatory values ​​of these two types of pathways is used as the local pathway conflict value to quantify the intensity of physiological antagonism induced by candidate genes through contradictory pathways in different tissues. A comprehensive analysis of the physiological antagonism intensity across all tissue types yields the tissue pathway conflict value. The larger this value, the more severe the pathway effect conflict induced by the candidate gene in different tissues, and the more significant its physiological antagonistic effect.

[0075] The antagonistic potential analysis module 150 is used to obtain the cross-tissue antagonistic potential value of candidate genes based on their relative expression levels in different tissue types, cross-tissue polarity divergence, network functional consistency value, and tissue pathway conflict value.

[0076] Based on the position of the expression characteristics of candidate genes in each tissue type relative to the expression characteristics of genes contained in that tissue type, the expression abundance weight of candidate genes is determined. The expression abundance weight reflects the relative expression level of candidate genes in each tissue type. In this embodiment, the calculation method includes:

[0077] The median expression level of the candidate gene was obtained from all samples in the same tissue type and used as the baseline expression level of the candidate gene in each tissue type.

[0078] The baseline expression levels of all candidate genes included in each tissue type were sorted in ascending order to obtain the baseline expression sequence of the corresponding tissue type;

[0079] The ratio of the number of elements in the baseline expression sequence of each tissue type that are less than or equal to the baseline expression level of the candidate gene in that tissue type to the total number of elements in that baseline expression sequence is used as the expression rank of the candidate gene in that tissue type.

[0080] The mean of the expression rank ratio of the candidate gene across all tissue types is calculated as the expression abundance weight of the candidate gene;

[0081] The expression rank ratio, obtained by percentile transformation of baseline expression levels, eliminates dimensional interference caused by differences in expression abundance among different genes. The expression rank ratio characterizes the relative expression activity of candidate genes in the microenvironment of each tissue type; the closer the value is to 1, the more significant the abundance expression of the candidate gene in that tissue type. An expression abundance weight, ranging from 0 to 1, is obtained through mean operation. The larger this value, the higher the relative expression level of the candidate gene in each tissue type, and the more sufficient the material basis for exerting biological functions.

[0082] Negative correlation mapping is performed on the network functional consistency value of candidate genes, and the normalized result of the product of the negative correlation mapping, cross-organizational polarity divergence, organizational pathway conflict value and the expression abundance weight is used as the cross-organizational antagonistic potential value of the candidate gene.

[0083] The true antagonistic potential of a gene requires the simultaneous fulfillment of four necessary conditions: First, after transcriptional regulation correction, its functional tendencies are opposite in different tissues, indicating significant cross-tissue regulatory divergence and a high degree of cross-tissue polarity divergence. Second, this opposite functional tendency is supported by the protein-protein interaction network, meaning that the functional tendencies of its interaction partners also show inconsistency across different tissues, indicating a low network consistency value. Third, the signaling pathways involved by this gene exhibit clear functional direction conflicts across different tissues, indicating a high tissue pathway conflict value. Fourth, the gene possesses sufficient material basis for exerting its biological function in each tissue, meaning its expression abundance is sufficiently high, indicating a high expression abundance weight. These four conditions together constitute a comprehensive basis for determining gene antagonistic potential. Therefore, network functional consistency value is negatively correlated with cross-tissue antagonistic potential value, while cross-tissue polarity divergence, tissue pathway conflict value, and expression abundance weight are all positively correlated with cross-tissue antagonistic potential value.

[0084] In this embodiment of the invention, since the network functional consistency value ranges from 0 to 1, the difference between the constant 1 and the network functional consistency value is used to perform a negative correlation mapping on the network functional consistency value. The product of the mapping result, cross-organizational polarity divergence, organizational pathway conflict value and expression abundance weight is calculated and recorded as the initial antagonistic potential value of the candidate gene. Based on the maximum and minimum values ​​among the initial antagonistic potential values ​​of all candidate genes, the initial antagonistic potential value of a single candidate gene is normalized using the minimax normalization method to obtain the cross-organizational antagonistic potential value of the candidate gene.

[0085] Priority ranking module 160 is used to rank all candidate genes according to the cross-tissue antagonistic potential value to obtain a target priority sequence, including:

[0086] The cross-tissue antagonistic potential values ​​of each candidate gene are input into the pre-trained machine learning model to obtain the drug target confidence of the corresponding candidate gene.

[0087] The drug target confidence of all candidate genes is sorted in descending order to obtain the target priority sequence.

[0088] Here, the construction and training process of the machine learning model is as follows: Genes targeting insulin resistance improvement drugs that have been experimentally validated are collected from drug target databases such as DrugBank and TTD, serving as positive samples with a label of 1; genes involving metabolic processes or insulin (such as insulin receptor signaling pathway, insulin secretion, etc.) in the annotation entries are removed from the gene ontology database, and genes not directly related to metabolism and insulin function are randomly selected as negative samples with a label of 0; all positive and negative samples are then divided into training and validation sets in a 4:1 ratio. A label of 1 indicates that the gene is a known effective target for insulin resistance improvement drugs, and a label of 0 indicates that the gene is not related to metabolism and insulin function.

[0089] The feature vectors of each sample are extracted, including two types of features: the first type is traditional genomic features, which specifically include the expression level of each tissue type, the risk association degree of genome-wide association studies such as the inverse of the p-value or effect size, and the differential expression fold; the second type is the network functional consistency value, cross-organizational polarity divergence, tissue pathway conflict value and expression abundance weight, and cross-organizational antagonistic potential value calculated using the same method as this scheme.

[0090] The machine learning model is trained by using the feature vector as input and the corresponding label as output, and supervised learning is employed.

[0091] The machine learning model in this scheme can be a classification or regression model such as random forest, support vector machine or deep neural network. During training, the objective function is binary log loss or mean squared error. The hyperparameters are optimized through cross-validation until the model's performance converges on the validation set, thus obtaining the pre-trained machine learning model.

[0092] In this embodiment, the priority ranking module 160 prioritizes all candidate genes based on cross-tissue antagonistic potential values, which can break the limitations of traditional manual experience-based qualitative screening and give higher priority to high-potential targets with significant cross-tissue specificity advantages.

[0093] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the functional propensity value includes: obtaining an improvement reference gene set and a deterioration reference gene set for each tissue type, denoted as the reference gene set; arranging the expression levels of candidate genes in all samples of each tissue type to obtain expression vectors; for each reference gene set, calculating the correlation coefficient between the candidate gene and the expression vectors of all genes in the reference gene set, and taking the arithmetic mean of the normalized results of all correlation coefficients as the functional similarity between the candidate gene and the reference gene set; calculating the difference and sum of the functional similarity between the candidate gene and the improvement reference gene set and the deterioration reference gene set for each tissue type, and taking the ratio of the difference to the sum as the functional propensity value of the candidate gene in that tissue type.

[0094] In this embodiment of the invention, the correlation coefficient is the Pearson correlation coefficient. In other embodiments, it can also be the Spearman correlation coefficient or the Kendall rank correlation coefficient, etc.

[0095] In addition, this solution also provides a machine learning-based method for prioritizing drug targets for improving insulin resistance, including the following steps:

[0096] Step 1: Obtain the expression levels of each candidate gene in samples from different tissue types;

[0097] Step 2: Based on the similarity between the expression level of the candidate gene in each tissue type and the expression pattern of the gene that has a bidirectional regulatory effect on insulin resistance in that tissue type, obtain the functional tendency value of the candidate gene in that tissue type; based on the degree of difference between the functional tendency values ​​of the candidate gene in different tissue types after being corrected by upstream transcriptional regulatory features, determine the cross-tissue polarity divergence of the candidate gene.

[0098] Step 3: Obtain the network functional consistency value of the candidate gene based on the consistency of the functional tendency distribution of the candidate gene and its interacting genes in different tissue types.

[0099] Step 4: Based on the degree of regulatory differences between the functional regulatory directions of the signaling pathways in which the candidate genes participate in different tissue types, obtain the tissue pathway conflict value of the candidate genes.

[0100] Step 5: Obtain the cross-tissue antagonistic potential value of candidate genes based on their relative expression levels in different tissue types, cross-tissue polarity divergence, network functional consistency value, and tissue pathway conflict value.

[0101] Step 6: Use a machine learning network to rank all candidate genes according to their cross-tissue antagonistic potential values ​​to obtain a target priority sequence.

[0102] Combination Figure 3This solution also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps:

[0103] Step 1: Obtain the expression levels of each candidate gene in samples from different tissue types;

[0104] Step 2: Based on the similarity between the expression level of the candidate gene in each tissue type and the expression pattern of the gene that has a bidirectional regulatory effect on insulin resistance in that tissue type, obtain the functional tendency value of the candidate gene in that tissue type; based on the degree of difference between the functional tendency values ​​of the candidate gene in different tissue types after being corrected by upstream transcriptional regulatory features, determine the cross-tissue polarity divergence of the candidate gene.

[0105] Step 3: Obtain the network functional consistency value of the candidate gene based on the consistency of the functional tendency distribution of the candidate gene and its interacting genes in different tissue types.

[0106] Step 4: Based on the degree of regulatory differences between the functional regulatory directions of the signaling pathways in which the candidate genes participate in different tissue types, obtain the tissue pathway conflict value of the candidate genes.

[0107] Step 5: Obtain the cross-tissue antagonistic potential value of candidate genes based on their relative expression levels in different tissue types, cross-tissue polarity divergence, network functional consistency value, and tissue pathway conflict value.

[0108] Step 6: Use a machine learning network to rank all candidate genes according to their cross-tissue antagonistic potential values ​​to obtain a target priority sequence.

[0109] In this embodiment, the computer device includes: a memory 201, a processor 202, and a computer program 203 stored in the memory 201 and running on the processor 202. When the processor 202 executes the computer program 203, the computer device can implement any of the aforementioned machine learning-based drug target priority ranking methods for improving insulin resistance.

[0110] The embodiments described above are merely one implementation method of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A machine learning-based drug target prioritization system for improving insulin resistance, characterized in that, Includes the following modules: The data acquisition module is used to obtain the expression levels of each candidate gene in samples from different tissue types; The tissue divergence quantification module is used to obtain the functional tendency value of the candidate gene in each tissue type based on the similarity between the expression level of the candidate gene in each tissue type and the expression pattern that has a bidirectional regulatory effect on insulin resistance in that tissue type. Based on the degree of difference in the functional tendency values ​​of candidate genes after upstream transcriptional regulation characteristics in different tissue types, the cross-tissue polarity divergence of candidate genes is determined. The network consistency assessment module is used to obtain the network functional consistency value of a candidate gene based on the degree of consistency between the candidate gene and the functional tendency distribution of its interacting genes in different tissue types. The pathway conflict analysis module is used to obtain the tissue pathway conflict value of candidate genes based on the degree of regulatory differences between the functional regulatory directions of the signaling pathways in which candidate genes participate in different tissue types. The antagonistic potential analysis module is used to obtain the cross-tissue antagonistic potential value of candidate genes based on the relative expression level of candidate genes in various tissue types, cross-tissue polarity divergence, network functional consistency value, and tissue pathway conflict value. The priority ranking module is used to arrange all candidate genes according to the cross-tissue antagonistic potential value to obtain the target priority sequence.

2. The machine learning-based drug target priority ranking system for improving insulin resistance according to claim 1, characterized in that, The process of obtaining the functional predisposition values ​​of candidate genes in this tissue type includes: Obtain the reference gene set for improvement and the reference gene set for deterioration for each tissue type, and denote them as the reference gene set; The expression levels of candidate genes in all samples of each tissue type are ranked to obtain expression vectors; For each reference gene set, the correlation coefficient between the candidate gene and the expression vector of all genes in the reference gene set is calculated, and the arithmetic mean of the normalized results of all correlation coefficients is taken as the functional similarity between the candidate gene and the reference gene set. Calculate the difference and sum of the functional similarity between the candidate gene and the improvement reference gene set and the deterioration reference gene set for each tissue type, and use the ratio of the difference to the sum as the functional tendency value of the candidate gene in that tissue type.

3. The machine learning-based drug target priority ranking system for improving insulin resistance according to claim 2, characterized in that, The determination of the cross-tissue polarity divergence of candidate genes specifically involves: Obtain the tissue regulatory transcription factor set of candidate genes in each tissue type, and calculate the functional tendency value of each transcription factor in the tissue regulatory transcription factor set; The correlation coefficients between the expression vectors of all transcription factors in the tissue regulatory transcription factor set of each tissue type and the expression vector of the corresponding candidate gene are calculated, which are used as the regulatory association strength between the corresponding transcription factor and the candidate gene. From the regulatory association strength between the candidate gene and all transcription factors in the tissue regulatory transcription factor set of each tissue type, the transcription factors corresponding to the regulatory association strength greater than the preset regulatory threshold are selected as the positive regulatory factors of the candidate gene in each tissue type. Using the strength of the regulatory association between the candidate gene and its positive regulatory factor in each tissue type as the weight, the functional tendency value of each positive regulatory transcription factor is normalized and weighted to obtain the transcriptional regulatory tendency value of the candidate gene in each tissue type. The functional propensity value and the transcriptional regulation propensity value are weighted and summed to obtain the corrected propensity value of the candidate gene in each tissue type. Determine whether the signs of the corrected tendency values ​​of the candidate gene are the same across all tissue types. If so, assign the cross-tissue polarity divergence of the candidate gene to zero; otherwise, calculate the range of the corrected tendency values ​​of the candidate gene across all tissue types as the cross-tissue polarity divergence of the candidate gene.

4. The machine learning-based drug target priority ranking system for improving insulin resistance according to claim 3, characterized in that, The process of obtaining network functional consistency values ​​for candidate genes includes: Obtain the set of interacting genes that are associated with the candidate gene in all tissue types; The interaction gene set of candidate genes is arranged according to the corrected propensity values ​​of all interacting genes in each tissue type to obtain the interaction network propensity vector of that tissue type. The network functional consistency value of the candidate gene is obtained by averaging the normalized cosine similarity of the interaction network tendency vectors of every two tissue types across all tissue types.

5. The machine learning-based drug target priority ranking system for improving insulin resistance according to claim 1, characterized in that, The acquisition of the tissue pathway conflict value of candidate genes specifically involves: Obtain a set of insulin-related pathways consisting of several signaling pathways. Each signaling pathway contains several genes and their interconnections, and the regulatory direction of the genes and the type of tissue they act on are pre-labeled. The regulatory direction labels include an improvement direction value of 1 and a deterioration direction value of -1. For each signaling pathway in the insulin-related pathway set, the number of genes in the signaling pathway that are directly related to the candidate gene is counted as the node connectivity. The ratio of the node connectivity to the total number of genes in the signaling pathway is used as the importance weight of the candidate gene in the signaling pathway. Two tissue types are randomly selected and designated as the first tissue type and the second tissue type, respectively; the candidate participation pathway set is composed of all signaling pathways containing candidate genes in the insulin-related pathway set; In the candidate pathway set, the signaling pathway with the first tissue type and the regulation direction identifier being the improvement direction value is selected as the first tissue-matched functional pathway, and the signaling pathway with the second tissue type and the regulation direction identifier being the deterioration direction value is selected as the second tissue-matched functional pathway. Calculate the product of the importance weight of the candidate gene in each tissue matching functional pathway and the regulatory direction identifier. The sum of the corresponding products of the candidate gene in the same tissue matching functional pathway is used as the functional pathway weighted regulation value of each tissue matching functional pathway. The absolute difference between the functional pathway weighted regulation values ​​of the first tissue matching functional pathway and the second tissue matching functional pathway is used as the local pathway conflict value between the first tissue type and the second tissue type. The sum of the local pathway conflict values ​​between all pairs of tissue types is calculated as the tissue pathway conflict value of the candidate gene.

6. The machine learning-based drug target priority ranking system for improving insulin resistance according to claim 1, characterized in that, The specific steps for obtaining the cross-tissue antagonistic potential value of candidate genes are as follows: The expression abundance weight of candidate genes is determined based on the position of the expression characteristics of candidate genes in each tissue type relative to the expression characteristics of genes contained in that tissue type. Negative correlation mapping is performed on the network functional consistency value of candidate genes, and the normalized result of the product of the negative correlation mapping, cross-organizational polarity divergence, organizational pathway conflict value and the expression abundance weight is used as the cross-organizational antagonistic potential value of the candidate gene.

7. The machine learning-based drug target priority ranking system for improving insulin resistance according to claim 6, characterized in that, The determination of the expression abundance weights of candidate genes specifically involves: The median expression level of the candidate gene was obtained across all samples in the same tissue type and used as the baseline expression level of the candidate gene in each tissue type. The baseline expression levels of all candidate genes included in each tissue type were sorted in ascending order to obtain the baseline expression sequence of the corresponding tissue type; The ratio of the number of elements in the baseline expression sequence of each tissue type that are less than or equal to the baseline expression level of the candidate gene in that tissue type to the total number of elements in that baseline expression sequence is used as the expression rank of the candidate gene in that tissue type. The mean of the expression rank of the candidate gene across all tissue types is calculated and used as the expression abundance weight of the candidate gene.

8. The machine learning-based drug target priority ranking system for improving insulin resistance according to claim 1, characterized in that, The obtained target priority sequence is specifically as follows: The cross-tissue antagonistic potential values ​​of each candidate gene are input into the pre-trained machine learning model to obtain the drug target confidence of the corresponding candidate gene. The drug target confidence of all candidate genes is sorted in descending order to obtain the target priority sequence.

9. A machine learning-based method for prioritizing drug targets for improving insulin resistance, characterized in that, Includes the following steps: Step 1: Obtain the expression levels of each candidate gene in samples from different tissue types; Step 2: Based on the similarity between the expression level of the candidate gene in each tissue type and the expression pattern of the gene that has a bidirectional regulatory effect on insulin resistance in that tissue type, obtain the functional tendency value of the candidate gene in that tissue type. Based on the degree of difference in the functional tendency values ​​of candidate genes after upstream transcriptional regulation characteristics in different tissue types, the cross-tissue polarity divergence of candidate genes is determined. Step 3: Obtain the network functional consistency value of the candidate gene based on the degree of consistency between the functional tendency distribution of the candidate gene and the genes that interact with it in different tissue types. Step 4: Based on the degree of regulatory differences between the functional regulatory directions of the signaling pathways in which the candidate genes participate in different tissue types, obtain the tissue pathway conflict value of the candidate genes. Step 5: Obtain the cross-tissue antagonistic potential value of candidate genes based on their relative expression levels in different tissue types, cross-tissue polarity divergence, network functional consistency value, and tissue pathway conflict value. Step 6: Use a machine learning network to rank all candidate genes according to their cross-tissue antagonistic potential values ​​to obtain a target priority sequence.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method of claim 9.

Images