Gene regulatory network consensus inference method based on adaptive feature analysis

Abstract

The invention belongs to the field of bioinformatics, and relates to a gene regulatory network consensus inference method based on adaptive feature analysis. The method comprises the following steps of: firstly, inputting a multi-algorithm inference result and preprocessing, automatically extracting five key features including a gene number, an edge number, a weight mean value, a weight variance and network density by a system, and calculating a consistency score among the algorithms; secondly, dynamically distributing an integrated weight of each algorithm based on features and consistency scores, fusing by adopting a weighted Borda counting method to generate a consensus ranking, and adaptively determining an optimal screening threshold according to a gene number and a network density; and finally, performing quantitative quality evaluation on the obtained consensus network, and outputting an evaluation report including the coverage rate, the stability and the consensus strength and a final network. According to the method, full-automatic and data-driven consensus inference is realized, the dependence on external parameters and priori knowledge is remarkably reduced, and a reliable and explainable calculation tool is provided for accurately identifying the gene regulation and control relationship.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of bioinformatics, and particularly relates to a gene regulatory network consensus inference method based on adaptive feature analysis. The method can automatically analyze the features of input data, dynamically adjust the integrated parameters and output high-quality consensus networks. BACKGROUND

[0002] Gene regulatory network inference is the key to understanding cell function and disease mechanism. The results of single inference algorithm, such as GENIE3, GRNBOOST2, etc., are often unstable and biased. Therefore, the consensus method COFFEE uses a fixed weight Borda counting strategy to fuse the edge ranking results of multiple algorithms to improve the robustness of the results. However, this method still has the following problems. ① Lack of adaptability: fixed weights and thresholds cannot be dynamically adjusted according to the characteristics of the input data, such as network size and density; ② Result distortion: the use of key features such as network density and weight distribution of algorithm output is insufficient, resulting in integrated results deviating from the true biological background; ③ Lack of evaluation: there is no system for quantitative quality evaluation of the output consensus network, and the result reliability is questionable. Therefore, there is an urgent need in the field for an adaptive consensus inference method that can perceive data features, make dynamic decisions, and provide objective evaluation. SUMMARY

[0003] The present application proposes a gene regulatory network consensus inference method based on adaptive feature analysis, aiming to solve the problems of strong parameter dependence, unstable results and lack of objective evaluation in the prior art, and to realize data-driven fully automatic consensus network inference. The specific technical solution includes the following 5 steps: Step 1, preprocessing and feature extraction: read the sorted edge list generated by multiple gene regulatory network inference algorithms, and perform data cleaning. Then, automatically extract the multi-dimensional feature vector of the input data, including the number of genes, the number of edges, the overall mean, the weight variance, and the core network density calculated based on high-confidence edges.

[0004] Step 2, dynamic weight allocation: calculate the consistency score between each input algorithm and other algorithm results. Based on the preset baseline weight of each algorithm, combined with its consistency score and weight distribution characteristics, dynamically calculate the maximum weight of each algorithm in consensus integration.

[0005] Step 3, weighted consensus integration: use the dynamic weight obtained in step 2 to improve the traditional Borda counting method. The score of each edge in each algorithm will be multiplied by its weight, and the weighted scores of all algorithms will be summed to obtain the total consensus score of each edge, and a consensus ranking will be generated accordingly.

[0006] Step 4, intelligent threshold selection and quality assessment. According to the number of genes and network density features extracted in step 1, the best threshold for filtering the final consensus network is automatically calculated by a nonlinear mapping function. At the same time, the generated consensus network is quantitatively evaluated, and three indicators of coverage, stability and consensus strength are output.

[0007] Step 5, result output. When the output meets the threshold condition, the final consensus network file will be output, accompanied by a detailed quality assessment report.

[0008] A gene regulatory network consensus inference method based on adaptive feature analysis, the implementation process of step 1 is as follows: the system reads all algorithm result files from the directory specified by the user, the files are all tab-separated text files, and must contain three columns: Gene1, Gene2, Edge_weight; perform data cleaning operation: first, remove all self-loop edges with Gene1 and Gene2 same; second, check the Edge_weight column to ensure it is a numerical type; finally, for each algorithm's edge list, sort it in descending order according to the Edge_weight column to form an initial sorted list; then, automatically extract the multi-dimensional feature vector of the input data, including the number of genes, the number of edges, the mean weight, the weight variance and the core network density.

[0009] A gene regulatory network consensus inference method based on adaptive feature analysis, the implementation process of step 2 is as follows: Calculate the consistency score vector between algorithms; for each algorithm, take the first 100 edges in its sorted list to form a set, and for any two algorithms, calculate their Jaccard similarity; the consistency score of each algorithm is the arithmetic mean of the similarity with all other algorithms; then, perform dynamic weight assignment, and the final adaptive weight is calculated by the formula: .

[0010] A gene regulatory network consensus inference method based on adaptive feature analysis, the implementation process of step 3 is as follows: Perform weighted Borda consensus integration; for each edge, its score in the algorithm is: ; the total consensus score of the edge is the sum of its scores in all algorithms that appear; the system sorts all edges in descending order according to the total consensus score to generate the final consensus ranking.

[0011] A gene regulatory network consensus inference method based on adaptive feature analysis, the implementation process of step 4 is as follows: the intelligent threshold selection function is: ; at the same time, the consensus network is quantitatively evaluated, and three indicators of coverage, stability and consensus strength are calculated.

[0012] A gene regulatory network consensus inference method based on adaptive feature analysis, the implementation process of step 5 is as follows: The system writes the edges filtered by the intelligent threshold in step 4, and the corresponding total consensus score, normalized score and calculated confidence, into the final consensus network file. The consensus network file is a tab-delimited text file containing Gene1, Gene2, BordaScore, NormalizedScore and Confidence columns. At the same time, the system generates a quality assessment report, which is also a tab-delimited text file, containing but not limited to: the dynamic threshold used in this running instance, the adaptive weight assigned to each algorithm, the feature vector extracted from the input data, and the specific values of the coverage, stability and consensus strength quality indicators calculated in step 4. Finally, the system generates the adaptive_consensus_network.tsv consensus network file and the quality_report.tsv quality assessment report in the specified output directory. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 It is a general architecture diagram of a gene regulatory network consensus inference method based on adaptive feature analysis.

[0014] Figure 2 It is a flowchart of the adaptive feature extraction module.

[0015] Figure 3 It is a flowchart of the dynamic weight allocation and weighted consensus integration module.

[0016] Figure 4 It is a flowchart of the intelligent threshold selection and quality assessment module. DETAILED DESCRIPTION

[0017] The present application will be described in detail below in conjunction with the drawings and examples.

[0018] Preprocessing and feature extraction. As shown in Figure 1 and Figure 2 , this step includes two stages: data input and feature calculation. In the data input stage, the system reads all algorithm result files from the user-specified directory. The files are tab-delimited text files and must contain three columns: Gene1: regulatory gene, Gene2: regulated gene, Edge_weight : regulation strength, which is a non-negative real number. The system uses the pandas library to read all files in parallel and performs data cleaning operations: first, remove all self-loop edges with the same Gene1 and Gene2; second, check Edge_weightColumns, ensuring they are of numeric type; finally, for each algorithm's separate edge list, sort by... Edge_weight Sort the columns in descending order to form an initial sorted list. Through the above processing, the system will obtain four sorted lists of edges. In the feature extraction stage, the system automatically extracts multidimensional feature vectors from the input data. Among them, the number of genes Merge the Gene1 and Gene2 columns from all edge lists, and use the pandas.unique() function to get the set of unique genes, the number of which is 1. In this example, =1000; Number of edges Calculate the length of the edge list for each algorithm and take its arithmetic mean. In this example, =45000; Weighted mean With weighted variance : Merge all edge lists Edge_weight The column uses the NumPy library to calculate the mean and variance. In this example, =0.004, =0.0001; Core network density The system first extracts the top 100 edges by weight from the sorted list of each algorithm, forming a high-confidence edge set. E_top100 Here, 100 is a fixed parameter designed to capture the most precise regulatory relationship for each algorithm, and then calculate... = | E_top100 | / [ * ( - 1) / 2]. Among them, | E_ top100 |for sets E_top100 The actual number of edges in the middle may vary depending on the algorithm, as the first 100 edges may be repeated. The denominator is the theoretically maximum number of edges. In this example, | E_top100 | = 5970, therefore = 5970 / (1000*999 / 2) = 0.012. Finally, the system generates a feature vector. F = [1000, 45000, 0.004, 0.0001, 0.012].

[0019] Dynamic weight allocation and consensus integration. For example... Figure 3 As shown, the algorithm consistency is first calculated, followed by dynamic weight allocation, and finally weighted consensus integration is performed. First, the consistency score vector among the algorithms is calculated. For each algorithm i Take the first 100 edges from its sorted list to form a set. For any two algorithms i and j Calculate their Jaccard similarity = | ∩ | / | ∪ |. Each algorithm i Consistency score It is the arithmetic mean of its similarity to all other algorithms, i.e. = mean( ) for all j ≠ i In this embodiment, the calculation is as follows: C_ GENIE3 = 0.65, C_GRNBOOST2 = 0.58, C_PPCOR = 0.45, C_PIDC = 0.28. Subsequently, dynamic weight allocation is performed. The system pre-stores the baseline weights for each algorithm. : B_GENIE3 = 1.2, B_GRNBOOST2 = 1.15, B_PPCOR = 1.1, B_PIDC = 1.0; Final adaptive weights Calculated using the following formula: , where mean( C ) for all The mean, For the algorithm i The standard deviation of the weighted column, The mean of the standard deviations of the weights of all algorithms, and the adjustment coefficient. = 0.4, = 0.3. Taking GENIE3 as an example, substituting the numerical values, its final weight W_GENIE3 = 1.27 is obtained. Similarly, the weights of other algorithms can be obtained. Finally, weighted Borda consensus integration is performed. Let the length of the sorting list for a single algorithm be... N For each edge e Its algorithm i The score is: ( e ) = ( N - r ),in r For this edge in the algorithm i In the ranking, 0 is the highest. (Side) e The total consensus score is the sum of the scores of all the algorithms that appeared in it: TotalScore(e ) = ∑ ( e ). The system ranks all edges in descending order according to TotalScore( e ), and generates the final consensus ranking.

[0020] Intelligent threshold selection and quality assessment. As shown in Figure 4 , this step automatically determines the screening threshold and quantitatively assesses the consensus network. The intelligent threshold selection function is: , where limits the function x to the interval . The characteristic value = 1000, = 0.012 of the present example is brought into calculation: = 0.7645. The system will retain the edges of the normalized consensus score to form the final consensus network, while the system performs quantitative quality assessment, covering coverage, stability and consensus strength. Among them, the coverage: , is the top i = 100 edges of the algorithm K , and is the edge in the final consensus network. This index measures the degree to which the consensus network covers the top-level predictions of each algorithm; stability: , a plurality of consensus networks are generated by Bootstrap resampling 100 times, and the ranking set ranks( e ) of each edge e in these networks is calculated, and then its stability is calculated; consensus strength: , the average score of all edges in the final consensus network is calculated, and finally the index is written into the quality_report.tsv file.

[0021] The baseline weight described in the present application is determined in advance by network search and regression analysis on multiple gold standard data sets such as DREAM4 and DREAM5. To verify the effect of the present application, the DREAM5 benchmark data set is tested. Compared with the traditional COFFE method, the F-score value of the present application is improved from 0.721 to 0.744, with an improvement of 2%, which proves the effectiveness and superiority of the present application.

[0022] The above is a further detailed description of the present application in combination with a specific preferred embodiment, and the specific implementation of the present application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be regarded as falling within the protection scope of the present application.

Claims

1. A consensus inference method for gene regulatory networks based on adaptive feature analysis, characterized in that, Includes the following steps: Step 1, Preprocessing and Feature Extraction: Read the sorted edge list generated by multiple gene regulation network inference algorithms and perform data cleaning; then, automatically extract the multidimensional feature vector of the input data, including the number of genes, the number of edges, the mean weight, the variance of the weight, and the core network density calculated based on high confidence edges; Step 2, Dynamic Weight Allocation: Calculate the consistency score between each input algorithm and the results of other algorithms; Based on the preset baseline weights of each algorithm, and combined with their consensus scores and weight distribution characteristics, the final weight of each algorithm in consensus integration is dynamically calculated. Step 3, Weighted Consensus Integration: Using the dynamic weights obtained in Step 2, the traditional Borda counting method is improved by weighting; the score of each edge in each algorithm is multiplied by its weight, and finally the weighted scores of all algorithms are summed to obtain the total consensus score of each edge, and a consensus ranking is generated accordingly. Step 4, Intelligent Threshold Selection and Quality Assessment: Based on the number of genes and network density features extracted in Step 1, the optimal threshold for screening the final consensus network is automatically calculated through a non-linear mapping function; at the same time, the generated consensus network is quantitatively assessed, and three indicators are output: coverage, stability, and consensus strength. Step 5, Output Results: Output the final consensus network file that meets the threshold conditions, along with a detailed quality assessment report.

2. The method according to claim 1, characterized in that, In step 1, the data cleaning operation includes: first, removing all self-loop edges that are identical between Gene1 and Gene2; second, checking... Edge_weight Columns, ensuring they are of numeric type; finally, for each algorithm's separate edge list, sort by... Edge_weight Sort the columns in descending order to form an initial sorted list. .

3. The method according to claim 1, characterized in that, In step 2, the final adaptive weights Calculated using the following formula: ,in, For baseline weights, The consistency score of the algorithm is mean(C), which represents all... The mean, For the algorithm i The standard deviation of the weight column is given by σ, where mean(σ) is the mean of the standard deviations of the weights for all algorithms. and This is the adjustment coefficient.

4. The method according to claim 1, characterized in that, In step 4, the nonlinear mapping function is: ,in For the number of genes, Given the network density, the clamp(x, a, b) function restricts x to the interval [a, b].

Images