Enhancer promoter regulation and control network prediction model construction method

Abstract

The invention provides an enhancer-promoter regulation and control network prediction model construction method, which comprises the following steps: S1, obtaining an original data set which comprises a plurality of enhancer-promoter pair data of a plurality of biological samples, and dividing the original data set into a plurality of subsets, dividing all enhancer promoter pairs on the same chromosome pair into the same subset; s2, the subsets are preprocessed, each subset comprises a plurality of data samples, and each data sample is an enhancer promoter pair, the feature vector of each data sample is a feature vector formed by splicing a sequence feature of a corresponding enhancer promoter pair, a distance feature between the enhancer promoter pairs and a chromatin open feature corresponding to the enhancer promoter pairs, and the tag of each data sample indicates whether interaction exists between the corresponding enhancer promoter pairs; and S3, iteratively constructing a plurality of symmetric decision trees to form a prediction model based on all the preprocessed subsets by adopting a category type feature gradient lifting mode.

Description

Technical Field

[0001] The present invention relates to bioinformatics technology, specifically, to the field of gene regulation genomics in bioinformatics, and more specifically, to a method for constructing an enhancer promoter regulatory network prediction model. Background Art

[0002] Eukaryotic gene expression is determined by the coding region of the gene sequence and the non-coding region that regulates gene expression. Among them, cis-acting elements are sequences in the non-coding region that exist in the gene-adjacent sequence and can affect gene expression. Cis-acting elements include promoters, enhancers, etc., and their function is to participate in the regulation of gene expression. Accurately identifying the interactions between genomic elements is a core task in deciphering transcriptional regulation and human disease risk.

[0003] Under the existing technology, researchers have proposed a variety of schemes to calculate and predict the interactions between candidate cis-regulatory elements. According to the technical means adopted by different schemes, the existing technologies can be divided into three categories:

[0004] The first type is to train prediction models based on sequence features of enhancers and promoters. There is no requirement for experimental data. For example, the schemes adopted in references [1], [2], and [3] all use the extracted sequence features of enhancers and promoters to infer the interaction between enhancers and promoters. Among them, the schemes in references [1] and [2] extract the sequence features of enhancers and promoters respectively through convolutional neural networks, and then splice the two and input them into recurrent neural networks to predict whether there is an interaction; the scheme in reference [3] converts the sequence features into k-mer sentences with a step size of 1, and then embeds each sentence into a vector through paragraph vectors. After splicing all vectors, they are classified by normalized exponential function. It should be noted that the inherent mechanism of sequence-based prediction is that the interaction between enhancers and promoters depends on the binding of transcription factors with clear sequence preferences. Although the first type of scheme is based on sequence and is user-friendly because only sequence data needs to be input, this type of method ignores information closely related to gene regulation such as chromatin openness, which makes the prediction of enhancer-promoter interaction limited.

[0005] The second type uses the sequence features of enhancers and promoters and various types of experimental data as features, such as enhancer and promoter sequences, protein sequencing, chromatin open region sequencing, methylation sequencing, and transcriptome sequencing, for example, the schemes in references 0 and 0. Among them, the scheme in reference [4] uses a gradient boosting tree classifier to classify whether enhancers and promoters interact based on features, and the scheme in reference [5] uses a random forest algorithm to predict whether enhancers and promoters interact based on features. The second type of scheme needs to use various types of experimental data as features, needs to input various omics data, and needs to obtain various omics data through experiments. It has high data requirements and is prone to overfitting problems.

[0006] The third type uses the combined features of enhancer and promoter sequences and chromatin open region sequencing for prediction, which can reduce the requirements for experimental data and take into account convenience, such as the scheme in reference [6]. Among them, the scheme in reference 0 uses the omics data signals of chromatin open region data and the sequence features of enhancers and promoters for learning, and jointly infers the interaction between enhancer-promoter pairs. In the scheme of reference [6], the sequence features and chromatin open region features of enhancers and promoters are extracted respectively by convolutional neural networks, and then these features are spliced ​​and input into recurrent neural networks to predict whether there is an interaction. Although the third type of scheme comprehensively considers the sequence and chromatin open region information, it increases the demand for chromatin open region sequencing on the basis of sequence data and increases the information related to regulation, but the requirements for the type of experimental sequencing are relatively low, which not only reduces the requirements for data but also takes into account convenience.

[0007] These algorithms have made progress in the problem of enhancer-promoter interaction prediction and helped to decipher gene regulatory networks and disease mechanisms. Among them, the most typical sequence features are represented as k-mers, that is, substrings of length k contained in biological sequences. The frequency features of k-mers have been shown to be very effective in predicting a variety of biological interactions, such as lncRNA classification prediction in reference 0, RBP binding prediction in reference 0, etc., which can effectively characterize sequence features. However, these schemes still have some problems. For example, these previous algorithms have been affected by inappropriate cross-validation schemes, that is, when the training set and the test set contain the same enhancer and promoter genomic sites, by effectively memorizing the average activity associated with each site during the training process, the generated model may perform well by mistake, causing information leakage of enhancer-promoter pairs in the validation and test sets, resulting in defective data and erroneous high precision during testing.

[0008] It should be noted that this background technology is only used to introduce the relevant information of the present invention to help understand the technical solution of the present invention, but it does not mean that the relevant information is necessarily the prior art. If there is no evidence that the relevant information has been disclosed before the application date of the present invention, the relevant information shall not be regarded as the prior art.

[0009] References:

[0010] [1]Singh S,Yang Y,Póczos B,et al.Predicting enhancer-promoterinteraction from genomic sequence with deep neural networks[J].QuantitativeBiology,2019,7(2):122-137.

[0011] [2]Hong Z, Zeng

[0012] [3]Zeng W,Wu M,iang R.Prediction of enhancer-promoter interactionsvia natural language processing[J].BMC genomics,2018,19:13-22.

[0013] [4]Whalen S,Truty RM,Pollard K S.Enhancer–promoter interactions areencoded by complex genomic signatures on looping chromatin[J].Naturegenetics,2016,48(5):488-496.

[0014] [5]Cao Q, Anyansi C, Hu X, et al. Reconstruction of enhancer–target networks in 935samples of human primary cells, tissues and cell lines [J]. Nature genetics, 2017, 49(10): 1428-1436.

[0015] [6] Li W, Wong WH, Jiang R. DeepTACT: predicting 3D chromatin contacts via bootstrapping deep learning [J]. Nucleic acids research, 2019, 47(10): e60-e60.

[0016] [7]Kirk JM,Kim SO,Inoue K,et al.Functional classification of longnon-coding RNAs by k-mer content[J].Nature genetics,2018,50(10):1474-1482.

[0017] [8] Bressin A, Schulte-Sasse R, Figini D, et al. TriPepSVM: de novoprediction of RNA-binding proteins based on short amino acid motifs[J]. Nucleic acids research, 2019, 47(9): 4406-4417. Summary of the invention

[0018] Therefore, the purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a solution for constructing a predictive model of enhancer promoter regulatory network.

[0019] The objective of the present invention is achieved through the following technical solutions:

[0020] According to a first aspect of the present invention, a method for constructing a prediction model of an enhancer-promoter regulatory network is provided, wherein the prediction model is used to predict whether there is interaction between enhancers and promoters, and the method comprises: S1, obtaining an original data set, wherein the original data set contains data of multiple enhancer-promoter pairs of multiple biological samples, and dividing the original data set into multiple subsets based on chromosome pairs, wherein all enhancer-promoter pairs on the same chromosome pair are divided into the same subset; S2, preprocessing all subsets obtained in step S1, and obtaining a feature vector of each sample based on each enhancer-promoter pair, wherein each subset contains multiple data samples, and each data sample is an enhancer. The feature vector of each data sample is a feature vector formed by splicing the sequence features of the corresponding enhancer promoter pair, the distance features between the enhancer promoter pair, and the chromatin openness features corresponding to the enhancer promoter pair. The label of each data sample is whether there is interaction between the corresponding enhancer promoter pair; S3, based on all subsets after preprocessing, the category feature gradient boosting method is adopted, with the feature vector of the data sample as input and whether the enhancer promoter pair corresponding to the sample interacts as output to iteratively construct multiple symmetric decision trees to form a prediction model, wherein in each iteration, any one subset is used as a validation set and other subsets are used as training sets to train the symmetric decision tree.

[0021] Preferably, in step S1, subsets are divided as follows: S11, obtaining the number of enhancers and promoters on each chromosome pair in the data set and counting the number of enhancer-promoter pairs on each chromosome pair; wherein, enhancer-promoter pairs are enhancer-promoter pairs that interact with each other; S12, sorting all chromosome pairs according to the number of enhancer-promoter pairs thereon, and performing complementary combinations on all chromosome pairs in pairs so that the number of enhancer-promoter pairs in each combination is balanced, wherein when performing complementary combinations, the chromosome pairs at both ends of the sorting are simultaneously moved toward the middle, and each time the chromosome pairs are selected as a combination, and if there is a single chromosome pair in the middle, it forms a combination by itself; S13, taking a chromosome pair combination as a subset, and constructing multiple subsets.

[0022] Preferably, in step S2, each enhancer-promoter pair in each subset is preprocessed in the following manner: S21, obtaining basic information of enhancers and promoters: the chromosome where the enhancer is located, the enhancer position, the chromosome where the promoter is located, and the promoter position; S22, extracting enhancer sequences and promoter sequences from the genome based on the position information of enhancers and promoters; S23, obtaining k-mer sequence features corresponding to enhancer sequences and promoter sequences based on enhancer sequences and promoter sequences, respectively, thereby constructing sequence features corresponding to enhancers and promoters, respectively; S24, obtaining the center position of the enhancer and the center position of the promoter based on the position information of the enhancer and the promoter, and calculating the distance feature of the enhancer and the promoter based on the center position of the enhancer and the center position of the promoter; S25, obtaining the chromatin openness data corresponding to the enhancer and promoter regions based on the position information of the enhancer and the promoter, and calculating the chromatin openness features of the enhancer and the promoter based on this; S26, splicing the sequence features of enhancers and promoters, the distance features of enhancers and promoters, and the chromatin openness features of enhancers and promoters to obtain feature vectors.

[0023] Preferably, in step S22, the enhancer sequence and the promoter sequence are extracted in the following manner: a human gene chromosome corresponding to the chromosome where the enhancer-promoter pair is located is found in the human genome, and the center position of the enhancer and the promoter is obtained from the human gene chromosome, and the enhancer sequence is extracted after the center position of the enhancer is extended by a first preset length on the human gene chromosome, and the promoter sequence is extracted after the center position of the promoter is extended by a second preset length on the human gene chromosome.

[0024] Preferably, the first preset length is 3000 bp and the second preset length is 2000 bp.

[0025] Preferably, in step S23, the k-mer sequence features of the enhancer sequence and the promoter sequence are obtained in the following manner: the percentage frequency of each k-mer in the enhancer sequence and the promoter sequence is converted into a k-mer feature vector, and all k-mer feature vectors in each sequence constitute a multidimensional feature vector corresponding to the sequence, thereby obtaining the sequence features corresponding to the enhancer sequence and the promoter sequence, respectively.

[0026] Preferably, in step S25, the chromatin openness characteristics of the enhancer are calculated in the following manner:

[0027]

[0028] Among them, end en and start enThe DHS represents the end and start positions of the enhancer, respectively. en represents the signal of the enhancer region, m represents the number of signals in the enhancer region on the chromosome; and the chromatin openness characteristics of the promoter are calculated as follows:

[0029]

[0030] Among them, end pro and start pro Indicate the end and start positions of the promoter, DHS pro represents the signal of the promoter region, and m′ represents the number of signals in the promoter region on the chromosome.

[0031] Preferably, in step S3, the CatBoost method is used to train the symmetric decision tree.

[0032] Preferably, in step S3, the following loss is used to update the parameters of the decision tree:

[0033]

[0034] Among them, y i is the true label of the i-th sample in the training set, p i is the probability that the i-th sample is predicted as a positive sample.

[0035] Compared with the prior art, the advantages of the present invention are that the present invention has high accuracy, low experimental requirements for input data, is user-friendly, avoids the information leakage problem existing in the evaluation of most existing methods through effective data partitioning, and effectively solves the problem of extreme imbalance of positive and negative samples in prediction through weighted gradient boosting symmetric decision tree, and can achieve efficient and accurate enhancer-promoter interaction prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:

[0037] Figure 1 A schematic diagram of the process of constructing an enhancer promoter regulatory network model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0039] As described in the background technology, the solutions under the prior art still have the following problems, among which a typical common problem is the information leakage problem caused by cross-validation, that is, when the training set and the test set contain the same enhancer and promoter genomic sites, the good performance of the generated model may be wrong by effectively remembering the average signal associated with each site during the training process, because the enhancer-promoter pair information is leaked in the validation and test sets, resulting in defective data and erroneous high precision during testing.

[0040] In order to solve some problems existing in the existing solutions, the inventors proposed a new enhancer promoter regulatory network prediction model construction scheme. In the scheme of the present invention, the distance characteristics between enhancer promoters and the chromatin openness characteristics of enhancer promoters are introduced on the basis of enhancer and promoter sequence characteristics to improve the characterization ability of the characteristics, and the symmetric decision tree is trained by gradient boosting to obtain the prediction model. Among them, the inventors studied the prior art and found that k-mers is a simple and effective method in the existing methods for representing enhancer and promoter sequence characteristics. K-mer is a substring of length k contained in a biological sequence. The frequency characteristics of k-mers have been proven to be very effective in predicting a variety of biological interactions and can effectively characterize sequence characteristics. Therefore, k-mers are still used in the scheme of the present invention to characterize enhancer-promoter sequence characteristics.

[0041] In summary, according to one embodiment of the present invention, the present invention provides a method for constructing an enhancer promoter regulatory network prediction model, the prediction model is used to predict whether there is interaction between enhancers and promoters, the method comprising: S1, obtaining an original data set, the original data set containing interaction pair data of multiple chromosomes of the same biological sample, and dividing the original data set into multiple subsets based on enhancer promoter pairs, wherein all enhancer promoter pairs on the same chromosome pair are divided into the same subset; S2, preprocessing all subsets obtained in step S1 to obtain a feature vector of each data sample based on each enhancer promoter pair, wherein each subset contains multiple data samples, and each data The sample is an enhancer promoter pair, and the feature vector of each data sample is a feature vector formed by splicing the sequence features of the corresponding enhancer promoter pair, the distance features between the enhancer promoter pair, and the chromatin openness features corresponding to the enhancer promoter pair. The label of each data sample is whether there is interaction between the corresponding enhancer promoter pair; S3, based on all subsets after preprocessing, the category feature gradient boosting method is adopted, with the feature vector of the sample as input and whether the enhancer promoter pair corresponding to the sample interacts as output to iteratively construct multiple symmetric decision trees to form a prediction model, wherein in each iteration, any one subset is used as a validation set and other subsets are used as training sets to train the symmetric decision tree.

[0042] In order to better understand the solution of the present invention, the present invention is described in detail from three aspects: subset division, subset preprocessing, and model training, in conjunction with the accompanying drawings and embodiments.

[0043] 1. Subset Division

[0044] As introduced in the background technology, among the existing solutions, those that only use sequence data as model input ignore information closely related to gene regulation such as chromatin openness, which limits the prediction of enhancer-promoter interactions. Solutions that use multiple types of experimental data as features have high requirements for experimental data, and most solutions are affected by inappropriate cross-validation schemes. There is information leakage in the test and validation sets, which leads to defective data and false high precision during testing, thus affecting the final model effect.

[0045] In order to better understand the concept of information leakage, a simple example is used here to illustrate. According to whether there is interaction, enhancer-promoter pairs are divided into positive pairs and negative pairs, where positive pairs: enhancer-promoter pairs have interaction; negative pairs: enhancer-promoter pairs do not have interaction. Taking enhancer A as an example, the positive pair of enhancer A is enhancer A-promoter B, and the negative pairs are enhancer A-promoter C, enhancer A-promoter D, and enhancer A-promoter E. When data is divided during training, if the positive pair enhancer A-promoter B is divided into the training set, and the negative pairs are enhancer A-promoter C, enhancer A-promoter D, and enhancer A-promoter E are divided into the test set, information leakage will occur, making all enhancer-promoter pairs related to enhancer A in the test set more likely to be predicted as negative pairs.

[0046] Therefore, in order to avoid information leakage, in the scheme of the present invention, when performing cross-validation division, the positive pairs and negative pairs related to the same enhancer or promoter are divided into a cross-validation subset. According to one embodiment of the present invention, in the scheme of the present invention, in order to avoid inappropriate cross-validation problems, the existing data collected for training the model are re-divided, and the positive pairs and negative pairs of enhancer-promoter pairs in the same chromosome are always assigned to the same subset according to the chromosome position, so as to ensure that the positive pairs and negative pairs related to the same enhancer or promoter will not appear in different subsets at the same time to cause data leakage.

[0047] In addition, since only a small part of all possible enhancer and promoter pairs consisting of all enhancers and all promoters are truly interactive, the number of interacting enhancer and promoter pairs (positive pairs) is far less than the number of non-interacting pairs (negative pairs), and machine learning tasks are prone to typical imbalance problems (there are many enhancer-promoter pairs in the data set, and the interactions of these pairs are known, that is, whether there is an interaction between the two is known, and the number of interacting enhancer-promoter pairs is much less than the number of non-interacting enhancer-promoter pairs). In order to solve the imbalance problem, bootstrapping and oversampling are the preferred solutions of existing methods. However, these methods artificially change the true distribution of interacting samples and non-interacting samples, which not only deviates from the original situation, but also increases memory and time costs. Therefore, in the scheme of the present invention, by pairing large chromosomes with small chromosomes to maintain similar sizes of different subsets, the imbalance problem is avoided without changing the true distribution of samples.

[0048] According to one embodiment of the present invention, in the scheme of the present invention, subsets are divided as follows: the number of enhancers and promoters on each chromosome pair in the data set is obtained and the number of enhancer-promoter pairs on each chromosome pair is counted; wherein, enhancer-promoter pairs are enhancer-promoter pairs that interact with each other; all chromosome pairs are sorted according to the number of enhancer-promoter pairs thereon, and all chromosome pairs are complementarily combined in pairs so that the number of enhancer-promoter pairs in each combination is balanced, wherein, when performing complementary combinations, the chromosome pairs at both ends of the sorting are simultaneously moved toward the middle, and each time the chromosome pairs are selected as a combination, and if there is a single chromosome pair in the middle, it forms a combination by itself; a chromosome pair combination is taken as a subset, and multiple subsets are constructed.

[0049] According to one embodiment of the present invention, the present invention collects experimental data from TargetFinder and BENGI, including Hi-C, pcHi-C, CHIA-PET, eQTLs and CRISPR / Cas9 and other chromatin openness feature data sets, each of which contains multiple sub-data sets: 1. TargetFinder data set includes GM12878 (B lymphocytes), HUVEC (umbilical vein endothelial cells), HeLa-S3 (ectodermal cervical cancer cells), IMR90 (human lung fibroblasts), K562 (mesodermal leukemia cells), NHEK (epidermal keratinocytes); 2. BENGI data set includes CD34.CHiC (CD34 cells-CHiC), GM12878.CHiC (B lymphocytes-CHiC), GM12 878.CTCF-ChIAPET (B lymphocyte-ChIAPET), GM12878.HiC (B lymphocyte-HiC), GM12878.RNAPII-ChIAPET (B lymphocyte-ChIAPET), HeLa.CTCF-ChIAPET (ectodermal cervical cancer cell-ChIAPET), HeLa.HiC (ectodermal cervical cancer cell-HiC), HeLa.RNAPII-ChIAPET (ectodermal cervical cancer cell-ChIAPET), HMEC.HiC (umbilical vein endothelial cell-HiC), 'IMR90.HiC' (human lung fibroblast-HiC), K562.HiC (mesoderm leukemia cell-HiC), NHEK.HiC (epidermal keratinocyte-HiC). Among them, a sub-dataset contains multiple biological samples, and each cell type is a biological sample.

[0050] Taking the K562 cell line in the K562 subdataset as an example, there are 23 pairs of chromosomes, one of which is a sex chromosome (chr1-22+chrX). Each chromosome contains genes, which are DNA molecular fragments with genetic effects. They are the basic genetic units that control biological traits and record and transmit genetic information. All genes are composed of sequences composed of 4 bases. Each chromosome has many enhancers and promoters to regulate gene transcription. Both enhancers and promoters are regulatory elements that affect gene transcription, but their locations and modes of action are different. Among them, the promoter is located upstream of the gene coding region, where RNA polymerase II binds and starts transcription, while the enhancer is located upstream or downstream of the promoter and can exist at a long distance. They can not only increase the rate of transcription, but also adjust the timing and location of transcription. The above-mentioned sequencing technologies such as HiC, ChiC, RNAPⅡ-ChIAPET and CTCF-ChIAPET can obtain the positions of promoters and enhancers in chromosomes (such as promoter 128745bp, enhancer 168754bp) and whether the promoter and enhancer interact (the labels can be set to 0 / 1, 0 means no interaction, 1 means interaction). First, the number of enhancer-promoter positive pairs on the same pair of chromosomes measured by the corresponding experimental technology is counted, and then the chromosome pairs are sorted in ascending or descending order according to the number of positive pairs. Then, the top and bottom ranked chromosome pairs are set as complementary groups, and the chromosome pairs of complementary sizes are assigned to the same subset (for example, the chromosome pairs ranked 1 and 23 are assigned to the same group, the chromosome pairs ranked 2 and 22 are assigned to the same group, and so on, the chromosome pairs ranked 12 are a group by themselves), so that the positive pairs and negative pairs on the same chromosome are assigned to the same subset so that these groups contain approximately the same number of positive pairs.

[0051] 2. Preprocessing of Subsets

[0052] The inventors have found through research that enhancer and promoter sequence data, the distance between enhancers and promoters, and the chromatin accessibility of enhancers and promoters are closely related to gene regulation and can be used to study the interaction between enhancers and promoters. Therefore, in the scheme of the present invention, the distance characteristics of enhancers and promoters and the chromatin accessibility characteristics are added at the feature level, that is, the sequence characteristics, distance characteristics and chromatin accessibility characteristics of enhancers and promoters are used to predict whether enhancer-promoter pairs interact. Among them, as the name implies, the distance feature refers to the distance between enhancers and promoters. Chromatin accessibility reveals the degree of openness or accessibility of chromatin, and this degree of difference may have an important impact on the transcriptional activity of genes. The chromatin accessibility of promoters and enhancers indicates the degree to which promoters and enhancers can approach.

[0053] In each subset, before preprocessing, the content of each data sample includes: 1) the chromosome where the promoter is located; 2) the location of the promoter; 3) the chromosome where the enhancer is located; 4) the location of the enhancer. The sample label is whether the promoter and enhancer interact, marked as 0 / 1. Based on this basic information, each subset is preprocessed next:

[0054] First, the enhancer / promoter sequence and its upstream and downstream sequences were extracted, specifically, Figure 1 As shown, the corresponding chromosome and the central position of the corresponding enhancer and promoter in the chromosome are found in the human reference genome (Hg19 or Hg38), the enhancer is padded to 3000bp in length, and the promoter is padded to 2000bp in length, and then the genome data analysis tool bedtools is used to extract the sequences of the enhancer and promoter locations from the human genome.

[0055] Then, the enhancer sequence and the promoter sequence are processed to obtain sequence features. Specifically, in the scheme of the present invention, reference is still made to Figure 1 , the percentage frequency of each k-mer (k-mer is a sequence fragment of length k) in the enhancer and promoter sequence is calculated as a k-mer feature vector. Taking k as 3 as an example, a substring of length 3 has four possibilities of ATGC at each position, that is, the total possibility is 4 to the power of 3, 64 possible substrings, such as AAA, AAT, AAG, AAC, ATA, ATT, ATG, ATC, etc. In the sequence region of the enhancer or promoter, the percentage frequency of each substring appearing in this short sequence can be counted, thereby encoding the sequence feature into a 64-dimensional vector, each dimension corresponds to a substring, and the value of each dimension is the percentage frequency of the substring. Similarly, if k is 4, and the length of the substring is 4, there are 4 possibilities of ATGC at each position, that is, the total possibility is 4 to the fourth power, 256 possible substrings, such as AAAA, AATA, AAGA, AACA, ATAA, ATTA, ATGA, ATCA, etc. In the sequence region of the enhancer or promoter, the percentage frequency of each substring in this short sequence can be counted, thereby encoding the sequence features into a 256-dimensional vector, each dimension corresponds to a substring, and the value of each dimension is the percentage frequency of the substring. The same is true for substrings of other lengths, which will not be repeated here. The k-mer feature vectors of enhancers and promoters are normalized to obtain sequence features.

[0056] Next, the chromatin accessibility of enhancers and promoters is calculated respectively, that is, the DNase (chromatin openness feature) signals of enhancers and promoters are obtained and added to the feature vector, wherein the chromatin openness feature of enhancers is calculated in the following way:

[0057]

[0058] Among them, end en and start en The DHS represents the end and start positions of the enhancer, respectively. en represents the signal of the enhancer region, m represents the number of signals in the enhancer region on the chromosome; and the chromatin openness characteristics of the promoter are calculated as follows:

[0059]

[0060] Among them, end pro and start pro Indicate the end and start positions of the promoter, DHS pro represents the signal of the promoter region, and m′ represents the number of signals in the promoter region on the chromosome.

[0061] Next, the distance between the center positions of the enhancer and the promoter is added to the feature vector, wherein the center position of the enhancer and the center position of the promoter are obtained based on the basic information of the enhancer and the promoter, respectively, and the distance feature of the enhancer and the promoter is calculated based on the center position of the enhancer and the center position of the promoter.

[0062] Finally, the two percentage-normalized k-mer feature vectors, namely the sequence features of enhancers and promoters, the distance features of enhancers and promoters, and the chromatin openness features of enhancers and promoters, are concatenated to obtain the feature vector.

[0063] 3. Model Training

[0064] According to one embodiment of the present invention, the present invention adopts a symmetric decision tree based on gradient boosting to predict whether enhancer-promoter pairs interact with each other, and uses a categorical feature boosting algorithm package (CatBoost) to achieve this. Because the inventors found in the study of enhancer-promoter interactions that compared with the traditional gradient boosting decision tree algorithm, CatBoost has made a number of innovative optimizations at the model training level. In particular, when processing categorical features, CatBoost introduces an order-based target encoding method to avoid the common high-dimensional problems and overfitting risks of categorical features. CatBoost also has a unique ability to handle data partial order problems, and can effectively prevent target leakage in training data. In addition, CatBoost uses a symmetric tree structure, which makes model training and prediction more efficient while reducing memory consumption.

[0065] Still reference Figure 1 , by inputting the feature vector into the classifier of the categorical feature boosting algorithm package (CatBoost), the mutual and non-mutual enhancer and promoter pairs are distinguished by learning the weighted gradient boosting symmetric decision tree. CatBoost is a supervised machine learning method and a gradient boosting decision tree algorithm specifically for categorical feature optimization. CatBoost has two main functions. It is suitable for categorical data (Category) and uses gradient boosting (Boost). It is a gradient boosting framework based on symmetric decision trees (oblivious trees) as the base learner with fewer parameters, support for categorical variables and high accuracy. The decision tree model is essentially a tree composed of multiple judgment nodes. Parameter judgment is made at each node of the tree, and then the last branch (leaf node) of the tree can make the best judgment on the value of the variable of concern. Usually, a decision tree contains a root node, several internal nodes and several leaf nodes, and the leaf nodes correspond to the decision classification results. The branches make judgments and the leaves draw conclusions. In the symmetric decision tree, the judgment conditions of each node in each layer are the same, and the fitting mode is relatively simple, which can improve the prediction speed and achieve fast prediction. The core idea of ​​the gradient boosting algorithm is to iteratively build and combine multiple decision trees to gradually reduce the prediction residual, thereby improving the overall performance of the model. In the process of iteratively building decision trees, each subsequent tree will improve the results of the previous tree to obtain better results. For example, the fitted decision tree may make decisions based on the distance feature in a certain layer of nodes. If the distance is greater than a certain value, it will be decided as the left node of the next layer, and if it is less than or equal to a certain value, it will be decided as the right node of the next layer.

[0066] According to one embodiment of the present invention, in the present invention, multiple subsets of the same biological sample after preprocessing are divided into multiple cross-validation groups, one cross-validation group is used for testing, and the other cross-validation groups are used to build decision trees; multiple decision trees are iteratively constructed, and each decision tree corrects the predicted value of the previous decision tree. During the training process, the number of iterations is 1000, the depth of the tree is 10, and the learning rate is 0.1. Since the technical theory of CatBoost is well known to those skilled in the art, CatBoost will not be described in detail here.

[0067] In the solution of the present invention, in order to solve the imbalance problem, by setting the parameter 'scale_pos_weight' in the CatBoost algorithm package to the initial imbalance ratio, the weighted decision tree is adjusted to help the model focus on a small number of data samples, solve the imbalance problem without changing the distribution of the data itself, and achieve fast and accurate prediction. The loss function is the logarithmic loss function (Logloss), that is, the log-likelihood loss, and the expression is:

[0068]

[0069] where y i is the true label of the i-th data sample, p i is the probability that the i-th sample is predicted to be a positive sample.

[0070] It can be seen from the above embodiments that the present invention aims to solve the limitations of existing enhancer-promoter interaction prediction technologies in terms of high requirements for experimental data in terms of input features, ignoring epigenetic information such as chromatin openness, and information omissions in cross-validation schemes, and proposes a new scheme that comprehensively considers the sequence characteristics, distance characteristics and chromatin accessibility characteristics of enhancers and promoters, and uses a weighted gradient boosting symmetric decision tree algorithm to solve the imbalance problem and redistribute subsets, thereby improving the accuracy and reliability of enhancer-promoter interaction prediction.

[0071] Compared with the prior art, the scheme of the present invention has the following beneficial effects: 1. Enhancer-promoter interaction prediction is performed by comprehensively considering the sequence features, distance features and chromatin accessibility features of enhancers and promoters; distance features and chromatin accessibility features improve the prediction accuracy of promoter-enhancer interaction, and the acquisition method of the three types of features is user-friendly and does not require high types of experimental data; 2. Positive and negative pairs of enhancer-promoter interactions on the same chromosome are assigned to the same subset by chromosome position, and large chromosomes are paired with small chromosomes to control the similar sizes of different subsets, effectively avoiding information leakage of verification and test sets; 3. A category-based feature boosting algorithm package (CatBoost, Category Boosting) is used to implement a weighted gradient boosting symmetric decision tree, and a penalty term is set to make the model pay more attention to data samples that are classified into the wrong category; the problem of extremely unbalanced positive and negative samples in the prediction of enhancer-promoter interaction is effectively solved, and the prediction accuracy of promoter-enhancer interaction is improved.

[0072] In order to verify the effect of the present invention, the present invention compared the performance of the model constructed by the scheme of the present invention with other models in different biological samples on different data sets, and used the area under the precision-recall curve (AUPR) as a performance indicator for evaluating the present invention, wherein AUPR is a classic indicator for evaluating model performance on imbalanced data.

[0073] Among them, Table 1 is a comparison of the AUPR of the model constructed by the solution of the present invention (represented by EPnet) with the TargetFinder, EP2vec, SPEID, and EPIVAN models in different biological samples on the TargetFinder dataset. It can be seen that the average AUPR of EPnet on the TargetFinder dataset is 0.81, which is higher than the comparison methods TargetFinder, EP2vec, SPEID, and EPIVAN.

[0074] Table 1

[0075]

[0076]

[0077] Table 2 is a comparison of the AUPR of the model constructed by the solution of the present invention (represented by EPnet) with the TargetFinder, EP2vec, SPEID, and EPIVAN models in different biological samples on the BENGI dataset. It can be seen that the average AUPR of EPnet on BENGI is 0.5841, which is also higher than the comparison method.

[0078] Table 2

[0079]

[0080] It should be noted that although the above describes the various steps in a specific order, it does not mean that the various steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order as long as the required functions can be achieved.

[0081] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0082] Computer readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. Computer readable storage medium can include, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, for example, a punch card or a protruding structure in a groove on which instructions are stored, and any suitable combination thereof.

[0083] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for constructing a prediction model of an enhancer-promoter regulatory network, wherein the prediction model is used to predict whether there is interaction between enhancers and promoters, characterized in that: The method comprises: S1. Obtain an original data set, wherein the original data set contains data of multiple enhancer-promoter pairs of multiple biological samples, and divide the original data set into multiple subsets based on chromosome pairs, wherein all enhancer-promoter pairs on the same chromosome pair are divided into the same subset; S2. Preprocess all subsets obtained in step S1, and obtain a feature vector for each sample based on each enhancer-promoter pair, wherein each subset contains multiple data samples, each data sample is an enhancer-promoter pair, and the feature vector of each data sample is a feature vector formed by splicing the sequence feature of the corresponding enhancer-promoter pair, the distance feature between the enhancer-promoter pair, and the chromatin openness feature corresponding to the enhancer-promoter pair, and the label of each data sample is whether there is interaction between the corresponding enhancer-promoter pair; S3. Based on all the preprocessed subsets, the categorical feature gradient boosting method is used, with the feature vector of the data sample as input and whether the enhancer-promoter pair corresponding to the sample interacts as output to iteratively construct multiple symmetric decision trees to form a prediction model. In each iteration, any one subset is used as a validation set and the other subsets are used as training sets to train the symmetric decision tree.

2. The method according to claim 1, characterized in that In step S1, the subsets are divided as follows: S11, obtaining the number of enhancers and promoters on each chromosome pair in the data set and counting the number of enhancer-promoter pairs on each chromosome pair; wherein enhancer-promoter pairs are enhancer-promoter pairs that interact with each other; S12, sorting all chromosome pairs according to the number of enhancer promoter pairs on them, and performing complementary combinations on all chromosome pairs in pairs so that the number of enhancer promoter pairs in each combination is balanced, wherein when performing complementary combinations, the chromosome pairs at both ends of the sorting are simultaneously moved closer to the middle, and each time the chromosome pairs at both ends are selected as a combination, and if there is a single chromosome pair in the middle, it will form a combination by itself; S13. Take a chromosome pair combination as a subset and construct multiple subsets.

3. The method according to claim 2, characterized in that In step S2, each enhancer-promoter pair in each subset is preprocessed as follows: S21. Obtain basic information of enhancers and promoters: chromosome where the enhancer is located, enhancer position, chromosome where the promoter is located, promoter position; S22, extracting enhancer sequences and promoter sequences from the genome based on the position information of enhancers and promoters; S23, respectively obtaining k-mer sequence features corresponding to the enhancer sequence and the promoter sequence based on the enhancer sequence and the promoter sequence, thereby respectively constructing sequence features corresponding to the enhancer and the promoter; S24, obtaining the center position of the enhancer and the center position of the promoter respectively based on the position information of the enhancer and the promoter, and calculating the distance feature between the enhancer and the promoter based on the center position of the enhancer and the center position of the promoter; S25, based on the position information of the enhancer and the promoter, respectively obtaining the chromatin openness data corresponding to the enhancer and the promoter region, and based on this, respectively calculating the chromatin openness characteristics of the enhancer and the promoter; S26. The sequence features of enhancers and promoters, the distance features of enhancers and promoters, and the chromatin openness features of enhancers and promoters are spliced ​​to obtain a feature vector.

4. The method according to claim 3, characterized in that In step S22, the enhancer sequence and the promoter sequence are extracted in the following manner: A human gene chromosome corresponding to the chromosome where the enhancer-promoter pair is located is found in the human genome, and the center positions of the enhancer and the promoter are obtained from the human gene chromosome. The enhancer sequence is extracted after the first preset length is extended forward and backward on the human gene chromosome based on the center position of the enhancer. The promoter sequence is extracted after the second preset length is extended forward and backward on the human gene chromosome based on the center position of the promoter.

5. The method according to claim 4, characterized in that The first preset length is 3000 bp, and the second preset length is 2000 bp.

6. The method according to claim 3, characterized in that In step S23, the k-mer sequence features of the enhancer sequence and the promoter sequence are obtained in the following manner: The percentage frequency of each k-mer in the enhancer sequence and the promoter sequence is converted into a k-mer feature vector, and all the k-mer feature vectors in each sequence constitute the multidimensional feature vector corresponding to the sequence, thereby obtaining the sequence features corresponding to the enhancer sequence and the promoter sequence, respectively.

7. The method according to claim 3, characterized in that In step S25, the chromatin openness characteristics of the enhancer are calculated in the following manner: Among them, end en and start en The DHS represents the end and start positions of the enhancer, respectively. en represents the signal of the enhancer region, and m represents the number of signals in the enhancer region on the chromosome; And the chromatin openness characteristics of the promoter are calculated as follows: Among them, end pro and start pro Indicate the end and start positions of the promoter, DHS pro represents the signal of the promoter region, and m′ represents the number of signals in the promoter region on the chromosome.

8. The method according to claim 3, characterized in that In step S3, the CatBoost method is used to train the symmetric decision tree.

9. The method according to claim 8, characterized in that In step S3, the following loss is used to update the parameters of the decision tree: Among them, y i is the true label of the i-th sample in the training set, p i is the probability that the i-th sample is predicted as a positive sample.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the method according to any one of claims 1 to 9.

11. An electronic device, characterized in that: include: one or more processors; as well as A memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method of any one of claims 1 to 9 by executing the executable instructions.

Images