Methods of predicting cell type specific enhancer-promoter interactions

Abstract

The invention provides a method for predicting cell type specific enhancer-promoter interaction, which solves the problems of predicting cell type specific interaction and the like, and comprises the following steps: S1, downloading data; s2, data preprocessing; s3, random forest prediction; s4, feature importance analysis; s5, evaluating indexes; s6, carrying out cross validation in the cell line; s7, performing cell line inspection; and S8, feature importance sorting. The method has the advantages of good cell type specificity prediction effect, high applicability and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical information processing, and in particular relates to a method for predicting cell type-specific enhancer-promoter interactions. Background Art

[0002] Existing researchers have explained the possible molecular mechanisms of chromatin interaction formation from different perspectives and developed computational models to predict long-range interactions on chromatin. The algorithms proposed by predecessors can be divided into three categories according to the different features used: the first category is the method of identifying enhancer-promoter interactions based on genomic sequence features, such as the computational method SPEID of deep learning models, the algorithm PEP of gradient boosting decision trees, and the attention neural network model EPIANN; the second category is the method of identifying enhancer-promoter interactions based on genomic epigenetic modification features, such as the supervised machine learning framework RPPLE, EPIP, the enhanced tree model TargetFinder, and the integrated machine learning model LoopPredictor; the third category is the method of identifying enhancer-promoter interactions based on correlation, such as JEME and FOCS.

[0003] Studies have shown that enhancer-promoter interactions have stronger cell type specificity, that is, the same pair of enhancers and promoters exist in the form of chromatin loops in some cell lines, but not in other cell lines. Due to the high specificity of this interaction, among the three prediction methods mentioned above, the first method only considers sequence information, ignores the influence of epigenetic modifications, and does not explain the tissue specificity of the interaction enough, resulting in the prediction model established in a certain cell line being difficult to accurately predict on other cell line data. The second method has richer input features, so it generally has better prediction performance, but it is difficult to generalize due to the large number of constraints on input features (few cell lines can obtain enough characteristic signals). The third method first analyzes various correlations between enhancers and promoters (gene expression, chromatin openness, histone modification, etc.) based on almost all available cell line samples to obtain putative enhancer-promoter pairs, and then uses specific cell line-specific epigenetic modifications or expression signals to predict cell type-specific enhancer-promoter interactions.

[0004] But existing computational methods are suboptimal at predicting cell-type-specific interactions. Summary of the invention

[0005] The object of the present invention is to provide a method for predicting cell type-specific enhancer-promoter interactions with a reasonable design and good prediction effect of cell type-specific interactions in order to solve the above problems.

[0006] To achieve the above object, the present invention adopts the following technical scheme: The method for predicting cell type-specific enhancer-promoter interactions comprises the following steps:

[0007] S1: Data download;

[0008] S2: data preprocessing;

[0009] S3: Random Forest Prediction;

[0010] S4: Feature importance analysis;

[0011] S5: Evaluation indicators;

[0012] S6: Cross-validation within cell lines;

[0013] S7: Inter-cell line testing;

[0014] S8: Feature importance ranking.

[0015] In the above method for predicting cell type-specific enhancer-promoter interactions, step S1 includes Neu, Ery, Mac0, Mac1, Mac2, Mon, Mk, nCD4, nCD8 human blood cell Promter Capture Hi-C data and DNA methylation and histone modification H3K4me1, H3K4me3, H3K9me3, H3K27me3, H3K36me3, H3K27ac data.

[0016] In the above method for predicting cell type-specific enhancer-promoter interactions, step S2 comprises the following steps:

[0017] S21: According to Ensembl v.75, the region 0.5 kb upstream and 2 kb downstream of the transcription start point was defined as the promoter region; enhancers were defined based on the Regulatory Build and trimmed or extended to 2.5 kb while keeping the original center unchanged;

[0018] S22: Match promoters and enhancers with the two ends of the loop in the original dataset. If and only if one end of the loop overlaps with the promoter of a gene and the other end overlaps with an annotated enhancer, retain it as a candidate loop;

[0019] S23: The enhancer and promoter regions were divided into five fragments of 0.5 kb each, and the histone modifications H3K4me1, H3K4me3, H3K9me3, H3K27me3, H3K27ac, H3K36me3 and DNA methylation modification signals on each genomic fragment were quantified.

[0020] In the above method for predicting cell type-specific enhancer-promoter interactions, step S3 comprises the following steps:

[0021] S31: Apply the Bagging algorithm to the data set to perform multiple random samplings without removing the samples that have been sampled, obtain a new training data set with the same sample size as the original data, and use it to train a single decision tree;

[0022] S32: Randomly select some features in the new training data set and split each node of the CART decision tree according to the Gini coefficient until the Gini index is zero or there are no features to calculate and the node can no longer be split;

[0023] S33: repeating steps S31 and S32, and finally forming a random forest consisting of a specified number of decision trees, where each tree in the forest grows freely without pruning;

[0024] S34: Input new test data. Each decision tree classifies the new sample and gives a prediction result. The results given by each tree in the forest are counted, and finally the category with the highest frequency is used as the final prediction result.

[0025] In the above method for predicting cell type-specific enhancer-promoter interactions, in step S32, the CART algorithm calculates and compares the Gini index for each selected feature, and then selects the feature with the smallest Gini index to split the node; the training data set after random data and feature screening has a total of S samples, and the samples have s 1 and 2 There are two categories, representing interaction and no interaction, assuming that the sample is s 1 The probability of is p, then the Gini index at node k is:

[0026]

[0027] If the probability is unknown, the Gini index is:

[0028]

[0029] in, and are the number of samples of the two child nodes respectively;

[0030] When using feature A j Split the node and take its possible value A according to the sample point j = a test, because it is a binary tree, the data set S is divided into S according to the response yes or no to the value. 1 and S 1 Two parts, through A j=aThe Gini index of the data S after splitting is:

[0031]

[0032] Among them, Gini (S 1 ) and Gini(S 2 ) are the Gini indices of the two new nodes.

[0033] In the above method for predicting cell type-specific enhancer-promoter interactions, step S4 applies the Gini index to an indicator for estimating the importance of a single feature;

[0034] When using feature A j When splitting at node k, the importance of the feature at node k is defined by the change in the Gini index before and after the split:

[0035]

[0036] Assume that feature A j It is used K times in the i-th tree, and the importance of this feature in the n-th tree is:

[0037]

[0038] From this we can define feature A j Importance of n trees in the entire forest:

[0039]

[0040] In the above method for predicting cell type-specific enhancer-promoter interactions, step S5 uses accuracy Acc, recall Recall, precision Precision, F1-score, and area under the ROC curve Area Under the Curve as evaluation indicators;

[0041] Acc is the proportion of correctly classified samples among all samples:

[0042]

[0043] Recall is the proportion of correctly classified samples in the positive samples to the actual positive samples:

[0044]

[0045] Precision is the ratio of the number of correctly classified positive samples to the total number of samples predicted as positive samples by the model:

[0046]

[0047] F1-score is the harmonic mean of Precision and Recall:

[0048]

[0049] Among them, TP represents the number of samples in which the positive class is correctly classified, FP represents the number of samples in which the negative class is misclassified, TN represents the number of samples in which the negative class is correctly classified, and FN represents the number of samples in which the positive class is misclassified;

[0050] Weighted Acc is a weighted average of the Acc of the interacting and non-interacting categories, and the weight is the proportion of each category in the total sample:

[0051] Acc w =β*Acc + +(1-β)*Acc - ;

[0052] Among them, β is the proportion of interaction samples in the total samples, Acc w is the weighted average accuracy, Acc + To predict the accuracy of the interaction, Acc - To predict the accuracy of non-interaction;

[0053] Weighted Recall is a weighted average of the Recalls of each category, with the weight being the proportion of each category in the sample:

[0054] R w =β*R + +(1-β)*R - ;

[0055] Among them, R w is the recall rate after weighted average, R + The accuracy of predicting interactions, R - To predict the accuracy of non-interaction;

[0056] Weighted Precision is the result of weighted average of Precision in each category:

[0057] P w =β*P + +(1-β)*P - ;

[0058] Among them, P w is the recall rate after weighted average, P + The accuracy of predicting interactions, P - To predict the accuracy of non-interaction;

[0059] The weighted F1-score is the weighted average of the F1-scores of each category:

[0060] F1 w =β*F1 + +(1-β)*F1 - ;

[0061] Among them, F1 w is the recall rate after weighted average, F1 + F1 is the accuracy of predicting interactions. - is the accuracy of predicting non-interactions.

[0062] In the above-mentioned method for predicting cell type-specific enhancer-promoter interactions, step S6 uses random forest to perform a 10-fold cross-validation when predicting enhancer-promoter interactions in each cell line; the data of a certain cell line is randomly divided into 10 parts, 1 part of the data is used for testing, and the remaining data is used for training. The 10 parts of the data are traversed, and finally the results of the 10 tests are averaged, and the average value is used as the estimation of the prediction accuracy of the algorithm for the cell line; during the cross-validation process, the cross-validation results in each cell line are evaluated by calculating AUC, Accw, F1w, Rw, and Pw.

[0063] In the above-mentioned method for predicting cell type-specific enhancer-promoter interactions, enhancer-promoter interactions between different cell lines are predicted in step S7, and all data of each cell line are used as training sets to train the model in turn, and the remaining 8 cell line data are used as test sets to predict the performance of the model; AUC, Accw, F1w, Rw, and Pw are used to evaluate the model established for each cell line.

[0064] In the above method for predicting cell type-specific enhancer-promoter interactions, step S8 comprises the following steps:

[0065] S81: Train the model separately using each cell line data, then generate and rank the top 10 important features for each cell line;

[0066] S82: Pearson correlation coefficients were calculated for each category and each feature in each cell line to obtain the correlation between the feature and the category;

[0067] S83: Pearson correlation coefficients were calculated between all features in each cell line and the features were clustered.

[0068] Compared with the existing technologies, the advantages of the present invention are: it excludes the influence of genomic sequence, only focuses on the influence of epigenetic modification status on the interaction, and has a better prediction effect on cell type-specific interactions; the Gini index of random forest ranks the features by importance, enhancing the interpretability of the model; and can realize the prediction of enhancer-promoter interactions across cell lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a schematic diagram of the 10-fold cross-check evaluation index in each cell line of the present invention.

[0070] Figure 2 It is a schematic diagram of cross-checking between cell lines of the present invention.

[0071] Figure 3 It is a schematic diagram of the importance ranking of features in each cell line of the present invention.

[0072] Figure 4 It is a schematic diagram of the correlation between features and categories of the present invention.

[0073] Figure 5 It is a schematic diagram of correlation and clustering among features in the Neu cell line of the present invention. DETAILED DESCRIPTION

[0074] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] like Figure 1-5 As shown, the method for predicting cell type-specific enhancer-promoter interactions realizes the prediction of enhancer-promoter interactions across cell lines based on six epigenetic modification signals, which specifically includes the following steps:

[0076] S1: Data download;

[0077] S2: data preprocessing;

[0078] S3: Random Forest Prediction;

[0079] S4: Feature importance analysis;

[0080] S5: Evaluation indicators;

[0081] S6: Cross-validation within cell lines;

[0082] S7: Inter-cell line testing;

[0083] S8: Feature importance ranking.

[0084] Specifically, step S1 includes Neu (Neutrophils), Ery (Erythroblasts), Mac0 (Macrophages M0), Mac1 (Macrophages M1), Mac2 (Macrophages M2), Mon (Monocytes), Mk (Megakaryocytes), nCD4 (Naive CD4+ T cells), nCD8 (Naive CD8+ T cells) human blood cell Promter Capture Hi-C data and DNA methylation and histone modification H3K4me1, H3K4me3, H3K9me3, H3K27me3, H3K36me3, H3K27ac data. Among them, the PCHi-C data of 9 types of human blood cells were downloaded from Open Science Framework. The processed data sets of 9 types of cells, including histone modification ChIP-seq data, DNA methylation data and gene expression RNA-seq data, were downloaded from BLUEPRINT.

[0085] In-depth, step S2 uses the CHiCAGO method based on the convolutional background model to distinguish true interactions from background noise in order to identify enhancer-promoter loops in each cell line. The original dataset defines that interactions with CHiCAGO score>5 are of high confidence. It specifically includes the following steps:

[0086] S21: According to Ensembl v.75, the region 0.5 kb upstream and 2 kb downstream of the transcription start point was defined as the promoter region; enhancers were defined based on the Regulatory Build and trimmed or extended to 2.5 kb while keeping the original center unchanged;

[0087] S22: Match promoters and enhancers with the two ends of the loop in the original dataset. If and only if one end of the loop overlaps with the promoter of a gene and the other end overlaps with an annotated enhancer, retain it as a candidate loop;

[0088] S23: The enhancer and promoter regions were divided into five fragments of 0.5 kb each, and the histone modifications H3K4me1, H3K4me3, H3K9me3, H3K27me3, H3K27ac, H3K36me3 and DNA methylation modification signals on each genomic fragment were quantified.

[0089] After the above processing, the data with CHiCAGO scores>=10 in each cell line were selected as the positive set representing enhancer-promoter interaction and added with the label "1", and the data with CHiCAGO scores=0 were selected as the negative set without interaction and added with the label "-1". The number of interaction samples, non-interaction samples and total samples in each cell line are listed in the following table:

[0090] Cell lines Interaction sample size Non-interaction sample size Total sample size Neu 4183 13476 17659 Ery 3843 9557 13400 Mac0 5556 7516 13072 Mac1 4953 9056 14029 Mac2 6179 8296 14475 Mk 5139 11322 16461 Mon 4513 7938 12451 nCD4 6747 7374 14121 nCD8 6874 6289 13163

[0091] Furthermore, step S3 uses the random forest algorithm to perform prediction analysis, including the following steps:

[0092] S31: Apply the Bagging algorithm to the data set to perform multiple random samplings without removing the samples that have been sampled, obtain a new training data set with the same sample size as the original data, and use it to train a single decision tree;

[0093] S32: Randomly select some features in the new training data set and split each node of the CART decision tree according to the Gini coefficient until the Gini index is zero or there are no features to calculate and the node can no longer be split;

[0094] S33: repeating steps S31 and S32, and finally forming a random forest consisting of a specified number of decision trees, where each tree in the forest grows freely without pruning;

[0095] S34: Input new test data. Each decision tree classifies the new sample and gives a prediction result. The results given by each tree in the forest are counted, and finally the category with the highest frequency is used as the final prediction result.

[0096] Furthermore, in step S32, the CART algorithm calculates the Gini index for each selected feature and compares them, and then selects the feature with the smallest Gini index to split the node;

[0097] The Gini index is used to measure the purity of the classification result obtained after splitting a certain feature, that is, the similarity of the categories in the two nodes after the split. The smaller the Gini index, the better the classification result. Assuming there are s categories in total, the Gini coefficient is:

[0098]

[0099] Among them, p i is the probability that the sample belongs to the i-th category;

[0100] The binary classification problem in this embodiment has a total of S samples in the training data set after random data and feature screening, and the samples have s 1 and 2There are two categories, representing interaction and no interaction, assuming that the sample is s 1 The probability of is p, then the Gini index at node k is:

[0101]

[0102] If the probability is unknown, the Gini index is:

[0103]

[0104] in, and are the number of samples of the two child nodes respectively;

[0105] When using feature A j Split the node and take its possible value A according to the sample point j = a test, because it is a binary tree, the data set S is divided into S according to the response yes or no to the value. 1 and S 1 Two parts, through A j =aThe Gini index of the data S after splitting is:

[0106]

[0107] Among them, Gini (S 1 ) and Gini(S 2 ) is the Gini index of the two new nodes. According to the above principle, a complete decision tree can be established using the CART algorithm.

[0108] In addition, when the sample features are very large, it is hoped to find features that contribute more to the classification and accurately predict the data with a small number of features. This requires calculating the contribution value of each feature to the classification and ranking their importance to select the required number of features. Step S4 uses the Gini index as an indicator to estimate the importance of a single feature;

[0109] When using feature A j When splitting at node k, the importance of the feature at node k is defined by the change in the Gini index before and after the split:

[0110]

[0111] Since quantitative features are selected from all features at each node for calculation when building a tree, a feature may be used multiple times in a tree. Suppose feature A j It is used K times in the i-th tree, and the importance of this feature in the n-th tree is:

[0112]

[0113] From this we can define feature A j Importance of n trees in the entire forest:

[0114]

[0115] At the same time, step S5 uses accuracy (Acc), recall (R), precision (P), F1-score, and area under the ROC curve (AUC) as evaluation indicators;

[0116] Acc is the proportion of correctly classified samples in all samples:

[0117]

[0118] Recall is the proportion of correctly classified samples in the positive samples to the actual positive samples:

[0119]

[0120] Precision is the ratio of the number of correctly classified positive samples to the total number of samples predicted as positive samples by the model:

[0121]

[0122] F1-score is the harmonic mean of Precision and Recall:

[0123]

[0124] Among them, TP (true positives) represents the number of positive classes correctly classified in the sample, FP (false positives) represents the number of negative classes incorrectly classified in the sample, TN (true negatives) represents the number of negative classes correctly classified in the sample, and FN (false negatives) represents the number of positive classes incorrectly classified in the sample;

[0125] Since the selected data is unbalanced between positive and negative sets, a weighted average is performed on each evaluation indicator to obtain the corresponding weighted score. The weighted Acc is the weighted average of the Acc of the two categories of interaction and non-interaction, and the weight is the proportion of each category in the total sample:

[0126] Acc w =β*Acc + +(1-β)*Acc - ;

[0127] Among them, β is the proportion of interaction samples in the total samples, Acc w is the weighted average accuracy, Acc + To predict the accuracy of the interaction, Acc - To predict the accuracy of non-interaction;

[0128] Weighted Recall is a weighted average of the Recalls of each category, with the weight being the proportion of each category in the sample:

[0129] R w =β*R + +(I-β)*R - ;

[0130] Among them, R w is the recall rate after weighted average, R + The accuracy of predicting interactions, R - To predict the accuracy of non-interaction;

[0131] Weighted Precision is the result of weighted average of Precision in each category:

[0132] P w =β*P + +(1-β)*P - ;

[0133] Among them, P w is the recall rate after weighted average, P + The accuracy of predicting interactions, P - To predict the accuracy of non-interaction;

[0134] The weighted F1-score is the weighted average of the F1-scores of each category:

[0135] F1 w =β*F1 + +(1-β)*F1 - ;

[0136] Among them, F1 w is the recall rate after weighted average, F1 + F1 is the accuracy of predicting interactions. - is the accuracy of predicting non-interactions.

[0137] It can be seen that step S6 uses random forest to perform a 10-fold cross-validation when predicting enhancer-promoter interactions in each cell line; the data of a certain cell line is randomly divided into 10 parts, 1 part of the data is used for testing, and the remaining data is used for training. The 10 parts of the data are traversed, and finally the results of the 10 tests are averaged, and this average value is used as the estimation of the prediction accuracy of the algorithm for the cell line; during the cross-validation process, the cross-validation results in each cell line are evaluated by calculating AUC, Accw, F1w, Rw, and Pw.

[0138] like Figure 1 As shown in Figure 2, in the cross-validation prediction of interactions in each cell line, each evaluation parameter was the highest in the Neu cell line, such as R w R w The value is 0.791, which means that the correct classification ratio of positive examples in this cell line is high, and the classification is more accurate than other cell lines. However, it performs poorly in the cell line Mac2, with low scores in various indicators.

[0139] Obviously, in step S7, enhancer-promoter interactions between different cell lines are predicted. All data of each cell line are used as training sets to train the model in turn, and the remaining 8 cell line data are used as test sets to predict the performance of the model; AUC, Accw, F1w, Rw, and Pw are used to evaluate the model established for each cell line.

[0140] Figure 2 As shown, the AUC and Acc generated by comparing the intra-cell line cross-validation and inter-cell line validation w 、F1 w The results of cross-validation within cell lines were significantly better than those between cell lines, and the cross-validation results within different cell lines were also quite different. This suggests that enhancer-promoter interactions, especially cell type-specific interactions, are complex.

[0141] Preferably, step S8 includes the following steps:

[0142] S81: Train the model separately using each cell line data, then generate and rank the top 10 important features for each cell line; Figure 3 As shown, four histone modifications H3K4me1, H3K4me3, H3K27ac, and H3K27me3 have high importance scores in each cell line and contribute more to the identification of interactions. In cell lines nCD4 and nCD8, the histone modification H3K9me3 also plays an important role in classification;

[0143] S82: Pearson correlation coefficients were calculated for each category and each feature in each cell line to obtain the correlation between the feature and the category; Figure 4 It can be seen that H3K4me1 and H3K27ac are highly correlated with the positive category, which is consistent with previous studies. These two categories are enhancer-specific modifications and are necessary for activating target genes; while H3K9me3 is highly correlated with the negative category, because H3K9me3 is a repressive histone modification, which is mainly related to the formation of heterochromatin. Its enrichment in enhancer and promoter regions can lead to the inactivation of regulatory elements, thus failing to form a ring structure; among them, H3K4me1 has a relatively higher correlation with the category than other features in various cell lines, indicating that this feature is better for identifying enhancer-promoter interactions in these cell lines; and in the Neu cell line, the correlation between category and feature is more obvious, and the histone modification H3K4me1 has the strongest correlation with category, indicating that this feature can be used as one of the important indicators for identifying interactions in this cell line; the correlation between DNA methylation modification and category is very weak, and its contribution to classification is small;

[0144] In addition, from Figure 4 It can also be seen that the correlations of each epigenetic modification in each cell line are very different. For example, H3K4me3_T4 and H3K27ac_T4 modifications are positively correlated with the category in Ery, Mac0, Mac1, Mac2, and Mk cell lines, but negatively correlated in Neu, Mon, nCD4, and nCD8 cell lines, which is also the reason for the poor cross-test results between cell lines;

[0145] S83: The Pearson correlation coefficients between all features were calculated in each cell line and the features were clustered. Different features at different sites in the nine cell lines had different effects on identifying enhancer-promoter interactions, but some features showed synergistic regulation of enhancer-promoter interactions. Therefore, clustering the features was beneficial for discovering synergistic regulation and reducing the amount of redundant features.

[0146] Depend on Figure 5 It can be seen that there is information redundancy in each epigenetic modification signal in their respective adjacent regions. There is a clear correlation between the two histone modifications H3K27ac and H3K4me1 in the enhancer region and the three modifications H3K27ac, H3K4me1 and H3K4me3 in the promoter region, which indicates that they have a synergistic effect on remodeling chromatin.

[0147] In summary, the principle of this embodiment is to use random forests to train on 9 types of human primary blood cells using DNA methylation and 6 histone modifications H3K4me1, H3K4me3, H3K9me3, H3K27me3, H3K36me3, and H3K27ac data. A 10-fold cross-validation is performed in each cell line, and cell type-specific enhancer-promoter interactions are predicted between cell lines. The performance of the established prediction model is then evaluated by generating a series of evaluation parameters, and the Pearson correlation coefficient is used in each cell line to generate a feature and category, feature correlation matrix, and the Gini index is used to rank the features by importance, to explore which features have certain correlations in modification and which features play an important role in identifying enhancer-promoter interactions, which will further promote the understanding of the formation mechanism of the ring structure.

[0148] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0149] Although cell line and other terms are used more frequently in this article, the possibility of using other terms is not excluded. These terms are used only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional limitation is contrary to the spirit of the present invention.

Claims

1. A method for predicting cell type-specific enhancer-promoter interactions, characterized in that The steps include: S1: Data download; S2: data preprocessing; S3: Random Forest Prediction; S4: Feature importance analysis; S5: Evaluation indicators; S6: Cross-validation within cell lines; S7: Inter-cell line testing; S8: Feature importance ranking.

2. The method for predicting cell type-specific enhancer-promoter interactions according to claim 1, characterized in that The step S1 includes Neu, Ery, Mac0, Mac 1, Mac2, Mon, Mk, nCD4, nCD8 human blood cell Promter Capture Hi-C data and DNA methylation and histone modification H3K4me1, H3K4me3, H3K9me3, H3K27me3, H3K36me3, H3K27ac data.

3. The method for predicting cell type-specific enhancer-promoter interactions according to claim 1, characterized in that The step S2 comprises the following steps: S21: According to Ensembl v.75, the region 0.5 kb upstream and 2 kb downstream of the transcription start point was defined as the promoter region; enhancers were defined based on the Regulatory Build and trimmed or extended to 2.5 kb while keeping the original center unchanged; S22: Match promoters and enhancers with the two ends of the loop in the original dataset. If and only if one end of the loop overlaps with the promoter of a gene and the other end overlaps with an annotated enhancer, retain it as a candidate loop; S23: The enhancer and promoter regions were divided into five fragments of 0.5 kb each, and the histone modifications H3K4me1, H3K4me3, H3K9me3, H3K27me3, H3K27ac, H3K36me3 and DNA methylation modification signals on each genomic fragment were quantified.

4. The method for predicting cell type-specific enhancer-promoter interactions according to claim 1, characterized in that The step S3 comprises the following steps: S31: Apply the Bagging algorithm to the data set to perform multiple random samplings without removing the samples that have been sampled, obtain a new training data set with the same sample size as the original data, and use it to train a single decision tree; S32: Randomly select some features in the new training data set and split each node of the CART decision tree according to the Gini coefficient until the Gini index is zero or there are no features to calculate and the node can no longer be split; S33: repeating steps S31 and S32, and finally forming a random forest consisting of a specified number of decision trees, where each tree in the forest grows freely without pruning; S34: Input new test data. Each decision tree classifies the new sample and gives a prediction result. The results given by each tree in the forest are counted, and finally the category with the highest frequency is used as the final prediction result.

5. The method for predicting cell type-specific enhancer-promoter interactions according to claim 4, characterized in that In step S32, the CART algorithm calculates and compares the Gini index for each selected feature, and then selects the feature with the smallest Gini index to split the node; the training data set after random data and feature screening has a total of S samples, and the samples have two categories, s1 and s2, representing interaction and no interaction, respectively. Assuming that the probability that the sample is s1 is p, the Gini index at node k is: If the probability is unknown, the Gini index is: in, and are the number of samples of the two child nodes respectively; When the feature Aj is used to split the node, its possible value A is calculated according to the sample point. j = a test, because it is a binary tree, so according to the response to the value is yes or no, the data set S is divided into two parts, S1 and S2, after A j =aThe Gini index of the data S after splitting is: Among them, Gini(S1) and Gini(S2) are the Gini indexes of the two new nodes.

6. The method for predicting cell type-specific enhancer-promoter interactions according to claim 1, characterized in that The step S4 uses the Gini index as an indicator for estimating the importance of a single feature; When using feature A j When splitting at node k, the importance of the feature at node k is defined by the change in the Gini index before and after the split: Assume that feature A j It is used K times in the i-th tree, and the importance of this feature in the n-th tree is: From this we can define feature A j Importance of n trees in the entire forest:

7. The method for predicting cell type-specific enhancer-promoter interactions according to claim 1, characterized in that The step S5 uses the accuracy Acc, recall Recall, precision Precision, F1-s core, and area under the ROC curve Area Under the Curve as evaluation indicators; Acc is the proportion of correctly classified samples among all samples: Re call is the proportion of correctly classified samples in the positive samples to the actual positive samples: Precision is the ratio of the number of correctly classified positive samples to the total number of samples predicted as positive samples by the model: F1-score is the harmonic mean of Precision and Recall: Among them, TP represents the number of samples in which the positive class is correctly classified, FP represents the number of samples in which the negative class is misclassified, TN represents the number of samples in which the negative class is correctly classified, and FN represents the number of samples in which the positive class is misclassified; Weighted Acc is a weighted average of the Acc of the interacting and non-interacting categories, and the weight is the proportion of each category in the total sample: Acc w =β*Acc + +(1-β)*Acc - ; Among them, β is the proportion of interaction samples in the total samples, Acc w is the weighted average accuracy, Acc + To predict the accuracy of the interaction, Acc - To predict the accuracy of non-interaction; Weighted Recall is a weighted average of the Recalls of each category, with the weight being the proportion of each category in the sample: R w =β*R + +(1-β)*R - ; Among them, R w is the recall rate after weighted average, R + The accuracy of predicting interactions, R - To predict the accuracy of non-interaction; Weighted Precision is the result of weighted average of Precision in each category: P w =β*P + +(1-β)*P - ; Among them, P w is the recall rate after weighted average, P + The accuracy of predicting interactions, P - To predict the accuracy of non-interaction; The weighted F1-score is the weighted average of the F1-scores of each category: F1 w =β*F1 + +(1-β)*F1 - ; Among them, F1 w is the recall rate after weighted average, F1 + F1 is the accuracy of predicting interactions. - is the accuracy of predicting non-interactions.

8. The method for predicting cell type-specific enhancer-promoter interactions according to claim 1, characterized in that: The step S6 uses random forest to perform a 10-fold cross-validation when predicting enhancer-promoter interactions in each cell line; the data of a certain cell line is randomly divided into 10 parts, 1 part of the data is used for testing, and the remaining data is used for training. The 10 parts of the data are traversed, and finally the results of the 10 tests are averaged, and the average value is used as the estimation of the prediction accuracy of the algorithm for the cell line; during the cross-validation process, the cross-validation results in each cell line are evaluated by calculating AUC, Accw, F1w, Rw, and Pw.

9. The method for predicting cell type-specific enhancer-promoter interactions according to claim 1, characterized in that: In the step S7, enhancer-promoter interactions between different cell lines are predicted, and all the data of each cell line are used as training sets to train the model in turn, and the remaining 8 cell line data are used as test sets to predict the performance of the model; AUC, Accw, F1w, Rw, and Pw are used to evaluate the model established for each cell line.

10. The method for predicting cell type-specific enhancer-promoter interactions according to claim 1, characterized in that: The step S8 includes the following steps: S81: Train the model separately using each cell line data, then generate and rank the top 10 important features for each cell line; S82: Pearson correlation coefficients were calculated for each category and each feature in each cell line to obtain the correlation between the feature and the category; S83: Pearson correlation coefficients were calculated between all features in each cell line and the features were clustered.

Images