Method for predicting CTCF chromatin loop on basis of transcription factor sequence in loop extrusion model

Abstract

A method for predicting a CTCF chromatin loop on the basis of a transcription factor sequence in a loop extrusion model, comprising: constructing a data set; converting a key transcription factor sequence; establishing a CTCF chromatin loop classification model on the basis of the key transcription factor sequence; and evaluating the model. In the method, a transcription factor binding order is encoded into a transcription factor sequence, and the transcription factor sequence is decoded by utilizing a pre-trained natural language BERT model, thereby achieving the prediction of the CTCF chromatin loop. The method not only improves the prediction accuracy but also enhances the interpretability of a predictive model.

Description

Method for predicting CTCF chromatin loop based on transcription factor sequence in loop extrusion model TECHNICAL FIELD

[0001] The present application relates to the technical field of healthcare informatics, in particular to a method for predicting CTCF chromatin loop based on transcription factor sequence in loop extrusion model. BACKGROUND

[0002] Existing methods for predicting CTCF chromatin loop include machine learning methods based on sample genomic sequences and epigenomic features, and mathematical modeling methods (LC model). Based on complex genomic sequences and epigenomic features as input, simply relying on the learning ability of machine learning or deep learning technology, there is a problem of black box, and the interpretability is limited. Based on a simple mathematical model, it is necessary to manually select and construct features, which may miss important information, resulting in inaccurate prediction results. SUMMARY

[0003] The purpose of the present application is to provide a method for predicting CTCF chromatin loop based on transcription factor sequence in loop extrusion model, aiming at the chromatin loop prediction method with poor feature pattern interpretability at present.

[0004] To achieve the above purpose, the following technical scheme is adopted in the present application: the method for predicting CTCF chromatin loop based on transcription factor sequence in loop extrusion model comprises the following steps:

[0005] S1, constructing a data set;

[0006] S2, key transcription factor sequence conversion;

[0007] S3, establishing a CTCF chromatin loop classification model based on key transcription factor sequence;

[0008] S4, model evaluation.

[0009] In step S1, chromatin loops are detected by using Hi-C experimental data, and the coordinates of transcription factors in the genome are obtained by ChIP-seq data of transcription factors in cell differentiation stages.

[0010] In the above method, step S1 comprises:

[0011] S11, obtaining all CTCF combination patterns within 1Mb;

[0012] S12, screening chromatin loops verified by Hi-C experiments in all CTCF pairs as positive set;

[0013] S13, removing the positive set and trash data, and the remaining is the negative set pool.

[0014] In the above method, step S11 comprises:

[0015] S111, first combine all CTCF within each chromosome pairwise within the SC-βorg stage of pancreatic cell differentiation;

[0016] In the above method, step S12 comprises:

[0017] S121, remove all CTCF pairs outside the 1Mb range, keep the start coordinate of the latter CTCF-end coordinate of the former CTCF, and call this dataset 1mb_ctcf_pairs, then overlap all 1mb_ctcf_pairs with Hi-C data to find the region that contains both 1mb_ctcf_pairs and has overlap in Hi-C data;

[0018] S122, select the CTCF binding site with the highest peak in the overlapping region, defined as the positive sample;

[0019] In the above method, step S13 comprises:

[0020] S131, the remaining overlapping region data is put into the trash database;

[0021] S132, remove the positive set and trash data from all 1mb_ctcf_pairs, and the remaining is the negative set pool.

[0022] In the above method, step S2 comprises:

[0023] S21, download the motif of CTCF: MA0139.1.transfac on the JASPAR database;

[0024] S22, obtain directional data by combining the FASTA file of CTCF genomic sequence;

[0025] S23, integrate the narrowpeak files of 5 transcription factors with Hi-C loop data through bedtools tool to generate a file containing transcription factor information and overlapping regions;

[0026] S24, sort and process the data to construct the arrangement of key transcription factors within the chromatin loop.

[0027] In step S3, when establishing the classification model, 80% of the samples are randomly selected as the training set and 20% as the independent test set, and the pre-trained model BERT-Base-Uncased is used for binary classification.

[0028] In the above method, the BERT is a natural language processing model, and the BERT-Base-Uncased is a specific variant of the BERT model.

[0029] In the above method, the model establishment comprises:

[0030] S31, a pre-processing function is defined, tokenization is performed on the sequence, and fixed-length padding is used.

[0031] S32, training parameters and evaluation indexes are defined.

[0032] Through the above steps, a CTCF chromatin loop classification model based on the BERT-Base-Uncased is established, and the key transcription factor sequence in the chromatin loop is classified through the classification model.

[0033] In the above method, step S4 comprises:

[0034] S41, the prediction probability and the real label of the LC model and the BERT-Base-Uncased are obtained, and the AUROC and AUPRC values are calculated;

[0035] S42, the logic regression of TPR and FPR is calculated based on different threshold values.

[0036] S43, the FPR is taken as the x-axis, and the TPR is taken as the y-axis, and the ROC curve is drawn.

[0037] S44, the performance of the classifier is measured, and the accuracy of the classification result is judged.

[0038] In step S41, the formula for calculating the AUROC and AUPRC values is:

[0039] Wherein, TP, FP, TN and FN represent the quantity of true positive, false positive, true negative and false negative respectively, TPR is the true positive rate, that is, the proportion of positive samples correctly identified as positive to all positive samples, and FPR is the false positive rate, that is, the proportion of negative samples incorrectly identified as positive to all negative samples.

[0040] Compared with the prior art, the advantages of the present application are:

[0041] 1. Only the key transcription factor sequence is used as input, and the prediction of the CTCF chromatin loop can be realized. It can be popularized to other chromatin structure and transcription factor research fields, and provide new tools and ideas for genomics and epigenomics research.

[0042] 2. By the innovative feature representation method, combining the pre-training model of natural language for fine-tuning, the accuracy of CTCF chromatin loop prediction is significantly improved.

[0043] 3. Innovatively encode the transcription factor binding sequence into the transcription factor sequence, use the pre-training BERT model of natural language to decode the transcription factor sequence, realize the prediction of CTCF chromatin loop, improve the prediction accuracy, and enhance the interpretability of the prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0044] Fig. 1 is a schematic diagram of the loop extrusion model in the present application;

[0045] Fig. 2 is a data processing flowchart in the present application;

[0046] Fig. 3 is a schematic diagram of the results of testing the Bert-Base-Uncased model parameters of the training set fine-tuning on the independent test set. DETAILED DESCRIPTION

[0047] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0048] The method for predicting CTCF chromatin loop based on the transcription factor sequence in the loop extrusion model comprises the following steps:

[0049] S1, constructing a data set;

[0050] S2, key transcription factor sequence conversion;

[0051] S3, establishing a CTCF chromatin loop classification model based on the key transcription factor sequence;

[0052] S4, model evaluation.

[0053] In step S1, chromatin loops are obtained by using Hi-C experimental data detection, and the location coordinates of transcription factors in the genome are obtained by ChIP-seq data of transcription factors in the SC-beta org stage of pancreatic cell differentiation.

[0054] Loop extrusion model is a widely recognized model for explaining the molecular mechanism of cohesin-mediated chromatin interaction. According to this model, specific protein complexes such as CTCF and cohesin bind to specific sites of chromatin and actively extrude chromatin to form a loop structure, and finally WAPL is responsible for releasing cohesin from chromatin, thus ending the formation of chromatin loop.

[0055] As shown in FIG. 1, according to the loop extrusion model, chromatin architectural proteins CTCF and cohesin bind at specific sites of the genome, actively extrude chromatin to form loop structures with the participation of ATP, and finally the unloading protein WAPL is responsible for releasing cohesin from chromatin, ending the extrusion of the chromatin loop.

[0056] The data processing flow is shown in FIG. 2:

[0057] a CTCF is combined in pairs to form a CTCF pairs data set;

[0058] b Obtain all CTCF combination patterns within a 1Mb range;

[0059] c Distribution of positive and negative sets;

[0060] d Source of key transcription factor sequences: Hi-C loop and narrowpeak file of key transcription factors.

[0061] e After the positive and negative sets (unbalanced data sets) are input into the Bert-Base-Uncased model, AUROC and AUPRC are used as evaluation indicators.

[0062] Step S1 construction of positive and negative sets in the data set includes the following steps:

[0063] S11, obtain all CTCF combination patterns within a 1Mb range;

[0064] S111, combine all CTCFs in the pancreatic cell differentiation SC-βorg stage in pairs within each chromosome;

[0065] S12, screen chromatin loops verified by Hi-C experiments among all CTCF pairs as positive sets;

[0066] S121, remove all CTCF pairs outside the 1Mb range, keep the start coordinate of the latter CTCF-the end coordinate of the former CTCF, and call this data set 1mb_ctcf_pairs, then perform overlap analysis on all 1mb_ctcf_pairs and Hi-C data to find regions that are both contained in 1mb_ctcf_pairs and have overlap in Hi-C data;

[0067] S122, select the CTCF binding site with the highest peak value in the overlapping region as the positive sample;

[0068] S13, remove the positive set and the trash data, and the remaining is the negative set pool;

[0069] S131, the remaining overlapping region data is put into the Bin Set trash database;

[0070] S132, remove positive set and trash data from all 1mb_ctcf_pairs, the remaining is the negative set pool.

[0071] The sample size of the above data set is shown in Table 1:

[0072] Table 1 - Sample size summary

[0073] As shown in Table 2, each transcription factor has a corresponding letter representation, where CTCF+ and CTCF- represent the forward and reverse of CTCF binding sites, which plays an important role in the formation of chromatin loops, affecting the three-dimensional structure of the genome and gene regulation network.

[0074] Table 2 - Letter representation of key transcription factors

[0075] The operation of obtaining the direction of CTCF using the Strom tool is as follows:

[0076] S21, download the motif of CTCF on the JASPAR database: MA0139.1.transfac;

[0077] S22, obtain directionality data combined with the FASTA file of the CTCF genomic sequence;

[0078] S23, integrate the narrowpeak files of the five transcription factors with the Hi-C loop data using the bedtools tool to generate a file containing transcription factor information and overlapping regions;

[0079] S24, sort and process the data to construct the arrangement of key transcription factors within the chromatin loop.

[0080] In step S3, when establishing the classification model, 80% of the samples are randomly selected as the training set and 20% as the independent test set, and the pre-trained model of BERT-Base-Uncased is used for binary classification.

[0081] Wherein, BERT is a natural language processing model, and BERT-Base-Uncased is a specific variant of the BERT model.

[0082] Model establishment includes:

[0083] S31, define a preprocessing function to tokenize the sequence and use fixed length padding;

[0084] S32, define training parameters and evaluation metrics.

[0085] Through the above steps, a CTCF chromatin loop classification model based on BERT-Base-Uncased is established, and the key transcription factor sequence in the chromatin loop is classified through the classification model.

[0086] Step S4 includes:

[0087] S41, calculate the AUROC and AUPRC values of the prediction probability and true label of the LC model and the BERT-Base-Uncased model;

[0088] S42, calculate the TPR and FPR of the logistic regression based on different thresholds;

[0089] S43, take FPR as the x-axis and TPR as the y-axis to draw the ROC curve;

[0090] S44, measure the performance of the classifier to determine the accuracy of the classification result.

[0091] As shown in FIG. 3, the left graph is the ROC curve, the horizontal axis is the false positive rate, and the vertical axis is the true positive rate. The blue curve in the figure represents the LC model, and the orange curve represents the Bert-base-uncased model. A higher AUROC indicates that the classification performance of the Bert-base-uncased model is better than that of the LC model. The right graph is the PRC curve, the horizontal axis is the recall rate, and the vertical axis is the precision. A higher AUPRC also indicates that the Bert-base-uncased model performs better when dealing with an unbalanced data set.

[0092] The area under the ROC curve is called AUROC, which can measure the performance of the classifier and determine the accuracy of the classification result. Due to the unbalanced data set of this experiment, the positive and negative set ratio is 1:65, and the positive samples are much less than the negative samples. The PRC curve is used to evaluate the model performance, and the horizontal axis of the PRC curve is the recall rate and the vertical axis is the precision.

[0093] In step S41, the formula for calculating AUROC and AUPRC values is:

[0094] Where TP, FP, TN and FN represent the number of true positives, false positives, true negatives and false negatives, respectively. TPR is the true positive rate, i.e. the proportion of positive samples correctly identified as positive to all positive samples. FPR is the false positive rate, i.e. the proportion of negative samples incorrectly identified as positive to all negative samples.

[0095] The principle of the embodiment is that: by collecting 5 transcription factors in the SC-beta org stage of pancreatic cell differentiation, namely: CTCF, NIBPL, WAPL, YY1, RAD21, wherein CTCF has directionality, ChIP-seq data of CTCF+ and CTCF- are recorded, the distribution of different transcription factors in the genome is described, each transcription factor is represented by different letters, and the transcription factor sequence is coded according to the binding position; whether any two CTCF binding sites form a chromatin loop is marked by using Hi-C experimental data; and a Bert pre-training model is used to predict the CTCF chromatin loop.

Claims

1. A method for predicting CTCF chromatin loops based on transcription factor sequences in loop extrusion model, characterized in that, The method comprises: S1, constructing a data set; S2, key transcription factor sequence transformation; S3, establishing a CTCF chromatin loop classification model based on the key transcription factor sequence; S4, model evaluation.

2. The method of claim 1, wherein, In step S1, chromatin loops are detected using Hi-C experimental data, and the location coordinates of transcription factors in the genome are obtained through ChIP-seq data of transcription factors in the SC-betaorg stage of pancreatic cell differentiation.

3. The method of claim 1, wherein, Step S1 includes: S11, obtaining all CTCF combination patterns within a 1Mb range; S12, screening chromatin loops verified by Hi-C experiments among all CTCF pairs as a positive set; S13, removing the positive set and trash data, and the remaining negative set pool.

4. The method of claim 3, wherein, Step S11 includes: S111, combining all CTCFs in the SC-betaorg stage of pancreatic cell differentiation in pairs within each chromosome; Step S12 includes: S121, removing all CTCF pairs outside the 1Mb range, retaining the start coordinates of the latter CTCF and the end coordinates of the former CTCF, and calling this data set 1mb_ctcf_pairs, then performing overlap analysis on all 1mb_ctcf_pairs and Hi-C data to find overlapping regions that are both in 1mb_ctcf_pairs and in Hi-C data; S122, selecting the CTCF binding site with the highest peak in the overlapping region as the positive sample; Step S13 includes the following steps: S131, the remaining overlapping region data is put into the trash database; S132, removing the positive set and trash data from all 1mb_ctcf_pairs, and the remaining negative set pool.

5. The method of claim 4, wherein, Step S2 includes: S21, downloading the CTCF motif MA0139.1.transfac on the JASPAR database; S22, obtaining directional data in combination with the FASTA file of the CTCF genomic sequence; S23, integrating the narrowpeak files of the five transcription factors with the Hi-C loop data through the bedtools tool to generate a file containing transcription factor information and overlapping regions; S24, sorting and processing the data to construct the arrangement of key transcription factors in the chromatin loop.

6. The method of claim 1, wherein, In step S3, when establishing the classification model, 80% of the samples are randomly selected as the training set and 20% as the independent test set, and a pre-trained model of BERT-Base-Uncased is used for binary classification.

7. The method of claim 6, wherein, BERT is a natural language processing model, and BERT-Base-Uncased is a specific variant of the BERT model.

8. The method of claim 7, wherein, Model establishment includes: S31, define a preprocessing function to tokenize the sequence and use fixed-length padding; S32, define training parameters and evaluation indicators; Through the above steps, a CTCF chromatin loop classification model based on BERT-Base-Uncased is established, and the key transcription factor sequence in the chromatin loop is classified through the classification model.

9. The method of claim 1, wherein, Step S4 includes: S41, calculate AUROC and AUPRC values by obtaining the prediction probability and real label of the two models of LC model and BERT-Base-Uncased; S42, calculate the logic regression of TPR and FPR based on different threshold values; S43, draw the ROC curve by taking FPR as the x-axis and TPR as the y-axis; S44, measure the performance of the classifier and determine the accuracy of the classification result.

10. The method of claim 9, wherein, In step S41, the formula for calculating the AUROC and AUPRC values is: Among them, TP, FP, TN and FN represent the quantity of true positive, false positive, true negative and false negative respectively, TPR is the true positive rate, that is, the proportion of positive samples correctly identified as positive to all positive samples; FPR is the false positive rate, that is, the proportion of negative samples incorrectly identified as positive to all negative samples.

Images