阅读说明：本技术 一种融合dna形状特征的转录因子结合位点预测方法 (Transcription factor binding site prediction method fusing DNA shape characteristics ) 是由 李阳阳 魏志强 刘昊 闫金盟 于 2021-08-06 设计创作，主要内容包括：本发明涉及一种融合DNA形状特征的转录因子结合位点预测方法,属于生物信息学领域,所述方法结合了结构生物学和基因组学以及深度学习神经网络的知识提出了一个使用CNN结合DNA序列与形状特征信息预测转录因子结合位点的新模型。同时,构建了包含DNA形状特征及DNA序列信息的特殊数据集,在传统转录因子预测的数据集基础之上添加了对应的DNA形状信息。从而提高DNA转录因子结合位点预测的准确性。(The invention relates to a transcription factor binding site prediction method fusing DNA shape characteristics, belonging to the field of bioinformatics. Meanwhile, a special data set containing DNA shape characteristics and DNA sequence information is constructed, and corresponding DNA shape information is added on the basis of a data set predicted by a traditional transcription factor. Thereby improving the accuracy of the prediction of the DNA transcription factor binding site.)

1. A prediction method of transcription factor binding sites fused with DNA shape characteristics is characterized by comprising the following specific steps:

1) designing and constructing a special data set with DNA shape characteristic data and DNA sequence information according to the information disclosed in the prior art, and predicting various important structural characteristics of DNA by adopting an HT-MC method aiming at the acquisition of the DNA shape characteristic, wherein the predicted characteristics comprise small groove width, rolling, propeller twisting and spiral twisting;

2) DNA sequence motif data and DNA shape feature data preprocessing

3D DNA shape characteristics are predicted using a pentamer-based model that is built based on full-atom Monte Carlo simulations of DNA structures; the input data is divided into two parts, namely a sequence and a shape; for the DNA sequence portion, the input is a 4 × L matrix, where L is the length of the sequence, and each base pair A, C, T, G in the sequence is represented as four unique heat codes [1,0,0,0], [0,1,0,0], [0,0,1,0] and [0,0,0,1 ]; for the shape feature portion of DNA, the input is a 4 × L matrix, where L is the length of the sequence, and the shape features of the DNA sequence are described as one channel vector for each nucleotide position, respectively;

3) novel transcription factor binding site prediction model based on CNN fusion DNA shape characteristics

After the DNA sequence, the DNA shape characteristics, the label data and the coding characteristics of each sample are collected, determining that a model of training data is a sequence + DSS model, and combining the sequence + DSS model with two types of data of the sequence and the DSS to form a comprehensive model for prediction; the sequence + DSS model is based on a convolution neural network in deep learning, a double-input parallel convolution architecture is adopted, two 4 xL matrixes are input and are respectively a sequence information matrix and a shape information matrix of a gene, then convolution and global maximum pooling are respectively carried out, the number of convolution kernels is 128, the size of a convolution window is 1 x 24, finally pooling results aiming at the two types of data are connected and serve as input of a full connection layer, the number of neurons is 32 or 64, a dropout method is used, parameters are set to be 0.1,0.5 and 0.75, the number of neurons in a final output layer is 2, and an activation function used in an output stage is softmax regression;

4) training the new prediction model in the step 3) by using the data preprocessed in the step 2).

2. The method according to claim 1, characterized in that cross entropy is used as a loss function in the training of the model, and the model is trained using standard error back propagation algorithm and AdaDetla method, with batch _ size set to 100, and the model is verified after each epoch, and then the training is stopped using early stopping techniques.

Technical Field

The invention belongs to the field of bioinformatics, and relates to a new method for predicting transcription factor binding sites by combining knowledge design of structure biology and genomics and realizing a set of DNA shape characteristics.

Background

Transcription Factors (TFs) can coordinate the expression of many genes by binding to genomic regions that regulate transcription. Cellular mechanisms utilize these primary regulators to regulate key cellular processes and to adapt to environmental stimuli. Indeed, alterations in the sequence or number of TF may be a major cause of genetic diseases, complex diseases, autoimmune deficiencies and cancer. How TF binds to specific DNA regulatory sequences (called TF binding sites, or TFBS for short, such as promoters, enhancers) to coordinate regulation of gene transcription and protein synthesis is a very important process that plays a key role in many biological processes. In the last decade, a large amount of immunoprecipitation and its high-throughput sequencing (ChIP-seq) data has been generated and used to study the mechanisms behind these regulatory processes, but because this method is TF-specific, i.e. specific for a certain TF to determine the binding site sequence on its DNA sequence, and its high experimental cost, it is not possible to analyze every TF binding map in all cell types, and therefore an accurate computational method is required to decode the underlying binding rules. Of course, how to predict TFBS in DNA sequences is a fundamental problem in bioinformatics.

The DNA binding specificity of transcription factors is a key component of the gene regulatory process, but the underlying mechanism of highly specific binding of TF to its genomic target site is poorly understood. In early studies, we hypothesized that the binding site for a DNA transcription factor was completely defined by the base sequence. Position Weight Matrix (PWM) based methods have enjoyed great success in modeling DNA-protein binding processes. Later, gkm-SVM (i.e., notched k-mers and support vector machines) showed advantages over PWM-based methods. In recent years, convolutional neural networks, coupled with the single-hot-coded format of DNA sequences, have attracted great interest in predicting TFBS. However, prediction or insertion of TFBS using only primary DNA sequences has proven insufficient to adequately model its underlying binding rules. Obviously, if the prediction accuracy is really improved, the underlying modeling mode needs to be improved, and the process is an important guarantee for subsequent prediction work.

Indeed, technological advances over the past decade have facilitated the discovery and study of the characterization of DNA binding preferences for many TFs. Recent high-throughput studies highlight that TF-DNA binding is not solely dependent on nucleotide sequence preference, and a number of relevant factors have been identified. Increasing evidence supports a broad contribution of sequence context, including flanking sequences and DNA shape, in regulating sequence recognition. Interacting cofactors and TF can also alter sequence preference. In addition, some cell-type specific information, here mainly including chromatin accessibility and histone modifications, also have a large influence on binding of TFs to their target sites.

In this context, more and more research tends to model DNA motifs in a manner that combines them with other features, including histone modification, chromatin accessibility, and cell type, among others. And attempts have been made to different approaches. For example, there are methods to use uncontrolled methods, such as hierarchical mixed models or hidden markov models, to identify transcription factor footprints using chromatin accessibility data. They used sequence motif scores to attribute footprints to different transcription factors. More recent approaches use Matrix Completion (Matrix Completion) to accomplish transcription factor binding prediction, i.e., inferring TF binding using a 3-mode tensor that represents genomic location, cell type and TF binding. This approach does not rely on sequence specificity, but only predicts TF binding in well studied cell types with many ChIP-seq datasets. Of interest, sequences are combined with epigenetic genome data in convolutional neural network models to predict transcription factor binding sites. Its prediction process uses histone modification and chromatin accessibility information in addition to DNA sequence. Although an improvement over the same model considering only motif information, this approach was trained and validated using only 15 cell types of standardized DNase-seq data and 5 specific core histone modifications, which may lead to a good prediction of the trained model only for TF binding preference in this specific cell environment. In summary, few attempts have been made to incorporate three-dimensional structural features of DNA into predictive modeling of TFBS.

Disclosure of Invention

The technical problem to be solved by the invention is to provide a transcription factor binding site prediction method fusing DNA shape characteristics, and the method comprises the steps of firstly constructing a special data set which simultaneously contains DNA sequence motif information and DNA three-dimensional shape information and can be used for transcription factor set site prediction; then, a novel transcription factor binding site prediction model which can simultaneously fuse the DNA shape characteristics and the sequence information is provided, and the model can combine the structural characteristics of the DNA with the DNA sequence information, so that the accuracy of the prediction of the DNA transcription factor binding site is improved.

The invention is realized by the following technical scheme:

a prediction method of transcription factor binding sites fused with DNA shape characteristics comprises the following specific steps:

1) designing and constructing a special data set with DNA shape characteristic data and DNA sequence information according to the information disclosed by the prior art, and predicting various important structural characteristics of DNA by adopting an HT-MC method aiming at the acquisition of the DNA shape characteristic, wherein the predicted characteristics comprise small Groove Width (MGW), rolling (Roll), Propeller Twist (ProT) and Helix Twist (Helix Twist, HelT);

2) DNA sequence motif data and DNA shape feature data preprocessing

3D DNA shape characteristics are predicted using a pentamer-based model that is built based on full-atom Monte Carlo simulations of DNA structures; the input data is divided into two parts, namely a sequence and a shape; for the DNA sequence portion, the input is a 4 × L matrix, where L is the length of the sequence, and each base pair A, C, T, G in the sequence is represented as four unique heat codes [1,0,0,0], [0,1,0,0], [0,0,1,0] and [0,0,0,1 ]; for the shape features part of the DNA, the input is a 4 × L matrix, where L is the length of the sequence and the shape features of the DNA sequence (MGW, Roll, ProT, HelT) are described as one channel vector for each nucleotide position, respectively;

3) novel transcription factor binding site prediction model based on CNN fusion DNA shape characteristics

After the DNA sequence, DNA shape characteristics (DSS), label data and coding characteristics of each sample are collected, determining that a model of training data is a sequence + DSS model, and combining the sequence + DSS model with two types of data of the sequence and the DSS to form a comprehensive model for prediction; the sequence + DSS model is based on a convolution neural network in deep learning, a double-input parallel convolution architecture is adopted, two 4 xL matrixes are input and are respectively a sequence information matrix and a shape information matrix of a gene, then convolution and global maximum pooling are respectively carried out, the number of convolution kernels is 128, the size of a convolution window is 1 x 24, finally pooling results aiming at the two types of data are connected and serve as input of a full connection layer, the number of neurons is 32 or 64, a dropout method is used, parameters are set to be 0.1,0.5 and 0.75, the number of neurons in a final output layer is 2, and an activation function used in an output stage is softmax regression;

4) training the new prediction model in the step 3) by using the data preprocessed in the step 2).

As a preferential technical scheme, cross entropy is used as a loss function in the training process of the model, the model is trained by using a standard error back propagation algorithm and an AdaDella method, the batch _ size is set to be 100, the model is verified after each epoch, and then the training is stopped by using an early stopping skill.

Compared with the prior art, the invention has the beneficial effects that:

1. a special data set containing DNA shape characteristics and DNA sequence information is constructed, and corresponding DNA shape information is added on the basis of a data set predicted by a traditional transcription factor.

The data set is based on the conventional method only comprising sequence motif and label information, and DNA shape characteristic information corresponding to the original sequence information is added. We have conducted intensive research and study on the aspects of DNA shape feature information acquisition methods, and processing of such information to adapt to CNN models, etc. The formed universal data set can be used for other researches for predicting the binding site of the transcription factor by combining DNA shape information and sequence information.

2. A new model for predicting transcription factor binding sites using CNN binding DNA sequences and shape data was designed and implemented.

The model adopts a novel CNN fusion framework, and the result proves that the model successfully learns the shape information of DNA and fuses the shape information into the task of predicting the binding site of the transcription factor. Compared with other existing deep learning models fusing DNA shape characteristics, the model is low in design complexity, short in training time and high in usability, and compared with the traditional mathematical model fusing DNA shape characteristics, the model is high in prediction accuracy.

Drawings

FIG. 1 is a schematic diagram of the type of DNA shape feature of the present invention;

FIG. 2 is a unified framework for predicting TFBS using CNN binding DNA sequences and shape information according to the present invention;

FIG. 3 is a diagram of a dual-input parallel convolution architecture of the convolutional neural network based on deep learning;

FIG. 4 is a frame diagram of a TFBS prediction model based on the shape characteristics of Keras fused DNA;

FIG. 5 comparison of sequence-based models and model experimental data distributions based on sequence and shape information.

Detailed Description

The technical solution of the present invention is further explained by the following embodiments with reference to the attached drawings, but the scope of the present invention is not limited in any way by the embodiments.

Example 1

A prediction method of transcription factor binding sites fused with DNA shape characteristics comprises the following specific steps:

1. construction of data sets

The method comprises the steps of deeply researching related research progress of a bottom layer mechanism of protein-DNA combination, summarizing research progress and current situation of the current DNA transcription factor binding site prediction, and collecting and investigating source information of a mainstream data set in the field. Secondly, the relevant progress of the DNA shape characteristic acquisition method and the construction method of the DNA transcription factor binding site prediction relevant data set are deeply researched aiming at a transcription factor binding site prediction model fusing the DNA shape characteristic and the sequence information, and a special data set with the DNA shape characteristic data and the DNA sequence information is designed and constructed.

The acquisition of DNA shape characteristics was performed using the HT-MC method, and previous studies improved the efficiency of conformational sampling by reducing the degrees of freedom in the system. The high throughput methods herein are intended to predict various important structural features of DNA and can accommodate sequences of essentially any length or number. The method can improve accuracy. In view of their importance in DNA shape reading, predicted features include small Groove Width (MGW), Roll (Roll), Propeller Twist (ProT) and Helix Twist (Helix Twist, HelT), as shown in fig. 1.

Extensive validation of extensive experimental and computational data demonstrates the robustness of the HT-MC method, and high-throughput methods under DNA shape web servers can be used to accomplish nucleotide-resolution DNA structural feature predictions of the entire yeast genome in less than 1 minute on a single processor.

In this example, ChIP-seq experimental data for 69 sets of transcription factors were obtained from ENCODE (http:// hgdownload. cse. ucsc. edu/goldenPath/hg19/encodeDCC/wgEncodeAwgTfbsUniform /). Each set of experimental data is in a FASTA format and is divided into a training data set and a testing data set. In the dataset, the DNA sequences and their corresponding labeling information are given. The positive and negative samples had the same GC number and sequence length (101 bp). The DNA shape features (DSS) used in this example (including MGW, Roll, ProT, HelT) were then generated based on an existing method based on a pentamer look-up table from thousands of full-atom Monte Carlo simulations and validated by X-ray and NMR structures.

And extracting the signal value of the corresponding position from the DSS according to the position of the sample in the sequence data set. Where each nucleotide position can be considered to have a corresponding value characterizing the shape of each DNA. Thus, TFBS and non-TFBS are described as two types of features: (1) for the one-hot characterization of DNA sequence information; (2) DSS characterization for DNA shape information. For each data set, 70% of the samples were used for training, 10% for verification, and 20% for testing.

2. And preprocessing DNA sequence motif data and DNA shape characteristic data.

The 3D DNA shape characteristics were predicted using a pentamer-based model (HT-MC) that was built based on an all-atom monte carlo simulation of the DNA structure. Four different shape features, including Minor Groove Width (MGW), rolling (Roll), propeller twist (ProT), and helix twist (HelT), have been shown to play important roles in protein-DNA binding site recognition under specific circumstances.

The convolutional neural network architecture of the evaluation is shown in fig. 3, where the input is divided into two part sequences and shapes. For the DNA sequence portion, the input is a 4 × L matrix. Where L is the length of the sequence, which in this example is 101 bp. Each base pair A, C, T, G in the sequence is represented as four unique codes [1,0,0,0], [0,1,0,0], [0,0,1,0] and [0,0,0,1], respectively. For the shape-characterizing portion of DNA, the input is a 4 × L matrix, where L is the length of the sequence. The shape characteristics of the DNA sequences (MGW, Roll, ProT, HelT) are described as a channel vector for each nucleotide position. In this example, using 101bp DSS data, the carrier size of the sample was 1 × 101, and since this example uses four types of DNA shape features, the size was 4 × 101. DSS is a continuous attribute that describes the apparent 3D characteristics of DNA, possibly related to the binding of a particular TF. The DNA shape characteristics used in this example are data of single base resolution.

From a data perspective, to combine DSS and sequence features in a unified deep learning framework, after collecting DNA sequences, DSS data, tag data and coding features for each sample, this example first implements two different models: (1) sequence CNN model, using DNA sequence as a feature; (2) DSS _ CNN model, using DSS data as features. In this embodiment, the CNN is composed of an input layer, a convolutional layer, a max pooling layer, a full connection layer, a dropout layer, and an output layer. For the CNN model, the number of kernels is 128 in this embodiment, the size of the kernel window is 1 × 24, and the number of neurons in the fully connected layer is set to 64 to achieve the best effect of the model. As shown in fig. 3, the model is based on a Convolutional Neural Network (CNN) in deep learning, and adopts a dual-input parallel Convolutional architecture, and fig. 4 shows a TFBS prediction model framework diagram based on the shape characteristics of the fusion DNA of Keras, first, two 4 × 101 matrices are input, which are respectively a sequence information matrix and a shape information matrix of a gene. Then convolution is carried out respectively (the number of convolution kernels is 128, the convolution window size is 1 x 24) and global maximum pooling is carried out, finally, the pooling results for the two types of data are connected as the input of a full connection layer (the number of neurons is 64 here), meanwhile, a dropout method is used, parameters are set to be (0.1,0.5 and 0.75), the number of neurons in the final output layer is 2, and the activation function used by the output stage is softmax regression.

After setting the appropriate model for each type of data (here, two types of DNA sequence data and DNA shape data), the example compared the performance of two different models: (1) a sequence model, using only DNA sequence data as features; (2) the sequence + DSS model combines two types of data, sequence and DSS, into a comprehensive model as a feature.

For the training process, the present embodiment uses cross entropy as a loss function. In view of the selected loss function and different hyper-parameters, a standard error back propagation algorithm and an AdaDella method are used to train the model. The number of iterations for each model (nb _ epoch) is set to 100, the batch _ size is set to 100, and the model is verified after each epoch. Early stopping techniques are then used to stop the training, as sometimes the error rate may fluctuate to a later stage. The best model is selected based on the accuracy of the verification phase.

The results of the two model experiments are compared as shown in fig. 5.

This example records 69 sets of AUC curve data for both models and compares them. As shown in fig. 5, the new model predicts TFBS using a deep learning integration framework that combines DNA sequence data with DNA shape data. Experimental evaluation showed that the integration framework has better performance and accuracy (the accuracy is relatively higher for higher values of AUC) than the model based on the primary DNA sequence.