Prediction method for identifying super enhancer and typical enhancer in DNA sequence

Abstract

The invention discloses a prediction method for identifying a super enhancer and a typical enhancer in a DNA sequence, and the method comprises the steps: carrying out the coding through employing an integer coding and k-mer coding scheme, extracting the local features of the sequence through employing a residual error connection convolutional neural network, and extracting the multi-scale global features of the sequence through employing two bidirectional long-short-term memory networks. And performing feature fusion on the local features and the global features by using an attention mechanism, and inputting the fused features into a full-connection neural network to obtain a prediction probability. Evaluating the performance of the model by adopting a five-fold cross validation evaluation mode, and measuring the model by using five evaluation indexes such as ACC (Accuracy), REC (Recall Rate), PREE (Precision Reference Element), F1 score (F1) and AUC (Area under an ROC curve); experimental results on an independent test set show that compared with the most advanced method, the model has remarkable performance improvement in recognition of super enhancers and typical enhancers.

Description

Technical Field

[0001] The present invention relates to the technical field of prediction of functional sites in sequence analysis in bioinformatics, and in particular to a prediction method for identifying super enhancers and typical enhancers in DNA sequences. Background Art

[0002] Super enhancers are regulatory regions on the human genome with high enrichment of transcription factors and cofactor binding, responsible for defining cell identity and controlling gene transcription regulation. They usually splice together multiple typical enhancers based on specific splicing distances, thus spanning a large genomic domain. The typical enhancers that make up super enhancers are similar to normal enhancers elsewhere on the genome and are hypothesized to work synergistically in participating in gene transcription regulation. Accurate prediction of super enhancer sequences can help study mutations associated with various diseases such as cancer and Alzheimer's disease. Therefore, it is of great significance to explore the identification of super enhancers and typical enhancers based on DNA sequences.

[0003] For the identification of super enhancers and typical enhancers, for a long time in the past, the prediction of enhancers relied on biological experimental techniques. However, it remains challenging to determine super enhancers by experimental methods for two main reasons. First, it is currently time-consuming and laborious to perform ChIP analysis in all known tissues / cell types to map super enhancers. Second, specific antibodies for super enhancer-related markers may be lacking in some cell types, limiting the further application of super enhancer identification experimental methods. Some computational methods use various genomic data to discover super enhancers from typical enhancers. However, the vast majority of features of these methods rely on expensive and time-consuming wet experimental data, which leads to the problem that we cannot use it at the genome level when experimental data is insufficient, but DNA sequence information is easily available. With the rapid development of machine learning and deep learning, many prediction models have been established that can distinguish super enhancers from typical enhancers based solely on DNA sequence information. These methods usually use DNA2vec as feature embedding, but due to the inherent limitations of DNA2Vec, these methods cannot fully obtain long-range dependencies in the sequence. In addition, although these methods can identify super enhancers, the accuracy of the models needs to be further improved, and it should be possible to develop better models using new deep learning frameworks. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of the accuracy of the existing prediction of enhancers and their strengths, and to provide a predictor based on contextual language embedding for identifying enhancers and their strengths. The prediction method can predict enhancers and classify enhancer strengths, and at the same time construct an optimal two-layer classification model to improve the accuracy of predicting enhancers and their strengths. The following is a technical solution to achieve the purpose of the present invention, comprising the following steps: 1. Collect sample data sets: obtain super enhancer sequences and typical enhancer sequences for training sets and independent test sets; 2. Feature encoding: Integer encoding and k-mer encoding are used to encode DNA sequences, where the length of kmer is set to 4 and 6 respectively; 3. Model building: Use a residual-connected convolutional neural network to extract local features encoded by integers, and use two bidirectional long short-term memory networks to extract multi-scale global features encoded by 4-mer and 6-mer. Use the attention mechanism to fuse these local features with the global features, and then input them into the fully connected neural network to obtain the prediction probability; 4. Model tuning: Use batch normalization and dropout methods to optimize the prediction model; 5. Model evaluation: Model evaluation is performed based on 5-fold cross validation and independent testing. Specifically, five evaluation indicators, including accuracy (ACC), recall (REC), precision (PRE), F1 score (F1), and area under the ROC curve (AUC), are used to measure performance; 6. Cross-species experiments: Conduct cross-species experiments to verify the generalization ability of the model, train on the human enhancer dataset and test on the mouse enhancer dataset, and train on the mouse enhancer dataset and test on the human enhancer dataset;

[0005] Beneficial Effects This model uses integer coding and k-mer coding schemes for encoding, uses a residual-connected convolutional neural network to extract local features after integer coding, and uses two bidirectional long short-term memory networks to extract multi-scale global features after 4-mer and 6-mer encoding. The attention mechanism is used to fuse these local and global features, and then input them into the fully connected neural network to obtain the prediction probability. The experimental results show that we have achieved significant performance improvement compared with existing methods. The ACC and AUC values ​​on all datasets are higher than the most advanced methods, which shows that the super enhancer prediction model we proposed has strong generalization ability while having high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 is a model diagram of a prediction method for identifying super enhancers and typical enhancers; Figure 2 shows the ROC curve of the model on the human dataset under 5-fold cross validation; Figure 3 is the ROC curve of the model on the mouse dataset under 5-fold cross validation; Figure 4 is a table of standard data sets; Figure 5 is a comparison of the identification of super enhancers and typical enhancers with the most advanced methods under an independent test set; Figure 6 shows the cross-species validation in humans and mice. DETAILED DESCRIPTION

[0007] The specific implementation modes of the present invention are explained below in conjunction with the accompanying drawings. The accompanying drawings are for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0008] As shown in Figure 1, Figure 1 shows the workflow of this work. Integer encoding and k-mer encoding schemes are used for encoding, residual-connected convolutional neural networks are used to extract local features of the sequence, and two bidirectional long short-term memory networks are used to extract multi-scale global features of the sequence. These local features and global features are fused using the attention mechanism and then input into the fully connected neural network to obtain the prediction probability.

[0009] To build and evaluate the model, we used two datasets for model training, namely the human enhancer dataset and the mouse enhancer dataset. The detailed information of the dataset is as follows: Figure 4 shown.

[0010] We used 10 rounds of 5-fold cross validation on the training set to test the effectiveness of our proposed model, and used an independent dataset to verify the generalization ability of the predictor. To evaluate the predictor, we used the following five evaluation indicators: accuracy (ACC), recall (REC), precision (PRE), F1 score (F1), and area under the ROC curve (AUC) to measure performance.

[0011] Our method is then compared with existing state-of-the-art methods. The comparison results of independent tests are as follows: Figure 5 As shown in the figure, from the comparison results, we can see that on the human enhancer dataset and the mouse enhancer dataset, our model is significantly better than other methods in all five evaluation indicators.

[0012] Finally, we conducted cross-species experiments on the model to verify its generalization ability. Specifically, we set up two types of experiments: training on a human enhancer training set and testing on a mouse enhancer test set; training on a mouse enhancer training set and testing on a human enhancer test set. Figure 6 As shown in the figure, our method also shows superior performance in cross-species experiments, with both AUC and ACC values ​​higher than the existing optimal method MuSE.

Claims

1. A prediction method for identifying super enhancers and typical enhancers in DNA sequences, the process comprising the following steps: 1) Collect sample data sets: obtain super enhancer sequences and typical enhancer sequences for training sets and independent test sets; 2) Feature encoding: Integer encoding and k-mer encoding are used to encode the DNA sequence, where the length of k-mer is set to 4 and 6 respectively; 3) Model construction: Use a residual-connected convolutional neural network to extract local features encoded by integers, and use two bidirectional long short-term memory networks to extract multi-scale global features encoded by 4-mer and 6-mer. Use the attention mechanism to fuse these local and global features, and then input them into the fully connected neural network to obtain the prediction probability; 4) Model tuning: Use batch normalization and dropout methods to optimize the prediction model; 5) Model evaluation: Model evaluation was performed based on 5-fold cross validation and independent testing, and performance was measured using five evaluation indicators: accuracy (ACC), recall (REC), precision (PRE), F1 score (F1), and area under the ROC curve (AUC); 6) Cross-species experiments: Cross-species experiments were conducted to verify the generalization ability of the model, training on the human enhancer dataset and testing on the mouse enhancer dataset, as well as training on the mouse enhancer dataset and testing on the human enhancer dataset.

2. This method uses integer coding and k-mer coding schemes for encoding, uses residual-connected convolutional neural networks to extract local features of the sequence, and uses two bidirectional long short-term memory networks to extract multi-scale global features of the sequence. The attention mechanism is used to fuse these local features with the global features, and then input into the fully connected neural network to obtain the prediction probability.

3. A prediction method for identifying super enhancers and typical enhancers according to claim 1, characterized in that The feature encoding scheme and model construction described in step 2) and step 3) achieve high performance prediction results by integrating local and multi-scale global features.

Images