A DHSs Recognition and Prediction Method Based on Dual-Path Parallel Integrated Decision

Abstract

The present invention discloses a DHSs identification and prediction method based on dual-path parallel integrated decision-making, comprising the following steps: first, processing a data set to obtain a training set and a test set; second, establishing an iDHS-DPPE model; third, uniformly amplifying the training set and the test set; fourth, extracting features from the amplified training set sequence features as input; fifth, performing dual-path parallel training to obtain a trained iDHS-DPPE model; sixth, performing model testing using the amplified test set; and seventh, performing model identification and prediction and outputting results. The present invention can capture important information at different levels and reduce information loss through dual-model training of a BiGRU-Attention model and an MSFRN model. The dual-path iDHS-DPPE model identification and prediction constructed can effectively improve generalization ability and robustness, and can effectively and accurately identify DHSs.

Description

Technical Field

[0001] The present invention relates to the technical field of site prediction, and in particular to a DHSs identification and prediction method based on dual-path parallel integrated decision-making. Background Art

[0002] In genetics, specific chromatin regions that are hypersensitive to cleavage by DNase I endonuclease are called DNase I hypersensitive sites (DHSs). DHSs can provide important information about transcriptional regulatory elements and chromatin states in mammals. For many years, DHSs have served as reliable landmarks for determining the precise locations of transcriptional regulatory elements such as promoters, enhancers, repressors, and genomic control regions. In addition, genome-wide association studies (GWAS) have identified many non-coding variants associated with diseases and traits such as cancer and cardiovascular disease. These variants are concentrated in regulatory DNA marked by DHSs. Therefore, accurately identifying DHSs will help further explore the regulation of genes and the pathogenic mechanisms of common diseases.

[0003] DNase I hypersensitive sites (DHSs) are chromatin regions that show hypersensitivity to cleavage by the DNase I enzyme. They are universal markers of regulatory DNA and are associated with genetic variation in a variety of diseases and phenotypic traits. However, conventional experimental methods for identifying DHSs are time-consuming and expensive. DHSs on genes at different developmental stages of mice change dynamically, and sequences in different tissues also vary greatly. A single model often has difficulty accurately identifying multiple types of data. Therefore, the present invention proposes a DHSs identification and prediction method based on dual-path parallel integrated decision-making to address the problems existing in the prior art. Summary of the Invention

[0004] In response to the above problems, the purpose of the present invention is to propose a DHSs identification and prediction method based on dual-path parallel integrated decision-making. The DHSs identification and prediction method based on dual-path parallel integrated decision-making can capture important information at different levels through dual-model training of BiGRU-Attention model and MSFRN model, reduce information loss, and the dual-path iDHS-DPPE model identification and prediction can effectively improve the generalization ability and robustness, can adapt to the changes of DHS in different tissues and developmental stages of mice, and can effectively and accurately identify DHSs.

[0005] To achieve the purpose of the present invention, the present invention is implemented by the following technical solution: a DHSs identification and prediction method based on dual-path parallel integrated decision-making, comprising the following steps:

[0006] Step 1: Dataset processing: obtain the DHS sequences of the mouse genome in the iDHS-Deep benchmark dataset, process them using CD-HIT, and divide them into training and test sets;

[0007] Step 2: Model establishment: Establish an iDHS-DPPE model that includes a feature extraction module, a dual-path integration module, and an output module. The dual-path integration module includes a BiGRU-Attention model and an MSFRN model.

[0008] Step 3: Unify the training set and test set by expanding the sequence lengths in the training set and test set to obtain sequences of uniform length and split them into 2-mers according to the sliding window to obtain the expanded training set and test set;

[0009] Step 4: Feature extraction: The amplified training set sequence features are used as input to the feature extraction module of the iDHS-DPPE model to extract features and obtain extracted features;

[0010] Step 5: Dual-path parallel training: The extracted features are imported as input into the dual-path integration module containing the BiGRU-Attention model and the MSFRN model for training, and the trained iDHS-DPPE model is obtained by combining the feature extraction module and the output module;

[0011] Step 6: Model testing: The expanded test set is also subjected to the feature extraction process of step 4, and the extracted features are input into the trained iDHS-DPPE model for model testing to obtain the tested iDHS-DPPE model;

[0012] Step 7: Model recognition and prediction: the data to be tested is input into the tested iDHS-DPPE model after being processed in steps 3 and 4, and the iDHS-DPPE model outputs the recognition and prediction results.

[0013] Further improvements are as follows: the DHS sequence of the mouse genome obtained in the step 1 includes a comprehensive map of DHS in different tissues and developmental stages of mice, DHS sequences with a length of 50 to 301 bp are selected as positive samples, specific DNA fragments are selected as negative samples, and then CD-HIT is used to eliminate sequences with a similarity threshold of 0.8. Finally, 70% of the samples of different cell types and developmental stages are used as training sets, and 30% are used as independent test sets.

[0014] A further improvement is that the feature extraction module in step 2 includes an embedding layer, a convolution layer and a pooling layer, and the output module integrates the decisions output by the dual-path integration module to obtain a comprehensive evaluation result and identify DHSs.

[0015] Further improvements are as follows: the BiGRU-Attention model in step 2 includes BiGRU and a multi-head attention layer, BiGRU is used to capture long-range dependencies, and the multi-head attention layer is used to select key feature information. The MSFRN model includes six cascaded convolutional layers of different scales with a convolution kernel of 64 and a convolution layer with a convolution kernel of 192 for residual connection fusion of features.

[0016] A further improvement is that the expansion process in step 3 is specifically to add "N" to the end of the sequence when the sequence length is less than 301bp, and the insufficient length sequence is uniformly expanded to the maximum length of 301bp, and then split to obtain 2-mers, and then different nucleotides are mapped to different values in the range of 0 to 20. The expanded sequence is then encoded into a digital vector of length 300 as the input of the feature extraction module.

[0017] A further improvement is that the output of the model training in step 5 uses the G value to measure the confidence. The calculation formula of the G value is as follows:

[0018] G = abs(2cp-1), cp∈[0,1]

[0019] Where cp is the confidence probability. The higher the confidence probability cp, the larger G is, and the more reliable the judgment of the model is. When cp ≥ 0.5, the model judges that the input sequence is DHSs, otherwise the input sequence is not DHSs.

[0020] The beneficial effects of the present invention are as follows: the present invention can capture important information at different levels and reduce information loss through the dual-model training of the BiGRU-Attention model and the MSFRN model. The dual-path iDHS-DPPE model recognition and prediction constructed can effectively improve the generalization ability and robustness, can adapt to the changes in DHS in different tissues and developmental stages of mice, and can effectively and accurately identify DHSs, providing an efficient and reliable method for DHSs identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of the method of Example 1 of the present invention.

[0022] Figure 2 This is the overall architecture diagram of the iDHS-DPPE model in Example 1 of the present invention.

[0023] Figure 3 This is the BiGRU-Attention model architecture diagram in Example 1 of the present invention.

[0024] Figure 4 This is a diagram of the MSFRN model architecture in Example 1 of the present invention.

[0025] Figure 5 This is a line chart comparing the performance of different feature encoding methods in Example 2 of the present invention.

[0026] Figure 6 This is a performance comparison diagram of Example 2 of the present invention and different machine learning classifiers.

[0027] Figure 7 This is a comparison result diagram of various neural networks in Example 2 of the present invention.

[0028] Figure 8 1 is a diagram showing the comparison results between Example 2 of the present invention and various existing methods. DETAILED DESCRIPTION

[0029] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the examples. The examples are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0030] Example 1

[0031] according to Figure 1-Figure 4 As shown, this embodiment provides a DHSs identification and prediction method based on dual-path parallel integrated decision-making, including the following steps:

[0032] Step 1: Dataset processing: obtain the DHS sequences of the mouse genome in the iDHS-Deep benchmark dataset, process them using CD-HIT, and divide them into training and test sets;

[0033] The obtained DHS sequences of the mouse genome contain a comprehensive map of DHSs in different mouse tissues and developmental stages. DHS sequences with a length of 50 to 301 bp were selected as positive samples, and specific DNA fragments were selected as negative samples. Then, CD-HIT was used to eliminate sequences with a similarity threshold of 0.8. Finally, 70% of the samples of different cell types and developmental stages were used as training sets, and 30% were used as independent test sets, as shown in Table 1.

[0034] Table 1 Dataset details

[0035]

[0036]

[0037] Step 2: Model establishment: Establish an iDHS-DPPE model that includes a feature extraction module, a dual-path integration module, and an output module. The dual-path integration module includes a BiGRU-Attention model and an MSFRN model, as shown in the attached manual. Figure 2 As shown;

[0038] The feature extraction module includes an embedding layer, a convolution layer, and a pooling layer. The output module integrates the decisions output by the dual-path integration module to obtain a comprehensive evaluation result and perform DHSs identification.

[0039] Specifically, the DNA sequence is mapped from a 2-mer encoding scheme to an embedding layer, a convolutional layer, and a pooling layer to further extract feature information. A dual-path BiGRU-Attention model and an MSFRN model are used to capture local information and long-range dependencies, respectively, to obtain a probabilistic representation. Finally, the decisions of these two models are integrated to obtain a comprehensive evaluation output to identify DHSs.

[0040] The BiGRU-Attention model includes BiGRU and multi-head attention layers. BiGRU is used to capture long-range dependencies. BiGRU concatenates the generated forward and reverse sequences and passes the resulting output to the downstream network, enabling it to fully learn the contextual information in the sequence. The BiGRU output is set to 35 dimensions, and the final two outputs in opposite directions are concatenated into 70 dimensions.

[0041] The multi-head attention layer is used to select key feature information. The multi-head attention layer makes the model training process focus more on important features and ignore unimportant information. Different weights are assigned to the output of BiGRU, and an intermediate representation is obtained in each head. The results of each head are then connected and mapped back to the original dimension, as shown in the attached manual. Figure 3 As shown;

[0042] The MSFRN model includes six convolutional layers of different scales with a convolution kernel of 64 and a convolutional layer with a convolution kernel of 192 for feature residual connection fusion, organizing information across channels, as shown in the appendix of the manual. Figure 4 As shown in the figure, the cascaded six convolution layers with different scales and a convolution kernel of 64 can obtain information of different scales of the sequence, reduce information loss, and realize the fusion of hierarchical features;

[0043] The MSFRN model extracts the local relationship between adjacent features by fusing features of different scales, thereby improving the model's expressiveness.

[0044] Step 3: Unify the training and test sets. Since convolutional neural networks require fixed-size inputs, and the training and test sets have different sequence lengths, the sequences in the training and test sets are expanded to obtain uniform-length sequences. These sequences are then split into 2-mers using a sliding window to obtain the expanded training and test sets.

[0045] The specific expansion process is to add "N" to the end of the sequence when the sequence length is less than 301bp, and uniformly expand the insufficient length sequence to the maximum length of 301bp, and then split it to obtain 2-mers. Then, different nucleotides are mapped to different values in the range of 0 to 20. The expanded sequence is encoded into a digital vector of length 300 as the input of the feature extraction module. The generated sequence features are imported into the embedding layer matrix and further feature extraction is performed through the convolution layer and the maximum pooling layer.

[0046] Step 4: Feature extraction: The amplified training set sequence features are used as input to the feature extraction module of the iDHS-DPPE model to extract features and obtain extracted features.

[0047] Step 5: Dual-path parallel training: The extracted features are imported as input into the dual-path integration module containing the BiGRU-Attention model and the MSFRN model for training, and the trained iDHS-DPPE model is obtained by combining the feature extraction module and the output module;

[0048] The output of model training uses G value to measure confidence. The calculation formula of G value is as follows

[0049] G = abs(2cp-1), cp∈[0,1]

[0050] Where cp is the confidence probability. The higher the confidence probability cp, the larger G is, and the more reliable the judgment of the model is. When cp ≥ 0.5, the model judges that the input sequence is DHSs, otherwise the input sequence is not DHSs.

[0051] Step 6: Model testing: The expanded test set is also subjected to the feature extraction process of step 4, and the extracted features are input into the trained iDHS-DPPE model for model testing to obtain the tested iDHS-DPPE model.

[0052] Step 7: Model recognition and prediction. The data to be tested, after being processed in steps 3 and 4, is input into the tested iDHS-DPPE model, which outputs the recognition and prediction results. The output results are the same as the judgment method in step 5. When cp ≥ 0.5, the model judges that the input sequence is DHSs, otherwise the input sequence is not DHSs.

[0053] Example 2

[0054] according to Figure 5-Figure 8 As shown, this embodiment provides a comparison and discussion of DHSs identification and prediction methods based on dual-path parallel integration decision-making.

[0055] 1. Evaluation Criteria

[0056] Four common indicators including sensitivity (Sn), specificity (Sp), accuracy (Acc) and Matthew correlation coefficient (Mcc) are used to evaluate the prediction performance of the model. The calculation formula is as follows:

[0057]

[0058]

[0059]

[0060]

[0061] Where TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives, respectively. The area under the curve (AUROC) and the area under the precision-recall curve (AUPR) are also used to evaluate the model.

[0062] 2. Comparison with Different Feature Representations

[0063] Initial features have a significant impact on the training results of deep learning models. Experiments were conducted to determine the optimal length of nucleotide mapping. Due to the limitation of sequence length, only cases where k = 1, 2, or 3 are considered for k in k-mers. In addition, the onehot encoding scheme commonly used in the biological field is also added. The AUROC values obtained by 5-fold cross-validation for all datasets are shown in the appendix of the manual. Figure 5 middle.

[0064] Both 1-mers and one-hot encodings extract information about a single base, but may overlook the connections between bases. 2-mers, on the other hand, treat two bases as a whole, making better use of the long-term dependencies between adjacent nucleotides. 3-mers can also capture the sequential information of multiple base pairs, resulting in excellent prediction performance. However, higher-order encodings often result in higher computational costs and are prone to overfitting. However, in some smaller datasets, 2-mers may be more suitable than 3-mers.

[0065] The results demonstrated that the 2-mers encoding scheme can fully extract sequence information and is effective for identifying DHSs.

[0066] 3. Comparison with Different Machine Learning Classifiers

[0067] iDHS-DPPE was compared with four common machine learning classifiers, including SVM, KNN, RF, and XGBoost. Among them, SVM, KNN, RF, and XGBoost all used default parameter settings. To be fair, these methods, like iDHS-DPPE, all use 2-mers encoded feature vectors as input. AUROC results are shown in the attached manual. Figure 6 As shown in Figure 3, iDHS-DPPE is clearly the best, indicating that building an integrated deep learning model is advantageous and feasible for predicting DHSs.

[0068] 4. Comparison with different neural network structures

[0069] The feature vectors were input into a variety of different neural networks to verify the effectiveness of the dual-path parallel integration strategy and compared with three structures: CNN-LSTM, BiGRU-Attention model, and MSFRN model; the CNN-LSTM network structure was the same as that in iDHS-Deep.

[0070] The average AUROC values of CNN-LSTM, BiGRU-Attention, MSFRN and iDHS-DPPE are 0.910, 0.926, 0.928 and 0.931 respectively. Figure 7 As shown in the figure, iDHS-DPPE achieves the best results among all methods. Comparison with CNN-LSTM shows that the network structure of iDHS-DPPE can express features more accurately. Comparison with BiGRU-Attention and MSFRN verifies the effectiveness of the dual-path parallel integration strategy, which enhances the prediction performance of the model.

[0071] 5. Comparison with Existing Methods

[0072] To further demonstrate the superiority of the iDHS-DPPE model in Example 1, it was compared with the following published methods on the neural tube benchmark dataset, including SVM-RevcKmer, iDHS-EL, iDHSs-PseTNC, iDHS-DSAMS, iDHS-DXG, and iDHS-Deep. iDHS-Deep is a prediction tool specifically for DHSs in the mouse genome, while the other methods are used to predict human DHSs. The experimental results are shown in the attached manual. Figure 8 As shown in Figure 3, we can see that iDHS-DPPE is the best in all evaluation indicators.

[0073] The average results of iDHS-Deep and iDHS-DPPE on all independent test sets are shown in Table 2. Specifically, the average values of Sn, Sp, Acc, Mcc, and AUROC of iDHS-DPPE are 0.9%, 3.8%, 2.1%, 4.3%, and 2% higher than those of the state-of-the-art method iDHS-Deep, respectively. Clearly, iDHS-DPPE is more powerful and robust in predicting and distinguishing mouse DHSs.

[0074] Table 2 Comparison with the average results of iDHS-Deep independent tests

[0075] method Sn Sp Acc Mcc AUROC iDHS-Dee 0.777 0.869 0.828 0.659 0.912 iDHS-DPP 0.786 0.907 0.849 0.702 0.932

[0076] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A DHSs identification and prediction method based on dual-path parallel integrated decision-making, characterized by: The following steps are involved: Step 1: Dataset processing: obtain the DHS sequences of the mouse genome in the iDHS-Deep benchmark dataset, process them using CD-HIT, and divide them into training and test sets; Step 2: Model establishment: Establish an iDHS-DPPE model that includes a feature extraction module, a dual-path integration module, and an output module. The dual-path integration module includes a BiGRU-Attention model and an MSFRN model. Step 3: Unify the training set and test set by expanding the sequence lengths in the training set and test set to obtain sequences of uniform length and split them into 2-mers according to the sliding window to obtain the expanded training set and test set; Step 4: Feature extraction: The amplified training set sequence features are used as input to the feature extraction module of the iDHS-DPPE model to extract features and obtain extracted features; Step 5: Dual-path parallel training: The extracted features are imported as input into the dual-path integration module containing the BiGRU-Attention model and the MSFRN model for training, and the trained iDHS-DPPE model is obtained by combining the feature extraction module and the output module; Step 6: Model testing: The expanded test set is also subjected to the feature extraction process of step 4, and the extracted features are input into the trained iDHS-DPPE model for model testing to obtain the tested iDHS-DPPE model; Step 7: Model recognition and prediction: the data to be tested is input into the tested iDHS-DPPE model after being processed in steps 3 and 4, and the iDHS-DPPE model outputs the recognition and prediction results.

2. The DHSs identification and prediction method based on dual-path parallel integrated decision-making according to claim 1 is characterized by: The DHS sequence of the mouse genome obtained in the first step includes a comprehensive map of DHSs in different tissues and developmental stages of mice. DHS sequences with a length of 50 to 301 bp are selected as positive samples, and specific DNA fragments are selected as negative samples. Then, CD-HIT is used to eliminate sequences with a similarity threshold of 0.

8. Finally, 70% of the samples of different cell types and developmental stages are used as training sets, and 30% are used as independent test sets.

3. The DHSs identification and prediction method based on dual-path parallel integrated decision-making according to claim 1 is characterized by: The feature extraction module in step 2 includes an embedding layer, a convolution layer, and a pooling layer. The output module integrates the decisions output by the dual-path integration module to obtain a comprehensive evaluation result output and performs DHSs identification.

4. The DHSs identification and prediction method based on dual-path parallel integrated decision-making according to claim 1 is characterized by: The BiGRU-Attention model in step 2 includes BiGRU and a multi-head attention layer. BiGRU is used to capture long-range dependencies, and the multi-head attention layer is used to select key feature information. The MSFRN model includes six cascaded convolutional layers of different scales with a convolution kernel of 64 and a convolution kernel of 192 for residual connection fusion of features.

5. The DHSs identification and prediction method based on dual-path parallel integrated decision-making according to claim 1 is characterized by: The expansion process in step 3 is specifically to add "N" to the end of the sequence when the sequence length is less than 301bp, uniformly expand the insufficient length sequence to the maximum length of 301bp, and then split it to obtain 2-mers. Then, different nucleotides are mapped to different values in the range of 0 to 20. The expanded sequence is then encoded into a digital vector of length 300 as the input of the feature extraction module.

6. The DHSs identification and prediction method based on dual-path parallel integrated decision-making according to claim 1 is characterized by: The output of the model training in step 5 uses the G value to measure the confidence. The calculation formula of the G value is as follows G = abs(2cp-1), cp∈[0,1] Where cp is the confidence probability. The higher the confidence probability cp, the larger G is, and the more reliable the judgment of the model is. When cp ≥ 0.5, the model judges that the input sequence is DHSs, otherwise the input sequence is not DHSs.

Images