Structural variation detection method based on deep learning

Abstract

The invention relates to a deep learning-based structural variation detection method, which specifically comprises the following steps: S1, data uploading: uploading structural variation feature sequence data, comparing sequencing data to a reference genome sequence or a reference main transcript sequence, and checking an uploaded format; and S2, data preprocessing: carrying out data conversion and data cleaning on the uploaded structural variation feature sequence data, and then storing the uploaded data. According to the structural variation detection method based on deep learning, a detection method capable of automatically learning and analyzing complex biological signals is constructed, so that structural variation can be detected and positioned more accurately. Through deep learning, various signal sources can be automatically integrated, and the recognition precision of structural variation is improved.

Description

Technical Field

[0001] The present invention relates to the field of precision medicine technology, and specifically to a structural variation detection method based on deep learning. Background Art

[0002] Various algorithms have been developed to detect SVs from whole genome sequencing (WGS) data. In general, these methods can be divided into four categories: read depth (RD)-based methods, paired-end (PE) mapping-based methods, sorted read (SR)-based methods, and de novo assembly (AS)-based methods. RD-based methods first divide the reference genome into non-overlapping bins and calculate the RD of each bin. Duplications (DUPs) and deletions (DELs) are detected based on the abnormal RD changes of adjacent bins. PE-based methods use the span and direction of paired-end reads to detect more SV types, such as inversions (INVs). For SR-based methods, if an adjacent read is aligned to different coordinates of the reference genome, the read can be used to determine the SV breakpoint at single-nucleotide resolution. AS-based methods assemble reads into larger alleles and use the assembled alleles to detect SVs.

[0003] However, for short-read sequencing data, assembly itself has a high computational cost. In addition to the above four categories of methods, there are also some methods that integrate more than one method. Recent benchmark studies have shown that no single SV caller can accurately and sensitively detect all types and sizes of SVs.

[0004] At present, it is still difficult for humans to interpret genomic data, and biologists' analysis of bioinformatics data is still a very time-consuming process. If a computer program can be used to speed up this process, it can directly generate SV call results, improve the accuracy of SV detection, reduce the workload of auditors, and reduce the cost and uncertainty of manual review, which will greatly promote the development of new drugs. Based on this, the present invention proposes a structural variation detection method based on deep learning to improve the detection efficiency and accuracy of SV-related diseases. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In order to solve the above technical problems, the present invention provides a structural variation detection method based on deep learning.

[0007] (II) Technical solution

[0008] Based on this, the present invention provides the following technical solution: a structural variation detection method based on deep learning, the specific steps are as follows:

[0009] Step S1: Data upload: upload the structural variation feature sequence data, align the sequencing data to the reference genome sequence or reference master transcript sequence, and check the uploaded format;

[0010] Step S2: Data preprocessing, converting and cleaning the uploaded structural variation feature sequence data, and then storing the uploaded data;

[0011] Step S3: Structural variation prediction, read the stored data, encode the structural variation feature sequence and the reference genome sequence using image multi-channel encoding, convert them into images, obtain the comparison image of the sample and the reference genome sequence, use the pre-trained structural loading model to identify all the variation regions in the structural variation image, obtain the structural variation fragments with single-base resolution, and map the variation information of the corresponding position of the image back to the original locus, load and predict the image;

[0012] Step S4: The backend service stores the predicted results and the converted images in a structured manner and provides structured queries;

[0013] Step S5: Front-end display, UI display of the results.

[0014] Preferably, the model loaded in step S3 as described above is a convolutional hourglass network model, and the structural variation region in the image is detected by the convolutional hourglass network model to obtain image segmentation information. The pixel category in the image indicates whether a structural variation occurs at the corresponding base position, thereby obtaining single-base resolution breakpoint information of the structural variation.

[0015] Preferably, the convolutional hourglass network model is mainly composed of a convolutional layer, an hourglass module, and an addition module. The convolutional hourglass network model is provided with two groups of connection modes, one group is: the right end of the addition module is connected to the hourglass module, the right end of the hourglass module is connected to two groups of convolutional layers, and the right end of the convolutional layer is connected to the hourglass module through the addition module; the other group is: the right end of the addition module is connected to the hourglass module, the right end of the hourglass module is connected to two groups of convolutional layers, and an intermediate supervision output module is added at the connection between the two groups of convolutional layers, and the right end of the convolutional layer is connected to the hourglass module through the addition module, and one group of the addition modules is connected to the other group of the addition modules in a residual connection manner.

[0016] Preferably, the convolution and maximum pooling layers in the convolutional hourglass network model first process the image features to a very low resolution. After reaching the lowest resolution, the network begins to continuously upsample and combine features across scales, bringing together information from two adjacent resolutions, performing nearest neighbor upsampling on the lower resolution image, and adding the two sets of features element by element. The hourglass module structure is symmetrical, and for each downward layer, there is a corresponding upward layer. After reaching the output resolution of the network, two consecutive rounds of 1x1 convolutions are applied to produce the final network prediction.

[0017] Preferably, the calculation formula of a single hourglass module is:

[0018]

[0019] intermediate_o i =conv(o i );

[0020]

[0021] where x input represents the input of the i-th hourglass network, o represents the output after unet network encoding, and intermediate_o represents the output of the i-th layer of intermediate supervision; the loss function of the hourglass network is defined as: The intermediate supervision output of each layer is subjected to cross entropy loss with the ground truth. The final loss is the sum of the losses of the n-layer hourglass network. The output of the final model is the hourglass network output of the last layer. Channel attention can improve the quality of the representation produced by the network by explicitly modeling the interdependence between its convolutional feature channels. It can learn to use global information to selectively emphasize important features and suppress less useful features by recalibrating the network. Using the channel attention mechanism can not only reduce the computational complexity, but also automatically search for the optimal subnetwork in the original network to achieve better results.

[0022] Preferably, the three-channel encoding method in step S3 described above has the following specific steps:

[0023] Step Q1: squeeze operation, pooling the global spatial features of each channel as the representation of the channel to form a channel description feature map, i.e., RGB image, calculated as follows:

[0024]

[0025] Step Q2: excitation operation, using a fully connected neural network, performs a nonlinear transformation on the result after sequence, outputs the degree of dependence s on each channel, and then adjusts the feature map according to the degree of dependence. The calculation method is as follows:

[0026] s=F ex (z, W)=σ(g(z, W))=σ(W2δ(W1z));

[0027]

[0028] Where δ represents the ReLU activation function, σ represents the sigmoid activation function, and W is the weight of the neural network. The design of the above algorithm provides a basis for the system to process data. The algorithm will be presented to users in the form of interactive software. Users can upload data, and the system will detect mutations based on the algorithm in the background.

[0029] (III) Beneficial effects

[0030] Compared with the prior art, the present invention provides a structural variation detection method based on deep learning, which has the following beneficial effects:

[0031] 1. This deep learning-based structural variation detection method constructs a detection method that can automatically learn and analyze complex biological signals to more accurately detect and locate structural variations. Deep learning can automatically integrate multiple signal sources to improve the recognition accuracy of structural variations. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a specific flow chart of the detection method of the present invention;

[0033] Figure 2 Schematic diagram of the convolutional hourglass network structure of the present invention;

[0034] Figure 3 It is a schematic diagram of the three-channel encoding method flow of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] Example 1

[0037] See also Figure 1-Figure 3 , a structural variation detection method based on deep learning, the specific steps are as follows:

[0038] 1. Upload data: upload the structural variation feature sequence data, align the sequencing data to the reference genome sequence or reference master transcript sequence, and check the uploaded format;

[0039] 2. Data preprocessing: convert and clean the uploaded structural variation feature sequence data, and then store the uploaded data;

[0040] 3. Structural variation prediction: Read the stored data, encode the structural variation feature sequence and the reference genome sequence using image multi-channel encoding, convert them into images, obtain the comparison image of the sample and the reference genome sequence, use the pre-trained structural loading model to identify all the variation regions in the structural variation image, obtain the structural variation fragments with single-base resolution, and map the variation information at the corresponding position of the image back to the original locus, load and predict the image;

[0041] 4. Backend service, which stores the predicted results and converted images in a structured manner and provides structured queries;

[0042] 5. Front-end display, UI display of the results.

[0043] In some embodiments, the loading model is a convolutional hourglass network model, and the structural variation area in the image is detected by the convolutional hourglass network model to obtain image segmentation information. The pixel category in the image indicates whether the corresponding base position has a structural variation, thereby obtaining the single-base resolution breakpoint information of the structural variation. The convolutional hourglass network model is mainly composed of a convolutional layer, an hourglass module, and an addition module. The convolutional hourglass network model is provided with two groups of connection modes, one group is: the right end of the addition module is connected to the hourglass module, the right end of the hourglass module is connected to two groups of convolutional layers, and the right end of the convolutional layer is connected to the hourglass module through the addition module; the other group is: the right end of the addition module is connected to the hourglass module, the right end of the hourglass module is connected to two groups of convolutional layers, and the connection between the two groups of convolutional layers is increased. An intermediate supervision output module is provided, and the right end of the convolution layer is connected to the hourglass module through an addition module. One group of the addition modules is connected to another group of the addition modules by a residual connection. The convolution and maximum pooling layers in the convolution hourglass network model first process the image features to a very low resolution. After reaching the lowest resolution, the network begins to continuously upsample and combine features across scales, bringing together information from two adjacent resolutions, performing nearest neighbor upsampling on the lower resolution image, and adding the two sets of features element by element. The hourglass module structure is symmetrical. For each downward layer, there is a corresponding upward layer. After reaching the output resolution of the network, two consecutive rounds of 1x1 convolutions are applied to generate the final network prediction.

[0044] In this application, the calculation formula for a single hourglass module is:

[0045]

[0046] intermediate_o i =conv(o i );

[0047]

[0048] where x input represents the input of the i-th hourglass network, o represents the output after unet network encoding, and intermediate_o represents the output of the i-th layer of intermediate supervision; the loss function of the hourglass network is defined as: The intermediate supervision output of each layer is subjected to cross entropy loss with the ground truth. The final loss is the sum of the losses of the n-layer hourglass network. The output of the final model is the hourglass network output of the last layer. Channel attention can improve the quality of the representation produced by the network by explicitly modeling the interdependence between its convolutional feature channels. It can learn to use global information to selectively emphasize important features and suppress less useful features by recalibrating the network. Using the channel attention mechanism can not only reduce the computational complexity, but also automatically search for the optimal subnetwork in the original network to achieve better results.

[0049] 9. In some embodiments, the three-channel encoding method has the following specific steps:

[0050] Step Q1: squeeze operation, pooling the global spatial features of each channel as the representation of the channel to form a channel description feature map, i.e., RGB image, calculated as follows:

[0051]

[0052] Step Q2: excitation operation, using a fully connected neural network, performs a nonlinear transformation on the result after sequence, outputs the degree of dependence s on each channel, and then adjusts the feature map according to the degree of dependence. The calculation method is as follows:

[0053] s=F ex (z, W)=σ(g(z, W))=σ(W2δ(W1z));

[0054]

[0055] Where δ represents the ReLU activation function, σ represents the sigmoid activation function, and W is the weight of the neural network. The design of the above algorithm provides a basis for the system to process data. The algorithm will be presented to users in the form of interactive software. Users can upload data, and the system will detect mutations based on the algorithm in the background.

[0056] Example 2

[0057] In this embodiment, the experimental effect comparison between the convolutional neural network and Unet after adding new signals is described in detail.

[0058] Overall, both networks have improved their performance after adding RD, DRP, SR, and Mapability signals, and Unet is better than convolutional neural networks. Unet effectively solves the problems of information loss and resolution reduction in traditional convolutional neural networks in image segmentation tasks by introducing skip connections and upsampling paths, so it has more advantages in practical applications.

[0059] This article compares the prediction effects of CNN and Unet under different signal combinations. The experiment uses a five-fold cross validation. It can be seen that the additional added signal can achieve improved results under different structures. Therefore, the effect of the above signal enhancement is universal. Among them, the SR signal has the most obvious improvement on the model prediction effect. After the fusion of RD, DRP, SR, and Mapability, the final score on Unet is 0.869dice. The comparison table of the prediction effect of different signal inputs on Unet is as follows:

[0060]

[0061] The comparison table of the prediction effect of different signal inputs on CNN is as follows:

[0062]

[0063]

[0064] In summary, the convolutional hourglass network builds a deeper network structure by stacking multiple hourglass modules. This design can better learn complex feature expressions and improve the model's expressiveness. In addition, the convolutional hourglass network also uses the intermediate supervised learning method and adds supervision signals to the middle layer of the network, which can help the network converge faster and prevent the gradient vanishing problem, thereby improving the learning efficiency and performance of the model on complex tasks. The convolutional hourglass network has a more powerful feature extraction capability. The comparison experiment between it and the baseline is as follows:

[0065]

[0066]

[0067] After adding channel attention, the network can pay more attention to the most critical information by weighted processing of different channels. The following table shows that the performance of the hourglass network structure has been significantly improved on this dataset:

[0068]

[0069] In addition, both the hourglass network and Unet can maintain good results with a small amount of data set, which is suitable for the actual application scenario of structural variation detection, and provides an effective solution for processing a small amount of data in practical applications. The following table shows the comparison of the decline in the effect of the hourglass network and CNN after the training data set is reduced. After the proportion of training data drops from 80% to 20%, the dice score of CNN drops by 0.04, while the hourglass network only drops by 0.017.

[0070] The comparison table of the effect of the hourglass network after the data set is reduced is as follows:

[0071]

[0072]

[0073] The comparison of the effects of CNN after the data set is reduced is as follows:

[0074]

[0075] In the research and optimization of structural variation detection methods based on multiple signal fusion, this paper plans to build a system that can automatically learn and analyze complex biological signals to more accurately detect and locate structural variations. Deep learning can automatically integrate multiple signal sources to improve the recognition accuracy of structural variations. This paper will study how to effectively integrate these signals and how to use the back-propagation mechanism of deep learning to optimize the performance of the model. In addition, this paper also sets up a variety of comparative experiments to explore the impact of different signal type combinations on model performance in order to determine the optimal signal combination strategy, which can also help bioinformatics personnel understand the role of various signals in structural variation detection.

[0076] In the research and optimization of image segmentation algorithms based on convolutional hourglass networks, this paper will deeply analyze the design principles of deep learning network structures and how to improve the accuracy of image segmentation by improving algorithms and adjusting parameters. In addition, this paper will compare and explore convolutional hourglass networks with other mainstream deep learning models (such as U-Net and CNN) to prove the effectiveness of the image segmentation algorithm constructed in this paper.

[0077] The design and implementation of the hospital-oriented structural variation detection system includes the design of data acquisition module, preprocessing module, structural variation algorithm model prediction module, back-end service processing and front-end display interface. The data acquisition module will design a unified data upload interface to obtain gene sequence data publicly available on the Internet or in hospitals, and the preprocessing module will clean and standardize the data; next, the structural variation algorithm model will be used to predict the data and identify possible structural variations; the back-end service processes these results and stores and manages them; the front-end display interface adopts a modern UI design to present all information in a user-friendly graphical visualization to improve the user's understanding of the system.

[0078] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A structural variation detection method based on deep learning, characterized by: The specific steps are as follows: Step S1: Data upload: upload the structural variation feature sequence data, align the sequencing data to the reference genome sequence or reference master transcript sequence, and check the uploaded format; Step S2: Data preprocessing, converting and cleaning the uploaded structural variation feature sequence data, and then storing the uploaded data; Step S3: Structural variation prediction, read the stored data, encode the structural variation feature sequence and the reference genome sequence using image multi-channel encoding, convert them into images, obtain the comparison image of the sample and the reference genome sequence, use the pre-trained structural loading model to identify all the variation regions in the structural variation image, obtain the structural variation fragments with single-base resolution, and map the variation information of the corresponding position of the image back to the original locus, load and predict the image; Step S4: The backend service stores the predicted results and the converted images in a structured manner and provides structured queries; Step S5: Front-end display, UI display of the results.

2. The method for detecting structural variation based on deep learning according to claim 1, characterized in that: As described above, the model loaded in step S3 is a convolutional hourglass network model. The convolutional hourglass network model is used to detect the structural variation area in the image to obtain image segmentation information. The pixel category in the image indicates whether a structural variation occurs at the corresponding base position, thereby obtaining the single-base resolution breakpoint information of the structural variation.

3. The method for detecting structural variation based on deep learning according to claim 2, characterized in that: The convolutional hourglass network model is mainly composed of a convolutional layer, an hourglass module, and an addition module. The convolutional hourglass network model is provided with two groups of connection modes, one of which is: the right end of the addition module is connected to the hourglass module, the right end of the hourglass module is connected to two groups of convolutional layers, and the right end of the convolutional layer is connected to the hourglass module through the addition module; The other group is: the right end of the addition module is connected to the hourglass module, the right end of the hourglass module is connected to two groups of convolutional layers, and an intermediate supervision output module is added at the connection between the two groups of convolutional layers, and the right end of the convolutional layer is connected to the hourglass module through the addition module.

4. The method for detecting structural variation based on deep learning according to claim 3, characterized in that: One group of the addition modules is connected to another group of the addition modules by using a residual connection.

5. The method for detecting structural variation based on deep learning according to claim 2, characterized in that: The convolution and maximum pooling layers in the convolutional hourglass network model first process the image features to a very low resolution. After reaching the lowest resolution, the network begins to continuously upsample and combine features across scales, bringing together information from two adjacent resolutions, performing nearest neighbor upsampling on the lower resolution image, and adding the two sets of features element by element.

6. The method for detecting structural variation based on deep learning according to claim 3, characterized in that: The hourglass module structure is symmetrical, and for each downward layer, there is a corresponding upward layer. After reaching the output resolution of the network, two consecutive rounds of 1x1 convolutions are applied to produce the final network prediction.

7. The method for detecting structural variation based on deep learning according to claim 6, characterized in that: The calculation formula for a single hourglass module is: intermediate_o i =conv(o i ); where x input represents the input of the i-th hourglass network, o represents the output after unet network encoding, and intermediate_o represents the output of the i-th layer of intermediate supervision; the loss function of the hourglass network is defined as: The intermediate supervision output of each layer performs cross entropy loss with the ground truth. The final loss is the sum of the losses of n layers of hourglass networks. The output of the final model is the output of the hourglass network of the last layer.

8. The method for detecting structural variation based on deep learning according to claim 1, characterized in that: The specific steps of the three-channel encoding method in step S3 are as follows: Step Q1: squeeze operation, pooling the global spatial features of each channel as the representation of the channel to form a channel description feature map, i.e., RGB image, calculated as follows: Step Q2: excitation operation, using a fully connected neural network, performs a nonlinear transformation on the result after sequence, outputs the degree of dependence s on each channel, and then adjusts the feature map according to the degree of dependence. The calculation method is as follows: s=F ex (z, W)=σ(g(z, W))=σ(W2δ(W1z)); Where δ represents the ReLU activation function, σ represents the sigmoid activation function, and W is the weight of the neural network.

Images