CN118862971A - A method for constructing a neural network model for predicting LOH regions and its application - Google Patents

Abstract

The present invention relates to the field of sequencing, and in particular to a method for constructing a neural network model for predicting LOH regions and its application. The present invention discloses a method for screening LOH regions, constructing a neural network model through clinical sample historical data and known identified LOH regions, and performing final LOH region determination through data calibration and threshold selection. The present invention constructs a neural network model for identifying LOH regions, which can quickly and accurately identify LOH regions by using sequencing data. The present invention is integrated into sequencing detection products and can be used for genetic diagnosis of imprinted gene-related diseases.

Description

A method for constructing a neural network model for predicting LOH regions and its application

Technical Field

The present invention relates to the technical field of sequencing, and in particular to a method for constructing a neural network model for predicting LOH regions and an application thereof.

Background Art

Copy number variation, loss of heterozygosity and uniparental disomy are large genomic variations that can lead to many common genetic diseases. Loss of heterozygosity (LOH) refers to the situation where one of the two alleles (or part of the nucleotide fragments) of the same locus on a pair of homologous chromosomes is lost, while it still exists on the paired chromosome. In 1929, the genetic mechanism of LOH was first explained by studying the X-ray-induced mutation sites of Drosophila melanogaster. LOH is common in cancer. Studies have shown that LOH can lead to the inactivation of suppressor genes, thereby affecting the occurrence and progression of cancer. There are three main mechanisms for the occurrence of LOH: chromosome loss, partial chromosome loss and gene conversion, among which chromosome loss is the main mechanism for the formation of loss of heterozygosity. The presence of LOH in chromosomes indicates the possible existence of uniparental disomy (UPD). When UPD appears on a specific chromosome, it will cause related diseases due to the genetic imprinting effect. A large number of studies have shown that the risk of Mendelian recessive genetic diseases in the LOH region is significantly increased.

Existing methods for detecting LOH regions include short tandem repeats (STR), methylation detection, chromosome microarray analysis (CMA), etc., but STR detection requires the selection of highly polymorphic STR markers based on the detection purpose and genomic location, which limits the detection method to a certain extent and has a high detection cost; methylation detection has a high time cost; the most ideal technology for detecting LOH is chromosome microarray (CMA), but CMA is a high-throughput and high-resolution screening technology. Under the premise of ensuring data accuracy, the LOH information obtained is very large, and different thresholds need to be set for screening according to different purposes. At the same time, it is necessary to consult a large amount of literature or databases to annotate the data based on the screened information in order to finally obtain a reasonable result report.

In recent years, with the continuous development of high-throughput sequencing technology, whole-exome sequencing (WES) has been widely used in the field of disease prevention and treatment, such as genetic diseases, rare syndromes and complex diseases. In the clinical testing process, since most functional variations are concentrated in exon sequences and exon sequencing is more likely to detect rare variations, a large number of historical samples or samples with unknown information can obtain high-depth functional mutation data through this technology.

At present, after obtaining WES data, the bioinformatics software PLINK is often used to analyze the data. This method uses a fixed-size sliding window to scan each chromosome to find continuous homozygous SNPs. PLINK first calculates the proportion of completely homozygous sliding windows containing a certain SNP. If the proportion exceeds a pre-set threshold, then the SNP is considered to be in a segment of LOH. A certain number of missing or heterozygous SNPs can be specified in each sliding window to include genotyping errors, failures, or rare variations. Finally, if the number of continuous homozygous SNPs in a certain segment exceeds a number or distance threshold (the number of SNPs or the distance of chromosomes), then the segment can be determined to be LOH. This method is not accurate enough for clinical data, and there are cases of missed detection and false positives. The essential reason is that the set threshold cannot meet the classification of all different WES clinical data.

Therefore, there is an urgent need for an analysis method to accurately develop LOH based on sequencing data. The present invention constructs a neural network classification model based on a large amount of clinical data and realizes accurate identification of LOH regions by reviewing historical data.

Summary of the invention

The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

To this end, the present invention provides a method for constructing a neural network model for predicting LOH regions, the method comprising:

(A) obtaining sequencing data of an LOH standard, aligning the sequencing data with a reference genome, and obtaining an alignment result carrying variation information after alignment, wherein the LOH standard is a sample with known LOH region location information;

(B) setting a sliding window on each chromosome information of the comparison result, taking one end of the sliding step in the sliding window as a point, and obtaining SNP information of each point in the sliding window during sliding;

(C) based on the SNP information of each point in the sliding window, counting the homozygosity rate of the points within 1.5 to 3 MB before each point in the sliding window, and constructing a two-dimensional matrix based on the homozygosity rate information;

(D) constructing a feature matrix according to whether the point in step (C) is within the LOH region of the standard;

(E) constructing a recurrent neural network model for predicting the LOH region based on the feature matrix, wherein the input layer of the recurrent neural network model for predicting the LOH region is the two-dimensional matrix, and the output layer is the probability value of the point in the LOH region.

The present invention constructs a neural network model for identifying LOH regions, which can quickly and accurately identify LOH regions by using sequencing data. The present invention is integrated into sequencing detection products and related products for clinical LOH detection, which can be used for genetic diagnosis of imprinted gene-related diseases, quickly locate LOH regions, improve interpretation efficiency, and reduce the turnaround time and cost of orthogonal validation tests.

According to some embodiments of the present invention, the LOH standard in step (A) comprises at least 3 LOH regions.

According to some embodiments of the present invention, the length of each LOH region is not less than 5 MB.

According to some embodiments of the invention, the total length of the LOH regions contained in the LOH standard is no less than 15 MB.

According to some embodiments of the present invention, in step (A), the sequencing data of the LOH standard includes any one of whole exome sequencing data, whole genome sequencing data or panel sequencing data.

According to some embodiments of the present invention, the depth of the panel sequencing is greater than 30X.

According to some embodiments of the present invention, the reference genome comprises any one of CHM13, hg19, hg38, GRCh37, GRCh38, b37, and hs375d.

According to some embodiments of the present invention, in step (A), the comparison result is obtained using a variation detection tool.

According to some embodiments of the present invention, the variation detection tool comprises at least one of GATK, Samtools, and Deepvariant.

According to some embodiments of the present invention, in step (B), the sliding step size is 50-150 kb.

According to some embodiments of the present invention, the sliding step size is 100 kb.

According to some embodiments of the present invention, in step (B), the sliding window size is 2-5 MB.

According to some embodiments of the present invention, the sliding window size is 3.5 MB.

According to some embodiments of the present invention, in step (B), the SNP information includes the number of SNPs, the number of homozygous SNPs and the SNP density.

According to some embodiments of the present invention, in step (C), the homozygosity rate of points within 2MB before each point in the sliding window is counted.

According to some embodiments of the present invention, the neural network model includes any one of a recurrent neural network model, a convolutional neural network model, and a radial basis neural network model.

According to some embodiments of the present invention, the neural network model is a recurrent neural network model, and the recurrent neural network model includes any one of a long short-term memory model, a bidirectional long short-term memory model, and a gated recurrent unit model.

According to some embodiments of the present invention, the neural network model has no less than 3 hidden layers.

According to some embodiments of the present invention, the neural network model activation function includes any one of a tanh function, a Sigmoid function, and a ReLU function.

Another aspect of the present invention provides a method for predicting the LOH region in a sample to be tested, the method comprising:

(1) obtaining sequencing data of a sample to be tested, comparing the sequencing data with a reference genome, and obtaining a comparison result carrying variation information after comparison;

(2) setting a sliding window on each chromosome information of the comparison result, taking one end of the sliding step in the sliding window as a point, and obtaining SNP information of each point in the sliding window during sliding;

(3) based on the SNP information of each point in the sliding window, the homozygosity rate of the points within 1.5 to 3 MB before each point in the sliding window is counted, and a two-dimensional matrix is constructed based on the homozygosity rate information;

(4) using the two-dimensional matrix in step (3) as an input layer, inputting it into the recurrent neural network model for predicting the LOH region constructed by the above construction method, and obtaining an output result, wherein the output result is a probability value of the point in the LOH region, wherein each probability value corresponds to each point in the sliding window in step (3);

(5) Counting the output results corresponding to the continuous points within at least 2MB in the sliding window, wherein when each output probability value corresponding to the continuous points is greater than 0.6, the area formed by the continuous points is determined to be the LOH area of the sample to be tested.

According to some embodiments of the present invention, in step (1), the sequencing data includes any one of whole exome sequencing data, whole genome sequencing data, and panel sequencing data.

According to some embodiments of the present invention, the depth of the panel sequencing is greater than 30X.

According to some embodiments of the present invention, the reference genome comprises any one of CHM13, hg19, hg38, GRCh37, GRCh38, b37, and hs375d.

According to some embodiments of the present invention, the comparison result is obtained using a variation detection tool.

According to some embodiments of the present invention, the variation detection tool comprises at least one of GATK, Samtools, and Deepvariant.

According to some embodiments of the present invention, in step (2), the sliding step size is 50-150 kb.

According to some embodiments of the present invention, the sliding step size is 100 kb.

According to some embodiments of the present invention, in step (2), the sliding window size is 2-5 MB.

According to some embodiments of the present invention, the sliding window size is 3.5 MB.

According to some embodiments of the present invention, the SNP information in step (2) includes the number of SNPs, the number of homozygous SNPs and the SNP density.

According to some embodiments of the present invention, in step (3), the homozygous rate of points within 2MB before each point in the sliding window is counted.

According to some embodiments of the present invention, the method further comprises: establishing a calibration model, obtaining a Z-score threshold based on the calibration model, and screening the LOH region of the sample to be tested outputted in step (4) based on the Z-score threshold and the SNP density threshold to determine the final LOH region of the sample to be tested.

According to some embodiments of the present invention, the threshold of the Z-score is obtained by the following method:

1) replacing the sequencing data of the sample to be tested in step (1) with historical data of clinical samples, and obtaining SNP information and LOH regions of the clinical samples through steps (1) to (4), wherein the historical data includes whole exome sequencing data of clinical samples;

2) Calculate the mean μ and standard deviation δ of the homozygosity rate of the clinical sample in the LOH region, and obtain the homozygosity rate Z-Score of the sample according to the following formula:

Z-Score = (X-μ)/(δ),

Wherein, X is the homozygosity rate of LOH of a single sample in the clinical sample, μ is the mean of the homozygosity rate of the overall data, δ is the standard deviation of the homozygosity rate of the overall data,

3) Based on multiple clinical sample historical data, 0.5% of the area after the standard normal distribution is retained, and the corresponding Z-Score value is the threshold of the Z-score.

According to some embodiments of the present invention, the density threshold of the SNP is obtained by the following method:

4) replacing the sequencing data of the sample to be tested in step (1) with the historical data of the clinical sample, and obtaining the SNP information and LOH region of the clinical sample through steps (1) to (4), wherein the historical data includes the whole exome sequencing data of the clinical sample;

5) Calculate the mean value μ and standard deviation σ of the SNP density of the clinical sample in the LOH region, and select the value of μ-3σ as the threshold of the SNP density according to the 3σ rule.

According to some embodiments of the present invention, the number of clinical samples is no less than 50.

According to some embodiments of the present invention, the number of clinical samples is no less than 200.

According to some embodiments of the present invention, the screening criteria for determining the final LOH region of the sample to be tested are:

The Z-Score is not less than 2.56 and the SNP density is not less than 1.25.

In another aspect, the present invention provides a system for predicting LOH regions in a sample to be tested, the system comprising:

A data comparison module, which is used to compare the sequencing data of the sample to be tested with the reference genome data to obtain a comparison result;

A data statistics module, the data statistics module is connected to the data comparison module, and the data statistics module is used to set a sliding window on each chromosome information of the comparison result, take one end of the sliding step in the sliding window as a point, and count the SNP information of each point in the sliding window during sliding;

A matrix construction module, the matrix construction module is connected to the data statistics module, and the matrix construction module is used to count the homozygosity rate of the points within 1.5 to 3MB before each point in the sliding window based on the SNP information of each point in the sliding window, and to construct a two-dimensional matrix based on the homozygosity rate information;

A data output module, the data output module is connected to the matrix construction module, the data output module includes a data model module, the data output module is used to use the matrix construction module as an input layer and input into the data model module, wherein the output layer of the output module is a probability value of the point in the LOH region, each probability value corresponds to each point in the sliding window in step (3), wherein the neural network model for predicting the LOH region contained in the data model module is constructed by the above-mentioned construction method;

A data determination module is connected to the data output module, and is used to count the output results corresponding to continuous points within at least 2MB in the sliding window, wherein when each output probability value corresponding to the continuous points is greater than 0.6, the area composed of the continuous points is determined to be the LOH area of the sample to be tested.

According to some embodiments of the present invention, in the data comparison module, the sequencing data includes whole exome sequencing data.

According to some embodiments of the present invention, in the data comparison module, the reference genome includes any one of CHM13, hg19, hg38, GRCh37, GRCh38, b37, and hs375d.

According to some embodiments of the present invention, in the data comparison module, the comparison result is obtained using a variation detection tool.

According to some embodiments of the present invention, in the data comparison module, the variation detection tool includes at least one of GATK, Samtools, and Deep variant.

According to some embodiments of the present invention, in the data statistics module, the sliding step size is 50-150 kb.

According to some embodiments of the present invention, in the data statistics module, the sliding step size is 100 kb.

According to some embodiments of the present invention, in the data statistics module, the sliding window size is 2-5 MB.

According to some embodiments of the present invention, in the data statistics module, the sliding window size is 3.5 MB.

According to some embodiments of the present invention, in the data statistics module, the SNP information includes the number of SNPs, the number of homozygous SNPs and the SNP density.

According to some embodiments of the present invention, in the matrix construction module, the homozygosity rate of the points within 2MB before each point in the sliding window is counted.

According to some embodiments of the present invention, the method further includes: a calibration model module, the calibration model module is connected to the data output module, the calibration model module obtains a Z-score threshold based on the calibration model, and screens the LOH region of the sample to be tested output by the data output module based on the Z-score threshold and the SNP density threshold to determine the final LOH region of the sample to be tested.

According to some embodiments of the present invention, the threshold of the Z-score is obtained by the following method:

1) replacing the sequencing data of the sample to be tested in the data comparison module with the historical data of the clinical sample, and obtaining the SNP information and LOH region of the clinical sample through the data comparison module to the data output module, wherein the historical data includes the whole exome sequencing data of the clinical sample;

2) Calculate the mean μ and standard deviation δ of the homozygosity rate of the clinical sample in the LOH region, and obtain the homozygosity rate Z-Score of the sample according to the following formula:

Z-Score = (X-μ)/(δ),

Wherein, X is the homozygosity rate of LOH of a single sample in the clinical sample, μ is the mean of the homozygosity rate of the overall data, δ is the standard deviation of the homozygosity rate of the overall data,

3) Based on multiple clinical sample historical data, 0.5% of the area after the standard normal distribution is retained, and the corresponding Z-Score value is the threshold of the Z-score.

According to some embodiments of the present invention, the density threshold of the SNP is obtained by the following method:

4) replacing the sequencing data of the sample to be tested in step (1) with the historical data of the clinical sample, and obtaining the SNP information and LOH region of the clinical sample through steps (1) to (4), wherein the historical data includes the whole exome sequencing data of the clinical sample;

5) Calculate the mean value μ and standard deviation σ of the SNP density of the clinical sample in the LOH region, and select the value of μ-3σ as the threshold of the SNP density according to the 3σ rule.

According to some embodiments of the present invention, the number of clinical samples is no less than 50.

According to some embodiments of the present invention, the number of clinical samples is no less than 200.

According to some embodiments of the present invention, the screening criteria for determining the LOH region of the final sample to be tested are: the Z-Score is not less than 2.56 and the SNP density is not less than 1.25.

Another aspect of the present invention provides an electronic device for predicting LOH regions in a sample, comprising a memory, a processor;

The processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to implement the above-mentioned method for predicting the LOH region in the sample.

In yet another aspect, the present invention provides a computer-readable storage medium storing a computer program, wherein the program, when executed by a processor, implements the above-mentioned method for predicting LOH regions in a sample.

Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS

The above and/or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

FIG1 shows a schematic flow chart of a method for constructing a neural network model for predicting LOH regions in one embodiment of the present invention;

FIG. 2 shows overall feature extraction and modeling based on a neural network model in one embodiment of the present invention.

DETAILED DESCRIPTION

The embodiments of the present invention are described in detail below. The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. Further, in the description of the present invention, unless otherwise specified, the meaning of "plurality" is two or more.

The endpoints and any values of the ranges disclosed in the present invention are not limited to the precise ranges or values, and these ranges or values should be understood to include values close to these ranges or values. For numerical ranges, the endpoint values of each range, the endpoint values of each range and the individual point values, and the individual point values can be combined with each other to obtain one or more new numerical ranges, and these numerical ranges should be considered as specifically disclosed in this article.

In order to make the present invention more easily understood, certain technical and scientific terms are specifically defined below. Unless otherwise clearly defined elsewhere in the present invention, all other technical and scientific terms used in the present invention have the meanings commonly understood by those skilled in the art to which the present invention belongs.

In the present invention, the terms "comprise" or "include" are open expressions, that is, including the contents specified in the present invention, but not excluding other contents.

In the present invention, the terms "optionally", "optional" or "optionally" generally mean that the subsequently described event or situation can but does not necessarily occur, and the description includes cases where the event or situation occurs and cases where it does not occur.

In the present invention, the term "whole exome sequencing" refers to a genome analysis method that uses sequence capture technology to capture and enrich DNA in the exon region of the whole genome and then performs high-throughput sequencing.

In the present invention, the term "Uniparental Disomy (UPD)" refers to a situation where a chromosome segment from one parent is replaced by a homologous segment from the other parent, or two homologous chromosomes of an individual are from the same parent. The former is called segmental uniparental disomy.

In the present invention, the term "short tandem repeats (STR)", also known as microsatellite DNA, is usually a DNA repeat sequence consisting of 1 to 6 base units in the genome. Since the number of core unit repeats is highly variable between individuals and the number is abundant, it constitutes the genetic polymorphism of the STR locus. It is generally believed that there is an STR locus every 15kb in the human genome on average.

In the present invention, the term "chromosomal microarray analysis" refers to a method for detecting DNA duplications and deletions in the human genome. It is a high-resolution whole-genome screening that can identify major chromosomal aneuploidies and the location and type of specific genetic changes that cannot be detected by conventional karyotype analysis.

In the present invention, the term "VCF file" is used to describe text files of SNP (single base variation), INDEL (insertion and deletion markers) and SV (structural variation site) results.

In the present invention, the term "CNV" refers to abnormal DNA copy number changes, which is an important molecular mechanism for many human diseases (such as cancer, hereditary diseases, and cardiovascular diseases). As a biomarker of disease, changes such as deletion and amplification at the chromosome level have become a hot topic in the study of many diseases. The main operation method is to detect the DNA copy number changes (CNV) of the sample genome relative to the control genome by co-hybridizing samples (case samples and control samples) labeled with different fluorescent substances on a chip. It is often used for whole genome CNV detection of tumors or hereditary diseases, and intuitively shows the deletion or amplification of tumor and hereditary disease genomic DNA in the entire chromosome group. For tumors, the deleted fragments may contain tumor suppressor genes, while the amplified fragments may contain oncogenes.

In the present invention, the term "Recurrent Neural Network model" (RNN) refers to a type of recursive neural network that takes sequence data as input, performs recursion in the direction of sequence evolution, and all nodes (recurrent units) are connected in a chain. The study of recurrent neural networks began in the 1980s and 1990s, and developed into one of the deep learning algorithms in the early 21st century. Among them, bidirectional recurrent neural networks (Bi-RNN) and long short-term memory networks (LSTM) are common recurrent neural networks.

In the present invention, the term "Z-Score", Z score, also called standard score, is the process of dividing the difference between a number and the mean by the standard deviation. In statistics, the standard score is the number of standard deviations above the mean of the observed or measured values.

In the present invention, the term "3σ rule" is also called the three-sigma rule, that is, samples with values distributed in (μ-3σ, μ+3σ) are considered to be the majority of samples, and samples outside three times the standard deviation are considered to be minority samples. μ represents the average value of the SNP density of all LOH intervals, and σ represents the standard deviation of the SNP density of all LOH intervals.

A method for constructing a neural network model for predicting LOH regions

According to some embodiments of the present invention, the present invention proposes a method for constructing a neural network model for predicting LOH regions, as shown in FIG1 , comprising:

S110, obtaining sequencing data of an LOH standard, aligning the sequencing data with a reference genome, and obtaining an alignment result carrying variation information after alignment, wherein the LOH standard is a sample with known LOH region location information;

S120, setting a sliding window on each chromosome information of the comparison result, taking one end of the sliding step in the sliding window as a point, and obtaining SNP information of each point in the sliding window during sliding;

S130, based on the SNP information of each point in the sliding window, counting the homozygosity rate of the points within 1.5 to 3 MB before each point in the sliding window, and constructing a two-dimensional matrix based on the homozygosity rate information;

S140, constructing a characteristic matrix according to whether the point in step S130 is within the LOH region of the standard;

S150. Constructing a recurrent neural network model for predicting the LOH region according to the feature matrix, wherein the input layer of the recurrent neural network model for predicting the LOH region is the two-dimensional matrix, and the output layer is the probability value of the point in the LOH region.

According to a specific implementation scheme of the present invention, the output layer is any value between [0, 1], wherein each value corresponds to the point in step S140, and the closer the value is to 1, the greater the possibility that the point is judged to exist in the LOH region.

According to some specific embodiments of the present invention, the neural network model can be any one of the recurrent neural network model, convolutional neural network model, and radial basis neural network model known in the art, preferably a recurrent neural network model.

According to some embodiments of the present invention, the recurrent neural network model includes any one of a long short-term memory model, a bidirectional long short-term memory model, and a gated recurrent unit model.

According to a more specific embodiment of the present invention, the present invention proposes a method for constructing a recurrent neural network (RNN) model for predicting LOH regions, comprising:

1) Get VCF file

The present invention first obtains the WES genome data of 5 known clinical samples, compares them with the human reference genome, identifies the variants and outputs the VCF file;

2) Obtaining the LOH region coordinates of known clinical samples

The known clinical samples in step 1) are subjected to chromosome microarray (CMA) technical analysis. Specifically, the analysis is performed using the CytoScan750K [550,000 CNV markers and 200,000 single nucleotide polymorphism (SNP) markers] SNP microarray chip (SNP750K) produced by Affymetrix, USA to detect the samples, and the genomic DNA digestion, amplification, purification, fragmentation, labeling, chip hybridization, washing and scanning, and data analysis are performed strictly in accordance with the standard experimental operation procedures provided by Affymetrix to obtain the LOH segment coordinates of the known clinical samples;

3) Model feature acquisition

As shown in Figure 2, first, from the VCF file, the number of SNPs, the number of homozygous SNPs and the SNP density of each point in the window are obtained through a sliding window with a size of 3.5MB and a step size of 100KB. Then, starting from the 2MB position of each chromosome, the homozygous rate (number of homozygous SNPs/total number of SNPs) of each point in the sliding window within the first 2MB of the point is obtained to construct a 1*20 two-dimensional matrix. Then, according to whether the point is contained in the known LOH region, all the points on the chromosome are divided into two categories (whether there is and the LOH region), and an N*20 matrix (N is the number of points that exist/do not exist in the LOH region) is constructed as a feature matrix. During the modeling process, whether the site is in the LOH region is coded as 0 or 1, and the presence in the LOH region is coded as 1, otherwise it is 0;

4) Construct a recurrent neural network (RNN) model based on the feature matrix

The input layer of the recurrent neural network (RNN) model is a 1*20 two-dimensional matrix, which passes through a hidden layer with 5 neurons to an output layer of neurons. The output layer is an arbitrary value in [0,1], wherein each value corresponds to each point in the corresponding output step (3). The closer the value is to 1, the greater the possibility that the point exists in the LOH area.

The activation function of the recurrent neural network (RNN) model is the tanh function (Formula 1), and the RNN model selects the classic long short-term memory model (LSTM). The overall feature extraction and RNN modeling are shown in Figure 2;

Formula 1:

5) Determination of suspected LOH area

The region consisting of continuous sites with a screening length greater than 2MB and output values greater than 0.6 was considered a suspected LOH region;

6) Establishment of calibration model

The clinical sample historical data was used as a reference data set to estimate the mean μ and standard deviation δ of the homozygosity rate of the overall sample in the suspected LOH region, and the homozygosity rate Z value of the sample was obtained according to Formula 2:

Formula 2:

Where x is the observed value of the individual, μ is the mean of the homozygosity rate of the overall data, and δ is the standard deviation of the homozygosity rate of the overall data;

7) Threshold screening and LOH region determination

After Z-Score, the area in the first 0.5% of the standard normal distribution is retained. When the clinical sample historical data is 300 normal clinical WES samples, the area in the first 0.5% of the standard normal distribution is the area with Z-Score>2.56;

The SNP density information of samples located in the LOH region was statistically analyzed. According to the 3σ law, the value of μ-3σ was selected as the threshold of SNP density, where μ represents the mean value of SNP density and σ represents the standard deviation of SNP density. The region with density greater than 1.25 was selected as the final LOH region.

According to a specific embodiment of the present invention, the activation function is used to convert the output into a probability value between 0 and 1. These probability values can be interpreted as the confidence or probability that a given input sample belongs to a certain category. If it is close to 1, it is more likely to be judged as a LOH region, and vice versa.

According to some embodiments of the present invention, in step 1), the LOH of the known clinical samples is known, and the LOH regions of the known clinical samples are different.

According to some embodiments of the present invention, in step 1), the number of the known clinical samples is no less than 3, preferably 3-10, and more preferably 5. When the number of the known clinical samples is 3-7, the prediction of the LOH region is more accurate.

According to some embodiments of the present invention, in step 2), the LOH region coordinates of the known clinical sample can be obtained by any technique known in the art including chromosome microarray (CMA) technical analysis, Mutagenically separated PCR, and whole genome SNP microarray chip (SNP array).

According to some embodiments of the present invention, in step 1), the human reference genome includes any one of CHM13, hg19, hg38, GRCh37, GRCh38, b37, and hs375d.

According to some embodiments of the present invention, the size of the sliding window in step 3) is between 2-5 MB, preferably 3.5 MB. When the sliding window is within this range, the prediction of the LOH region is more accurate.

According to some embodiments of the present invention, the sliding window step size in step 3) is between 50-150 kb, preferably 100 kb. When the sliding step size is within this range, the prediction of the LOH region is more accurate.

According to some embodiments of the present invention, the statistics of the homozygosity rate in step 3) can start from any site between 1.5-3MB of the chromosome, preferably the 2MB site. When the site is within this interval, the prediction of the LOH region is more accurate.

According to some embodiments of the present invention, the number of hidden layers in step 4) is not less than 3, preferably between 3 and 7, and more preferably 5. When the number of hidden layers is within the range of 3 to 7, the prediction of the LOH region is more accurate.

According to some embodiments of the present invention, the selection of the SNP density region is based on the statistics of the SNP density located in the LOH region of historical clinical samples. When the SNP density is greater than 1.25, the prediction of the LOH region is more accurate.

According to some embodiments of the present invention, the baseline may be a region that occupies more than the first 1% of the total population after standard normal distribution, preferably a region that occupies 0.5%. When the baseline Z-Score is within this interval, the prediction of the LOH region is more accurate.

According to some embodiments of the present invention, the activation function in step 4) includes any one of a tanh function, a sigmoid function, and a ReLU function.

The present invention constructs a neural network model for identifying LOH regions, which can quickly and accurately identify LOH regions by using sequencing data. The present invention is integrated into sequencing detection products and related products for clinical LOH detection, which can be used for genetic diagnosis of imprinted gene-related diseases, quickly locate LOH regions, improve interpretation efficiency, and reduce the turnaround time and cost of orthogonal validation tests.

Application of Neural Network Model in Detecting LOH Area Products

According to some specific embodiments of the present invention, the present invention provides a method for applying a neural network model in detecting LOH area products, the method comprising:

1) Compare the WES genome data of the sample to be tested with the human reference genome, identify the variation and output the VCF file;

2) importing the VCF file into the RNN model for predicting the LOH region to obtain the suspected LOH region of the sample to be tested;

3) Bringing the clinical sample historical data into the RNN model for predicting the LOH region, obtaining the mean μ and standard deviation δ of the homozygosity rate of the clinical sample historical data in the suspected LOH region, and calculating the homozygosity rate Z value of the clinical sample historical data according to the method for screening suspected LOH regions based on the RNN model according to Formula 2, and retaining the region after 0.5% of the standard normal distribution, that is, the region with Z-Score>2.56,

Formula 2:

Where x is the observed value of the individual, μ is the mean of the homozygosity rate of the overall data, and δ is the standard deviation of the homozygosity rate of the overall data;

4) Finally, the SNP density information of the samples located in the LOH region was counted. According to the 3σ law, the μ-3σ value was selected as the SNP density threshold, that is, the region with a density greater than 1.25 was the final LOH region, where μ was the mean SNP density and σ was the standard deviation of the SNP density;

5) Substitute the suspected LOH region of the sample to be tested in step 2) into formula 2, calculate the homozygosity rate Z value of the sample to be tested, and screen the region with Z-Score>2.56 and SNP density greater than 1.25 as the baseline, and determine the part above the baseline as the final LOH region.

According to some embodiments of the present invention, the selection of the SNP density region is based on the statistics of the SNP density located in the LOH region of historical clinical samples. When the SNP density is greater than 1.25, the prediction of the LOH region is more accurate.

According to some embodiments of the present invention, the baseline may be a region that occupies more than the first 1% of the total population after standard normal distribution, preferably a region that occupies 0.5%. When the baseline Z-Score is within this interval, the prediction of the LOH region is more accurate.

The scheme of the present invention will be explained below in conjunction with embodiments. It will be understood by those skilled in the art that the following embodiments are only used to illustrate the present invention and should not be regarded as limiting the scope of the present invention.

Example 1 Construction of a neural network model for predicting LOH regions

(1) Acquisition of variant sites

The whole exome sequencing data of 5 clinical samples were aligned with the human reference genome (hg19), and GATK was used to identify variants and output VCF files;

(2) Acquisition of known LOH regions

The five clinical samples in step (1) were subjected to chromosome microarray (CMA) analysis. The samples were tested using the CytoScan750K [550,000 CNV markers and 200,000 single nucleotide polymorphism (SNP) markers] SNP microarray chip (SNP750K) produced by Affymetrix, USA. The genomic DNA digestion, amplification, purification, fragmentation, labeling, chip hybridization, washing and scanning, and data analysis were performed strictly in accordance with the standard experimental operation procedures provided by Affymetrix to obtain the LOH segment coordinates of the five clinical samples. As shown in Table 1, a total of 10 LOH regions were identified, with a total region length of approximately 265.67 MB.

Table 1

(3) Model feature acquisition

First, from the VCF file in step (1), the number of SNPs, the number of homozygous SNPs and the SNP density of each point in the window are obtained by sliding the window with a size of 3.5MB and a step size of 100KB. Then, starting from the 2MB position of each chromosome, the homozygous rate (number of homozygous SNPs/total number of SNPs) of each point in the sliding window within the first 2MB of the point is obtained to construct a 1*20 two-dimensional matrix.

According to whether the point is contained in the known LOH region in step (2), all points on the chromosome are divided into two categories (whether there is and the LOH region). Points that exist in the LOH region are marked as 1, that is, points in the LOH region; the remaining points on the chromosome are marked as 0, indicating normal sites that do not exist in the LOH region. An N*20 matrix (N is the number of points that exist/do not exist in the LOH region) is constructed as the feature matrix.

(4) RNN model establishment

According to the feature matrix, a recurrent neural network (RNN) model is constructed, whose input layer is a 1*20 two-dimensional matrix, passing through a hidden layer with 5 neurons to an output layer of one neuron, and the output layer is any decimal in [0,1].

During the modeling process, whether the site is in the LOH region is coded as 0 or 1. If it is in the LOH region, it is coded as 1, otherwise it is 0. Therefore, after the RNN model, each point corresponds to an output of any decimal in [0,1]. The closer its value is to 1, the greater the possibility that the point is in the LOH region.

The RNN model uses the classic long short-term memory model (LSTM), and the activation function uses the tanh function (Formula 1).

Formula 1:

(5) Determination of suspected LOH areas

The region consisting of continuous sites with output values greater than 0.6 within 2 MB was regarded as a suspected LOH region.

Example 2 Validation of the Neural Network Model for Predicting LOH Regions

The verification case uses a total of 6 test samples, and the test samples have a total of 36 known LOH regions. The known LOH regions are obtained by referring to step (2) in Example 1.

(1) Obtaining VCF files of 6 samples to be tested

The six samples to be tested were subjected to whole exome sequencing, the sequencing data were aligned with the human reference genome (hg19), and GATK was used to identify variants and output VCF files;

(2) Obtaining the input layer of the neural network model

From the VCF file in step (1), the number of SNPs, the number of homozygous SNPs and the SNP density of each point in the window are obtained by sliding the window with a size of 3.5MB and a step size of 100KB. Then, starting from the 2MB position of each chromosome, the homozygous rate (number of homozygous SNPs/total number of SNPs) of each point in the sliding window within the first 2MB of the point is obtained to construct a 1*20 two-dimensional matrix;

(3) Acquisition of suspected LOH regions

The two-dimensional matrix described in step (2) is put into the neural network model for predicting LOH regions constructed in Example 1 to obtain the suspected LOH regions of the six samples to be tested;

(4) Calibration model (Z-SCORE) establishment

The 6 samples to be tested in step (1) were replaced with 300 clinical history samples for which WES had been obtained, and steps (1) to (3) were repeated to obtain the regions of suspected LOH in the clinical history samples.

Calculate the mean μ and standard deviation δ of the homozygosity rate in the suspected LOH region of 300 clinical history samples, as well as the mean μ and standard deviation σ of the SNP density;

The homozygosity rate Z value of the sample is obtained according to formula 2, where x is the observed value of the individual, μ is the mean of the homozygosity rate of the overall data, and δ is the standard deviation of the homozygosity rate of the overall data;

Formula 2:

According to the 3σ law, the value of μ-3σ was selected as the threshold of SNP density, where μ represented the average value of SNP density and σ represented the standard deviation of SNP density. The region with density greater than 1.25 was selected as the final LOH region.

After Z-Score, the regions in the first 0.5% of the standard normal distribution, that is, the regions with Z-Score>2.56 and SNP density greater than 1.25 were retained as the final determined LOH regions;

(5) Screening of LOH regions

Substitute the suspected LOH regions of the 6 samples obtained in step (3) into formula 2 in step (4) to obtain the Z-Score of the 6 samples, and screen the regions with Z-Score>2.56. The regions with SNP density greater than 1.25 are not finally determined as LOH regions.

The results show that, using the neural network model for predicting LOH regions provided by the present invention, 35 LOH regions were finally identified in 6 samples to be tested, with an identification accuracy of 97% and an average LOH region length error of no more than 1MB. As shown in Table 2, Table 2 only shows 10 regions of 5 samples to be tested, including the only unidentified region.

Table 2

The present invention constructs a neural network model for predicting LOH regions, which can quickly and accurately identify LOH regions by using sequencing data. The present invention is integrated into sequencing detection products and related products for clinical LOH detection, which can be used for genetic diagnosis of imprinted gene-related diseases, quickly locate LOH regions, improve interpretation efficiency, and reduce the turnaround time and cost of orthogonal validation tests.

In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", "some embodiments" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, substitute and vary the above embodiments within the scope of the present invention.

Claims (14)

1. A method for constructing a neural network model for predicting LOH regions, characterized in that the method comprises: (A) obtaining sequencing data of an LOH standard, aligning the sequencing data with a reference genome, and obtaining an alignment result carrying variation information after alignment, wherein the LOH standard is a sample with known LOH region location information; (B) setting a sliding window on each chromosome information of the comparison result, taking one end of the sliding step in the sliding window as a point, and obtaining SNP information of each point in the sliding window during sliding; (C) based on the SNP information of each point in the sliding window, counting the homozygosity rate of the points within 1.5 to 3 MB before each point in the sliding window, and constructing a two-dimensional matrix based on the homozygosity rate information; (D) constructing a feature matrix according to whether the point in step (C) is within the LOH region of the standard; (E) constructing a neural network model for predicting the LOH region based on the feature matrix, wherein the input layer of the neural network model for predicting the LOH region is the two-dimensional matrix, and the output layer is the probability value of the point in the LOH region. 2. The method according to claim 1, characterized in that in step (A), the LOH standard contains at least 3 LOH regions; Optionally, each LOH region is no less than 5 MB in length; Optionally, the total length of the LOH region contained in the LOH standard is not less than 15MB; Optionally, the sequencing data of the LOH standard includes any one of whole exome sequencing data, panel sequencing data, and whole genome sequencing data; Optionally, the reference genome includes any one of CHM13, hg19, hg38, GRCh37, GRCh38, b37, hs375d; Optionally, the alignment result is obtained using a variation detection tool; Optionally, the variation detection tool comprises at least one of GATK, Samtools, and Deep variant. 3. The method according to claim 1, characterized in that in step (B), the sliding step size is 50-150 kb; Optionally, the sliding step size is 100 kb; Optionally, the sliding window size is 2-5MB; Optionally, the sliding window size is 3.5MB; Optionally, the SNP information includes the number of SNPs, the number of homozygous SNPs and the SNP density. 4. The method according to claim 1 is characterized in that in step (C), the homozygous rate of points within 2MB before each point in the sliding window is counted. 5. The method according to claim 1, characterized in that the neural network model comprises any one of a recurrent neural network model, a convolutional neural network model, and a radial basis neural network model; Optionally, the neural network model is a recurrent neural network model, and the recurrent neural network model includes any one of a long short-term memory model, a bidirectional long short-term memory model, and a gated recurrent unit model; Optionally, the neural network model has no less than 3 hidden layers; Optionally, the neural network model activation function includes any one of a tanh function, a Sigmoid function, and a ReLU function. 6. A method for predicting LOH regions in a sample to be tested, characterized in that the method comprises: (1) obtaining sequencing data of a sample to be tested, comparing the sequencing data with a reference genome, and obtaining a comparison result carrying variation information after comparison; (2) setting a sliding window on each chromosome information of the comparison result, taking one end of the sliding step in the sliding window as a point, and obtaining SNP information of each point in the sliding window during sliding; (3) based on the SNP information of each point in the sliding window, the homozygosity rate of the points within 1.5 to 3 MB before each point in the sliding window is counted, and a two-dimensional matrix is constructed based on the homozygosity rate information; (4) using the two-dimensional matrix of step (3) as an input layer, inputting it into a neural network model for predicting the LOH region constructed by the method of any one of claims 1 to 5, and obtaining an output result, wherein the output result is a probability value of the point in the LOH region, wherein each probability value corresponds to each point in the sliding window of step (3); (5) Counting the output results corresponding to the continuous points within at least 2MB in the sliding window, wherein when each output probability value corresponding to the continuous points is greater than 0.6, the area formed by the continuous points is determined to be the LOH area of the sample to be tested. 7. The method according to claim 6, characterized in that in step (1), the sequencing data comprises any one of whole exome sequencing data, panel sequencing data, and whole genome sequencing data; Optionally, the reference genome includes any one of CHM13, hg19, hg38, GRCh37, GRCh38, b37, hs375d; Optionally, the alignment result is obtained using a variation detection tool; Optionally, the variation detection tool comprises at least one of GATK, Samtools, and Deep variant. 8. The method according to claim 6, characterized in that in step (2), the sliding step size is 50-150 kb; Optionally, the sliding step size is 100 kb; Optionally, the sliding window size is 2-5MB; Optionally, the sliding window size is 3.5MB; Optionally, the SNP information includes the number of SNPs, the number of homozygous SNPs and the SNP density. 9. The method according to claim 6 is characterized in that in step (3), the homozygous rate of points within 2MB before each point in the sliding window is counted. 10. The method according to claim 6 is characterized in that the method further comprises: establishing a calibration model, obtaining a Z-score threshold based on the calibration model, and screening the LOH region of the sample to be tested output in step (4) based on the Z-score threshold and the SNP density threshold to determine the final LOH region of the sample to be tested. 11. The method according to claim 10, characterized in that the threshold of the Z-score is obtained by the following method: 1) replacing the sequencing data of the sample to be tested in step (1) with historical data of clinical samples, and obtaining SNP information and LOH regions of the clinical samples through steps (1) to (4), wherein the historical data includes whole exome sequencing data of clinical samples; 2) Calculate the mean μ and standard deviation δ of the homozygosity rate of the clinical sample in the LOH region, and obtain the homozygosity rate Z-Score of the sample according to the following formula: Z-Score = (X-μ)/(δ), Wherein, X is the homozygosity rate of LOH of a single sample in the clinical sample, μ is the mean of the homozygosity rate of the overall data, δ is the standard deviation of the homozygosity rate of the overall data, 3) Based on multiple clinical sample historical data, 0.5% of the area after the standard normal distribution is retained, and the corresponding Z-Score value is the threshold of the Z-score. 12. The method according to claim 10, characterized in that the density threshold of the SNP is obtained by the following method: 4) replacing the sequencing data of the sample to be tested in step (1) with the historical data of the clinical sample, and obtaining the SNP information and LOH region of the clinical sample through steps (1) to (4), wherein the historical data includes the whole exome sequencing data of the clinical sample; 5) Calculate the mean value μ and standard deviation σ of the SNP density of the clinical sample in the LOH region, and select the value of μ-3σ as the threshold of the SNP density according to the 3σ rule. 13. The method according to claim 11 or 12, characterized in that the number of clinical samples is not less than 50; Optionally, the number of clinical samples is no less than 200. 14. The method according to any one of claims 10 to 12, characterized in that the screening criteria for determining the final LOH region of the sample to be tested are: The Z-Score is not less than 2.56 and the SNP density is not less than 1.25.