Method and system for detecting fetal chromosomal abnormalities - Patents.com

Abstract

The present invention relates to the field of biotechnology. A method and system for detecting fetal chromosomal abnormalities is disclosed. The method includes: (1) obtaining sequencing data and clinical phenotypic feature data of cell-free nucleic acid fragments from a target pregnant woman, the sequencing data including a plurality of read segments, and the clinical phenotypic feature data of the target pregnant woman forming a phenotypic feature vector of the target pregnant woman; (2) performing window division on at least a portion of a chromosomal sequence of a reference genome to obtain a plurality of sliding windows, counting the read segments included in the sliding windows, and generating a sequence feature matrix of the chromosomal sequence; (3) inputting the sequence feature matrix into a trained machine learning model to extract a sequence feature vector of the chromosomal sequence; and (4) combining the sequence feature vector and the phenotypic feature vector of the target pregnant woman to form a composite feature vector, and inputting the composite feature vector into a classification detection model to obtain the fetal chromosomal abnormality status of the target pregnant woman.

Description

[Technical field]

[0001] The present invention relates to the field of biotechnology, and more particularly to a method and system for detecting fetal chromosomal abnormalities. [Background technology]

[0002] Chromosomal aneuploidy refers to a type of serious genetic disease that affects normal gene expression by increasing or decreasing the number of individual chromosomes in the fetus. The main types include trisomy 21 syndrome, trisomy 18 syndrome, trisomy 13 syndrome, and 5p syndrome. Chromosomal aneuploidy carries a high risk of death and disability, and there is no effective treatment. Currently, prenatal screening and prenatal diagnosis are mainly used to reduce the birth rate of children with chromosomal aneuploidy.

[0003] Traditional chromosomal aneuploidy detection mainly includes non-invasive prenatal screening based on ultrasound diagnostic examination or serological screening, and prenatal diagnosis based on invasive sampling. Ultrasonography-based prenatal screening methods can determine whether fetal chromosomes are abnormal by checking the thickness of the fetal nuchal translucency (NT) at 10-14 weeks of gestation. It is generally believed that if the NT is greater than 3mm, the risk of fetal chromosomal aneuploidy is higher. Serology-based prenatal screening is performed by detecting the concentrations of alpha-fetoprotein (AFP) and human chorionic gonadotropin (HCG) in maternal serum at 13-16 weeks of gestation, and calculating the risk factors of fetal chromosomal abnormalities in combination with the expected delivery date and age of the pregnant woman and the gestational age at the time of blood collection. Invasive sampling-based prenatal diagnosis methods generally obtain fetal samples by amniocentesis, umbilical cord puncture or direct chorionic villus sampling at 16-24 weeks of gestation to detect whether the fetus has chromosomal abnormalities. The combined screening method based on ultrasound diagnostic examination and serological examination does not directly detect fetal chromosomes, but estimates the risk of fetal disease. The detection accuracy is 50%-95%, and the false positive rate is 3%-7%. [1,2]Methods based on invasive sampling are the "gold standard" for the detection and diagnosis of fetal chromosomal abnormalities, which can accurately and directly diagnose fetal aneuploidy. However, this method has a certain miscarriage rate (0.5%-2%), and pregnant women suffering from infectious diseases such as hepatitis B are not suitable for invasive sampling (e.g., amniocentesis) due to the risk of infection to the fetus. In addition, amniocentesis needs to be performed under the guidance of B-scan ultrasonography, which takes a long time and places high technical requirements on the operator.

[0004] With the discovery of fetal cell-free DNA (cfDNA) in maternal peripheral blood, the maturity of next-generation sequencing (NGS) technology, the significant reduction in sequencing costs, and the development of information analysis technology, non-invasive prenatal testing (NIPT) based on NGS technology is becoming the most widely used prenatal screening method for fetal chromosomal aneuploidy diseases. NIPT technology uses maternal peripheral blood, determines the sequence of cell-free DNA in maternal peripheral blood (containing cell-free fetal DNA) by NGS technology, and obtains fetal genetic information in combination with bioinformatics analysis to detect whether the fetus is affected by chromosomal abnormality diseases such as trisomy 21 syndrome (Down syndrome), trisomy 18 syndrome (Edwards syndrome), and trisomy 13 syndrome (Patau syndrome).

[0005] NIPT technology is widely used in clinical practice due to its high sensitivity and specificity (sensitivity of over 99% for T21, T18, and T13, respectively) and low false positive rate (<0.1%). [3-5] . NIPT technology can reduce the false positive rate of serological screening and avoid the risks of intrauterine infection and miscarriage of the fetus caused by invasive prenatal diagnostic procedures (such as amniocentesis and chorionic villus sampling). It is a non-invasive prenatal screening technology with high safety in the early and middle stages of pregnancy.

[0006] Conventional NIPT based on NGS technology calculates sequencing read counts and performs a baseline Z test. [6]The method detects fetal chromosomal abnormalities by using the NGS technique. The principle is as follows. First, a maternal peripheral blood sample is collected at 12 to 22 weeks of pregnancy, and the cell-free DNA in the peripheral blood sample is sequenced using NGS technology, the obtained sequencing read segments are aligned with the human reference genome sequence (and the GC content is simultaneously corrected for the read counts), the number of unique mapping reads for each chromosome is counted, and its ratio to the total number of unique mapping read counts for the chromosome in the sample is calculated, and the Z-score of the chromosome in the detection sample is obtained by subtracting the average value of the corresponding ratio of the unique mapping read counts of the chromosome in the control sample (i.e., normal sample) from the ratio of the unique mapping reads of the chromosome in the detection sample. Then, it is divided by the standard deviation of the corresponding ratio of the unique mapping read counts of the chromosome in the control sample. Finally, the Z-score is compared with a given threshold value, and if the Z-score is greater than the threshold value, it is determined that the risk of trisomy syndrome is high, and if not, it is determined that the risk of trisomy syndrome is low. Here, the average value of the unique mapping read counts of each chromosome in the normal samples of the control group is the baseline value. Therefore, the more normal samples present in the control group, the more accurate the mean value and standard deviation of the proportion of unique mapping reads can be obtained, and as a result, the more accurate the Z-score can be obtained. Here, the given threshold value of Z-score is generally 3, which is statistically defined, i.e., 99.9% deviation from normal expectation.

[0007] Different statistical hypothesis tests can be selected depending on different baseline values. For example, see Ref. [7]In this paper, correlation analysis and T-test are adopted, the median of the read counts of each chromosome in a fixed size window in the sample is used as a baseline value to represent the read count of this chromosome, the median of the total read counts of the chromosomes in the sample is used to represent the read count of the sample, the read count of each chromosome is divided by the read count of the sample to obtain the normalized read count of the corresponding chromosome, and finally, the normalized read counts of each chromosome of all the samples in the control group are used to calculate the confidence interval, and if the score of the sample to be detected is not included in the confidence interval, the sample is considered to be abnormal. [8] proposes to select a reference chromosome with a similar GC content to the chromosome of interest (e.g., chromosome 21) in a sample of known karyotype, and use the read count of the reference chromosome as a baseline value for the Z test, thereby maximizing the accuracy of detecting the chromosomal abnormality of interest in a sample of known karyotype. The reference chromosome serving as the baseline value is a so-called internal chromosome. As another example, a non-invasive fetal trisomy (NIFTY) detection method is described in Reference 1. [9] has been proposed in. In addition to comparing chromosomal read counts with those of normal control samples, this method also considers the proportion of cell-free fetal DNA. This method uses binary hypothesis testing, log-likelihood ratio, and the FCAPS binary partitioning algorithm to determine the detection results. NIFTY is a genome-wide approach. Although this method has been validated with high accuracy by a large population, the process is relatively complicated. The aforementioned statistical hypothesis testing (Z-test or T-test) method based on read counts is the key to current NIPT analysis.

[0008] The aforementioned statistical hypothesis tests (such as Z-tests) based on read counts are currently the mainstream NIPT analysis methods, but these analysis methods have obvious limitations, as follows: (1) Current NIPT analysis methods may bring deviations in the sequencing read segment distribution of individual samples, leading to variations in Z-score calculations in different situations, thereby affecting the final result judgment and related performance indicators. (2) Current NIPT analysis methods are highly dependent on the proportion of cell-free fetal DNA in maternal peripheral blood, and when the proportion of cell-free fetal DNA is excessively low (less than 4%), the risk of false-negative detection may increase due to the large individual differences between pregnant women. (3) Current NIPT analysis methods perform well in detecting trisomy 21 syndrome, but their accuracy in detecting trisomy 18 syndrome and trisomy 13 syndrome is low due to the individual differences between pregnant women and the deviation of GC content in different chromosomes. (4) Current NIPT analysis methods mainly detect common trisomy syndromes such as Down syndrome, but their clinical effectiveness in detecting chromosomal microdeletion and microduplication syndromes, which have a high overall incidence such as George syndrome and Prader-Willi syndrome, is limited.

[14] .

[0009] In addition, new techniques have been proposed to detect chromosomal abnormalities based on machine learning models using NIPT sequencing results. For example, refs.

[10] In

[2013] , a method is proposed to support NIPT decisions using support vector machines (SVMs). In this method, six different Z-score results are obtained by calculating different baseline values, and the clinical indication of the sample is also added to train the SVM model to determine chromosomal abnormalities. As another example, see Ref.

[11] In this paper, a Bayesian method for determining chromosomal abnormalities has been devised. This method utilizes prior information on the cell-free fetal DNA percentage, eliminates population-level interference and maternal CNVs using a hidden Markov model (HMM), performs GC content correction, and then calculates the Bayes factor by combining the likelihood value of the Z test and the estimated prior value of the cell-free fetal DNA percentage from the sex chromosome content. At the same time, multiple risk factors such as the age of the pregnant woman are incorporated into the prior probability to correct the Bayes factor, and the Z score and Bayes factor are integrated to evaluate whether a chromosome is abnormal. As another example, in Japanese Patent Publication No.

[12] It has been proposed to use NIPT sequencing results to train a simple convolutional neural network model to detect chromosome copy number variations and chromosomal aneuploidy abnormalities.

[13] It is proposed to first isolate cell-free fetal DNA and cell-free maternal DNA from peripheral blood samples, amplify various single nucleotide variant (SNV) loci from the isolated cell-free DNA, sequence the amplification products, and determine the gene sequence data or gene sequence data of multiple SNV loci.Then, based on these gene sequence data or gene sequence data, train an artificial neural network model to detect the ploidy normal state of individual chromosomes, tissue cancer state, or organ transplant rejection state.

[0010] In addition, the aforementioned methods for detecting chromosomal abnormalities based on machine learning models using NIPT sequencing results have the following limitations: Most of these methods calculate the desired features for model training based on the read counts of sequencing data, and most of these methods rely on calculating Z scores, which are too complicated to calculate (e.g., Ref.

[11] ) or the model design is too simple (e.g.,

[12] ) or gene sequencing data or gene sequence data based on SNV loci are required (e.g., patent publication

[13] ),which limits the prospects for clinical application, the,scalability and detection accuracy of the model,,so the detection accuracy needs to be improved. Summary of the Invention [Problem to be solved by the invention]

[0011] In view of the problems existing in the prior art in detecting chromosomal abnormalities, particularly aneuploidy, the present invention aims at least to further improve the detection accuracy of chromosomal abnormalities based on a deep hybrid model in order to detect chromosomal abnormalities more effectively.

[0012] Therefore, the method for detecting a fetal chromosomal abnormality according to the first aspect of the present invention comprises: (1) obtaining sequencing data of cell-free nucleic acid fragments and clinical phenotypic feature data from a subject pregnant woman, the sequencing data including a plurality of read segments, and the clinical phenotypic feature data of the subject pregnant woman forming a phenotypic feature vector of the subject pregnant woman; (2) performing window division on at least a part of the chromosome sequence of the reference genome to obtain a plurality of sliding windows, counting the read segments included in the sliding windows, and generating a sequence feature matrix of the chromosome sequence; (3) inputting the sequence feature matrix into a trained machine learning model to extract a sequence feature vector of the chromosome sequence; (4) combining the sequence feature vector and the phenotype feature vector of the pregnant woman to form a composite feature vector, and inputting the composite feature vector into a classification detection model to obtain a fetal chromosomal abnormality status of the pregnant woman to be detected; Includes.

[0013] In one embodiment, in step (1), the cell-free nucleic acid fragments are derived from the peripheral blood, liver, and / or placenta of the pregnant woman.

[0014] In one embodiment, in step (1), the cell-free nucleic acid fragments are cell-free DNA.

[0015] In one embodiment, in step (1), the sequencing data is derived from ultra-low depth sequencing, preferably the sequencing depth of the ultra-low depth sequencing is 1×, 0.1×, or 0.01×.

[0016] In one embodiment, in step (1), the read segments are aligned to the reference genome to obtain unique mapping reads (preferably GC content corrected), and preferably subsequent steps are performed using the unique mapping reads (preferably the read segments are GC content corrected).

[0017] In one embodiment, the GC content correction is performed as follows: First, randomly select m fragments of length l from chromosomes of a human reference genome; b. Number of fragments with GC content i, N i is calculated according to the following formula:

number

number

number

number

number

number

[0018] In one embodiment, in step (1), the clinical phenotypic characteristic data of the pregnant woman is selected from one or a combination of two or more of age, gestational age, height, weight, BMI, biochemical test results of prenatal testing, ultrasound diagnostic results, and cell-free fetal DNA concentration in plasma.

[0019] In one embodiment, in step (1), the clinical phenotype characteristic data of the pregnant woman is subjected to outlier treatment, missing value treatment, and / or null value treatment.

[0020] In one embodiment, in the step (1), the following records are recorded: ax age <10 or x age >80; bx GW <5 or x GW >50; cx height <40 or x height >300; dx weight <10 or x weight >200; is displayed, the phenotype data of the pregnant woman sample is determined to be abnormal values, and these abnormal values ​​are set as null values.

[0021] In one embodiment, the missing values ​​and the null values ​​are padded by a Miss Forest algorithm.

[0022] In one embodiment, in step (2), the chromosome is chromosome 21, chromosome 18, chromosome 13 and / or a sex chromosome.

[0023] In one embodiment, the step (2) is (2.1) Using a window of length b, overlap and slide the chromosome sequence of length L of the reference genome with a step size t to obtain a sliding window, where b is a positive integer, b=[10000,10000000], t is any positive integer, L is a positive integer, and L≧b; (2.2) counting the read segments included within each of the sliding windows and generating a sequence feature matrix of the chromosome sequence; Includes.

[0024] In one embodiment, in step (2), the sequence feature matrix includes the number of read segments within the sliding window, base quality, and mapping quality.

[0025] In one embodiment, the base quality comprises the mean, standard deviation, skewness, and / or kurtosis of the base quality.

[0026] In one embodiment, the mapping quality comprises a mean, a standard deviation, a skewness, and / or a kurtosis of the mapping quality.

[0027] In one embodiment, in step (2), the sequence feature matrix is ​​represented by the following formula:

number

[0028] In one embodiment, in step (3), the sequence feature matrix is ​​normalized.

[0029] In one embodiment, in step (3), the sequence feature matrix is ​​normalized using formula (I):

number

[0030] In one embodiment, in step (3), the trained machine learning model is a neural network model or an autoencoder model, preferably, the neural network model is a deep neural network model, more preferably, the neural network model is a 1D convolution-based deep neural network model.

[0031] In one embodiment, the structure of the deep neural network model comprises: an input layer receiving the sequence feature matrix; A pre-module connected to the input layer, which performs a first convolution and activation operation on the sequence feature matrix from the input layer to obtain a feature map; a core module connected with the pre-modules, which performs further abstraction and feature extraction of the feature maps from the pre-modules, effectively increasing the depth of the neural network model and thereby enhancing the representational capability of the neural network; a post module coupled to the core module that performs feature abstraction of the feature map from the core module; a first global average pooling layer connected to the post module, which vectorizes the feature map of the feature abstraction representation and outputs the sequence feature vector of the chromosome sequence; Includes.

[0032] In one embodiment, the pre-module comprises: (I) A 1D convolutional layer; (II) a batch normalization layer connected to the 1D convolution layer described in (I); (III) A ReLU activation layer connected to the batch normalization layer described in (II); Includes.

[0033] In one embodiment, the core module is composed of one or more residual sub-modules with the same structure, where the output of each residual module is the input of the next residual module.

[0034] In one embodiment, the residual sub-module comprises: (A) a pre-submodule of the core module, each of which includes a 1D convolution layer, a dropout layer connected to the 1D convolution layer, a batch normalization layer connected to the dropout layer, and a ReLU activation layer connected to the batch normalization layer; (B) a first 1D average pooling layer connected to a pre-submodule of the core module described in (A); (C) a Squeeze-Excite module (SE module) and / or a Spatial Squeeze-Excite module (sSE module) connected to the first 1D average pooling layer described in (B); (D) A first add layer (Add layer) connected to the Squeeze-Excite module and / or the Spatial Squeeze-Excite module described in (C); (E) a second 1D average pooling layer connected to the ReLU activation layer in the pre-module; and (F) A second add layer (Add layer) connected to the first add layer described in (D) and the second 1D average pooling layer described in (E); Includes.

[0035] In one embodiment, the SE module comprises: (a) A second global average pooling layer connected to the first 1D average pooling layer of the residual submodule described in (B); (b) a reshaping layer connected to the second global average pooling layer of (a), wherein an output feature map of the reshaping layer has a size of 1×f, where f is the number of 1D convolution kernels; (c) a first fully connected layer connected to the remodeled layer according to (b), the number of output neurons of the first fully connected layer being f / r SE where f is the number of 1D convolution kernels and r SE a first fully connected layer where is the decreasing speed of the Squeez-Excite module; (d) a second fully-connected layer connected to the first fully-connected layer of (c), the number of output neurons of the second fully-connected layer being f, where f is the number of 1D convolution kernels; and (e) a multiplication layer connected to the second fully connected layer described in (d) and the first 1D average pooling layer of the residual submodule described in (B); Includes.

[0036] In one embodiment, the sSE module comprises: a. a 1×1 1D convolutional layer connected to a first 1D average pooling layer according to (B) using a sigmoid function as an activation function; b. A multiplication layer connected to the first 1D average pooling layer described in (B) and the 1×1 1D convolution layer described in a; Includes.

[0037] In one embodiment, in step (4), the composite feature vector is obtained by combining the sequence feature vector and the phenotypic feature vector of the pregnant woman.

[0038] In one embodiment, in step (4), the composite feature vector x is normalized according to the following formula:

number

[0039] In one embodiment, in step (4), the classification detection model is an ensemble learning model.

[0040] In one embodiment, the ensemble learning model is a stacking or majority voting based ensemble learning model, preferably, the ensemble learning model is one or more of a support vector machine model, a naive Bayes classifier, a random forest classifier, XGBoost and logistic regression.

[0041] In one embodiment, the chromosomal abnormality comprises at least one or more of trisomy 21 syndrome, trisomy 18 syndrome, trisomy 13 syndrome, 5p syndrome, chromosomal microdeletion, and chromosomal microduplication.

[0042] The method for constructing a classification detection model for detecting fetal chromosomal abnormalities according to the second aspect of the present invention includes: (1) obtaining cell-free nucleic acid fragment sequencing data and clinical phenotypic feature data from a plurality of pregnant women, the sequencing data comprising a plurality of read segments, a fetal chromosomal status of each of the pregnant women being known, and the clinical phenotypic feature data of each of the pregnant women forming a phenotypic feature vector for the pregnant women; (2) for each of the pregnant women, performing window division on at least a part of the chromosomal sequence of the reference genome to obtain a plurality of sliding windows, counting the read segments included in the sliding windows, and generating a sequence feature matrix of the chromosomal sequence; (3) constructing a training data set for each of the pregnant women using the sequence feature matrix and the fetal chromosomal status, and training a machine learning model to extract sequence feature vectors of the chromosomal sequences; (4) combining the sequence feature vector and the phenotype feature vector of each of the pregnant women to form a composite feature vector, and training a classification model using the composite feature vector of the pregnant women and the fetal chromosomal status to obtain a trained classification detection model; Includes.

[0043] In one embodiment, said fetal chromosomal status of each pregnant woman is one or more of normal diploidy, chromosomal aneuploidy, partial monosomy syndrome, chromosomal microdeletion and chromosomal microduplication.

[0044] In one embodiment, the chromosomal aneuploidy comprises at least one or more of Trisomy 21 syndrome, Trisomy 18 syndrome, and Trisomy 13 syndrome.

[0045] In one embodiment, the partial monosomy syndrome comprises 5p syndrome.

[0046] In one embodiment, the number of pregnant women is more than 10, and the ratio of the number of normal diploid fetuses to the number of chromosomal aneuploid fetuses is 1 / 2 to 2.

[0047] In one embodiment, in step (3), the training data set is represented as follows:

number

number

[0048] Since the present invention has the same technical features as the first aspect of the present invention except for the trained machine learning model, the definition in the embodiment of the first aspect of the present invention also applies. In this aspect, the trained machine learning model includes an output layer. For example, the structure of the deep neural network model includes an output layer after the first global average pooling layer, and the output layer is a fully connected layer connected to the first global average pooling layer and has one output neuron used to output the chromosomal abnormality state.

[0049] The system for detecting fetal chromosomal abnormalities according to the third aspect of the present invention comprises: A data acquisition module for acquiring sequencing data of cell-free nucleic acid fragments and clinical phenotypic feature data from a sample of a pregnant woman to be detected, the sequencing data including a plurality of read segments, and the clinical phenotypic feature data of the sample of a pregnant woman to be detected forms a phenotypic feature vector of the pregnant woman; A sequence feature matrix generation module that performs window division on at least a part of a chromosome sequence of a reference genome to obtain a plurality of sliding windows, counts the read segments included in the sliding windows, and generates a sequence feature matrix of the chromosome sequence; a sequence feature vector extraction module that inputs the sequence feature matrix into a trained machine learning model to extract a sequence feature vector of the chromosome sequence; a classification detection module that combines the sequence feature vector and the phenotype feature vector of the pregnant woman to form a composite feature vector, and inputs the composite feature vector into a classification detection model to obtain the fetal chromosomal abnormality status of the target pregnant woman; Includes.

[0050] In one embodiment, the system further comprises an alignment module that aligns reads of the sequencing data to a reference genome to obtain unique mapping reads.

[0051] In one embodiment, in the data acquisition module, the cell-free nucleic acid fragments are derived from peripheral blood, liver, and / or placenta of the pregnant woman.

[0052] In one embodiment, in the data acquisition module, the cell-free nucleic acid fragments are cell-free DNA.

[0053] In one embodiment, in the data acquisition module, the sequencing data is derived from ultra-low depth sequencing, preferably the sequencing depth of the ultra-low depth sequencing is 1×, 0.1×, or 0.01×.

[0054] In one embodiment, in the data acquisition module, the read segments are aligned to the reference genome to obtain the unique mapping reads (preferably GC content corrected), and preferably subsequent steps are performed using the unique mapping reads (preferably the read segments are GC content corrected).

[0055] In one embodiment, the GC content correction is performed as follows: First, randomly select m fragments of length l from chromosomes of a human reference genome; b. Number of fragments with GC content i, N i is calculated according to the following formula:

number

number

number

number

number

number

[0056] In one embodiment, in the data acquisition module, the clinical phenotypic characteristic data of the pregnant woman is selected from one or a combination of two or more of age, gestational age, height, weight, BMI, prenatal biochemical test results, ultrasound diagnostic results, and cell-free fetal DNA concentration in plasma.

[0057] In one embodiment, in the data acquisition module, the clinical phenotypic characteristic data of the pregnant woman is outlier treated, missing value treated, and / or null value treated.

[0058] In one embodiment, in the data acquisition module, the following are recorded: ax age <10 or x age >80; bx GW <5 or x GW >50; cx height <40 or x height >300; dx weight <10 or x weight >200; is displayed, the phenotype data of the pregnant woman sample is determined to be outliers, and these outliers are set as null values.

[0059] In one embodiment, the missing values ​​and the null values ​​are padded by a Miss Forest algorithm.

[0060] In one embodiment, in the sequence feature matrix generation module, the chromosome is chromosome 21, chromosome 18, chromosome 13 and / or a sex chromosome.

[0061] In one embodiment, the sequence feature matrix generation module further comprises: (2.1) Using a window of length b, overlap and slide the chromosome sequence of length L of the reference genome with a step size t to obtain a sliding window, where b is a positive integer, b=[10000,10000000], t is any positive integer, L is a positive integer, and L≧b; (2.2) counting the read segments included within each of the sliding windows and generating a sequence feature matrix of the chromosome sequence; Do the following.

[0062] In one embodiment, in the sequence feature matrix generator, the sequence feature matrix includes the number of read segments within the sliding window, base quality, and mapping quality.

[0063] In one embodiment, the base quality comprises the mean, standard deviation, skewness, and / or kurtosis of the base quality.

[0064] In one embodiment, the mapping quality comprises a mean, a standard deviation, a skewness, and / or a kurtosis of the mapping quality.

[0065] In one embodiment, in the sequence feature matrix generation module, the sequence feature matrix is ​​expressed by the following formula:

number

[0066] In one embodiment, in the sequence feature vector extraction module, the sequence feature matrix is ​​normalized.

[0067] In one embodiment, in the sequence feature vector extraction module, the sequence feature matrix is ​​normalized using formula (I):

number

[0068] In one embodiment, in the sequence feature vector extraction module, the trained machine learning model is a neural network model or an autoencoder model, preferably the neural network model is a deep neural network model, more preferably the neural network model is A deep neural network model based on 1D convolution.

[0069] For the deep neural network model, the definition in the embodiment of the first aspect of the invention also applies.

[0070] In one embodiment, in the classification detection module, the composite feature vector is obtained by combining the sequence feature vector and the phenotype feature vector of the pregnant woman.

[0071] In one embodiment, in the classification detection module, the composite feature vector x is normalized according to the following formula:

number

[0072] In one embodiment, in the classification detection module, the classification detection model is an ensemble learning model.

[0073] In one embodiment, the ensemble learning model is a stacking or majority voting based ensemble learning model, preferably, the ensemble learning model is one or more of a support vector machine model, a naive Bayes classifier, a random forest classifier, XGBoost and logistic regression.

[0074] A system for constructing a classification detection model for detecting fetal chromosomal abnormalities according to a fourth aspect of the present invention includes: a data acquisition module for acquiring sequencing data of the cell free nucleic acid fragments and clinical phenotypic feature data from a pregnant woman, the sequencing data comprising a plurality of read segments, a fetal chromosomal status of the pregnant woman being known, and the clinical phenotypic feature data of the pregnant woman forming a phenotypic feature vector for the pregnant woman; A sequence feature matrix generation module that performs window division on at least a part of a chromosome sequence of a reference genome to obtain a plurality of sliding windows, counts the read segments included in the sliding windows, and generates a sequence feature matrix of the chromosome sequence; a sequence feature vector extraction module that uses the sequence feature matrix and the fetal chromosome status to construct a training dataset and train a machine learning model to extract sequence feature vectors of the chromosome sequences; a classification detection module that combines the sequence feature vector and the phenotype feature vector of the pregnant woman to form a composite feature vector, and trains the classification model using the composite feature vectors of a plurality of pregnant women and fetal chromosomal status to obtain a trained classification detection model; Includes.

[0075] In one embodiment, the system further comprises an alignment module that aligns read segments of the sequencing data to a reference genome to obtain the unique mapped reads.

[0076] Since the trained machine learning model has the same technical features as the third aspect of the present invention, the definition in the embodiment of the third aspect of the present invention also applies. In this aspect, the trained machine learning model includes an output layer. For example, the structure of the deep neural network model includes an output layer after the first global average pooling layer, and the output layer is a fully connected layer with the number of output neurons connected with the first global average pooling layer and used to output the chromosomal abnormality status being 1. The method and model of the present invention are based on an innovative algorithm of sequencing data instead of Z-test, which avoids the clinical problem that it is difficult to judge according to the threshold when the result score is included in the "gray area". In addition, as the number of samples (e.g., sequencing data of samples and corresponding phenotype data of pregnant women) increases, the hybrid model proposed by the present invention can be automatically upgraded and optimized to improve the detection accuracy. [Brief description of the drawings]

[0077] [Figure 1] 1 shows a flowchart of a method for detecting fetal chromosomal abnormalities based on a deep neural network hybrid model according to an embodiment of the present invention. [Diagram 2] 1 illustrates the computation of a feature matrix for sequencing data according to one embodiment of the present invention. [Diagram 3] 1 illustrates the structure of a deep neural network according to an embodiment of the present invention. [Figure 4] 1 illustrates a Squeeze-Excite module (SE module) according to an embodiment of the present invention. [Diagram 5] 1 illustrates a Spatial Squeeze-Excite (sSE) module according to an embodiment of the present invention. [Figure 6] 1 illustrates missing value padding of a phenotypic dataset according to an embodiment of the present invention. [Figure 7] 1 illustrates the structure of a stacking-based ensemble learning model according to an embodiment of the present invention. [Figure 8] 1 shows ROC curves of 5-fold cross-validation training results of a stacking-based ensemble learning model according to an embodiment of the present invention. [Figure 9] 1 shows a ROC curve evaluated by a model based on a test set according to an embodiment of the present invention. [Figure 10] 1 shows precision-recall curves evaluated by a model based on a test set according to an embodiment of the present invention. [Figure 11] FIG. 13 shows a confusion matrix diagram for the default decision threshold (i.e., 0.5) according to one embodiment of the present invention. [Figure 12] 13 illustrates a function of precision and recall as a threshold according to an embodiment of the present invention. [Figure 13] FIG. 1 shows a confusion matrix plot for a minimum recall of 0.95 (i.e., limiting Type II error) according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0078] In the present invention, the method for detecting fetal chromosomal abnormalities can be implemented by a system for detecting fetal chromosomal abnormalities, and the method for constructing a classification detection model for detecting fetal chromosomal abnormalities can be implemented by a system of a detection model for detecting fetal chromosomal abnormalities.

[0079] In the present invention, the data acquisition module is used to acquire sequencing data and clinical phenotypic feature data of cell-free nucleic acid fragments of a pregnant woman. The sequencing data includes a plurality of read segments, and the fetal chromosome status of the pregnant woman is known (training sample) or unknown (detection target sample), and the clinical phenotypic feature data of the pregnant woman forms a phenotypic feature vector of the pregnant woman. The data acquisition module may include a data receiving module for receiving the above data. The data acquisition module may further include a sequencing device that can obtain sequencing data by inputting the cell-free nucleic acid of the pregnant woman for sequencing. The sequencing may be high-throughput sequencing and ultra-low depth sequencing, and the sequencing depth of the ultra-low depth sequencing is 1×, 0.1×, or 0.01×. The cell-free nucleic acid may be derived from the peripheral blood, liver, and / or placenta of the pregnant woman. The clinical phenotypic characteristics of said pregnant women and the fetal chromosomal status of said pregnant women (training sample) are available in a database, and the fetal chromosomal status of said pregnant women may be a chromosomal aneuploidy, a microdeletion and / or a microduplication.

[0080] In the present invention, the alignment module is used to align the read segments to a reference genome to obtain the unique mapping reads. Application software for aligning the sequences to a reference genome can be obtained from open source developers, for example, from several online websites, or can be developed in-house.

[0081] In the present invention, the sequence feature matrix generation module is used to divide at least a part of the chromosome sequence of the reference genome into a window to obtain a sliding window, and the read segments included in the sliding window are counted to generate a sequence feature matrix of the chromosome sequence. This can be performed by sliding a window having a fixed length over the chromosome sequence, and the window having a fixed length may be 10k, 100k, 1M, or 10M, etc. The step size can be any length, and is generally set as half the length of the sliding window for convenience of calculation. The length of the chromosome sequence can be longer than the length of the sliding window, and can be 10k, 100k, 1M, 10M, or 100M... up to the length of the entire chromosome. The chromosome can be a target chromosome, such as chromosome 21 corresponding to the detection of trisomy 21 syndrome, chromosome 18 corresponding to the detection of trisomy 18 syndrome, chromosome 13 corresponding to the detection of trisomy 13 syndrome, chromosome XY corresponding to the detection of sex chromosome abnormalities, and all autosomes corresponding to the detection of chromosomal microdeletions / microduplications. For each window, parameters including number of reads, base quality (a measure of sequencing accuracy), and mapping quality (a measure of the accuracy of alignment of the read segment to the reference genome, the higher the mapping quality, the more unique the alignment position of the read segment to the reference genome) are counted, which can be done using computer software.

[0082] In the present invention, the sequence feature extraction module is used to extract sequence features of chromosome sequences. For training data sets, the sequence feature vector generation module uses the sequence feature matrix of the pregnant woman and the fetal chromosome status to construct the training data sets and train a machine learning model to extract the sequence feature vectors of the chromosome sequences. For test data, the sequence feature vector generation module uses the sequence feature matrix to construct a test data set and inputs it into a trained machine learning model, such as a deep neural network model, to extract the sequence feature vectors of the chromosome sequences.

[0083] In the present invention, for the training dataset, a classification detection module, such as a training module of the ensemble learning model, is used to train a classification detection model with the sequence feature vector and a composite feature vector formed by the phenotype feature vector of the pregnant woman and the chromosomal status of the fetus, to obtain the trained classification detection model.

[0084] For the test dataset, the classification detection module is used to combine the sequence feature vector with the phenotype feature vector of the pregnant woman to form a composite feature vector as input and utilize the trained classification detection model to detect chromosomal abnormality status.

[0085] The present invention proposes a completely innovative method for detecting chromosomal abnormalities such as aneuploidy, microdeletion or microduplication. Different from conventional methods, the present invention does not directly detect aneuploidy based on the number of read segments and Z-score, and does not require extra work of data preprocessing and feature extraction selection. Instead, the present invention automatically extracts sequence feature vectors from the sequence feature matrix generated from the sequencing data, combines the sequence feature vectors with the clinical phenotype features of the pregnant woman, detects using the classification detection model, and finally designs a machine learning model to obtain a prediction result of whether the fetal chromosome has a genetic abnormality.

[0086] In the present invention, the machine learning model is used to automatically extract the sequence feature vector from the sequencing data, thereby avoiding the shortcomings of traditional manual extraction of NIPT whole genome sequence features. The method of the present invention not only fully mines the sequencing data information, but also fully utilizes the clinical phenotype information of the pregnant woman (phenotype data information that can be added to the model includes birth age, gestational age, height, weight, BMI (body mass index), biochemical test results of prenatal testing, ultrasound diagnosis results such as NT value, etc.), and combines the extracted sequence feature vector with the phenotype feature vector of the pregnant woman to fully mine the rich feature data information contained in the NIPT sequencing data and the clinical phenotype results of the pregnant woman, ensuring the high reliability and validity of the detection results. The method of the present invention can not only be used to detect common trisomy syndromes, but also can be used to detect other chromosomal defects such as chromosome copy number variations, chromosomal microdeletions, and chromosomal microduplications.

[0087] In the present invention, the extraction of the sequence feature vector can also be performed by using a deep neural network model based on an autoencoder network or a variation autoencoder network, etc.

[0088] In the present invention, an ensemble learning model based on stacking or majority voting is trained to detect chromosomal abnormalities, and the detection of aneuploidy by different classifiers is fully utilized to greatly improve the accuracy of detecting aneuploidy.

[0089] In the present invention, the reference genome refers to a map of the human genome with normal diploid chromosomes produced by the Human Genome Project, such as hg38, hg19, etc. The reference genome may be one or more chromosomes, or may be a part of a chromosome. EXAMPLES

[0090] The present invention will be further described below with reference to specific examples, but the present invention is not limited to these examples.

[0091] Example 1: Example of detection model construction

[0092] In an exemplary embodiment, the process and steps of an exemplary model embodiment for constructing a detection model are described below.

[0093] 1. Obtaining NIPT sequencing data and alignment results

[0094] High-throughput sequencing platform BGIseq500 is used to sequence the training sample, i.e., cell-free nucleic acid fragments of pregnant women (SE35 is adopted, sequencing depth: 0.1×). The fetal chromosome status of said pregnant women is known. Sequencing data is aligned to the reference genome, and repeated alignment sequences are filtered to obtain unique mapping reads.

[0095] 2. Preprocess the unique mapping reads obtained in step 1 above, and re-correct the sequence coverage depth of each coverage region of the genome through the relationship between GC content and sequencing depth. The specific process is as follows (for details, see references):

[15] reference). First, randomly select m fragments of length l from a chromosome (e.g., chromosome 21) of the human reference genome; b. Number of fragments with GC content i, N i is calculated according to the following formula:

number

number

number

number

number

number

[0096] 3. Generate sequence feature matrix

[0097] The result of step 2 above is used to calculate the feature matrix. The calculation process is as follows (as shown in Figure 2): A sliding window of length b is used to slide the target chromosome of length L from the start site to the end site with a step size of t. The following features: a. Number of GC-corrected reads in the region b. Average base quality within the region c. Standard deviation (std) of base quality within the region d. Skewness of base quality within a region e. Kurtosis of base quality within a region f. Average mapping quality within the region g. Standard deviation (std) of mapping quality within the region h. Skewness of mapping quality within a region i. Kurtosis of mapping quality within a region for the region of length b covered by each sliding window, which gives us the sequence feature matrix:

number

number

[0098] Base quality is to quantitatively describe the accuracy of sequencing results.The mean, standard deviation, skewness and kurtosis of base quality refer to the mean, standard deviation, skewness and kurtosis of all base quality in sequencing reads, respectively.Map quality refers to the reliability of the alignment of a given sequencing read segment to a reference genome sequence, and the mean, standard deviation, skewness and kurtosis of map quality refer to the mean, standard deviation, skewness and kurtosis of the map quality of a given sequencing read segment, respectively.

[0099] 4. Building a Deep Neural Network Model

[0100] 4.1 Dataset construction

[0101] Use the results of step 3 to create a training set

number

number

number

[0102] 4.2 Building a deep neural network model

[0103] A deep neural network model is constructed, and its structure is shown in Figure 3. All convolution layers involved in the deep neural network model are 1D convolution operations. Unless otherwise specified, the parameters of the 1D convolution kernels (i.e., 1D filters) are the same, that is, the number of the 1D convolution kernels is f, the size of the 1D convolution kernel is k, the step size of the 1D convolution operation is s, the 1D convolution uses L2 regularization, and the regularization coefficient is r L2 where the initialization function of the 1D convolution kernel is g, the size of the output feature map of the 1D convolution operation is set to remain the same as the size of the input feature map, the size of the pooling kernel is p, and the size of the pooling step is p s It is.

[0104] The dropout ratios used in the dropout layers involved in the deep neural network model are the same and are set as d.

[0105] The structure of the deep neural network model includes: 4.2.1 Input Layer The input layer is used to generate a normalized array feature matrix Z of size h × w. (k) Receive. 4.2.2 Pre-modules A pre-module is used in connection with the input layer to perform a first convolution and activation operation of the input sequence feature matrix to obtain an abstract representation feature map, the pre-module including a 1D convolution layer, a batch normalization layer connected with the 1D convolution layer, and a ReLU activation layer connected with the batch normalization layer. 4.2.3 Core Modules A core module is used in conjunction with the pre-module to perform further abstraction and feature extraction of the feature map, effectively increasing the depth of the neural network model and thereby enhancing the representation capability of the neural network. The core module is composed of three repeated operations of residual modules with the same structure, and the output of each residual module is the input of the next residual module. Each residual module (A) a pre-submodule of the core module, which includes a 1D convolution layer, a dropout layer connected to the 1D convolution layer, a batch normalization layer connected to the dropout layer, and a ReLU activation layer connected to the batch normalization layer, and which is repeated twice, each having the same structure; (B) a first 1D average pooling layer connected to a second pre-submodule of the core module according to (A); (C) A Squeeze-Excite module (SE module) or a Spatial Squeeze-Excite module (sSE module) connected to the first 1D average pooling layer described in (B); Includes. First, let us define the rate of decrease of the SE module as r SEAs shown in FIG. 4, the structure of the SE module is as follows: (a) a second global average pooling layer connected to the first 1D average pooling layer of (B); (b) a reshaping layer connected to the second global average pooling layer of (a), where the size of the output feature map is 1×f, where f is the number of 1D convolution kernels; (c) a first fully connected layer connected to the remodeled layer according to (b), the number of output neurons of which is f / r SE where f is the number of 1D convolution kernels and r SE a first fully connected layer, where is the reduction rate of the SE module; (d) a second fully-connected layer connected to the first fully-connected layer of (c), the number of output neurons of which is f, where f is the number of 1D convolution kernels; and (e) A multiplication layer connected to the first 1D average pooling layer described in (B) and the second fully connected layer described in (d) (see reference 1 for details).

[16] reference). As shown in Figure 5, the structure of the sSE module is as follows: a. A 1×1 1D convolutional layer connected with a first 1D average pooling layer according to (B) using a sigmoid function as an activation function; b. a multiplication layer connected to the first 1D average pooling layer described in (B) and the 1×1 1D convolutional layer described in a; (D) a first add layer (Add layer) connected to the SE module and the sSE module described in (C); (E) a second 1D average pooling layer connected to the ReLU activation layer in the pre-module described in 4.2.2; (F) A second add layer (Add layer) connected to the first add layer described in (D) and the second 1D average pooling layer described in (E); (For more information, see References.)

[17] reference). The above (A) to (D) are the left branch of the residual module, and (E) is the right branch of the residual module. 4.2.4 Post Module The post module has the same structure as the pre module, the only difference being that the number of 1D convolution kernels in the post module is n out and a post module is used to abstract the feature map from the core module before outputting it. 4.2.5 First global average pooling layer A first global average pooling layer is used in conjunction with the post module to vectorize the feature maps of the feature abstraction representation. 4.2.6 Output layer The output layer is connected to the first global average pooling layer and is a fully connected layer with one output neuron, and the activation function is a sigmoid function used to output chromosomal abnormalities.

[0106] 5. Calculating sequence feature vectors

[0107] The training set is used to train a deep neural network model in step 4. The sequence feature vector of the sample is calculated using the trained deep neural network model. This process is as follows: (1) Calculate the normalized sequence feature vector for each sample according to 4.1 above. (2) The normalized sequence feature matrix obtained in (1) is input to the deep neural network model for calculation; (3) The output of the first global average pooling layer of the deep neural network model described in 4.2.5 is

number

[0108] 6. Obtaining phenotype results corresponding to pregnant samples

[0109] Obtaining phenotypic results corresponding to the pregnant samples;

number

[0110] 7. Preprocessing of phenotypic data

[0111] The maternal phenotype dataset is subjected to pre-processing, including outlier treatment and missing or null value treatment.

[0112] (1) Outlier Treatment

[0113] Recorded below: ax age <10 or x age >80; bx GW <5 or x GW >50; cx height <40 or x height >300; dx weight <10 or x weight >200; is displayed, the phenotypic data of the pregnant samples are determined to be outliers, and these outliers are set as null values.

[0114] (2) Handling Missing or Null Values

[0115] The phenotypic data matrix P is defined as follows:

number

[0116] We perform missing value padding using the Random Forest algorithm, a non-parametric missing value padding algorithm based on Random Forests (see references for details).

[18] ) The algorithm is as follows: [Table 1]

[0117] (3) Calculation of BMI

[0118] Using the phenotypic results after missing value padding,

number

[0119] (4) The result of (3) is added to the phenotype feature vector after missing value padding to obtain the final phenotype feature vector.

number

[0120] 8. Generate Composite Feature Vectors

[0121] The sequence feature vector described in step 5 and the final feature vector described in step 7 are combined to obtain a composite feature vector.

number

[0122] 9. Normalizing Composite Feature Vectors

[0123] The composite feature vector described in 8 is normalized by the following formula:

number

[0124] 10. Building an ensemble learning model based on stacking

[0125] Using the results from step 9,

number

number

[0126] We predict aneuploidy using a stacking-based ensemble training algorithm, which is as follows (see references for details):

[19] reference). [Table 2]

[0127] Example 2. Detection of chromosomal abnormalities

[0128] In an exemplary embodiment, the present invention proposes a method for detecting fetal chromosomal abnormality, which uses the nucleic acid sequencing results of non-invasive prenatal testing (NIPT) and the phenotype data of pregnant women together to predict whether genetic abnormality exists in fetal chromosomes.In a specific embodiment, the process and steps of the method for detecting fetal chromosomal abnormality are shown in Figure 1, and the specific process is described below.

[0129] 1. Obtaining NIPT sequencing data and alignment results

[0130] Use the high-throughput sequencing platform BGIseq500 to sequence the samples to be detected (SE35 is used, sequencing depth: 0.1×). Align the sequencing data to the reference genome and filter out repetitive alignment sequences to obtain unique mapping reads.

[0131] 2. Preprocess the unique mapping reads obtained in step 1 above, and re-correct the sequence coverage depth of each coverage region of the genome through the relationship between GC content and sequencing depth.For the specific process, see Example 1.

[0132] 3. Generate sequence feature matrix

[0133] The result of step 2 above is used to calculate the feature matrix. The calculation process is as follows (as shown in Figure 2): A window of length b is used to slide the target chromosome of length L from the start site to the end site with a step size of t. The following features: a. Number of GC-corrected reads in the region b. Average base quality within the region c. Standard deviation (std) of base quality within the region d. Skewness of base quality within a region e. Kurtosis of base quality within a region f. Average mapping quality within the region g. Standard deviation (std) of mapping quality within the region h. Skewness of mapping quality within a region i. Kurtosis of mapping quality within a region for the region of length b covered by each sliding window, which gives us the sequence feature matrix:

number

number

[0134] Base quality is to quantitatively describe the accuracy of sequencing results.The mean, standard deviation, skewness and kurtosis of base quality refer to the mean, standard deviation, skewness and kurtosis of all base quality in sequencing read segment, respectively.Map quality refers to the reliability of the alignment of a given sequencing segment to a reference genome sequence, and the mean, standard deviation, skewness and kurtosis of map quality refer to the mean, standard deviation, skewness and kurtosis of the map quality of a given sequencing read segment, respectively.

[0135] 4. Calculate the sequence feature vector of the sample using the trained deep neural network model in Example 1. The process is as follows: (1) Calculate the normalized sequence feature matrix of the sample according to 4.1 of Example 1. (2) The normalized sequence feature matrix obtained in (1) is input to a deep neural network model for calculation; (3) The output of the first global average pooling layer of the deep neural network model described in 4.2.5 of Example 1 is

number

[0136] 5. Obtaining phenotype results corresponding to the target pregnant sample

[0137] Obtaining a phenotype result corresponding to the target pregnant sample;

number

[0138] 6. Treatment of outliers in phenotypic data

[0139] Recorded below: ax age <10 or x age >80; bx GW <5 or x GW >50; cx height <40 or x height >300; dx weight <10 or x weight >200; is displayed, the phenotype data of the pregnant woman sample to be detected is determined to be an abnormal value, and these abnormal values ​​are set as null values.

[0140] 7. Combine the sequence feature vector from step 4 with the final feature vector from step 6 to obtain a composite feature vector.

number

[0141] 8. Normalizing Composite Feature Vectors

[0142] The composite feature vector described in 7 is normalized by the following formula:

number

[0143] 9. Input the composite feature vector into the stacking-based ensemble learning model of Example 1 to obtain the fetal chromosome status of the pregnant woman to be detected.

[0144] Example 3. Verification example

[0145] 1. Number of samples

[0146] In this example, 1205 samples of "trisomy 21 (T21)" are used as positive samples, and 1600 samples of normal chromosomes (diploid) are used as negative samples.

[0147] [Table 3]

[0148] 2. Preprocess the sequencing data of all positive and negative samples according to the steps described in 2. of Example 1. The number of random sampling fragments is m=50000000, and the fragment number is l=180.

[0149] 3. Generate sequence feature matrices for all positive and negative samples according to the steps described in step 3 of Example 1. The parameters are set as follows: Length of chromosome 21: L=46709983 Sliding window length: b=1000000 Sliding step size: t=500000

[0150] Therefore, the sequence feature matrix corresponding to each sample is obtained with a size of 9×93, i.e., w=9, h=93. Since the beginning of chromosome 21 does not have a mapping sequence in the reference genome, in this embodiment, the first 8 columns of the sequence feature matrix are filtered, i.e., the size of the sequence feature matrix actually used is 9×85.

[0151] 4. Based on the results of step 3, the feature matrix of the corresponding sequencing data in the training set is used to train the deep neural network model. (1) Normalize the feature matrix of the sequencing data of the training set according to 4.1 of Example 1 above and save the normalized model. (2) According to the deep neural network model described in (1), an input tensor of size 2524×85×9 is obtained and the deep neural network model is trained. (3) Train the deep neural network model according to the deep neural network model described in 4.2 of Example 1, and set the parameters of the deep neural network model as follows: Number of 1D convolution kernels: f=32, The size of the 1D convolution kernel is k=8. The step size of the 1D convolution operation is s=1, Regularization factor for l2 of 1D convolution kernel: r l2 = 0.0004, The initialization function g for the 1D convolution kernel is given in Ref.

[20] Using the "He normalization" initialization function described in The size of the output feature map of a 1D convolution operation is the same as the size of the input feature map. Pooling kernel size: p=2, Pooling step size: p s = 2, Dropout ratio of the dropout layer: r d = 0.5, SE module descending speed: r SE = 16, Number of 1D convolution kernels in the post module: n out =8. This embodiment is implemented based on the GPU version of Tensorflow (version=1.12.2) and keras (version=2.2.4). Tables 4-1 to 4-5 show the operation of each layer, the size of the output feature map, and the network connection in the deep neural network model based on the above parameters. [Table 4-1] [Table 4-2] [Table 4-3] [Table 4-4] [Table 4-5] (4) Train the deep neural network using 80% of the samples in the training set, and validate it using 20% ​​to calculate accuracy. (5) The training of the deep neural network is set as follows: iteration epoch=100 and sample batch mini_batch size=64. The gradient descent optimization algorithm is Adam algorithm (parameters β1=0.9, β2=0.999), and the initial training rate is set to 0.01. In the training process, if the accuracy rate is not improved after two consecutive iterations, the training rate is reduced by two times (i.e., multiplied by 0.5), and if the accuracy rate is not improved after 10 consecutive iterations, the training is stopped. (6) The class weight coefficients are introduced into the training process of the deep neural network model (by calculating the class weights using the compute_class_weight() function in the machine learning library scikit-learn (version=0.22.2) and assigning the class weights to the samples of the corresponding classes). (7) Save the trained deep neural network model.

[0152] 5. Calculate the sequence feature vector according to step 5 of Example 1 above. (1) Compute sequence feature matrices for all samples in the entire data set (including the training and test sets) according to step 3 in Example 1 above; (2) Normalizing the sequence feature matrix obtained in (1) above using the sequence normalization model obtained according to the sequence normalization model described in 4.1; (3) The result of (2) above is input to the deep neural network model obtained in 4, and the output layer of the model is modified to the global average pooling layer (i.e., the 65th layer in Table 4-5); (4) According to the process of (3), obtain sequence feature vectors of all samples in the entire dataset (including the training set and the test set).

[0153] 6. Obtain phenotypic features of all samples in the entire dataset (including the training and testing sets) and process outliers of the phenotypic features according to step 7 in Example 1 above.

[0154] 7. Missing value pad the phenotypic features in the training set according to step 7 in Example 1 above and save the missing value padded model.

[0155] 8. Follow step 7 in Example 1 above to calculate BMI for phenotypic features in the training set after missing value treatment, as shown in FIG. 6.

[0156] 9. Following step 8 of Example 1 above, combine the sequence feature vectors in the training set with the phenotypic feature vectors of the corresponding samples to obtain a composite feature vector.

[0157] 10. According to step 9 of Example 1 above, normalize the composite feature vector of each sample in the training set to obtain a normalized feature vector, and save the normalized model of the composite feature vector.

[0158] 11. According to the process of steps 7 to 10 above, perform missing value padding of the phenotypic features of each sample in the test set using the stored padding model for missing values, combine the sequence feature vector of the test set with the phenotypic feature vector of the corresponding sample to obtain a composite feature vector of the test set, and normalize the composite feature vector in the test set using the stored normalization model for the composite feature vector.

[0159] 12. Using the normalized feature vectors of the training set obtained in step 10 above, train an ensemble learning model based on stacking, as shown in Figure 7. This embodiment is implemented based on scikit-learn (version=0.22.2) machine training library, where class weight coefficients are introduced into each base classifier model and final meta-classifier model, and parameters are set as default unless otherwise specified. (1) According to step 10 of Example 1 above, the base classifier used in the example is · SVC with parameters C=0.5, kernel="rbf"; ν-svc with parameters ν=0.25, kernel=“rbf”; ·Gaussian NB (Gaussian Naive Bayes model), A random forest classifier with parameters: n_estimators=100, criterion="gini", max_depth=5, min_samples_leaf=1 and min_samples_split=2, An XGB classifier with parameters: n_estimators=100, min_child_weight=1, gamma=0.1, colsample_bytree=0.8, subsample=0.7, reg_alpha=0.01, max_depth=5, learning_rate=0.05, Logistic regression with parameter C=0.5, Includes. (2) Following step 10 of Example 1, the final meta-classifier is an Extra Trees classifier (a highly randomized tree classifier). The parameters involved in this classifier are set as follows: n_estimators=110, max_depth=6, min_samples_split=3, and min_samples_leaf=1, respectively. (3) Five-fold cross-validation training was performed on the stacking-based ensemble learning model, and the results are shown in Figure 8. It shows that the average AUC obtained by training the model using five-fold cross-validation training is 0.96.

[0160] 13. Validate the trained ensemble learning model based on stacking described in step 12 using the test set. (1) The ROC curve of the test results is shown in Figure 9. The AUC is 0.96. (2) The precision-recall curve of the test results is shown in Figure 10. AP=0.95. (3) The confusion matrix at the default decision threshold (i.e., 0.5) is shown in Figure 11. The recall and precision are 0.83 and 0.89, respectively. (4) Figure 12 shows the precision and recall functions as decision thresholds. (5) Set the minimum recall to 0.95 (i.e., limit the type II error). The obtained results are shown in Figure 12. The recall and precision are 0.96 and 0.70, respectively.

[0161] The present invention proposes to use a machine learning model (such as a deep neural network) to extract sequence feature vectors of NIPT sequencing data, combine the sequence feature vectors (features including but not limited to read counts, base quality, and mapping quality) with the phenotypic feature vectors of pregnant women (phenotypic features of pregnant women including but not limited to maternal age, gestational week, height, weight, BMI, biochemical test results of prenatal testing, and ultrasound diagnostic results such as NT values) to form a vector combination, and use a classification model such as a stacking-based ensemble learning model to obtain the final predicted aneuploidy. In the present invention, the extraction of the sequence feature vector is not limited to the method used herein, and methods including but not limited to autoencoder networks or variation autoencoder networks can also be used. The model structure proposed by the present invention is a hybrid model, i.e., the model includes two stages. In the first stage, a machine learning model (such as a deep neural network) is used to calculate the sequence feature vector. In the second stage, a classification model (such as a stacking-based ensemble learning model) is used to predict aneuploidy using a combination of the sequence feature vector and the phenotypic feature vector. Other ensemble learning models, such as models based on majority voting, can also be used.

[0162] Compared with other convolutional neural networks, the verified advanced deep neural network model used in the embodiments of the present invention has the following features in terms of network design and architecture: the deep neural network model used in the embodiments of the present invention is a deep neural network model based on a 1D convolutional model, the deep neural network model used in the embodiments of the present invention is a network model based on a residual network, and the SE module of the Squeeze-Excite network is introduced into the deep neural network model used in the embodiments of the present invention. Based on these designs, the neural network model used in the embodiments of the present invention has more layers (see Example 3), which effectively reduces the risk of gradient vanishing and overfitting in the process of training the model and improves stability, thus effectively improving the accuracy of the model prediction results.

[0163] The present invention can be implemented as a computer-readable storage medium on which a computer program is stored, and the steps of implementing the method of the present invention are performed when the computer program is executed by a processor. As an example, the computer program is distributed to several computer devices or processors coupled by a network, so that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single step / operation, or two or more steps / operations, can be performed by a single computer device or processor, or by two or more computer devices or processors. One or more steps / operations can be performed by one or more computer devices or processors, and one or more other steps / operations can be performed by one or more other computer devices or processors. One or more computer devices or processors can perform a single step / operation, or two or more steps / operations.

[0164] Those skilled in the art will understand that the division and order of steps in the method for detecting fetal chromosomal abnormalities of the present invention are merely illustrative and not limiting, and deletion, addition, substitution, modification and alteration may be made by those skilled in the art without departing from the technical idea or technical scope of the present invention and its equivalent technical solutions as described in the appended claims. The technical features of the embodiments of the present invention can be combined in any way, but for the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, if there is no contradiction in the combination of these technical features, it shall be within the scope of this specification.

[0165] Although the present invention has been described with reference to exemplary embodiments, it should be understood that the present invention is not limited to the configurations and methods of the above-described embodiments. Instead, the present invention is intended to encompass various modifications and equivalent configurations. In addition, while the various elements and method steps disclosed in the present invention are shown in various exemplary combinations and configurations, other combinations including more or fewer elements or methods are also within the scope of the present invention. References: [1] Evans, Mark I., Stephanie Andriole, and Shara M. Evans. "Genetics: update on prenatal screening and diagnosis." Obstetrics and Gynecology Clinics 42.2 (2015): 193-208. [2] Norwitz, Errol R., and Brynn Levy. "Noninvasive prenatal testing: the future is now." Reviews in obstetrics and gynecology 6.2 (2013): 48. [3] Norton, Mary E., et al. "Cell-free DNA analysis for noninvasive examination of trisomy." New England Journal of Medicine 372.17 (2015): 1589-1597. [4] Langlois, Sylvie, et al. "Current status in non-invasive prenatal detection of Down syndrome, trisomy 18, and trisomy 13 using cell-free DNA in maternal plasma." Journal of Obstetrics and Gynaecology Canada 35.2 (2013): 177-181. [5] Allyse, Megan, et al. "Non-invasive prenatal testing: a review of international implementation and challenges." International journal of women's health 7 (2015): 113. [6] Chiu, Rossa WK, et al. "Noninvasive prenatal diagnosis of fetal chromosomal aneuploidy by massively parallel genomic sequencing of DNA in maternal plasma." Proceedings of the National Academy of Sciences 105.51 (2008): 20458-20463. [7] Fan, H. Christina, et al. "Noninvasive diagnosis of fetal aneuploidy by shotgun sequencing DNA from maternal blood." Proceedings of the National Academy of Sciences 105.42 (2008): 16266-16271. [8] Lau, Tze Kin, et al. "Noninvasive prenatal diagnosis of common fetal chromosomal aneuploidies by maternal plasma DNA sequencing." The Journal of Maternal-Fetal & Neonatal Medicine 25.8 (2012): 1370-1374. [9] Jiang, Fuman, et al. "Noninvasive Fetal Trisomy (NIFTY) test: an advanced noninvasive prenatal diagnosis methodology for fetal autosomal and sex chromosomal aneuploidies." BMC medical genomics 5.1 (2012): 57.

[10] Yang, Jianfeng, Xiaofan Ding, and Weidong Zhu. "Improving the calling of non-invasive prenatal testing on 13- / 18- / 21-trisomy by support vector machine discrimination." BioRxiv (2017): 216689.

[11] Xu, Hanli, et al. "Informative priors on fetal fraction increase power of the noninvasive prenatal screen." Genetics in Medicine 20.8 (2018): 817-824.

[12] Ehrich, Mathias, et al. "Deep learning-based methods, devices, and systems for prenatal testing", Publication number: WO2019191319A1, Filing Date: 27 March 2019.

[13] Egilsson, Agust, et al. "Methods and systems for calling ploidy status using a neural network". Publication number: WO2020018522A1, Filing date: 16 July 2019.

[14] Petersen, Andrea K., et al. "Positive predictive value estimates for cell-free noninvasive prenatal screening from data of a large referral genetic diagnostic laboratory." American journal of obstetrics and gynecology 217.6 (2017): 691-e1.

[15] Benjamini, Yuval, and Terence P. Speed. "Summarizing and correcting the GC content bias in high-throughput sequencing." Nucleic acids research 40.10 (2012): e72-e72.

[16] Hu, Jie, Li Shen, and Gang Sun. "Squeeze-and-excitation networks." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.

[17] Roy, Abhijit Guha, Nassir Navab, and Christian Wachinger. "Concurrent spatial and channel ‘squeeze & excitation’in fully convolutional networks." International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2018.

[18] Stekhoven, Daniel J., and Peter Buhlmann. "MissForest-non-parametric missing value imputation for mixed-type data." Bioinformatics 28.1 (2012): 112-118.

[19] Tang, J., S. Alelyani, and H. Liu. "Data Classification: Algorithms and Applications." Data Mining and Knowledge Discovery Series, CRC Press (2015): pp. 498-500.

[20] He, Kaiming, et al. "Delving deep into rectifiers: Surpassing human-level performance on imagenet classification." Proceedings of the IEEE international conference on computer vision. 2015.

Claims

1. A method for detecting fetal chromosomal abnormalities, comprising the steps of: (1) obtaining sequencing data and clinical phenotype characteristic data of cell-free nucleic acid fragments from a pregnant woman to be detected, the sequencing data comprises a plurality of read segments; forming a phenotypic feature vector of the subject pregnant woman from said clinical phenotypic feature data; (2) obtaining a plurality of sliding windows by performing window division on at least a portion of a chromosome sequence of a reference genome; Counting the read segments that fall within the sliding window; generating a sequence feature matrix for the chromosomal sequences; (3) inputting the sequence feature matrix into a trained machine learning model to extract a sequence feature vector of the chromosome sequence; (4) combining the sequence feature vector and the phenotypic feature vector of the pregnant woman to form a composite feature vector; A process of obtaining the fetal chromosomal abnormality status of the subject pregnant woman by inputting the composite feature vector into a classification detection model.

2. In the step (1), the cell-free nucleic acid fragments are derived from peripheral blood, liver, and / or placenta of the pregnant woman; The method of claim 1 , wherein the cell-free nucleic acid fragments are cell-free DNA.

3. The method of claim 1 or 2, wherein in step (1), the sequencing data is derived from ultra-low depth sequencing.

4. In the step (1), the read segments are aligned to the reference genome to obtain unique mapping reads; The method of any one of claims 1 to 3, wherein a subsequent step is performed using the unique mapping reads.

5. The lead segment is corrected for GC content, 5. The method of claim 4, wherein the GC content correction is performed as follows: a. randomly selecting m fragments of length l from chromosomes of said human reference genome; b. The number of fragments of GC content i, N ｉ is calculated according to the following formula: [0010] During the ceremony, [0025] where f(k) is the GC content of fragment k, i represents the GC content (i=0%, 1%, ..., 100%); c. The number of sequencing read segments of GC content i, F ｉ is calculated according to the following formula: [0030] During the ceremony, c ｋ II ｉ (f(k)) represents the GC content of fragment k, F ｉ represents the number of sequencing read segments with GC content i and the same start site as the fragment, d. GC content observed-expected ratio λ ｉ is calculated by the following formula: [0045] During the ceremony, r is a global scale factor defined as: [0050] e. The number of sequencing read segments is corrected by the formula: [006] During the ceremony, R ｉ represents the expected number of sequencing read segments with corrected GC content i.

6. The method according to any one of claims 1 to 5, wherein in step (1), the clinical phenotype characteristic data of the pregnant woman is selected from one or a combination of two or more of age, gestational week, height, weight, BMI, biochemical test results of a prenatal test, ultrasound diagnosis results, and cell-free fetal DNA concentration in plasma.

7. In the step (1), The clinical phenotypic characteristic data of the pregnant woman is outlier-treated, missing-value-treated, and / or null-value-treated; The missing values ​​and the null values ​​are padded by a Miss Forest algorithm; Record of: a. x age < 10 or x age > 80; b. x GW < 5 or x GW > 50; c. x height < 40 or x height > 300; d. x weight < 10 or x weight > 200; is displayed, the phenotype data of the pregnant woman's sample is determined to be an abnormal value, The method according to any one of claims 1 to 6, wherein these outlier values ​​are set as null values.

8. The step (2) (2.1) Using a window of length b, overlapping and sliding a chromosome sequence of length L of the reference genome with a step size t to obtain a plurality of sliding windows, b is a positive integer, b=[10000, 10000000], t is any positive integer, L is a positive integer, and L≧b; (2.2) counting the read segments included within each of the sliding windows and generating a sequence feature matrix of the chromosome sequence; The method according to any one of claims 1 to 7, comprising:

9. The method of any one of claims 1 to 8, wherein in step (2), the sequence feature matrix comprises the number of read segments within the sliding window, base quality, and mapping quality.

10. In the step (2), the sequence feature matrix is ​​represented by the following formula: [0070] where h represents the number of sliding windows, w represents the number of sequence features in a single sliding window; x ｉｊ represents the j-th array eigenvalue in the i-th sliding window, The method according to any one of claims 1 to 9.

11. In the step (3), the sequence feature matrix is ​​normalized; The sequence feature matrix is ​​normalized using formula (I): [0080] where Z(k)i,j is the normalized sequence feature matrix of sample k; X(k)i,j denotes the jth array eigenvalue in the ith sliding window of sample k; The method according to any one of claims 1 to 10, wherein μ i,j and σ i,j respectively represent the mean and standard deviation of the jth sequence eigenvalue within the ith sliding window of all samples.

12. The method according to any one of claims 1 to 11, wherein in step (3), the trained machine learning model is a neural network model or an autoencoder model.

13. The neural network model is a deep neural network model, The structure of the deep neural network model is an input layer receiving the sequence feature matrix; A pre-module connected to the input layer, which performs a first convolution and activation operation on the sequence feature matrix from the input layer to obtain a feature map; a core module connected with the pre-modules, which performs further abstraction and feature extraction of the feature maps from the pre-modules, effectively increasing the depth of the neural network model and thereby enhancing the representational capability of the neural network; a post module coupled to the core module that performs feature abstraction of the feature map from the core module; a first global average pooling layer connected to the post module, which vectorizes the feature maps of the feature abstraction representation and outputs the sequence feature vectors of the chromosome sequences; The method of claim 12, comprising:

14. The pre-module is (I) a 1D convolutional layer; (II) a batch normalization layer connected to the 1D convolution layer described in (I); (III) a ReLU activation layer connected to the batch normalization layer described in (II); The method of claim 13, comprising:

15. The core module is composed of one or more residual sub-modules having the same structure; The output of each residual module is the input of the next residual module, The residual sub-module: (A) a pre-sub-module of the core module, each of which includes a 1D convolutional layer, a dropout layer connected to the 1D convolutional layer, a batch normalization layer connected to the dropout layer, and a ReLU activation layer connected to the batch normalization layer; (B) a first 1D average pooling layer connected to a pre-submodule of the core module described in (A); (C) a Squeeze-Excite module and / or a Spatial Squeeze-Excite module connected to the first 1D average pooling layer described in (B); (D) a first additional layer connected to the Squeeze-Excite module and / or the Spatial Squeeze-Excite module according to (C); (E) a second 1D average pooling layer connected to the ReLU activation layer in the pre-module; and (F) a second additional layer connected to the first additional layer described in (D) and the second 1D average pooling layer described in (E); The method of claim 13 or 14, comprising:

16. The Squeeze-Excite module comprises: (a) a second global average pooling layer connected to the first 1D average pooling layer of the residual submodule of (B); (b) a reconstruction layer connected to the second global average pooling layer of (a), the size of the output feature map of the reshaping layer is 1×f; a reshaping layer, where f is the number of 1D convolution kernels; (c) a first fully connected layer connected to the remodeled layer of (b), The number of output neurons in the first fully connected layer is f / r ＳＥ and f is the number of 1D convolution kernels, and r ＳＥ a first fully connected layer, where is the reduction rate of the Squeez-Excite module; (d) a second fully connected layer connected to the first fully connected layer of (c), the number of output neurons of the second fully connected layer is f; a second fully connected layer, where f is the number of 1D convolution kernels; (e) a multiplication layer connected to the second fully connected layer described in (d) and the first 1D average pooling layer of the residual submodule described in (B); 16. The method of claim 15, comprising:

17. The Spatial Squeeze-Excite module comprises: a. a 1×1 1D convolutional layer connected to a first 1D average pooling layer according to (B) using a sigmoid function as an activation function; b. a multiplication layer connected to the first 1D average pooling layer described in (B) and the 1×1 1D convolutional layer described in a; 17. The method of claim 15 or 16, comprising:

18. The method according to any one of claims 1 to 17, wherein in step (4), the composite feature vector is obtained by combining the sequence feature vector and the phenotypic feature vector of the pregnant woman.

19. In the step (4), the composite feature vector x is normalized by the following formula: [0090] In the formula, x ’ i is the i-th sequence eigenvalue of the normalized composite feature vector x, x ｉ is the i-th array eigenvalue of the composite feature vector x, μ ｉ is the average of the i-th array eigenvalue of the composite feature vector x, σ ｉ is the standard deviation of the i-th sequence eigenvalue of the composite feature vector x; 19. The method according to any one of claims 1 to 18.

20. The method according to any one of claims 1 to 19, wherein in the step (4), the classification detection model is an ensemble learning model, and the ensemble learning model is an ensemble learning model based on stacking or majority voting.

21. 2. The method of claim 1, wherein the chromosomal abnormality comprises at least one or more of trisomy 21 syndrome, trisomy 18 syndrome, trisomy 13 syndrome, 5p syndrome, chromosomal microdeletion, and chromosomal microduplication.

22. A method for constructing a classification detection model for detecting fetal chromosomal abnormalities, comprising the steps of: (1) obtaining cell-free nucleic acid fragment sequencing data and clinical phenotypic characteristic data from a plurality of pregnant women, the sequencing data comprises a plurality of read segments, a fetal chromosomal status for each of the pregnant women is known, and the clinical-phenotypic feature data for each of the pregnant women forms a phenotypic feature vector for the pregnant women; (2) for each of the pregnant women, performing window division on at least a portion of the chromosomal sequence of the reference genome to obtain a plurality of sliding windows; Counting the read segments that fall within the sliding window; generating a sequence feature matrix for the chromosomal sequences; (3) for each of the pregnant women, constructing a training dataset using the sequence feature matrix and the fetal chromosomal status, and training a machine learning model to extract sequence feature vectors of the chromosomal sequences; (4) combining the sequence feature vector and the phenotype feature vector of each of the pregnant women to form a composite feature vector, and training a classification model using the composite feature vector of the pregnant women and the fetal chromosomal status to obtain a trained classification detection model.

23. the fetal chromosomal status of each of the pregnant women is one or more of normal diploidy, chromosomal aneuploidy, partial monosomy syndrome, chromosomal microdeletion, and chromosomal microduplication; 23. The method of claim 22.

24. The method according to claim 22 or 23, wherein the number of pregnant women is more than 10 and the ratio of the number of normal diploid fetuses to the number of chromosomal aneuploid fetuses is 1 / 2 to 2.

25. In the step (3), the training data set is expressed as follows: [0089] ##EQU00011## where N represents the number of training samples, and N is an integer equal to or greater than 1; Z （ｋ） ｉ，ｊ is the normalized sequence feature matrix of training sample k, k ∈ [1, N], i is an integer greater than or equal to 1, and j is an integer greater than or equal to 1; 25. The method of any one of claims 22 to 24, wherein the chromosomal abnormality comprises at least one or more of trisomy 21 syndrome, trisomy 18 syndrome, trisomy 13 syndrome, 5p syndrome, chromosomal microdeletion and chromosomal microduplication.

Images