A method for non-invasive screening of preeclampsia based on deep sequencing 8bp oligonucleotide double-fragment signatures

Abstract

The present application relates to the field of noninvasive prenatal diagnosis, and specifically discloses a preeclampsia noninvasive screening method based on deep sequencing of 8bp oligonucleotide double-fragment features, comprising: collecting peripheral blood samples of preeclampsia and healthy pregnant women, extracting free DNA for high-throughput sequencing; extracting core 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' in 150-200bp long fragments through bioinformatics analysis; using integrated support vector machines, K-nearest neighbors, extreme gradient boosting, random forests and multilayer perceptron combined with logistic regression meta-classifier to construct a stacking model, normalizing the frequency of core sequences, performing machine learning analysis, and predicting the risk of preeclampsia. The present application effectively breaks through the limitations of traditional screening methods with low specificity and invasive detection by specifically screening two 8bp oligonucleotide feature fragments and combining a deep learning architecture with multi-model fusion.

Description

Technical Field

[0001] This invention relates to the field of non-invasive prenatal diagnostic technology, specifically to a non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides from deep sequencing. Background Technology

[0002] Preeclampsia (PE) is a complex pregnancy disorder that typically occurs after 20 weeks of gestation, manifesting as hypertension and at least one complication (such as proteinuria, maternal organ dysfunction, or placental dysfunction). Globally, it affects approximately 5% to 9% of pregnant women and results in 10% to 15% maternal mortality. The etiology of PE is complex, with the placenta playing a crucial role in its pathogenesis. In early pregnancy (10-12 weeks), significant changes in gene expression are observed in the placenta of women with PE. After delivery, defects in spiral artery remodeling and insufficient trophoblast infiltration lead to inadequate placental perfusion, resulting in placental hypoxia and ischemia. Although chorionic villus sampling can be used for early placental phenotypic analysis, its invasiveness and risk of miscarriage limit its application in screening. With delayed marriage and childbearing and changes in modern lifestyles, the incidence of PE is increasing annually, becoming a significant complication threatening the health of pregnant women and fetuses. PE causes more than 70,000 maternal deaths and 500,000 fetal deaths annually. Therefore, early screening and prevention of PE are of paramount importance.

[0003] Currently, the diagnosis of preeclampsia (PE) mainly relies on clinical indicators such as urine protein and blood pressure, or invasive procedures such as placental single-cell sequencing and amniocentesis. However, these methods have limitations: urine protein is not a reliable predictor, and invasive procedures carry the risk of miscarriage. In recent years, non-invasive prenatal testing (NIPT) based on cfDNA has gradually become a research hotspot. Existing studies have analyzed PE using cfDNA, focusing on methylation, fragment information, or multi-omics combined analysis, but efficient and specific non-invasive biomarkers are lacking. Therefore, developing an efficient and non-invasive diagnostic method for PE has significant clinical implications. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing. This method obtains cell-free DNA (cfDNA) sequence data from the peripheral blood of pregnant women through high-throughput sequencing technology, and combines bioinformatics analysis and machine learning techniques to achieve efficient and accurate PE diagnosis.

[0005] To achieve the above objectives, this invention proposes a non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides from deep sequencing, comprising:

[0006] S1. Sample collection, extraction and sequencing: Peripheral blood samples were collected from pregnant women with preeclampsia (PE) and healthy pregnant women. cfDNA was extracted and high-throughput whole-genome sequencing was performed at a sequencing depth of 20-30X to obtain raw sequencing data.

[0007] S2. Data alignment and sorting: The original sequencing reads are aligned with the human reference genome hg38 to generate an alignment result file, which is then sorted and deduplicated to obtain hundreds of billions of cfDNAs.

[0008] S3. Split the dataset: Divide the preprocessed data into a training set and a test set in a 7:3 ratio. The training set is used for model training and parameter tuning, while the test set is used for independent validation and performance evaluation of the finally trained model.

[0009] S4. Feature extraction of sequencing data: Extract the frequency of kmer features of 150-200bp fragments from the training set and the test set respectively, and normalize them respectively.

[0010] S5. Feature selection: LASSO regression was used to perform preliminary feature selection in the training set, and TURF plus single-factor incremental feature selection method was used to select two key 8-mer sequences.

[0011] S6. Analyze regulatory functions and detect changes in cfDNA frequency: Perform conservation and transcriptional regulatory function analysis on the core 8-mer sequences GTGCGCCC and GATGGGGT to obtain the specificity of abnormal frequency changes of the core 8-mer sequences in peripheral blood cfDNA, and use it to reflect the placental dysfunction status of preeclampsia.

[0012] S7. Build and tune the model: Build 5 basic models and use the numerical values ​​of the training set to tune the parameters of the 5 basic models, and build a stacked model based on the optimal parameters.

[0013] S8. Model Validation and Performance Evaluation;

[0014] S9. Calculate and output the predicted probability of preeclampsia.

[0015] Preferably, in S4, the specific steps for feature extraction from sequencing data are as follows:

[0016] S41. Extract fragments with a sequence length of 150-200bp from hundreds of billions of sequences;

[0017] S42. Extract the frequencies corresponding to the kmer features from the training set and the test set respectively, where k is 8, to obtain 65536 subsequences;

[0018] S43. Normalize the frequencies of subsequences in the training and test sets respectively.

[0019] Preferably, in S5, the specific steps for screening the two key 8-mer sequences are as follows:

[0020] S51. Use LASSO to perform preliminary feature filtering on the 8-mer subsequences in the training, and obtain 80 features;

[0021] S52. Using TURF combined with single-factor incremental feature screening, two core 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' were selected from 80 features of 8-mer.

[0022] Preferably, in S6, the sequence is analyzed for conservation and transcriptional regulatory function as follows:

[0023] S61. Sequence evolution conservation: Construct position-specific base frequency matrices for the core 8-mer sequences GTGCGCCC and GATGGGGT, and align them with the genomes of multiple vertebrate species. Calculate the evolution conservation score for each base position using the expectation-maximization algorithm.

[0024] S62, Transcription Factor Binding Function: The base frequency matrix generated in S61 is compared with the public database of vertebrate transcription factor binding sites for homology, and the high-homology results are screened using the motif matching algorithm.

[0025] Preferably, in S7, the specific process of building the model and tuning parameters is as follows:

[0026] S71. Label patients with preeclampsia as 1 and healthy pregnant women as 0.

[0027] S72. Extract the normalized frequency values ​​of the two 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' in the 150-200bp long segments of the training and test sets, respectively.

[0028] S73. Based on the core features 'GTGCGCCC' and 'GATGGGGT' of the normalized values ​​of the frequency of the selected 8-mer sequences, and based on the data of preeclampsia samples and healthy pregnant women, we constructed the basic models of support vector machine (SVM), K-nearest neighbor (KNN), extreme gradient boosting tree (XGBoost), random forest (RF) and multilayer perceptron (MLP).

[0029] S74. Perform hyperparameter tuning on the above basic models respectively, and select the optimal parameters to obtain 5 optimal basic models;

[0030] S75. Using the five optimal base models obtained in S74 as base classifiers, the Multilayer Perceptron (MLP) is built on Keras and includes three ReLU activation hidden layers with L2 regularization and Dropout, an output layer that uses the Sigmoid function to handle binary classification tasks, and a classifier that is custom-encapsulated and compatible with the scikit-learn interface. Logistic regression with default parameters is used as the meta-classifier to build a stacked model, resulting in a trained prediction model.

[0031] Preferably, in S8, the specific steps for model validation and performance evaluation are as follows: using a strictly isolated test set to evaluate the generalization performance and diagnostic capability of the trained prediction model obtained in step S75, wherein the indicators of generalization performance and diagnostic capability include computational accuracy, sensitivity, specificity, and AUC.

[0032] Preferably, in S9, the specific steps for calculating and outputting the predicted probability of preeclampsia from the screened new pregnant women samples are as follows:

[0033] S91. Collect peripheral blood samples according to steps S1-S2, extract cfDNA and perform high-throughput sequencing at a depth of 20-30X, and align to hg38.

[0034] S92. Following steps S3-S5, extract a 150-200bp fragment, calculate the frequency of the two core 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' and normalize them to obtain the feature vector of the sample.

[0035] S93. Input the feature vector obtained in S92 into the prediction model trained in S75.

[0036] S94. Call the model's predict_proba method to obtain the predicted probability P that the sample belongs to the preeclampsia category label 1;

[0037] S95. Output the predicted probability P as the predicted risk value of the pregnant woman for preeclampsia.

[0038] Therefore, this invention proposes a non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides from deep sequencing, with the following beneficial effects:

[0039] (1) This invention requires only two key sequences to efficiently distinguish preeclampsia, which significantly simplifies the feature complexity and enables early diagnosis of preeclampsia through non-invasive detection, which is helpful for early intervention and treatment.

[0040] (2) The diagnosis is performed using a machine learning model. The AUC values ​​of both the training and test sets are over 0.95, indicating high accuracy. It only requires the collection of peripheral blood from pregnant women, without any invasive procedures, making it safe, easy to apply in clinical practice, and more practical.

[0041] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0042] Figure 1 This is a flowchart of a non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing, according to the present invention.

[0043] Figure 2 This invention provides an evolutionarily conserved motif map based on the dual-fragment feature of an 8bp oligonucleotide from deep sequencing.

[0044] Figure 3 This invention relates to a transcription factor FOXO1 enriched based on the sequence enrichment of 8bp oligonucleotide dual-fragment features from deep sequencing.

[0045] Figure 4 This invention relates to a transcription factor EGR1 enriched based on the sequence enrichment of 8bp oligonucleotide dual-fragment features from deep sequencing.

[0046] Figure 5 This invention relates to a transcription factor KLF4 enriched based on the sequence enrichment of 8bp oligonucleotide dual-fragment features from deep sequencing.

[0047] Figure 6 This invention relates to a transcription factor GCMI based on the sequence enrichment of 8bp oligonucleotide dual-fragment features from deep sequencing.

[0048] Figure 7 This invention presents the ROC curves of five basic models for a non-invasive preeclampsia screening method based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing on the training and testing sets.

[0049] Figure 8 This is the ROC curve of the stacked model of a non-invasive preeclampsia screening method based on the dual fragment feature of 8bp oligonucleotides in deep sequencing in the training set and the test set of the present invention.

[0050] Figure 9 This is a flowchart illustrating an implementation example of a non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing, according to the present invention.

[0051] Figure 10 This is a diagram showing the independent validation results of real clinical samples for a non-invasive preeclampsia screening method based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing, as described in this invention. Detailed Implementation

[0052] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0053] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0054] like Figures 1-8 As shown, the non-invasive screening method for preeclampsia based on 8bp oligonucleotide dual-fragment features according to the present invention specifically includes the following steps:

[0055] S1. Sample collection, extraction and sequencing: Peripheral blood samples were collected from pregnant women with preeclampsia (PE) and healthy pregnant women. cfDNA was extracted and high-throughput whole-genome sequencing was performed at a sequencing depth of 20-30X to obtain raw sequencing data.

[0056] S2. Data alignment and sorting: The original sequencing reads are aligned with the human reference genome hg38 to generate an alignment result file, which is then sorted and deduplicated to obtain hundreds of billions of cfDNAs.

[0057] S3. Split the dataset: Divide the preprocessed data into a training set and a test set in a 7:3 ratio. The training set is used for model training and parameter tuning, while the test set is used for independent validation and performance evaluation of the finally trained model.

[0058] S4. Feature extraction from sequencing data: Extract the frequency of kmer features for 150-200bp fragments from both the training and test sets, and normalize them accordingly. The specific steps for feature extraction from sequencing data are as follows:

[0059] S41. Extract fragments with a sequence length of 150-200bp from hundreds of billions of sequences;

[0060] S42. Extract the frequencies corresponding to the kmer features from the training set and the test set respectively, where k is 8, to obtain 65536 subsequences;

[0061] S43. Normalize the frequencies of subsequences in the training and test sets respectively.

[0062] S5. Feature Selection: Initial feature selection was performed using LASSO regression on the training set. Two key 8-mer sequences were then selected using TURF combined with single-factor incremental feature selection. The specific steps for selecting the two key 8-mer sequences are as follows:

[0063] S51. Use LASSO to perform preliminary feature filtering on the 8-mer subsequences in the training, and obtain 80 features;

[0064] S52. Using TURF combined with single-factor incremental feature screening, two core 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' were selected from 80 features of 8-mer.

[0065] S6. Analyze regulatory functions and detect changes in cfDNA frequency: Perform conservation and transcriptional regulatory function analysis on the core 8-mer sequences GTGCGCCC and GATGGGGT to obtain the specificity of abnormal frequency changes of the core 8-mer sequences in peripheral blood cfDNA, and use it to reflect the placental dysfunction status in preeclampsia; the steps for sequence conservation and transcriptional regulatory function analysis are as follows:

[0066] S61. Sequence Evolutionary Conservation: Construct position-specific base frequency matrices for the core 8-mer sequences GTGCGCCC and GATGGGGT, and align them with the genomes of multiple vertebrate species. Calculate the evolutionary conservation score for each base position using the expectation-maximization algorithm; for example... Figure 2 As shown, the results indicate strong conservation of thymine (T) and guanine (G) bases at positions 2, 3, and 5, suggesting their key functions in mammalian placental development.

[0067] S62, Transcription Factor Binding Function: The base frequency matrix generated in S61 is compared with the public database of vertebrate transcription factor binding sites for homology, and the high-homology results are screened using the motif matching algorithm.

[0068] like Figures 3-6 As shown, the results indicate that the sequence is significantly enriched at transcription factor binding sites that regulate placental angiogenesis and trophoblast differentiation, including but not limited to: FOXO1 (which regulates placental cell apoptosis and oxidative stress response).

[0069] Among them, transcription factor dysfunction is directly related to the pathological features of placental hypoxia and ischemia in preeclampsia.

[0070] S7. Build and tune the model: Build 5 basic models and use the numerical values ​​of the training set to tune the parameters of the 5 basic models, and build a stacked model based on the optimal parameters.

[0071] The specific process of building the model and tuning its parameters is as follows:

[0072] S71. Label patients with preeclampsia as 1 and healthy pregnant women as 0.

[0073] S72. Extract the normalized frequency values ​​of the two 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' in the 150-200bp long segments of the training and test sets, respectively.

[0074] S73. Based on the core features 'GTGCGCCC' and 'GATGGGGT' of the normalized values ​​of the frequency of the selected 8-mer sequences, and based on the data of preeclampsia samples and healthy pregnant women, we constructed the basic models of support vector machine (SVM), K-nearest neighbor (KNN), extreme gradient boosting tree (XGBoost), random forest (RF) and multilayer perceptron (MLP).

[0075] S74. Perform hyperparameter tuning on the above basic models respectively, and select the optimal parameters to obtain 5 optimal basic models;

[0076] S75. Using the five optimal base models obtained in S74 as base classifiers, the Multilayer Perceptron (MLP) is built on Keras and includes three ReLU activation hidden layers with L2 regularization and Dropout, an output layer that uses the Sigmoid function to handle binary classification tasks, and a classifier that is custom-encapsulated and compatible with the scikit-learn interface. Logistic regression with default parameters is used as the meta-classifier to build a stacked model, resulting in a trained prediction model.

[0077] S8. Model Validation and Performance Evaluation;

[0078] The specific steps for model validation and performance evaluation are as follows: Use a strictly isolated test set to evaluate the generalization performance and diagnostic capability of the trained prediction model obtained in step S75. The metrics for generalization performance and diagnostic capability include computational accuracy, sensitivity, specificity, and AUC.

[0079] S9. Calculate and output the predicted probability of preeclampsia.

[0080] The specific steps for calculating and outputting the predicted probability of preeclampsia from newly screened pregnant women are as follows:

[0081] S91. Collect peripheral blood samples according to steps S1-S2, extract cfDNA and perform high-throughput sequencing at a depth of 20-30X, and align to hg38.

[0082] S92. Following steps S3-S5, extract a 150-200bp fragment, calculate the frequency of the two core 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' and normalize them to obtain the feature vector of the sample.

[0083] S93. Input the feature vector obtained in S92 into the prediction model trained in S75.

[0084] S94. Call the model's predict_proba method to obtain the predicted probability P that the sample belongs to the preeclampsia category label 1;

[0085] S95. Output the predicted probability P as the predicted risk value of the pregnant woman for preeclampsia.

[0086] In S7, the five basic models are as follows:

[0087] Model K-Nearest Neighbors: The model was optimized for hyperparameters using GridSearchCV. Five-fold cross-validation was used to evaluate the performance of different parameter combinations on the training set. Accuracy was used as the evaluation metric. The final parameters were determined to be the number of neighbors n_neighbors = 5 and the weighting method was uniform, meaning that all neighbors have the same voting weight.

[0088] Random Forest Model: The model uses GridSearchCV to fine-tune key hyperparameters and employs 5-fold cross-validation, with accuracy as the evaluation metric. The final parameters are: number of decision trees (n_estimators = 200), maximum tree depth (max_depth = None), minimum number of samples per internal node split (min_samples_split = 2), and minimum number of samples per leaf node (min_samples_leaf = 2).

[0089] Support Vector Machine (SVM) Model: Its hyperparameters were tuned using GridSearchCV. The model employed 5-fold cross-validation, with accuracy as the scoring criterion. The final parameters were set as follows: regularization strength C = 0.1 and radial basis function (RBF) kernel.

[0090] Extreme gradient boosting of the model: Grid SearchCV was used to fine-tune the key hyperparameters. Through 5-fold cross-validation, the parameters were finally determined to be the learning rate (learning_rate = 0.1), the number of trees (n_estimators = 500), and the maximum tree depth (max_depth = 6).

[0091] The model uses a multilayer perceptron: Built with Keras and encapsulated using KerasClassifier for integration with GridSearchCV, enabling hyperparameter tuning. Grid search employs 5-fold cross-validation, with accuracy as the scoring metric. The final parameters are determined as follows: number of hidden layer neurons (units = 8), dropout rate (dropout_rate = 0.3), L2 regularization strength (l2_reg = 0.01), optimizer (Adam), number of training epochs (256), and batch size (batch_size = 64).

[0092] The multilayer perceptron architecture consists of three hidden layers with ReLU activation functions. Each layer applies L2 regularization and Dropout to prevent overfitting. The output layer is a single neuron, using a sigmoid activation function to handle binary classification. The model uses binary cross-entropy as the loss function for optimization.

[0093] The stacked model is constructed by integrating five high-performance base classifiers, including Support Vector Machine (SVM), K Nearest Neighbors (KNN), Random Forest (RF), Extreme Gradient Boosting Tree (XGBoost), and Multilayer Perceptron (MLP).

[0094] The MLP is built on Keras and consists of three ReLU-activated hidden layers with L2 regularization and Dropout. The output layer uses the Sigmoid function to handle binary classification tasks and is custom-encapsulated into a classifier compatible with the scikit-learn interface. Each base classifier is initialized based on the optimal hyperparameters obtained from prior parameter tuning, and the stacked framework uses Logistic Regression as the meta-learner.

[0095] The specific process of constructing the stacked model is as follows:

[0096] (1) Basic classifier and hyperparameter tuning

[0097] KNN: GridSearchCV was used to perform a 5-fold CV on n_neighbors and weights, and finally n_neighbors = 16 and weights = 'uniform' were selected.

[0098] RF: Five-fold CV was performed on n_estimator, max_depth, min_samples_split, and min_samples_leaf, and n_estimators=4, max_depth=5, min_samples_split=2, and min_samples_leaf=2 were finally selected.

[0099] SVM: Perform grid search on regularization parameter C and kernel parameter γ, and select C=1 and kernel='rbf' after 5-fold CV;

[0100] XGBoost: Performed 5-fold CV on learning_rate, n_estimators, and max_depth, and finally determined learning_rate = 1, n_estimators = 10, and max_depth = 2;

[0101] MLP: A three-layer hidden layer (ReLU + L2 regularization + Dropout) is built based on Keras. KerasClassifier is used in conjunction with GridSearchCV. 5-fold CV is scored by accuracy. The parameter space includes units, dropout_rate, l2_reg, epochs, and batch_size. The final settings are units=16, dropout_rate=0.1, l2_reg=0.01, epochs=32, and batch_size=16.

[0102] (2) Cross-validation generates first-level features

[0103] Five-fold cross-validation was used on the training set: each time, a base model was trained with four folds, and the remaining one fold was used to predict the probability that the sample belonged to preeclampsia. After repeating five times, the predicted probabilities of all folds were collected to obtain a first-level probability feature vector that corresponds one-to-one with the training sample. There are a total of 5 dimensions, and each dimension corresponds to a base model.

[0104] (3) Stacking structure construction

[0105] Using the above-mentioned first-level probability feature matrix (number of samples × 5) and the true labels as input, a meta-classifier is trained using logistic regression (C = 1.0) to form the final StackingClassifier.

[0106] After training, the stacked model can weighted and fused the predicted probabilities of the five base models to output more accurate classification results.

[0107] (4) Final prediction process

[0108] For test or validation set samples: extract two types of 8-mer normalized frequency features; input the features into the pre-trained StackingClassifier at once; call stacking_classifier.predict_proba(X) to directly obtain the predicted probability P of label 1 (preeclampsia); use P as the "preeclampsia risk prediction probability" of the sample for clinical high-risk screening and early intervention.

[0109] After constructing the stacked model, the stacked framework uses Logistic Regression as a meta-learner to learn and synthesize the predicted probabilities output by each base model. The entire model training uses the training set for fitting and the test set for evaluating the generalization performance. The training set is used for model construction, and the test set is used for model evaluation.

[0110] In S3, dividing the samples into a training set and a test set in a 7:3 ratio is a key step in achieving model reliability and generalization ability.

[0111] 1) Training set (used in model building): used for parameter learning and hyperparameter tuning and 5-fold cross-validation of five basic classifiers (SVM, KNN, RF, XGBoost, MLP) and stacked models. By learning the normalized frequency features of two key 8-mers ('GTGCGCCC', 'GATGGGGT') in 150–200bp fragments, the model masters the mapping relationship of "feature → preeclampsia risk".

[0112] 2) Test set (completely independent of the training process): Used to objectively evaluate the predictive performance of the model on unseen data after training. On 570 test samples (246 cases of preeclampsia and 324 cases of normal pregnancy), the stacked model showed excellent performance in terms of AUC=0.983 (95% CI 0.902–0.983), accuracy=0.96, and precision=0.96, demonstrating that the model does not merely "memorize" the training data but has the ability to generalize to real clinical samples.

[0113] Therefore, sample partitioning is not only a common best practice in the field of machine learning, but also ensures that the predictive model of this invention is both effective and reliable in real clinical applications: the training set is responsible for "learning" the classification rules, and the test set is responsible for "testing" whether the rules can be generalized and applied.

[0114] like Figure 9 As shown, an embodiment of the present invention provides a non-invasive screening method for preeclampsia based on the dual-fragment characteristics of 8bp oligonucleotides from deep sequencing. This method utilizes a preeclampsia probability assessment system that predicts the probability of the candidate having preeclampsia based on the frequencies of two 8-mer sequences in the candidate, employs a stacking model, and consists of seven steps:

[0115] Step 1: Obtain peripheral blood samples from pregnant women, extract cfDNA and perform quantification, library construction and sequencing;

[0116] Step 2: Preprocessing the sequencing data by comparing, sorting, and deduplicating it with the human reference genome hg38;

[0117] Step 3: Screen for sequences in cfDNA with a fragment length of 150-200bp;

[0118] Step 4: Calculate the frequencies of 'GTGCGCCC' and 'GATGGGGT' in these sequences;

[0119] Step 5: Normalize the frequency;

[0120] Step 6: Use the normalized frequency values ​​of the two sequences as features and input them into the stacking model;

[0121] Step 7: Calculate the probability of preeclampsia in the individuals to be screened using the model.

[0122] As shown in Table 1, the diagnostic performance of six models—Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), Extreme Gradient Boosting Tree (XGBoost), Multilayer Perceptron (MLP), and the generated Stacking model—was validated on the training and test sets. The results show that the area under the curve (AUC) of each model in the test set exceeds 0.95, indicating that they can accurately identify patients and healthy individuals.

[0123] Table 1 Evaluation metrics for the six models on the training / test sets.

[0124]

[0125] like Figure 10 As shown, to verify the feasibility and practicality of the method of this invention, we conducted independent validation based on real clinical samples, mainly including the following two datasets:

[0126] Test set (570 cases): During the model development phase, a separate test set (30% of the total data, including 246 cases of preeclampsia and 324 cases of normal pregnancy) was reserved and not used for model training. This test set was used to objectively evaluate the model's predictive performance in distinguishing between preeclampsia and normal pregnancy. Results showed that all five base models and the final stacked model exhibited good generalization ability. The stacked model achieved an AUC of 0.983 (95% CI: 0.902–0.983) on the test set (see...). Figure 3 The accuracy rate was 0.96, the precision rate was 0.96, and other evaluation indicators also performed well (see Table 1), demonstrating a significant ability to identify preeclampsia samples.

[0127] Validation set (55 cases): In addition, we collected 55 real-world clinical samples (28 cases of preeclampsia and 27 cases of normal pregnancy) to further evaluate the applicability of the model in clinical scenarios. The results showed that the predicted probability of preeclampsia was concentrated around 0.9 for individuals with preeclampsia and around 0.05 for individuals with normal pregnancy, indicating that the model could effectively distinguish between the two groups.

[0128] In summary, the evaluation results of two independent sample sets preliminarily verify that the modeling method proposed in this project has good feasibility and application prospects in cfDNA data.

[0129] Therefore, this invention provides a non-invasive screening method for preeclampsia based on the dual-fragment features of 8bp oligonucleotides in deep sequencing. By analyzing specific 8bp oligonucleotide dual-fragment features in cell-free DNA from the peripheral blood of pregnant women, it achieves early diagnosis with high specificity and accuracy. This method only requires peripheral blood collection, involves no invasive procedures, and is highly safe. Furthermore, the diagnostic accuracy is significantly improved through machine learning models, with an area under the curve exceeding 0.95 on the test set. In addition, this method simplifies feature complexity, requiring only two key sequences for efficient differentiation, providing strong support for early intervention and treatment. Simultaneously, its strong practicality lays the foundation for the fabrication of targeted detection chips and has broad clinical application prospects.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides from deep sequencing, characterized in that, include: S1. Sample collection, extraction and sequencing: Peripheral blood samples were collected from pregnant women with preeclampsia (PE) and healthy pregnant women. cfDNA was extracted and high-throughput whole-genome sequencing was performed at a sequencing depth of 20-30X to obtain raw sequencing data. S2. Data alignment and sorting: The original sequencing reads are aligned with the human reference genome hg38 to generate an alignment result file, which is then sorted and deduplicated to obtain hundreds of billions of cfDNAs. S3. Split the dataset: Divide the preprocessed data into a training set and a test set in a 7:3 ratio. The training set is used for model training and parameter tuning, while the test set is used for independent validation and performance evaluation of the finally trained model. S4. Feature extraction of sequencing data: Extract the frequency of kmer features of 150-200bp fragments from the training set and the test set respectively, and normalize them respectively. S5. Feature selection: LASSO regression was used to perform preliminary feature selection in the training set, and TURF plus single-factor incremental feature selection method was used to select two key 8-mer sequences. S6. Analyze regulatory functions and detect changes in cfDNA frequency: Perform conservation and transcriptional regulatory function analysis on the core 8-mer sequences GTGCGCCC and GATGGGGT to obtain the specificity of abnormal frequency changes of core 8-mer sequences in peripheral blood cfDNA, and use it to reflect the placental dysfunction status of preeclampsia. S7. Build and tune the model: Build 5 basic models and use the numerical values ​​of the training set to tune the parameters of the 5 basic models, and build a stacked model based on the optimal parameters. S8. Model Validation and Performance Evaluation; S9. Calculate and output the predicted probability of preeclampsia.

2. The non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing according to claim 1, characterized in that, In S4, the specific steps for feature extraction from sequencing data are as follows: S41. Extract fragments with a sequence length of 150-200bp from hundreds of billions of sequences; S42. Extract the frequencies corresponding to the kmer features from the training set and the test set respectively, where k is 8, to obtain 65536 subsequences; S43. Normalize the frequencies of subsequences in the training and test sets respectively.

3. The non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing according to claim 1, characterized in that, In S5, the specific steps for screening the two key 8-mer sequences are as follows: S51. Use LASSO to perform preliminary feature filtering on the 8-mer subsequences in the training, and obtain 80 features; S52. Using TURF combined with single-factor incremental feature screening, two core 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' were selected from 80 features of 8-mer.

4. The non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing according to claim 1, characterized in that, In S6, the steps for analyzing the conservation and transcriptional regulatory function of the sequence are as follows: S61. Sequence evolution conservation: Construct position-specific base frequency matrices for the core 8-mer sequences GTGCGCCC and GATGGGGT, and align them with the genomes of multiple vertebrate species. Calculate the evolution conservation score for each base position using the expectation-maximization algorithm. S62, Transcription Factor Binding Function: The base frequency matrix generated in S61 is compared with the public database of vertebrate transcription factor binding sites for homology, and the high-homology results are screened using the motif matching algorithm.

5. The non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing according to claim 1, characterized in that, In S7, the specific process of building the model and tuning parameters is as follows: S71. Label patients with preeclampsia as 1 and healthy pregnant women as 0. S72. Extract the normalized frequency values ​​of the two 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' in the 150-200bp long segments of the training and test sets, respectively. S73. Based on the core features 'GTGCGCCC' and 'GATGGGGT' of the normalized values ​​of the frequency of the selected 8-mer sequences, and based on the data of preeclampsia samples and healthy pregnant women, we constructed the basic models of support vector machine (SVM), K-nearest neighbor (KNN), extreme gradient boosting tree (XGBoost), random forest (RF) and multilayer perceptron (MLP). S74. Perform hyperparameter tuning on the above basic models respectively, and select the optimal parameters to obtain 5 optimal basic models; S75. Using the five optimal base models obtained in S74 as base classifiers, the Multilayer Perceptron (MLP) is built on Keras and includes three ReLU activation hidden layers with L2 regularization and Dropout, an output layer that uses the Sigmoid function to handle binary classification tasks, and a classifier that is custom-encapsulated and compatible with the scikit-learn interface. Logistic regression with default parameters is used as the meta-classifier to build a stacked model, resulting in a trained prediction model.

6. The non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing according to claim 5, characterized in that, In S8, the specific steps for model validation and performance evaluation are as follows: the generalization performance and diagnostic capability of the trained prediction model obtained in step S75 are evaluated using a strictly isolated test set. The metrics for generalization performance and diagnostic capability include computational accuracy, sensitivity, specificity, and AUC.

7. A non-invasive screening method for preeclampsia based on the dual-fragment feature of 8bp oligonucleotides in deep sequencing as described in any one of claims 1-6, characterized in that, In S9, the specific steps for calculating and outputting the predicted probability of preeclampsia from newly screened pregnant women are as follows: S91. Collect peripheral blood samples according to steps S1-S2, extract cfDNA and perform high-throughput sequencing at a depth of 20-30X, and align to hg38. S92. Following steps S3-S5, extract a 150-200bp fragment, calculate the frequency of the two core 8-mer sequences 'GTGCGCCC' and 'GATGGGGT' and normalize them to obtain the feature vector of the sample. S93. Input the feature vector obtained in S92 into the prediction model trained in S75. S94. Call the model's predict_proba method to obtain the predicted probability P that the sample belongs to the preeclampsia category label 1; S95. Output the predicted probability P as the predicted risk value of the pregnant woman for preeclampsia.

Images