Semi-supervised deep learning method, system and device for predicting functional regulation variation, and storage medium

Abstract

The invention provides a semi-supervised learning method, system and device for predicting functional regulation variation and a storage medium. Comprising the following steps: mutating an intersection in a processed GWAS data set and an InuNexUT data set, and dividing mutated data into a training set, a verification set and a test set; a test set for experimental verification is constructed on the basis of HGMD and ClinVar data sets, a training set is used for pre-training a convolutional neural network model, and then a semi-supervised convolutional neural network model is constructed on the basis of variation which does not intersect with GWAS and InuNexUT. Finally, the processed data set is applied to the sscNOVA model, the availability of the sscNOVA model is proved, the method is the first prediction method capable of recognizing the functional regulation and control variation related to the autoimmune disease based on the sequence, and the ideal prediction effect is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and in particular relates to a semi-supervised deep learning method, system, and device for predicting functional regulatory variations. Background Art

[0002] Autoimmune disease (AD) is a type of disease in which the immune system mistakenly attacks the body's own tissues and organs, causing symptoms such as myocarditis, rash, and joint pain. There are currently about 80 known autoimmune diseases, such as asthma, type 1 diabetes, and systemic lupus erythematosus. Autoimmune diseases reflect the complex interaction between genetic variation and the environment and are an important system for the study of human disease genetics. Familial clustering of different autoimmune diseases suggests that genetic factors are the basis of common disease pathways and may affect the function of the immune system, thereby increasing the risk of certain autoimmune diseases.

[0003] In recent years, genome-wide association studies (GWAS) have been used as a standard method to identify susceptibility genes and genetic risk factors for autoimmune diseases, and hundreds of risk sites associated with autoimmune diseases have been identified. A close association between numerous single nucleotide polymorphisms (SNPs) and susceptibility to autoimmune diseases was found in GWAS data. Although the association between variants and diseases can be determined, only a few regulatory variants have been verified, and it is still difficult to identify functional regulatory variants in autoimmune diseases. Researchers need to use large-scale genome sequencing data, histone modification information, transcription factor binding sites and other data to integrate and analyze these data, and use gene editing technology, functional verification experiments and other means to reveal the impact of functional variants.

[0004] At present, there are some studies and methods on regulatory variations related to autoimmune diseases, including gene expression analysis, construction of protein interaction networks, and genetics and expression genetics research. However, the existing technology has certain defects in the following aspects:

[0005] (1) Lack of sufficient attention to autoimmune diseases. Most existing methods pay less attention to the specific regulatory mechanisms related to autoimmunity in the prediction of functional variants. This results in the application of these methods being limited in the study of autoimmune diseases and failing to provide sufficient predictions and understanding.

[0006] (2) Limitations on the types of autoimmune diseases and the number of functional variants. Although some methods involve autoimmune diseases, the types of autoimmune diseases they focus on are relatively limited. This limitation makes it difficult for these methods to be applied to a wide range of autoimmune diseases, limiting their applicability in practical applications. In addition, when predicting functional variants associated with autoimmune diseases, the number of regulatory variants used is relatively small. This leads to an insufficient understanding of the complexity of the immune system and limits a comprehensive understanding of the complex regulatory mechanisms of the immune system.

[0007] The above defects are mainly due to the insufficient understanding of the complexity of the immune system by existing technologies, the limited data sets and the neglect of key regulatory factors in the prediction methods. The significance of solving the above problems and defects is that with the development of deep learning technology, valuable information can be extracted from complex genomic data to understand the functional regulatory variations associated with autoimmune diseases. Using deep learning methods to study the impact of functional variations on autoimmune diseases can provide new ideas and directions for the study of autoimmune diseases, further understand the biological functions and functional characteristics of functional regions, and provide important references for further in-depth research on the pathogenesis of autoimmune diseases. Summary of the invention

[0008] The purpose of the present invention is to solve the defects of the above-mentioned prior art and provide a semi-supervised deep learning method, system and device for predicting functional regulatory variation.

[0009] The present invention is committed to solving the defects of the above-mentioned prior art, and solves the current problem of limited data sets for autoimmune diseases by integrating GWAS data and expression quantitative trait loci (eQTL) data for data amplification. Based on large-scale data processing technology, with the help of the superiority of semi-supervised learning methods, a new prediction method based on deep learning is proposed for the first time. Using this method, predictions based on variant sequences will no longer be limited by the type of autoimmune disease and the number of variants. This method can more accurately and comprehensively predict regulatory variations related to autoimmune diseases. Through this method, the present invention can better understand the regulatory mechanism of the immune system and provide a more reliable scientific basis for the treatment and prevention of autoimmune diseases in the future. The present invention adopts the following technical solutions:

[0010] Semi-supervised deep learning methods for predicting functional regulatory variants, including:

[0011] Step 1. Construct an autoimmune disease-related variation dataset based on GWAS Catalog and ImmuNexUT.

[0012] The intersection of the processed GWAS Catalog and ImmuNexUT variant data numbered E-GEAD-398 was taken as the positive variants of the training set and independent test set.

[0013] The corresponding negative mutations all come from the mutation data of ImmuNexUT numbered E-GEAD-420.

[0014] Step 2. Annotate and filter features related to autoimmune diseases based on autoimmune diseases and immune cell types.

[0015] The Feature Importance of Random Forest was used to screen out 141 features, the mutual_info_classif method of SelectKBest was used to obtain 150 features, and the Sei model was used to screen out 40 features.

[0016] The specific steps include:

[0017] 201. First, the features related to immune cells were selected from the features of Sei model. Secondly, 1000, 800, 600, 400, and 200 features were selected from the features related to immune cells using the mutual_ingo_classif and f_classif methods of SelectKBest, respectively. The features were visualized using TSNE to compare the classification effects.

[0018] 202. Continue to use the mutual_info_classif method to filter out 150, 100, and 50 features from 3102 immune cell-related features.

[0019] 203. Based on the Feature Importance of Random Forest method, 141 features were selected, and the TSNE (T-Distributed Stochastic Neighbor Embedding) method was used to visualize the features and 40 sequence class features of Sei model to compare the classification effects.

[0020] Step 3. Pre-train the convolutional neural network model based on the intersection data of GWAS and ImmuNexUT.

[0021] Randomly sample the mutated data in step 1 and divide the data into training set and test set at a ratio of 80% and 20%. The training set is input into the CNN model for pre-training using five-fold cross validation, and the test set is used to evaluate the model performance. The model with the largest AUC value is used as the corresponding pre-trained convolutional neural network model.

[0022] Step 3 specifically includes: 141 features obtained by Feature Importance based on the Random Forest method, 150 features obtained by the mutual_info_classif method of SelectKBest, and 40 sequence class features provided by the Sei model, using a convolutional neural network (CNN) model for five-fold cross validation, and evaluating the prediction effect of the model on the test set.

[0023] Step 4. Build a semi-supervised sscNOVA model based on GWAS data that has no intersection with ImmuNexUT data.

[0024] Use GWAS data that do not intersect with ImmuNexUT to construct a pseudo-labeled dataset, and pseudo-label the unlabeled GWAS data based on the threshold and t-test method. Merge the GWAS data with the training set in step 3, and after five-fold cross-validation training, the model with the highest AUC is called sscNOVA.

[0025] The pseudo-label dataset construction method includes:

[0026] First, the convolutional neural network model pre-trained in step 3 is used to make probability predictions for GWAS data that do not intersect with ImmuNexUT, and it is subjected to five-fold cross validation. For each variant, five probability values ​​are generated as predictions. Then a Student's t-test is performed to determine whether the difference between the five probability values ​​of each variant is statistically significant, that is, the P value is less than 0.05. If the P value of the variant is less than 0.05, the variant is retained, otherwise it is discarded. Find the optimal pseudo-label threshold for the variant and perform parameter search. Then use 0.5 as the reference threshold to create five sets of thresholds for all unlabeled variants, namely ±0, ±0.1, ±0.2, ±0.3, and ±0.4. Repeat the above process and apply it to all unlabeled variants. Next, merge the pseudo-labeled variants and the original training dataset, retrain the model, and compare the model performance.

[0027] Step 6. Evaluate the performance of the sscNOVA model based on the independent test set and the experimental validation test set.

[0028] The sscnNOVA model approach was most appropriate using 141 features when predicting functional variants in autoimmune diseases.

[0029] On the other hand, the present invention also proposes a semi-supervised learning system for predicting functional regulatory variation, comprising:

[0030] The data prediction module is used to accept the uploaded variant data, use the sscNOVA model in the above method to predict the uploaded variant data, and generate corresponding prediction results. The prediction results include the variant score, and the predicted file can be downloaded.

[0031] The Prediction Results Viewing module enables users to view the predicted variant data and displays the prediction results of each variant, including the score, association with autoimmune diseases, and information on the immune-related cells involved.

[0032] The dataset download module allows users to download datasets related to autoimmune diseases (including training sets, test sets, and experimentally verified test sets) to facilitate subsequent research and application.

[0033] On the other hand, a computer program product is also provided, comprising computer program instructions. When the computer program instructions are read and run by a processor, the sscNOVA model is executed to predict the uploaded variation data, and a corresponding prediction result is generated. The prediction result includes a variation score, and the predicted file can be downloaded.

[0034] An electronic device comprising a processor, a memory and a bus;

[0035] The processor and the memory communicate with each other via a bus;

[0036] The memory stores computer program instructions that can be executed by the processor. The processor calls the computer program instructions to execute the sscNOVA model to predict the uploaded variation data and generate corresponding prediction results, which include variation scores and can download the predicted files.

[0037] A computer-readable storage medium is used to store computer program instructions. When the computer program instructions are executed by a computer, the computer is used to execute an sscNOVA model to predict uploaded variation data and generate corresponding prediction results, the prediction results including variation scores, and the predicted files can be downloaded.

[0038] Beneficial effects of the present invention: The present invention processes the intersection variation in the GWAS and ImmuNexUT data sets, and divides the variation data into a training set, a validation set, and a test set. An experimentally validated test set is constructed based on the HGMD and ClinVar data sets. A convolutional neural network model is pre-trained using the training set, and then a semi-supervised convolutional neural network model (sscNOVA) is constructed based on variations that do not intersect with GWAS and ImmuNexUT. Finally, the processed data set of the present invention is applied to the sscNOVA model, which proves the usability of the sscNOVA model, and can identify functional regulatory variations related to autoimmune diseases, achieving an ideal prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic diagram of a method for predicting functional regulatory variation provided by an embodiment of the present invention.

[0040] Figure 2 A flow chart for predicting functional regulatory variations provided in an embodiment of the present invention.

[0041] Figure 3 A flowchart for constructing positive and negative variants in a training set, an independent test set, and an experimentally verified test set provided in an embodiment of the present invention, wherein A is a positive variant construction process for a training set and an independent test set, B is a negative variant construction process for a training set and an independent test set, C is a positive variant construction process for an experimentally verified test set, and D is a variant construction process for an experimentally verified test set.

[0042] Figure 4 A flowchart of variant feature annotation and screening provided in an embodiment of the present invention.

[0043] Figure 5 A result diagram of visualizing the screened variation features using TSNE provided in an embodiment of the present invention.

[0044] Figure 6 A result diagram of visualizing three feature screening methods using TSNE provided in an embodiment of the present invention.

[0045] Figure 7 A result diagram of using the Random Forest model to verify three feature screening methods provided in an embodiment of the present invention.

[0046] Figure 8 A flow chart for pseudo-labeling unlabeled variants provided by the present invention.

[0047] Fig. 9 A performance comparison chart of the semi-supervised model and the pre-trained model provided in an embodiment of the present invention on an independent test set.

[0048] Fig.10 A performance comparison chart of the sscNOVA provided in an embodiment of the present invention and other models or tools on an experimentally verified test set. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] Example

[0051] like Figure 1-Figure 2 As shown, the semi-supervised deep learning method of the present invention is used to predict functional regulatory variations.

[0052] Step 1. Construct an autoimmune disease-related variation dataset based on GWAS Catalog and ImmuNexUT.

[0053] The processed GWAS Catalog and ImmuNexUT variant data with the number E-GEAD-398 were intersected as positive variants in the training set and independent test set, such as Figure 3 As shown in A. A total of 3362 positive variants are included. The corresponding negative variants are all from the variant data of ImmuNexUT numbered E-GEAD-420. The variant data with a P value greater than 0.1 and an allele frequency (AF) greater than 0.3 are selected, as shown in Figure 3 As shown in B, a total of 3670 negative data are included.

[0054] Step 2. Annotate and filter features related to autoimmune diseases based on autoimmune diseases and immune cell types.

[0055] 21907 features from the Sei model were used to improve the prediction accuracy and generalization ability of the model while reducing the complexity and training time of the model, such as Figure 4 As shown, the present invention performs feature screening. First, 3102 features related to immune cells are screened out from the 21907 features of the Sei model. Secondly, the mutual_info_classif and f_classif methods of SelectKBest are used to screen out 1000, 800, 600, 400, and 200 features from the 3102 immune cell-related features, and the above features are visualized using TSNE. Figure 5From the visualization results, it can be seen that the classification effect of the 3102 immune cell-related features screened by the present invention is better than that of 21907 features, and the classification effect of the features screened by the mutual_info_classif method is better than that of the features screened by the f_classif method. Then, the present invention continues to use the mutual_info_classif method to screen out 150, 100, and 50 features from the 3102 immune cell-related features. In addition, the present invention screens out 141 features based on the Feature Importance of the Random Forest method, and uses the TSNE (T-Distributed Stochastic Neighbor Embedding) method to visualize these features and the 40 sequence class features provided by the Sei model. According to Figure 6 The visualization results show that the 141 features screened by the RandomForest method and the 200, 150, 100, and 50 features screened by the mutual_ingo_classif method have better classification effects than the 40 sequence features of the Sei model. In order to verify whether the conclusions of the feature screening methods are correct, the present invention uses the features obtained by the above different methods to perform a five-fold cross validation on the Random Forest model to evaluate the model performance. Figure 7 The experimental results shown in the figure show that the features screened by the Random Forest method obtained the highest AUC (Area Under The Receive Operating Characteristic) and AUPRC (Area Under The Precision-Recall Curve), the features screened by the mutual_info_classif method obtained the second highest AUC and AUPRC, and the 40 sequence class features of the Sei model obtained the lowest AUC and AUPRC.

[0056] Step 3. Pre-train the convolutional neural network model based on the intersection data of GWAS and ImmuNexUT.

[0057] Before model training, the variant data were randomly sampled and divided into training and test sets at a ratio of 80% and 20%. After the data set was divided, the features mentioned above were used, namely, 141 features obtained by Feature Importance based on the Random Forest method, 150 features obtained by the mutual_info_classif method of SelectKBest, and 40 sequence class features provided by the Sei model. Then, a convolutional neural network (CNN) model was used for five-fold cross validation, and the prediction effect of the model was evaluated on the test set. The model with the largest AUC value of the five-fold cross validation was used as the corresponding pre-training model.

[0058] Step 4. Build a semi-supervised sscNOVA model based on GWAS data that has no intersection with ImmuNexUT data.

[0059] Because when constructing the positive data set, the GWAS data set contains 31 types of variation data related to autoimmune diseases, while the ImmuNexUT data set only contains 10 types of variation data related to autoimmune diseases, using the variation data that coexists in the two data sets as the positive data set is not enough to fully represent the functional variation data in autoimmune diseases. At the same time, in order to improve the accuracy of the model in predicting functional variations in autoimmune diseases, expand the data set when there is less standard data on functional variations related to autoimmune diseases, and systematically study functional variations related to autoimmune diseases, the present invention uses GWAS data that have no intersection with ImmuNexUT to construct a pseudo-labeled data set, and pseudo-labels the unlabeled GWAS data based on the threshold and t-test method to achieve semi-supervision. The construction of the pseudo-labeled data set, such as Figure 8As shown. First, the present invention uses a pre-trained convolutional neural network model to predict the probability of GWAS data that does not intersect with ImmuNexUT, and performs a five-fold cross validation. For each variant, five probability values ​​are generated as predictions. Then, a Student's t test is performed to determine whether the difference between the five probability values ​​of each variant is statistically significant, that is, the Pvalue is less than 0.05. If the P value of this variant is less than 0.05, the variant is retained, otherwise it is discarded. In order to find the best pseudo-label threshold for this variant, a parameter search is performed, and then, with 0.5 as the reference threshold, five sets of thresholds are created for all unlabeled variants, namely ±0, ±0.1, ±0.2, ±0.3 and ±0.4. Following the same process, the above steps are applied to all unlabeled variants. Next, the variants with pseudo labels are merged with the original training data set, the model is retrained, and the performance of the model is compared. In this way, the present invention determines the optimal threshold for applying pseudo labels, wherein the pre-trained convolutional neural network model predicts that the probability greater than 0.9 is marked as a positive variant, and the probability less than 0.1 is marked as a negative variant. Finally, the present invention screened out 2759 positive variants and 626 negative variants from 6924 variant data, and discarded 3539 variants that did not meet the conditions.

[0060] The problem of insufficient standard data on regulatory variations related to autoimmune diseases was solved through a semi-supervised learning method, thereby improving the accuracy of the model in predicting functional regulatory variations in autoimmune diseases.

[0061] Step 5. Evaluate the performance of the sscNOVA model based on the independent test set and the experimental validation test set.

[0062] like Fig. 9 As shown, when 141 features were used, the sscNOVA model performed best, with AUC and AUPRC of 0.892 and 0.896, respectively. When 150 features were used, the AUC and AUPRC of the model were 0.883 and 0.881, respectively. In addition, when the model used 40 features, its AUC and AUPRC were 0.838 and 0.834, respectively, which was the worst performance. The experimental results show that the method based on the sscNOVA model proposed in the present invention is most suitable for predicting functional variations in autoimmune diseases using 141 features, which is better than the 150 features screened by the mutual_info_classif method of SelectKBest and the 40 sequence class features provided by the Sei model.

[0063] In order to compare the capabilities of the sscNOVA model with other models, the present invention constructed three comparative models based on support vector machine (SVM), random forest (RF) and Transfoemer algorithm (TF). Using the above three types of features, the present invention performed a five-fold cross validation on the training data set and evaluated the prediction performance of the CNN model and the three models on an independent test data set. Fig.10 As shown in Figure A, the rf_141 model obtained slightly higher AUC and AUPRC values, followed by the cnn_141 model. Subsequently, the present invention used a dataset containing pseudo-labeled data and trained four models using the same method. Fig.10 As shown in Figure B, although the AUC and AUPRC of the sscNOVA model on this dataset are slightly lower than those of rf_pseudo_141, the sscNOVA model still has the best Recall. This shows that the sscNOVA model performs better in identifying positive variants and can more accurately capture the features associated with positive variants, thereby reducing the risk of misjudgment and helping to ensure that actual positive variants can be effectively identified.

[0064] In order to further verify the model performance, the present invention uses an experimentally verified test set, and these variant data come from HGMD and ClinVar, such as Figure 3 As shown in C, a total of 118 positive variant data conforming to the VCF (Variant Call Format) format are obtained. Then, the present invention screens variant data within 1 kbp upstream and downstream of the chromosome position where the 118 positive variant data are located, and calculates the conservative values ​​of these variants, and only retains variants with conservative values ​​less than 0.5. At the same time, in order to be consistent with the construction of the negative data set, when screening negative variant data, only variant data with AF greater than 0.3 are screened, as shown in FIG. Figure 3 As shown in D, a total of 190 negative variation data were obtained, such as Fig.10 As shown in Figure C, the sscNOVA model shows excellent performance in AUC and AUPRC indicators, ranking first (AUC = 0.658, AUPRC = 0.580), which is significantly improved compared with the rf_141 model. These results show that the sscNOVA model exhibits better generalization ability, enabling it to better adapt to new samples and data distributions.

[0065] In addition, the present invention compares the ssNOVA model with existing tools for predicting regulatory variants in autoimmune diseases. The present invention evaluates the capabilities of sscNOVA, ExPecto, and deltaSVM on an experimentally validated test dataset. When comparing with Expecto, the present invention attempted two methods to calculate the Expecto score. The first method involves comparing the predicted variant labels of the Expecto model with the true labels and then calculating the evaluation metrics. The second method involves taking the absolute value of the predicted probability of the ExPecto model, normalizing it, and calculating the evaluation metrics based on the normalized probability and the true label. For calculating the deltaSVM score, the present invention tested the GM12878, K562, and HepG2 cell line models developed by deltaSVM. According to Fig.10 As shown in Figure 5D, the sscNOVA model showed better performance than existing methods in identifying regulatory variants in autoimmune diseases.

[0066] Furthermore, in step 4, the sscNOVA model takes the VCF file as input to generate the predicted probability of each variant as output. Among them, 80% of the corpus variants are used as the training data set (green solid box and arrows) for input to the CNN model (black solid box). A five-fold cross-validation training strategy is used for the pre-training process, in which 20% of the variants are used as an independent test data set to evaluate the model performance (AUC = 0.891, green curve).

[0067] Through the above method, the present invention solves the problem of insufficient data on regulatory variations related to autoimmune diseases and improves the accuracy and reliability of the model for functional regulatory variations in autoimmune diseases.

[0068] On the other hand, the present invention also proposes a semi-supervised learning system for predicting functional regulatory variation, comprising:

[0069] The data prediction module is used to accept the uploaded variant data, use the sscNOVA model in the above method to predict the uploaded variant data, and generate corresponding prediction results. The prediction results include the variant score, and the predicted file can be downloaded.

[0070] The Prediction Results Viewing module enables users to view the predicted variant data and displays the prediction results of each variant, including the score, association with autoimmune diseases, and information on the immune-related cells involved.

[0071] The dataset download module allows users to download datasets related to autoimmune diseases (including training sets, test sets, and experimentally verified test sets) to facilitate subsequent research and application.

[0072] On the other hand, a computer program product is also provided, comprising computer program instructions. When the computer program instructions are read and run by a processor, the sscNOVA model is executed to predict the uploaded variation data, and a corresponding prediction result is generated. The prediction result includes a variation score, and the predicted file can be downloaded.

[0073] An electronic device comprising a processor, a memory and a bus;

[0074] The processor and the memory communicate with each other via a bus;

[0075] The memory stores computer program instructions that can be executed by the processor. The processor calls the computer program instructions to execute the sscNOVA model to predict the uploaded variation data and generate corresponding prediction results, which include variation scores and can download the predicted files.

[0076] A computer-readable storage medium is used to store computer program instructions. When the computer program instructions are executed by a computer, the computer is used to execute an sscNOVA model to predict uploaded variation data and generate corresponding prediction results, the prediction results including variation scores, and the predicted files can be downloaded.

[0077] It should be noted that the embodiments of the present invention can be implemented by hardware, software or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. A person of ordinary skill in the art will appreciate that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and module of the present invention can be implemented by a programmable hardware device circuit such as a very large scale integrated circuit or gate array, a semiconductor such as a logic chip, a transistor, etc., or a field editable gate array, an editable logic device, or can be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semi-supervised deep learning method for predicting functional regulatory variation, characterized in that: include: Step 1. Construct an autoimmune disease-related variation dataset based on GWAS Catalog and ImmuNexUT: The intersection of the processed GWAS Catalog and ImmuNex UT variant data with the number E-GEAD-398 was taken as the positive variants of the training set and the independent test set, and the corresponding negative variants all came from the variant data with the number E-GEAD-420 of ImmuNex UT; Step 2. Annotate and filter features related to autoimmune diseases based on autoimmune diseases and immune cell types: The features from the Sei model were used to screen out the immune-related features. Then, 141 features were screened out using the FeatureImportance of Random Forest, 150 features were obtained using the mutual_info_classif method of SelectKBest, and 40 features were screened out using the Sei model. Step 3. Pre-train the convolutional neural network model based on the intersection data of GWAS and ImmuNeXUT: Randomly sample the mutated data in step 1 and divide the data into a training set and a test set at a ratio of 80% and 20%. The training set is input into the CNN model for pre-training using a five-fold cross-validation. The test set is used to evaluate the model performance. The model with the highest AUC value is used as the corresponding pre-trained convolutional neural network model. Step 4. Build a semi-supervised sscNOVA model based on GWAS data that has no intersection with ImmuNexUT data: Use the GWAS data that do not intersect with ImmuNexUT to build a pseudo-labeled dataset, and pseudo-label the unlabeled GWAS data according to the threshold and t-test method. Merge the GWAS data with the training set in step 3, and after five-fold cross-validation training, the model with the highest AUC is called sscNOVA. Among them, the pseudo-label dataset construction method includes: First, use the convolutional neural network model pre-trained in step 3 to make probability predictions for GWAS data that do not intersect with ImmuNexUT, and perform five-fold cross-validation on it. For each variant, five probability values ​​are generated as predictions. Then perform a Student's t test to determine whether the difference between the five probability values ​​of each variant is statistically significant, that is, the P value is less than 0.

05. If the P value of the variant is less than 0.05, the variant is retained, otherwise it is discarded. Find the best pseudo-label threshold for the variant, perform parameter search, and then use 0.5 as the reference threshold to create five sets of thresholds for all unlabeled variants, namely ±0, ±0.1, ±0.2, ±0.3, and ±0.

4. Repeat the above process and apply it to all unlabeled variants. Then, merge the pseudo-labeled variants with the original training data set, retrain the model, and compare the model performance; Step 6. Evaluate the performance of the sscNOVA model based on the independent test set and the experimental validation test set.

2. The method according to claim 1, characterized in that Step 2 specifically includes the following steps:

201. First, the features related to immune cells were selected from the features of Sei model. Then, 1000, 800, 600, 400, and 200 features were selected from the features related to immune cells using the mutual_ingo_classif and f_classif methods of SelectKBest, respectively. The features were visualized using TSNE to compare the classification effects.

202. Continue to use the mutual_info_classif method to filter out 150, 100, and 50 features from immune cell-related features; 203. Based on the Feature Importance of Random Forest method, 141 features were selected, and the TSNE method was used to visualize the features and 40 sequence features of Se i model to compare the classification results; 204. Perform five-fold cross validation on the Random Forest model to evaluate model performance.

3. The method according to claim 1, characterized in that Step 3 specifically includes: based on the 141 features obtained by the Feature Importance of the RandomForest method in step 2, the 150 features obtained by the mutual_info_classif method of SelectKBest, and the 40 sequence class features provided by the Sei model, a convolutional neural network model is used for five-fold cross validation, and the prediction effect of the model is evaluated on the test set.

4. The method according to claim 1, characterized in that: Step 6 included: The sscnNOVA model approach was suitable for predicting functional variants in autoimmune diseases using 141 features.

5. A semi-supervised learning system for predicting functional regulatory variation in autoimmune diseases, characterized in that: include: A data prediction module, used to accept the uploaded variant data, use the sscNOVA model in the semi-supervised deep learning method for predicting functional regulatory variation as claimed in any one of claims 1 to 4 to predict the uploaded variant data, generate corresponding prediction results, the prediction results include the variant score, and can download the predicted file; The Prediction Results Viewing module enables users to view the predicted variant data, displaying the prediction results of each variant, including the score, association with autoimmune diseases, and information about the immune-related cells involved; The dataset download module allows users to download datasets related to autoimmune diseases.

6. The system according to claim 5, characterized in that The dataset includes a training set, a test set, and a test set for experimental verification.

7. A computer program product, comprising computer program instructions, which, when read and executed by a processor, execute the sscNOVA model in the semi-supervised deep learning method for predicting functional regulatory variations as claimed in any one of claims 1 to 4 to predict the uploaded variation data, generate corresponding prediction results, the prediction results include variation scores, and can download the predicted files.

8. An electronic device comprising a processor, a memory and a bus; The processor and the memory communicate with each other via a bus; The memory stores computer program instructions that can be executed by the processor. The processor calls the computer program instructions to execute the sscNOVA model in the semi-supervised deep learning method for predicting functional regulatory variations as claimed in any one of claims 1 to 4 to predict the uploaded variation data, generate corresponding prediction results, the prediction results include variation scores, and can download the predicted files.

9. A computer-readable storage medium, wherein the computer-readable storage medium is used to store computer program instructions. When the computer program instructions are executed by a computer, the computer is used to execute the sscNOVA model in the semi-supervised deep learning method for predicting functional regulatory variations as claimed in any one of claims 1 to 4 to predict the uploaded variation data, generate corresponding prediction results, the prediction results include variation scores, and can download the predicted files.

Images