Method for pathogenicity prediction of novel variants

Abstract

The application discloses a pathogenicity prediction method for novel variation, comprising the following steps: obtaining novel variation site data to be analyzed, and generating a numerical feature vector containing evolutionary conservation, functional prediction score and quantitative clinical evidence; constructing a feature extractor through the numerical feature vector; inputting a training data set containing labeled samples and unlabeled samples into the feature extractor, updating the feature extractor and the network parameters of each head; applying the feature extractor and the supervised classification head after parameter updating to predict target variation, obtaining pathogenicity probability, and calculating the contribution value of each feature in the numerical feature vector to the pathogenicity probability. The technical scheme provided by the application outputs auxiliary decision information conforming to the diagnosis habit of doctors, greatly reduces the work load of artificial interpretation of unknown variation, and realizes end-to-end intelligent processing from raw sequencing data to clinical auxiliary diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological information data processing, and particularly relates to a pathogenicity prediction method for de novo mutations. BACKGROUND

[0002] De novo mutations (DNMs) refer to DNA sequence changes that first appear in the genome of an individual during the formation of parental germ cells or early embryonic development. Such mutations do not exist in the parental somatic cells, and therefore have a high degree of randomness and uniqueness. With the popularization of genome sequencing technology, research has confirmed that de novo mutations play a crucial role in the occurrence and development of various human genetic diseases and complex diseases, especially in the field of early-onset neurological diseases, such as autism spectrum disorders, intellectual disabilities, early-onset Parkinson's disease, and early-onset Alzheimer's disease. Since de novo mutations often disrupt the function of key genes, accurate identification and interpretation of these mutations have important clinical significance for early diagnosis, genetic counseling, and precision treatment of diseases.

[0003] However, the existing prediction technology for pathogenicity of gene variation has poor generalization performance, low prediction accuracy and insufficient coverage when applied to the scenario of new variation. The mainstream prediction methods CADD and DANN generally adopt a pure supervised learning paradigm, which highly depends on a large amount of high-quality and uniformly distributed labeled data for model training. However, in the context of research on new variations, the data presents natural label scarcity and extreme class imbalance characteristics, that is, the number of known pathogenic variants is much less than that of benign or unknown variants. Under the condition of insufficient training data and serious distribution imbalance, the existing model cannot fully learn the complex and nonlinear discriminant boundary between pathogenic variants and benign variants, and tends to strongly favor the benign mode, resulting in a sharp decline in the recognition sensitivity of rare pathogenic variants, and a large number of false negative results. In addition, for variations with severe feature loss or not included in the knowledge base, especially a large number of variations with unknown significance, the existing tools often rely too much on explicit complete features due to the lack of ability to infer from the underlying structure of the data, resulting in ineffective prediction and severely limiting their coverage and automation efficiency in real-world clinical testing. On the other hand, the existing technology has a serious problem of black box decision-making and disconnection with clinical standards in clinical application. Although some deep learning methods have improved numerical performance, their internal decision-making process is highly opaque, only outputting a summary score or pathogenic level, and cannot provide feature-level contribution analysis. This unexplainability makes it difficult for clinical experts to trace the biological evidence of model conclusions and establish confidence in high-risk decisions. More critically, the abstract scoring system of traditional functional prediction tools is disconnected from the structured evidence items in authoritative clinical interpretation guidelines such as ACMG / AMP, lacking a direct quantitative mapping relationship, resulting in model prediction results that cannot be seamlessly embedded into standardized clinical diagnosis reports, still requiring heavy manual review and secondary interpretation, making it difficult for advanced algorithms to be truly integrated into clinical workflows as reliable decision support tools. SUMMARY

[0004] The present application provides a pathogenicity prediction method for new variations to effectively utilize massive unlabeled data through a multi-task joint training framework of supervised, self-supervised and contrastive learning, overcoming the generalization performance bottleneck caused by data scarcity and class imbalance of new variations, and improving the prediction sensitivity of rare pathogenic variations. At the same time, by fusing local and global feature extraction and establishing a reverse mapping mechanism from feature contribution to clinical evidence items, the black box barrier of traditional algorithms is broken, and an interpretable report conforming to the specifications of clinical guidelines can be generated, achieving high-precision prediction of the pathogenicity of new variations.

[0005] To solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0006] A pathogenicity prediction method for novel variants is provided, comprising:

[0007] S1, obtaining novel variant site data to be analyzed, performing multi-source functional annotation and clinical evidence quantification, and generating a numerical feature vector containing evolutionary conservation, functional prediction score and quantified clinical evidence;

[0008] S2, constructing a feature extractor through the numerical feature vector, the feature extractor being connected with a parallel supervised classification head, a self-supervised reconstruction head and a contrastive learning head, and the feature extractor being internally connected with a one-dimensional convolution layer for extracting local features and a dilated convolution layer for extracting global dependent features;

[0009] S3, inputting a training data set containing labeled samples and unlabeled samples into the feature extractor, calculating the classification loss of the supervised classification head, the reconstruction loss of the self-supervised reconstruction head and the contrastive loss of the contrastive learning head, and jointly updating the network parameters of the feature extractor and each head based on the weighted sum of the total loss;

[0010] S4, applying the feature extractor and the supervised classification head with updated parameters to predict the target variant, obtaining the pathogenicity probability, calculating the contribution value of each feature in the numerical feature vector to the pathogenicity probability, and mapping the features with high contribution value to the corresponding clinical evidence items to generate an explanatory report containing pathogenicity level and evidence traceability.

[0011] Further, in step S1, the numerical feature vector containing evolutionary conservation, functional prediction score and quantified clinical evidence is generated, comprising:

[0012] S11, detecting the missing rate of each feature in the novel variant site data, and removing samples with missing rate exceeding a first preset threshold and feature dimensions with missing rate exceeding a second preset threshold;

[0013] S12, performing K-nearest neighbor imputation on the remaining missing values after removal, calculating the filling values according to the weighted Euclidean distance of samples in the feature space, and forming a complete original feature matrix;

[0014] S13, performing Min-Max normalization processing on the original feature matrix, and splicing the processed features of each dimension in a predetermined channel order to form the numerical feature vector.

[0015] Further, in step S1, the novel variant site data to be analyzed is obtained, and a hierarchical label system is constructed, comprising:

[0016] S14, retrieve ClinVar database, directly mark the variant sites with no-conflict multi-submitter records or expert panel review records as first-level pathogenic or benign labels;

[0017] S15, retrieve Gene4Denovo database, mark the variants with all phenotypes as control as benign, and mark the variants with all phenotypes as pathogenic and meeting the frequency threshold as pathogenic as second-level labels;

[0018] S16, combine the first-level pathogenic or benign labels and the second-level labels to form a labeled sample set, and divide the remaining variant sites without labels into an unlabeled sample set.

[0019] Further, in step S2, a feature extractor is constructed based on the numerical feature vector, including:

[0020] S21, configure the one-dimensional convolution layer to perform a sliding convolution operation on the numerical feature vector according to a small-size convolution kernel to generate a local feature map reflecting the correlation of adjacent features;

[0021] S22, configure the atrous convolution layer to cascade a plurality of atrous convolution units after the local feature map, set an expansion rate parameter exponentially increasing to expand the receptive field, and generate a global dependent feature map;

[0022] S23, flatten the global dependent feature map and map it through a fully connected layer to generate a low-dimensional latent feature vector as a shared input of the supervised classification head, the self-supervised reconstruction head and the contrastive learning head.

[0023] Further, in step S3, the total loss based on weighted summation is used to jointly update the network parameters of the feature extractor and each head, including:

[0024] S31, calculate the classification loss of the labeled samples in the supervised classification head according to the Focal Loss function, and reduce the weight of simple negative samples through a focusing parameter;

[0025] S32, calculate the reconstruction loss of all samples in the self-supervised reconstruction head according to the mean square error function, and calculate the contrastive loss of the enhanced view of all samples in the contrastive learning head according to the normalized temperature scale cross-entropy function;

[0026] S33, add the classification loss, the reconstruction loss and the contrastive loss according to a preset hyperparameter weight to obtain a target function, and update the network parameters in the gradient direction of the target function according to an optimizer.

[0027] Further, in step S4, the features with high contribution values are mapped back to the corresponding clinical evidence items, including:

[0028] S41, based on the principle of cooperative game theory, the marginal change of the model output probability before and after the replacement characteristic value is calculated, and the contribution value of each feature dimension in the numerical feature vector is quantified;

[0029] S42, the features are sorted according to the absolute value of the contribution value, and the top N key features with the largest contribution to the pathogenicity probability are screened out;

[0030] S43, the original description corresponding to the key feature is found according to the feature index table, and if the key feature belongs to the quantitative clinical evidence feature, the corresponding evidence code and pathogenicity support level are output.

[0031] According to the second aspect of the application, a semi-supervised hybrid convolutional network pathogenicity prediction device for novel variants is provided, comprising:

[0032] The generating module is configured to obtain novel variant site data to be analyzed, perform multi-source functional annotation and clinical evidence quantification on the data, and generate a numerical feature vector containing evolutionary conservation, functional prediction score and quantitative clinical evidence;

[0033] The constructing module is configured to construct a feature extractor through the numerical feature vector, wherein the feature extractor is connected with a parallel supervised classification head, a self-supervised reconstruction head and a contrast learning head, and the feature extractor is internally connected with a one-dimensional convolutional layer for extracting local features and a dilated convolutional layer for extracting global dependent features;

[0034] The computing module is configured to input a training data set containing labeled samples and unlabeled samples into the feature extractor, calculate a classification loss of the supervised classification head, a reconstruction loss of the self-supervised reconstruction head and a contrast loss of the contrast learning head, and jointly update network parameters of the feature extractor and each head based on a weighted sum total loss;

[0035] The prediction module is configured to apply the feature extractor and the supervised classification head with updated parameters to predict a target variant, obtain a pathogenicity probability, calculate a contribution value of each feature in the numerical feature vector to the pathogenicity probability, and map the feature with a high contribution value to a corresponding clinical evidence item in reverse to generate an explanatory report containing a pathogenicity level and evidence traceability.

[0036] The application also provides a pathogenicity prediction system for novel variants, comprising a processor and a computer readable storage medium connected to each other, wherein the computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the pathogenicity prediction method for novel variants.

[0037] The application also provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program, when executed by a processor, implements the steps of the pathogenicity prediction method for a new variant.

[0038] The application also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the pathogenicity prediction method for a new variant.

[0039] Compared with the prior art, the application has the following advantages:

[0040] The application effectively utilizes massive unlabeled data through a multi-task joint training framework of supervised, self-supervised and contrastive learning in parallel, overcomes the generalization performance bottleneck caused by the scarcity and class imbalance of new variant data, and improves the prediction sensitivity of rare pathogenic variants. At the same time, by fusing local and global feature extraction and establishing a reverse mapping mechanism from feature contribution to clinical evidence items, the black box barrier of traditional algorithms is broken, and an interpretable report conforming to the specification of clinical guidelines can be generated, realizing high-precision prediction of the pathogenicity of new variants. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A method flowchart is provided for the embodiments of the application;

[0042] Figure 2 A structure diagram of a semi-supervised hybrid convolutional network pathogenicity prediction device for a new variant is provided for the embodiments of the application;

[0043] Figure 3 A schematic diagram of coding region variant and feature screening result is provided for the embodiments of the application;

[0044] Figure 4 A schematic diagram of label determination logic is provided for the embodiments of the application;

[0045] Figure 5 A schematic diagram of the running flow of the training stage is provided for the embodiments of the application;

[0046] Figure 6 A schematic diagram of the running logic of the supervised classification module is provided for the embodiments of the application;

[0047] Figure 7 A schematic diagram of the running logic of the reconstruction module is provided for the embodiments of the application;

[0048] Figure 8 A schematic diagram of the running logic of the contrastive learning module is provided for the embodiments of the application;

[0049] Figure 9A schematic diagram of the running flow of the reasoning phase provided for the embodiments of the present application. DETAILED DESCRIPTION

[0050] The present application is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the scope of protection of the present application is not limited thereby.

[0051] The technical solutions adopted by the present embodiment, as shown in Figure 1 include the following steps:

[0052] S1, obtaining new variant site data to be analyzed, performing multi-source functional annotation and clinical evidence quantification thereon, and generating a numerical feature vector containing evolutionary conservation, functional prediction score and quantified clinical evidence;

[0053] S2, constructing a feature extractor through the numerical feature vector, the feature extractor being connected with a parallel supervised classification head, a self-supervised reconstruction head and a contrastive learning head, and the feature extractor being internally connected with a one-dimensional convolution layer for extracting local features and a dilated convolution layer for extracting global dependency features;

[0054] S3, inputting a training data set containing labeled samples and unlabeled samples into the feature extractor, respectively calculating the classification loss of the supervised classification head, the reconstruction loss of the self-supervised reconstruction head and the contrastive loss of the contrastive learning head, and jointly updating the network parameters of the feature extractor and each head based on the weighted sum of the total loss;

[0055] S4, applying the feature extractor and the supervised classification head whose parameters are updated to predict the target variation, obtaining the pathogenicity probability, calculating the contribution value of each feature in the numerical feature vector to the pathogenicity probability, and inversely mapping the features with high contribution value to the corresponding clinical evidence items to generate an explanatory report containing pathogenicity level and evidence traceability.

[0056] In one embodiment, based on the deep semi-supervised learning and multi-task joint optimization framework, the problems of low prediction accuracy and poor clinical interpretability of the prior art in the scenario of scarce genetic data can be solved. The method of the present embodiment can be run on a bioinformatics analysis server configured with a high-performance graphics processor.

[0057] Step S1 is performed to construct a high-dimensional feature space with biological significance and computational operability. The system receives new variant site data in VCF format to be analyzed, and performs multi-source functional annotation on the sites by tools such as Annovar or VEP. The annotation content covers population frequency database, evolutionary conservation score, protein structure prediction and biochemical property prediction. The clinical evidence quantification module predefines the mapping rules of standard evidence items and numerical vectors in ACMG / AMP guidelines. For example, for sites that meet the criterion of truncated variants causing loss of gene function, the system not only records the sequence changes, but also encodes the confidence of the corresponding PVS1 item as a numerical feature. Through normalization processing and missing value filling, the above heterogeneous information is integrated into a standardized numerical feature vector, which not only retains the underlying sequence physical features, but also embeds high-level clinical prior knowledge.

[0058] Step S2 is performed to construct a deep neural network model with a specific topological structure. The deep neural network model is mainly composed of a feature extractor, which adopts a series connection architecture. The front end is a one-dimensional convolutional layer (1D-CNN) configured with a small convolution kernel and assisted by a ReLU activation function, which is specially used to capture local mutation patterns of gene sequences and short-range correlation features of adjacent sites. The rear end is connected to a dilated convolution layer, which exponentially expands the receptive field without increasing the number of parameters by setting the dilation rate, thereby effectively extracting global dependence features between long-distance regulatory elements in the genomic sequence. The output end of the feature extractor is connected in parallel to three functionally heterogeneous task heads, namely a supervised classification head, a self-supervised reconstruction head and a contrastive learning head. The supervised classification head is composed of a fully connected layer and a Softmax activation function, which is used to output the probability distribution of pathogenicity / normality. The self-supervised reconstruction head adopts a decoder structure, which aims to restore the extracted abstract features to the original input vector, forcing the network to learn the principal components of the data. The contrastive learning head contains a nonlinear projection layer, which is used to map the features to the hyperspherical space to calculate the similarity distance between samples.

[0059] Step S3 is performed, i.e., a model joint training phase based on a multi-task loss function. In view of the problem that there are few labeled samples and many unlabeled samples in the new variant data, a mixed data flow input strategy is adopted to solve the problem. The training set is input into the network, which includes a small number of clinically diagnosed pathogenic / benign samples (labeled) and a large number of unclassified population variant samples (unlabeled). In the training process, the loss function is composed of three parts. First, the cross-entropy loss (Cross-Entropy Loss) of the supervised classification head is calculated only by the labeled samples, guiding the model to establish a clear discrimination boundary; second, the mean square error loss (MSE Loss) of the self-supervised reconstruction head is calculated by all samples, whether labeled or not, and the feature extractor is forced to capture the potential distribution rule and noise-resistant features of the data itself through the reconstruction task; and third, the contrast loss (such as NT-Xent Loss) of the contrast learning head is used to narrow the distance between similar or similar variants in the feature space and to push away the distance between different variants. The system adjusts the weight coefficient dynamically, for example, giving a higher weight to the reconstruction loss and the contrast loss in the early stage of training, and focusing on the classification loss in the later stage. The network parameters of the feature extractor and each head are updated jointly based on the back propagation algorithm. The model can still learn robust feature representation from unlabeled data even in the case of extremely scarce labels.

[0060] Step S4 is performed, i.e., a prediction inference and explainability report generation phase. The trained model is applied to forward inference of the target new variant, and the supervised classification head outputs the probability value of the variant belonging to pathogenic (Pathogenic), likely pathogenic (Likely Pathogenic), or benign (Benign). At the same time, the system starts the explainability analysis module, calculates the contribution value (Shapley Value) of each dimension in the input feature vector to the final prediction probability through the SHAP (Shapley Additive Explanations) algorithm or the integral gradient method (Integrated Gradients). The system automatically selects the top K key features with the highest contribution values and triggers the reverse mapping mechanism. For example, if the dimension corresponding to the PhyloP conservation score in the feature vector has the highest contribution, the system will backtrack to the mapping table in S1 to retrieve the corresponding clinical evidence description “PP3: The calculation predicts that the variant has a harmful effect and the site is highly conserved”.

[0061] In a preferred embodiment, the present application proposes a semi-supervised hybrid convolutional neural network (SH-CNN) system for coding region de novo mutation pathogenicity prediction. By integrating multi-source functional annotation features and clinical evidence features, combined with innovative network architecture and multi-task learning mechanism, the system achieves high-precision and high-coverage pathogenicity prediction under the condition of label scarcity.

[0062] The specific process is: data preprocessing → label construction → feature system construction → semi-supervised hybrid convolutional network training → explainable clinical output → prediction result output.

[0063] 1. Data preprocessing

[0064] 1.1 Dataset

[0065] The dataset comes from the Gene4De novo database, which integrates 741,866 de novo mutation (DNM) records. After removing entries lacking annotation information and duplicate mutations from different literature during data preprocessing, a total of 70,415 coding region mutations were retained.

[0066] In addition, to evaluate the independent generalization performance of the model, the Autism Spectrum Disorder (ASD) cohort (SPARK) was used as an independent test set. Based on the de novo mutation identification method proposed by Samocha et al. (paper: Samocha KE, Robinson EB, Sanders SJ, Stevens C, Sabo A, McGrath LM, et al. A framework for the interpretation of de novo mutation in human disease. NatGenet. 2014;46(9):944-50), a standardized data processing workflow was established, including three main steps: First, input the VCF file, pedigree file (.fam), and mutation list for each mutation to identify the target trio (proband-parent) combination; Second, set quality control parameters, including genotype quality (GQ) filtering, genotype consistency check based on Mendelian inheritance laws, and further exclude false positive sites using indicators such as mutation frequency, parent-offspring depth ratio, and allele balance; Finally, use the de novo mutation calling module to evaluate the de novo mutation status of the mutations in the trio. This process effectively distinguishes between real novel variants and artificial variants caused by sequencing or genotyping errors through rigorous statistical modeling and multiple filtering strategies, thereby ensuring the high reliability of the generated variant set.

[0067] After the above process and data preprocessing, 33,948 variants in the coding region were ultimately retained. It is important to note that this dataset is completely independent of the training data; there is no overlap in variants between them. It is used solely for independent testing and comparative evaluation of the model and is not used in training.

[0068] 1.2 Quality Control Process

[0069] Regarding the screening order, directly calculating the feature missing rate would be subject to interference from a large number of highly missing rows. Therefore, severely missing variants should be cleaned up first, followed by a precise assessment of feature missingness. In comparing multiple combination schemes, the following thresholds were ultimately selected because they achieved a better balance between missing data control and information preservation: significantly reducing the overall missing rate while preserving clinical labels and core functional features as much as possible. The coding region variants and feature screening results are attached. Figure 3 As shown. A dual threshold filtering strategy is employed to ensure data quality:

[0070] (1) Mutation level filtering: Remove variant samples with a feature missing rate of more than 40%.

[0071] (2) Feature-level filtering: Remove feature dimensions with a missing rate of more than 70% in variant samples.

[0072] 1.3 Missing value handling

[0073] Fill the remaining missing values with K-Nearest Neighbors imputation algorithm:

[0074] Select the nearest neighbor variation based on Euclidean distance.

[0075] The selection basis is that functionally similar variations tend to cluster in feature space, and KNN can effectively utilize local structure information to restore missing data.

[0076] (1) There is a strong correlation between pathogenicity-related features, and K-Nearest Neighbors imputation based on neighboring variations can provide more reliable estimates.

[0077] (2) K-Nearest Neighbors imputation can better preserve the local structure information of the data, avoiding the distribution bias that may be introduced by simple mean or median imputation. The imputation formula is as follows:

[0078]

[0079] The explanation of each symbol in the above formula is shown in Table 1.

[0080] Table 1

[0081]

[0082] 1.4 Feature standardization

[0083] Min-Max normalization is performed on all numerical features. Linear transformation of the original data maps the result value to [0-1], eliminates the dimensional difference between different features, promotes model convergence stability, and the conversion function is as follows:

[0084]

[0085] The explanation of each symbol in the above formula is shown in Table 2.

[0086] Table 2

[0087]

[0088] 1.5 Data division strategy

[0089] Training set / validation set / test set ratio: 6:2:2.

[0090] Stratified sampling: Ensure consistent label distribution in each data set.

[0091] Over-sampling: Use SMOTE over-sampling technology for the minority class (pathogenic variation) in the training set to alleviate class imbalance.

[0092] 2 Labeling system construction

[0093] 2.1 Hierarchical labeling strategy

[0094] Two-level labeling principle is adopted: "ClinVar high confidence first, Gene4Denovo phenotype supplement".

[0095] First level: ClinVar high confidence labeling (highest priority), screening criteria is ClinVar annotation as criteria_provided\multiple_submitters\no_conflicts (exact multiple non-conflicting submitters) or reviewed_by_expert_panel (reviewed by expert panel). Directly use the provided pathogenic / benign annotation as the final label.

[0096] Second level: Gene4Denovo phenotype (phenotype represents the pathogenic state of the variation) supplement labeling, the label determination logic is shown in the attached Figure 4 .

[0097] Benign label (label=0): All phenotypes are control (not pathogenic).

[0098] Pathogenic label (label=1): All phenotypes are pathogenic, and reported in ≥2 independent literature or ≥2 research subjects.

[0099] Unknown label (label=-1): Variations that do not meet the above conditions.

[0100] 2.2 Label conflict resolution mechanism

[0101] Basic principle: ClinVar annotation priority is higher than Gene4Denovo.

[0102] Conflict detection: When the same variation has opposite annotations in ClinVar and Gene4Denovo.

[0103] Solution: Take ClinVar high confidence annotation as the standard to ensure the authority of the label.

[0104] Final label distribution:

[0105] (1) Gene4Denovo dataset: Among the 70,415 coding region variations, there are 3956 pathogenic labels (813 from ClinVar, 3143 from Gene4Denovo), 2965 benign labels (1704 from ClinVar, 1261 from Gene4Denovo), and 63,494 unknown labels.

[0106] (2) Spark independent test set: Among 33,948 coding region variants, 517 were pathogenic (264 from ClinVar, 253 from Gene4Denovo), 8,614 were benign (260 from ClinVar, 8,354 from Gene4Denovo), and 24,817 were unknown.

[0107] 3. Feature system construction

[0108] Candidate DNMs were first annotated by ANNOVAR (version 2024.09.24) for basic functional annotation, and then integrated with multiple functional prediction tools based on the VarCards2 framework. Among them, MutPred2 was used to quantitatively evaluate the potential damage of amino acid substitution to protein function; ReVe integrated the results of REVEL and VEST3 to improve the discrimination performance of missense variants; SpliceAI was used to predict abnormal splicing events that may be triggered by the variant. All scoring results were standardized and aligned with the basic annotation information to form the model input feature set, covering key biological aspects such as protein structure constraints, functional damage, and splicing regulation.

[0109] Clinical evidence features come from ClinVar and GeneBe's automated ACMG / AMP scoring tools. The annotations in ClinVar are submitted by experts manually, while GeneBe is an online tool that can automatically extract the evidence strength of various evidence clauses (such as PVS1, PM1, PP3, BS1, etc.) based on the ACMG / AMP variant interpretation guidelines and output quantitative scores to reflect the support level of each variant in clinical interpretation. Incorporating these ACMG scores into the feature system can effectively bridge deep learning models with clinical interpretation standards, enhancing the interpretability and clinical value of the prediction results.

[0110] To facilitate downstream modeling, all features are divided into six categories: (1) evolutionary conservation; (2) coding region functional prediction; (3) cross-region functional prediction features; (4) splicing effect; (5) population allele frequency; (6) clinical evidence score.

[0111] 3.1 Evolutionary conservation features (12 dimensions)

[0112] GERP++, phyloP, phastCons, SiPhy, and LINSIGHT were used to measure the evolutionary constraints of sequences across different species, reflecting the potential functional importance of the variant.

[0113] 3.2 Coding region functional prediction (52 dimensions)

[0114] Integrate 52 mainstream pathogenicity prediction scores, covering biological impact assessment of amino acid substitution, sequence model prediction and integrated scores (such as SIFT, PolyPhen-2, Mutation Taster, REVEL, MutPred2, Alpha Missense, EVE, etc.), comprehensive assessment of potential pathogenicity of variants.

[0115] 3.3 Cross-region functional prediction features (25 dimensions)

[0116] Some functional prediction tools (such as Eigen, GenoCanyon) were originally designed to assess the regulatory potential of non-coding region variants, but their scoring models are based on whole genome features, and they can also output results for coding region variants. This study includes them as supplementary features in the model to reflect the potential impact of variants on transcriptional regulation, chromatin openness and epigenetic modification. These features help capture regulatory signals beyond traditional protein function scores in coding regions, providing a more comprehensive characterization of the functional effects of variants.

[0117] 3.4 SpliceAI splicing effect (9 dimensions)

[0118] SpliceAI predicts the potential splicing effect of each variant site, including four types of delta scores (DS_AG, DS_AL, DS_DG, DS_DL) and delta position (DP), which are used to describe the position of splicing events. Variant sites close to the end of the chromosome or limited by the reference sequence are considered missing values.

[0119] 3.5 Allele population frequency (63 dimensions)

[0120] Population allele frequency information from multiple authoritative databases is included, including gnomADv4.1 (exome and genome), ExAC and ESP, a total of 63 frequency features, covering overall, gender, ethnic group and rare allele filtering frequency (FAF).

[0121] 3.6 Clinical evidence score

[0122] There are mainly two sources: one is the pathogenicity determination submitted and reviewed by experts in the ClinVar database; the second is the automated scoring tool based on ACMG / AMP guidelines provided by GeneBe. The combination of the two ensures the authority of expert evidence and provides systematic and quantitative ACMG item assignments.

[0123] ​The judgment of PS2 / PM6 needs to be based on sufficient literature evidence. At present, the Gene4Denovo database has systematically collected supporting literature, while GeneBe cannot directly provide PS2 / PM6 evidence related to new variants. Therefore, based on the recommendations of the ClinGen SVI working group, the literature data in Gene4Denovo were integrated to construct a PS2 / PM6 scoring algorithm to support more accurate pathogenicity assessment: According to the parent-child verification, phenotype and gene consistency, and the number of observed DNM of each DNM, a scoring system was constructed, and it was converted into quantitative evidence strength according to the ClinGen guidelines (2 points for high specificity phenotype and parent-child verification; 1 point for consistent phenotype; 0.5 points for high genetic heterogeneity; 0 points for inconsistent phenotype; half the corresponding score for unverified parent-child). According to the total score, it is further mapped to the evidence level (0.5: support, 1: medium, 2: strong, 4: very strong). Finally, the evidence strength of all ACMG items is integrated into the model features in the form of onehot encoding, which is used to connect deep learning prediction and clinical standards.

[0124] 4. Semi-supervised hybrid convolutional network training

[0125] SH-CNN combines the advantages of convolutional neural network (CNN) and dilated convolutional network (IDCNN). The shallow CNN can capture the local patterns between adjacent features in the feature vector, such as the adjacent conservation score and functional prediction score, which may jointly indicate the importance of the variant. Dilated convolution can model the feature dependence across a longer distance without reducing the feature resolution by expanding the receptive field, so that the model can integrate the global information between different categories of features. For example, the splicing effect score and the conservation score are associated to identify the pathogenic signal they jointly represent. By integrating CNN and IDCNN, SH-CNN can not only extract local features in detail, but also integrate global information, providing a solid structural foundation for semi-supervised feature learning.

[0126] The model is composed of three types of functional modules: feature extractor, supervised classification head, and self-supervised task module (including reconstruction head and contrastive learning projection head).

[0127] 4.1 Feature extractor

[0128] Input: One-dimensional annotation feature vector of candidate DNM (1 x n, n is the feature dimension).

[0129] Method: Convolution + pooling to model local patterns → multi-layer one-dimensional dilated convolution (dilation rate 1, 2, 4, 8) for cross-scale modeling.

[0130] Output: 128-dimensional latent representation as the unified feature basis for three types of tasks (supervised classification, reconstruction, and contrastive learning).

[0131] 4.2 Multi-task head parallel architecture

[0132] (1) Supervised classification head---learning pathogenic patterns

[0133] Input: A potential representation of labeled mutations.

[0134] Method: The latent features of the mutation are subjected to Dropout regularization and linear transformation.

[0135] Objective: To learn a discrimination pattern that can distinguish between pathogenic and benign mutations.

[0136] Output: Binary classification probability distribution (used during the training phase to calculate supervised classification loss and update model parameters).

[0137] The operational logic of the supervised classification module is as follows: Figure 6 As shown. For example, variants with pathogenicity labels often exhibit a combination of features including "high conservation + extremely low population frequency + high functional prediction score," and the classification module will directly learn these patterns.

[0138] (2) Reconstruction module --- capturing the internal feature structure of the variant

[0139] Input: A potential representation of unlabeled mutations.

[0140] Method: The latent representation is mapped back to the original feature space through a two-layer fully connected network to minimize the reconstruction error (MSE).

[0141] Objective: To improve the fidelity of latent representations to original features and preserve collaborative features within individual variants.

[0142] Output: Reconstructed feature vectors with the same dimensions as the original input, used to calculate the reconstruction loss and update the model parameters.

[0143] The operating logic of the rebuild module is as follows: Figure 7 As shown. For example, if an unlabeled variant has a high conservation score in LINSIGHT and a low population frequency, the reconstruction module will learn the joint pattern of the two types of features, so that the latent representation retains this internal association.

[0144] (3) Comparative learning module---consistency of the same variant and distinguishability of different variants

[0145] Input: A potential representation of unlabeled mutations.

[0146] Method: Map latent representations to a 64-dimensional low-dimensional projection space, and use a positive-negative sample contrast mechanism based on the SimCLR framework: First, apply enhancement operations such as random noise, scaling, or feature mask to generate two enhanced views, aiming to encourage the model to focus on the core biological features of the variation rather than incidental perturbations. Then, use the normalized temperature scale cross-entropy (NT-Xent) loss to optimize the feature space distribution of positive and negative sample pairs: Enhanced views of the same variation (positive sample pair) are close in feature space, achieving representation stability; different variation samples (negative sample pair) maintain a relative distance even if they are similar in some features, thereby distinguishing the overall biological feature combination of different variations.

[0147] Objective: Exclude incidental noise or input order effects, maintain the stability of the same variation, and distinguish different variations.

[0148] Output: Projected low-dimensional vectors for computing contrast loss and updating model parameters.

[0149] The running logic of the contrast learning module is shown in the accompanying Figure 8 .

[0150] The outputs of these three tasks flow to different targets: the classification head outputs a binary classification probability distribution, which is used to calculate the supervised classification loss; the reconstruction head outputs a reconstructed feature vector of the same dimension as the original input, which is used to calculate the reconstruction error; the contrast learning head outputs a low-dimensional embedding vector, which is used to measure the similarity between enhanced views. The three types of output do not directly interact with each other, but each calculates a loss function, which is then combined into a total loss by weighting, and the parameters of the feature extractor and task head are updated through backpropagation. In this way, supervised and self-supervised signals are optimized together during the training phase, and the running process during the training phase is shown in the accompanying Figure 5 .

[0151] Training process:

[0152] Data loading: Load labeled and unlabeled variation data in batches;

[0153] Forward propagation: Three task heads are calculated in parallel;

[0154] Loss calculation: Weighted sum to get total loss;

[0155] Backpropagation: Calculate gradients and update network parameters;

[0156] Iterative optimization: Repeat the above process until convergence.

[0157] After training is complete, the running process diagram in the inference phase is shown in the accompanying Figure 9As shown, all inputs (whether labeled or not) are passed through a feature extractor with a supervised classification head. The convolution kernel weights, dilated convolution parameters, etc. of the feature extractor are jointly updated during the training phase through the three tasks, and the output is a binary classification probability distribution, which realizes the prediction of variant pathogenicity. At this stage, the self-supervised module no longer participates in the calculation independently, but the latent structure learned by it during the training process has been embedded in the feature representation, thereby enhancing the reliability and generalizability of the classification prediction.

[0158] 4.3 Activation function

[0159] The classification output layer uses a Softmax activation function to normalize the model output into a binary classification probability, which is used to directly reflect the pathogenicity risk of the variant. The hidden layer uses a Rectified Linear Unit (ReLU) function to alleviate the gradient vanishing and maintain computational efficiency, and the calculation formula is as follows:

[0160]

[0161]

[0162] 4.4 Loss function

[0163] The loss function follows the idea of multi-task learning and is composed of three parts: the supervised classification task uses FocalLoss to alleviate the training bias caused by class imbalance and improve the sensitivity to rare pathogenic mutations of minority classes; the self-supervised reconstruction task uses Mean Squared Error (MSE) to constrain the stability of the latent representation by minimizing the difference between the input and the reconstructed features; the contrastive learning task uses Normalized Temperature Scaling Cross Entropy (NT-Xent) to optimize the consistent representation between variants by bringing the representations of different augmented views of the same variant closer and distinguishing the representations of different variants. The three types of losses are combined with weights determined by experimental tuning: λ_recon=1.0, λ_contrast=0.5, to balance the supervised and self-supervised tasks. The FocalLoss function is as follows:

[0164]

[0165]

[0166] The FocalLoss function is for binary classification, where y is the true label and p is the predicted probability. α is the balance factor (usually 0.25), and γ is the focus parameter (usually 2).

[0167] The Mean Squared Error formula is as follows:

[0168]

[0169] The Mean Squared Error formula is the original feature, is the reconstructed feature, N is the number of features.

[0170] 4.5 Training optimization strategy

[0171] Optimizer: Adam (Adaptive Moment Estimation).

[0172] Parameter settings: β1=0.9, β2=0.999, initial learning rate 0.001.

[0173] Learning rate adjustment: Reduce LROn Plateau, dynamically adjusted based on the performance of the validation set.

[0174] Gradient clipping: threshold = 1.0, to prevent gradient explosion.

[0175] Finally, the training-test process is repeated 10 times independently under the same hyperparameters, and the mean and standard deviation are reported to ensure the robustness and repeatability of the results. Through this design, supervised tasks provide clear discriminative signals, and self-supervised tasks provide potential structural constraints, both of which significantly improve the generalization performance of the model under the condition of limited labels.

[0176] 4.6 Explainable clinical output

[0177] To enhance the transparency of medical applications, this study introduces SHAP (SHapley Additive exPlanations) into SH-CNN to quantify the contribution of each input feature to the prediction result.

[0178] Based on the Shapley value principle of cooperative game theory, the formula for calculating feature contribution is as follows:

[0179]

[0180] The explanation of the symbols in the above formula is shown in Table 3.

[0181] Table 3

[0182]

[0183] First, a baseline is set, usually the feature mean vector or background variation, representing the model output under "no valid feature information". Then, the SHAP method replaces the input features with the baseline one by one, and inputs the new feature vector into the SH-CNN to recalculate the output probability distribution of pathogenicity. By comparing the change in prediction probability before and after replacement, the marginal contribution of the feature is obtained. Further, based on the Shapley value principle, the contributions of all possible feature combinations are accumulated to obtain the global importance of each feature to the model prediction. For example, when the SpliceAI score is removed, the prediction probability decreases from 0.85 to 0.65, indicating that the feature has a positive contribution. By accumulating the marginal contributions of different features, SHAP can achieve both local explanation (single variation) and global explanation (overall feature importance ranking, such as clinical evidence one-hot > SpliceAI > population frequency).

[0184] The clinical evidence feature is encoded using one-hot encoding, which directly corresponds to the ACMG items. The SHAP output contribution can be accurately mapped to specific features and evidence, realizing the transformation from "black box prediction" to "traceable clinical explanation", enhancing the transparency and verifiability of the model, and providing methodological support for molecular diagnosis and genetic counseling.

[0185] 4.7 Prediction result output

[0186] All performance indicators are the mean ± standard deviation of 10 random repeated experiments to reflect the stability of the model results.

[0187] On the test set of the Gene4Denovo dataset, the performance of SH-CNN is better than that of existing methods, with the area under the ROC curve (AUC) of SH-CNN reaching 0.879 ± 0.012 and the accuracy (AP) reaching 0.887 ± 0.015, which is significantly higher than that of traditional prediction tools such as SIFT (AUC = 0.696, AP = 0.614), PolyPhen-2 (AUC = 0.703, AP = 0.618), and CADD (AUC = 0.733, AP = 0.755).

[0188] On the ROC and PR curves, SH-CNN shows higher overall curve position and steeper initial segment, indicating that it can maintain high sensitivity and accuracy at low false positive rate. It is worth noting that n in the figure represents the number of labeled valid variations that can be evaluated by the tool in the test set. Unlike other methods, SH-CNN can predict all known labeled variations, with a missing rate of 0%, while other tools have different degrees of missing due to limitations in scope or variation type. This feature makes SH-CNN have the advantages of precision and coverage in large-scale variation evaluation.

[0189] To verify the generalization ability of the model, an independent Spark cohort (Autism Spectrum Disorder, ASD) was evaluated. The results show that SH-CNN still maintains good overall discriminant performance (AUC=0.763) in this external data set, which is better than all control tools except CADD, showing good cross-cohort robustness. However, its average precision (AP=0.430) is relatively low. Further analysis shows that this result is mainly due to the highly unbalanced label distribution and the extreme proportion of benign variants.

[0190] The proportion of benign variants (label=0) in the Spark test set is significantly higher than that in the Gene4Denovo test set (8614 vs 3956), while the number of pathogenic variants (label=1) is only 517, accounting for less than 6%. In addition, there are a large number of unlabeled variants (label=-1, a total of 24817) in this data set, further diluting the proportion of effective positive samples. In contrast, the number of pathogenic and benign labels in the Gene4Denovo dataset is more balanced (2965 vs 3956), so the AP (0.887±0.015) is significantly higher than that of the Spark set.

[0191] The benign labels in the Spark dataset mainly come from the phenotype annotation of Gene4Denovo (accounting for 97.0%), rather than the high-confidence labels from ClinVar, which means that this part of the sample may have label noise or phenotype heterogeneity, reducing the confidence of the model in distinguishing between positive and negative samples. Accordingly, SH-CNN still maintains a high level of AUC, indicating that its overall ability to rank positive and negative samples is stable, but due to the extremely high proportion of benign samples and the compression of the threshold discrimination space, the average precision (AP) decreases.

[0192] The extremely unbalanced class distribution (benign is significantly more than pathogenic samples) and the low confidence of some labels in the Spark cohort are the main reasons for the relatively low AP. This result reflects the robustness of SH-CNN on real clinical external data. Even under the condition of extremely unbalanced class distribution, where benign samples are much more than pathogenic samples, the model is still better than most existing tools in overall discriminant performance (AUC) and precision ranking (AP), showing strong generalization ability and adaptability to the complexity of real data.

[0193] Based on the results of the Gene4Denovo dataset test set, the analysis results show:

[0194] Positive SHAP value: push the prediction towards pathogenicity;

[0195] Negative SHAP value: push the prediction towards benign;

[0196] Eigen-raw_coding had the highest feature importance in the model, indicating that the integrated score combining sequence and evolutionary conservation information played a core role in pathogenicity discrimination. This result was consistent with the biological significance of this feature reflecting that functional sites were under evolutionary selection pressure. HUVEC_confidence_value (functional confidence from endothelial cell lines) and ACMG evidence items ps4_{4} and ps4_{10} followed closely, suggesting that the model not only captured cell type-specific functional signals, but also made full use of case frequency-related clinical epidemiological evidence. In addition, functional prediction and conservation indicators such as VARITY_ER_LOO_score, Eigen-phred_coding, FATHMM_MKL_coding, and CADD_coding made significant contributions to the model decision, further verifying that the model could comprehensively utilize multi-level biological information. Notably, some cross-region features (such as RegBase_Can, Funseq2, and HUVEC_fitCons_score) derived from non-coding or regulatory annotation frameworks still showed significant contributions in coding region prediction, indicating that these features captured regulatory potential and chromatin openness signals that could serve as auxiliary discriminators, providing additional context information for the model.

[0197] The directionality of feature contribution was consistent with biological expectations, with high functional scores and high confidence values promoting the model to predict towards pathogenicity, while high allele frequencies (such as the gnomAD_AF feature) were associated with benign variants. More importantly, the high contribution of ACMG evidence items (especially the PS4 series) indicated that the model not only captured the effects of functional variations, but also effectively learned the weight structure of clinical evidence in pathogenicity assessment, bridging the gap from functional prediction to clinical interpretation.

[0198] The technical solutions in the embodiments can achieve the following technical effects.

[0199] (1) A semi-supervised hybrid deep learning framework is proposed to address the model generalization problem under label scarcity and class imbalance.

[0200] By introducing self-supervised reconstruction and contrastive learning tasks, the model is forced to learn the internal structure and context relationships between annotated features from a vast amount of unlabelled variations, thereby extracting more robust and discriminative feature representations, improving the generalization performance and sensitivity to rare pathogenic variants under limited label conditions.

[0201] (2) An interpretable deep fusion mechanism is designed to achieve traceability of the prediction results at the feature and evidence levels.

[0202] By integrating SHAP explainability analysis, the model's prediction probability is attributed to each input feature. In particular, for quantified ACMG clinical evidence features, the reverse mapping from model output to specific clinical evidence items is achieved, providing transparent and verifiable evidence support for each variant's prediction, enhancing clinical trust and usability.

[0203] (3) Build a unified feature engineering and modeling system that integrates functional annotation and clinical evidence, aiming to bridge algorithmic prediction and clinical interpretation.

[0204] By systematically quantifying and characterizing key evidence items in ACMG / AMP guidelines and jointly training the model with multi-source functional annotation features, the model not only learns the biological functional impact but also directly learns the weight logic of clinical evidence, making its output more closely aligned with the decision-making thinking of clinical experts, achieving a smooth transition from "functional prediction score" to "clinical significance interpretation".

[0205] (4) Achieve 100% coverage of end-to-end prediction system, aiming to eliminate prediction blind spots and improve clinical workflow efficiency.

[0206] Through the above semi-supervised learning to enhance feature representation and unified feature processing flow, the model can generate reliable prediction results for all input variants (whether their labels or evidence are known), achieving near 100% prediction coverage, supporting complete automation of large-scale variant screening and reducing manual intervention.

[0207] The beneficial effects in the embodiment are:

[0208] (1) In terms of performance, through semi-supervised multi-task learning, the structural information of unlabeled variants is fully utilized under limited label conditions, achieving AUC=0.879±0.012 and AP=0.887±0.015 on the Gene4Denovo test set, significantly better than existing tools.

[0209] (2) In terms of practicality, 100% variant coverage is achieved, solving the prediction missing problem caused by the scope limitation of existing tools, supporting complete workflow for large-scale genomic analysis.

[0210] (3) In terms of clinical value, through SHAP explainability analysis, the model prediction is mapped to specific ACMG evidence items, achieving the transformation from "black box prediction" to "traceable clinical interpretation", providing verifiable methodological support for molecular diagnosis.

[0211] The technical solution in the embodiment can automatically generate a structured explanatory report, unlike the mode of traditional tools that only output a single score, but is displayed through a hierarchical structure. The first level shows the comprehensive prediction conclusion, such as a pathogenicity probability of 0.98, and a determination of pathogenicity; the second level shows the core evidence chain, listing the high-contribution features supporting the determination and the corresponding ACMG evidence items in natural language, such as the main basis: PVS1 (loss of function), PM2 (low frequency in the population), and PP3 (bioinformatics prediction of harmfulness); and the third level shows the feature radar chart, which intuitively compares the similarity of the target variant and known pathogenic variants in each dimension. The embodiment not only solves the problem that the black box model cannot be trusted by clinicians, but also directly outputs the auxiliary decision information that conforms to the habits of doctors in diagnosis, greatly reducing the workload of manual interpretation of VUS (variants of uncertain significance), and realizing end-to-end intelligent processing from raw sequencing data to clinical auxiliary diagnosis.

[0212] Preferably, in step S1, a numerical feature vector containing evolutionary conservation, functional prediction score and quantitative clinical evidence is generated, including steps S11-S13 as follows:

[0213] S11, detecting the missing rate of each feature in the new variant site data, and removing samples with a missing rate exceeding a first preset threshold and features with a missing rate exceeding a second preset threshold;

[0214] S12, performing K-neighbor interpolation on the remaining missing values after removal, and calculating the filling values according to the Euclidean distance weighting of the samples in the feature space to form a complete original feature matrix;

[0215] S13, performing Min-Max normalization processing on the original feature matrix, and splicing the processed features in each dimension according to a predetermined channel order to form the numerical feature vector.

[0216] The process of generating the numerical feature vector in step S1 is not simply data splicing, but contains data quality control and completion mechanisms, which can solve the problems of sparsity and heterogeneity commonly existing in genomics data. By performing step S11, the system starts the missing value cleaning program and performs bidirectional scanning on the new variant site data matrix to be analyzed. Due to the complex source of new variants, some sites may lack specific functional annotations due to insufficient sequencing depth or being located in the non-coding region of the genome, such as lacking conservation scores for some species. The system presets filtering standards, sets the first preset threshold to 50%, that is, removes low-quality sample rows with a feature missing rate exceeding 50%; and sets the second preset threshold to 20%, that is, removes low signal-to-noise ratio feature columns with a missing rate exceeding 20% in all samples. This can effectively remove invalid noise from the data and retain a core feature set with statistical significance, thereby preventing a large number of null values from interfering with the convergence process of the subsequent model.

[0217] For the small amount of discrete missing values still existing in the feature matrix reserved after filtering, the K-Nearest Neighbor (KNN) imputation operation described in step S12 is performed. Unlike traditional mean filling or zero filling, the embodiment adopts an imputation algorithm based on the topological structure of the feature space. Specifically, the system calculates the Euclidean distance of the sample to be imputed and other complete samples in the feature space, and selects the K nearest neighbor samples, K is preferably set to an integer between 10 and 20. The system assigns weights according to the reciprocal of the Euclidean distance, and the closer the neighbor, the greater the contribution to the imputed value. The imputed value at the missing place is calculated by weighted average. Through the principle of local similarity of biological data, i.e. functionally similar variants often have similar feature distribution, the real distribution pattern of the data is restored to the greatest extent.

[0218] In step S13, the complete original feature matrix is standardized and vectorized. Due to the evolutionary conservation score, the score range is -10 to 10; the functional prediction score, the score range is 0 to 1; and the quantitative clinical evidence, the score is discrete integer coding. The above three scores have completely different dimensions and physical meanings, and direct input into the neural network will cause gradient update imbalance. Therefore, the system independently performs Min-Max normalization processing on each dimension feature of the matrix, and linearly maps all feature values to the closed interval [0, 1]. After processing, the system splices the evolutionary features, functional features and clinical evidence features according to the predetermined channel order strategy in a fixed order to construct a numerically structured feature vector.

[0219] Preferably, in step S1, the new variant site data to be analyzed is obtained, and a hierarchical label system is constructed, including steps S14-S16 as follows:

[0220] S14, retrieving the ClinVar database, directly marking the variant sites with non-conflicting multi-submitter records or expert group review records as first-level pathogenic or benign labels;

[0221] S15, retrieving the Gene4Denovo database, marking the variants with all control phenotypes as benign, and marking the variants with all pathogenic phenotypes and meeting the frequency threshold as pathogenic as the second-level labels;

[0222] S16, merging the first-level pathogenic or benign labels and the second-level labels to form a labeled sample set, and dividing the remaining variants without labels into an unlabeled sample set.

[0223] To solve the problem of the lack of labeled data and the uneven quality of the data in the field of novel variants, step S1 adopts a hierarchical label system construction strategy. Step S14 is performed to establish high-confidence first-level pathogenic or benign labels. The system retrieves the authoritative ClinVar database through an API interface or local mirroring, which collects variant interpretation records submitted by clinical laboratories around the world. To prevent false labeling from introducing model noise, the system sets a filtering method to only select sites with multiple submitter records without conflicts or expert group review records. The system identifies entries with a star rating of 2 stars or above in the ClinVar rating system, and if their clinical significance is clearly determined as pathogenic / possibly pathogenic or benign / possibly benign, they are directly labeled as first-level pathogenic or benign label samples. This ensures that the anchor data used as the basis for model training has high clinical effectiveness and gold standard properties, reducing the negative impact of label noise on supervised learning branches.

[0224] Given the limited number of high-confidence variants in ClinVar, which is not sufficient to support the sufficient training of deep neural networks, the system performs step S15 to expand the second-level labels based on phenotypic evidence through the Gene4Denovo database. Gene4Denovo focuses on collecting novel variant data, including a large number of sporadic cases and control samples. The system infers sites not included in ClinVar. For variants that only appear in healthy control groups and never appear in disease groups, they are labeled as benign. For those variants that only appear in disease groups, the system introduces a frequency threshold to constrain them. For example, the variant must appear at least 2 times or more independently in the disease group to be labeled as pathogenic. Through statistical rules and phenotype correlation inference, the second-level labels with a larger scale but slightly lower confidence than the first-level pathogenic or benign labels are constructed, alleviating the training bottleneck caused by the long-tail distribution of data.

[0225] Step S16 is performed to integrate and divide the data set. The system physically combines the first-level pathogenic or benign label samples and the second-level label samples, performs a deduplication operation to form a labeled set, and the labeled set is used for subsequent step S3 to calculate the cross-entropy loss of the supervised classification head. At the same time, for the remaining variation sites in the database that do not meet the ClinVar high-star standard or the Gene4Denovo phenotype inference rule, for example, conflicting interpretation results, insufficient evidence, or single occurrence of rare variations, the system divides them into an unlabeled set. The present application does not discard this part of seemingly useless data, but uses it as input for self-supervised reconstruction and contrastive learning, and uses its implicit genomic distribution characteristics to assist the model in learning the underlying structure of the variation, thereby achieving effective use of the vast amount of unknown variation data.

[0226] Preferably, in step S2, a feature extractor is constructed by the numerical feature vector, including steps S21-S23 as follows:

[0227] S21, configure the one-dimensional convolution layer, perform a sliding convolution operation on the numerical feature vector according to a small-size convolution kernel, and generate a local feature map reflecting the correlation of adjacent features;

[0228] S22, configure the atrous convolution layer, cascade a plurality of atrous convolution units after the local feature map, set an expansion rate parameter that grows exponentially to expand the receptive field, and generate a global dependence feature map;

[0229] S23, flatten the global dependence feature map and map it through a fully connected layer to generate a low-dimensional latent feature vector as a shared input of the supervised classification head, the self-supervised reconstruction head, and the contrastive learning head.

[0230] The feature extractor described in step S2 is the core of connecting the original data to the downstream multi-task head, and its construction process follows the deep learning paradigm of abstracting from local microscopic features to global macroscopic features layer by layer. Perform step S21, configure a one-dimensional convolution layer (1D-CNN) to extract local correlations. Considering that the input data is a spliced heterogeneous numerical feature vector, its adjacent elements may represent attributes with biological intrinsic relationship, for example, several consecutive dimensions represent different evolutionary conservation scores of the same site. The system sets the convolution kernel size to a small value (such as 3 or 5), and the stride to 1, and performs a sliding convolution operation on the feature vector. Through the processing of the nonlinear activation function, this one-dimensional convolution layer can sensitively capture the short-range coupling pattern between adjacent features, generate a first-level feature map reflecting the local feature correlation, and realize the search for specific short sequence motifs on the genomic sequence.

[0231] In order to overcome the problem of limited receptive field of traditional convolution layer, perform step S22, cascade a dilated convolution layer (Dilated Convolution Layer) after the local feature map. The dilated convolution layer is composed of multiple series of residual blocks (Residual Blocks), each block uses convolution units with different dilation rates (Dilation Rate) inside. The system sets the dilation rate parameter to an exponential growth sequence, so that the network can expand the receptive field at an exponential rate without pooling operation (Pooling) and thus without losing resolution. The judgment of pathogenicity often depends on the nonlinear interaction between elements far apart in the feature vector, for example, the long-range dependence between sequence features at the head of the vector and clinical evidence quantification features at the tail of the vector. The model can generate a global dependence feature map, effectively integrating cross-dimensional information flow.

[0232] Perform step S23 to complete the mapping from high-dimensional feature map to low-dimensional latent space. Since the above global dependence feature map still retains the spatial structure (channel number x feature length), it is not suitable for direct use in classification or contrast calculation, so a flattening operation or global average pooling is performed to convert it into a one-dimensional vector. Subsequently, a fully connected layer is used for linear projection, which compresses high-dimensional information into a compact low-dimensional latent feature vector, for example, with a dimension of 128 or 256. This latent feature vector not only removes redundant information, but also serves as a shared representation and is input to the subsequent parallel supervised classification head, self-supervised reconstruction head and contrast learning head. This ensures that the model can be simultaneously constrained by classification discrimination, data reconstruction ability and sample discrimination when updating parameters through backpropagation, thereby learning more robust and essential pathogenicity feature representations.

[0233] Preferably, in step S3, the total loss based on weighted sum jointly updates the network parameters of the feature extractor and each head, including steps S31-S33 as follows:

[0234] S31, calculating the classification loss of the labeled samples in the supervised classification head according to the Focal Loss function, and reducing the weight of simple negative samples by a focusing parameter;

[0235] S32, calculating the reconstruction loss of all samples in the self-supervised reconstruction head according to the mean square error function, and calculating the contrastive loss of the enhanced view of all samples in the contrastive learning head according to the normalized temperature scale cross-entropy function;

[0236] S33, adding the classification loss, the reconstruction loss and the contrastive loss according to a preset hyperparameter weight to obtain a target function, and synchronously updating the network parameters along the gradient direction of the target function according to an optimizer.

[0237] The joint updating process in step S3 is the key to the model acquiring high generalization ability, and a composite target function capable of balancing the supervised signal and the self-supervised signal can be constructed. In step S31, the supervised classification head discards the traditional cross-entropy loss function and instead uses the Focal Loss function to address the extreme class imbalance problem of the number of benign variants (negative samples) far exceeding that of pathogenic variants (positive samples) in the new variant data. For example, set to 2.0) and a balance parameter For example, set to 0.25). In the calculation process, the Focal Loss function reduces the weight of simple negative samples (i.e., benign variants with very high prediction probability) that are easy to classify, forcing the model to focus the center of gravity of gradient update on difficult samples (i.e., pathogenic variants at the classification boundary) that are difficult to distinguish, which can prevent the model from falling into a local optimum due to the dominance of a large number of simple negative samples, thereby improving the recall rate for rare pathogenic variants.

[0238] Meanwhile, in order to fully tap the potential value of the unlabeled data, step S32 is performed. The system calculates the loss of the two auxiliary tasks through all the input samples (including labeled data and unlabeled data). On the one hand, the self-supervised reconstruction head restores the latent features to the original input through the decoder, and calculates the difference between the reconstruction output and the original input by using the mean square error function (MSE Loss), i.e. the reconstruction loss. This loss constrains the feature extractor to retain the relevant information of the input data, and plays a role in regularization and noise reduction. On the other hand, the contrastive learning head generates two enhanced views for each sample by using a data enhancement strategy, and calculates the contrastive loss based on the normalized temperature scale cross-entropy function, where the data enhancement strategy is, for example, random masking or Gaussian noise injection. This loss can reduce the distance between different views of the same sample in the feature space, while increasing the distance between different samples, so that the model learns the structural features of the samples that are essential and have rotation and translation invariance.

[0239] Step S33 is performed to perform multi-task joint backpropagation. The system constructs a total objective function according to a preset hyperparameter strategy, which is expressed by the formula:

[0240]

[0241] wherein , and are the weight coefficients of the classification, reconstruction and contrastive losses, respectively. In the early stage of training, the system gives and larger weights to assist the model to initialize a good feature space through unsupervised learning in the absence of labels; as the number of training iterations increases, the weight of is gradually increased to optimize the discrimination performance. Based on the calculated weighted total loss, the system calculates the gradient using the AdamW or SGD optimizer, and synchronously updates the network parameters of the feature extractor (Backbone) and each task head (Heads) in the gradient descent direction. Through the end-to-end joint training mechanism, the model has both the discrimination of supervised learning and the generalization of self-supervised learning.

[0242] Preferably, as shown in Figure 6 , in step S4, the feature with a high contribution value is reversely mapped to the corresponding clinical evidence item, including the following steps S41-S43:

[0243] S41, based on the principle of cooperative game theory, the contribution value of each feature dimension in the numerical feature vector is quantified by calculating the marginal change of the model output probability before and after replacing the feature value;

[0244] S42, ranking the features according to the absolute value of the contribution value, and screening the top N key features with the largest contribution to the pathogenicity probability;

[0245] S43, searching the original description corresponding to the key features according to the feature index table, and outputting the evidence code and pathogenicity support level corresponding to the key features if the key features belong to the quantified clinical evidence features.

[0246] In order to break the black box barrier of deep learning model in clinical application, step S4 introduces a post-explanation mechanism based on cooperative game theory. Step S41 is performed, the system calls the interpreter module (such as integrated SHAP algorithm kernel) to perform attribution analysis on each specific prediction result. Each numerical feature (such as evolutionary conservation score, specific clinical evidence code) is regarded as a participant in the game, and the pathogenicity probability output by the model is regarded as the total benefit. The system estimates the contribution value of each feature dimension by calculating the marginal change of the model output probability under different feature combinations. Specifically, the difference in prediction probability when a certain feature exists or does not exist (or is replaced by a benchmark value) is simulated, and the independent contribution of the feature to push the result to pathogenic or benign is quantified by weighted average of all possible feature permutation combinations, thereby converting the complex nonlinear neural network weight into a feature importance score with linear additivity.

[0247] Step S42 is performed to screen and grade the feature contribution. Since the input feature vector dimension is high, usually containing hundreds of features, directly showing all dimensions of contribution value to the user will cause information overload. Therefore, the system ranks all features in descending order according to the absolute value of the calculated contribution value. The larger the absolute value, the stronger the influence of the feature on the final decision of the model. Among them, positive values represent pushing towards pathogenic direction, and negative values represent pushing towards benign direction. The system sets a screening threshold N to automatically intercept the top N key feature set, similar to the differential diagnosis thinking of clinical experts, that is, from the complex biological indicators, the most core pathogenic factors leading to disease occurrence are locked, and the interference of background noise and redundant information is eliminated.

[0248] Step S43 is performed to realize reverse mapping from mathematical features to clinical semantics. The system is preset with a feature index table which records in detail the original biological meaning and data source corresponding to each dimension index in the feature vector. For the key features screened out, the system queries the index table. If the key feature is identified as a quantitative clinical evidence feature, for example, derived from the coding of ACMG guideline items in S1, the system will trigger the decoding logic to extract the corresponding standard evidence code (such as “PVS1”, “PM2”) and its pathogenicity support level (such as “very strong”, “moderate”). The system dynamically fills these structured information into the explanatory report template to generate natural language descriptions such as “main pathogenic basis: PVS1 (detecting loss-of-function variant, contribution +0.45)”. Not only does it inform the doctor of the prediction result, but it also clearly explains the evidence traceability, thereby generating a complete diagnostic report in line with the clinical guideline specifications.

[0249] In one embodiment, Figure 2 is a semi-supervised hybrid convolutional network pathogenicity prediction device block diagram for novel variants according to an exemplary embodiment. As shown in Figure 2 The semi-supervised hybrid convolutional network pathogenicity prediction device for novel variants includes a generation module 71, a construction module 72, a calculation module 73, and a prediction module 74.

[0250] The generation module 71 is configured to obtain novel variant site data to be analyzed, perform multi-source functional annotation and clinical evidence quantification, and generate a numerical feature vector containing evolutionary conservation, functional prediction score, and quantitative clinical evidence.

[0251] The construction module 72 is configured to construct a feature extractor from the numerical feature vector, wherein the feature extractor is connected with a parallel supervised classification head, a self-supervised reconstruction head, and a contrastive learning head, and the feature extractor is internally connected with a one-dimensional convolutional layer for extracting local features and a dilated convolutional layer for extracting global dependency features.

[0252] The calculation module 73 is configured to input a training data set containing labeled samples and unlabeled samples into the feature extractor, calculate the classification loss of the supervised classification head, the reconstruction loss of the self-supervised reconstruction head, and the contrastive loss of the contrastive learning head, and jointly update the network parameters of the feature extractor and each head based on the weighted sum of the total loss.

[0253] The prediction module 74 is configured to apply the updated feature extractor and the supervised classification head to the target variant to obtain a pathogenicity probability, and calculate a contribution value of each feature in the numerical feature vector to the pathogenicity probability, and map the features with high contribution values to corresponding clinical evidence items to generate an explanatory report including a pathogenicity level and evidence traceability.

[0254] The generation module 71, the construction module 72, the calculation module 73, and the prediction module 74 included in the semi-supervised hybrid convolutional network pathogenicity prediction device for novel variants are controlled to perform the pathogenicity prediction method for novel variants described in any of the above embodiments.

[0255] The embodiment also provides a pathogenicity prediction system for novel variants, which comprises a processor and a computer readable storage medium connected to each other, and the computer readable storage medium stores a computer program which is executed by the processor to implement the steps of the pathogenicity prediction method for novel variants described above.

[0256] The embodiment also provides a computer readable storage medium, which stores a computer program which is executed by a processor to implement the steps of the pathogenicity prediction method for novel variants described above.

[0257] The embodiment also provides a computer program product, which comprises a computer program which is executed by a processor to implement the steps of the pathogenicity prediction method for novel variants described above.

[0258] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer readable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer usable program code. The present application is described with reference to flowcharts and / or block diagrams illustrating the architecture, the devices (systems), and the computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 Each flow or multiple flows and / or blocks Figure 1apparatus for performing the functions specified in the block or blocks. Such computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 flow or flows and / or blocks Figure 1 the functions specified in the block or blocks. Such computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 flow or flows and / or blocks Figure 1 the functions specified in the block or blocks.

Claims

1. A method for predicting the pathogenicity of emerging variants, characterized in that, include: S1. Obtain the data of newly discovered variant sites to be analyzed, perform multi-source functional annotation and quantify clinical evidence, and generate a numerical feature vector containing evolutionary conservation, functional prediction score and quantified clinical evidence. S2, a feature extractor is constructed using the numerical feature vector. The feature extractor is connected to a parallel supervised classification head, a self-supervised reconstruction head, and a contrastive learning head. The feature extractor also contains a one-dimensional convolutional layer for extracting local features and a dilated convolutional layer for extracting global dependent features. S3, input the training dataset containing labeled samples and unlabeled samples into the feature extractor, calculate the classification loss of the supervised classification head, the reconstruction loss of the self-supervised reconstruction head and the contrastive loss of the contrastive learning head respectively, and jointly update the network parameters of the feature extractor and each head based on the weighted summation of the total loss; S4, the feature extractor and the supervised classification head with updated application parameters are used to predict the target variant, obtain the pathogenicity probability, and calculate the contribution value of each feature in the numerical feature vector to the pathogenicity probability. Features with high contribution values ​​are back-mapped to the corresponding clinical evidence items to generate an explanatory report containing pathogenicity level and evidence source. In step S1, acquiring the newly discovered variant site data to be analyzed and constructing a hierarchical labeling system includes: Search the ClinVar database and directly label variant sites with conflict-free multi-submitter records or expert panel review records as Level 1 pathogenic or benign. The Gene4Denovo database was searched, and variants with all control phenotypes were marked as benign, while variants with all pathogenic phenotypes and meeting the frequency threshold were marked as pathogenic, serving as secondary labels. The first-level pathogenic or benign labels are combined with the second-level labels to form a labeled sample set, and the remaining unlabeled variant sites are divided into an unlabeled sample set. In step S4, the step of back-mapping high-contribution features to corresponding clinical evidence items includes: S41, based on the principle of cooperative game theory, the contribution value of each feature dimension in the numerical feature vector is quantified by calculating the marginal change in the model output probability before and after the feature value is replaced. S42, Sort the features according to the absolute value of the contribution value, and select the top N key features that contribute the most to the pathogenicity probability. S43, find the original description corresponding to the key feature according to the feature index table. If the key feature belongs to the quantitative clinical evidence feature, output its corresponding evidence code and pathogenicity support level.

2. The method of pathogenicity prediction for novel variants of claim 1, wherein, In step S1, generating a numerical feature vector containing evolutionary conservation, functional prediction scores, and quantified clinical evidence includes: S11, detect the missing rate of each feature in the newly emerged variant site data, and remove samples with a missing rate exceeding a first preset threshold and feature dimensions with a missing rate exceeding a second preset threshold. S12, perform K-nearest neighbor imputation on the remaining missing values ​​after removal, calculate the imputation value based on the weighted Euclidean distance of the sample in the feature space, and form a complete original feature matrix; S13, perform Min-Max normalization on the original feature matrix, and then concatenate the processed features of each dimension in a predetermined channel order to form the numerical feature vector.

3. The method of pathogenicity prediction for novel variants of claim 1, wherein, In step S2, a feature extractor is constructed using the numerical feature vector, including: S21, Configure the one-dimensional convolutional layer to perform a sliding convolution operation on the numerical feature vector according to the small-sized convolutional kernel, and generate a local feature map that reflects the relationship between adjacent features; S22, Configure the dilated convolutional layer, cascade multiple dilated convolutional units after the local feature map, set an exponentially increasing dilation rate parameter to expand the receptive field, and generate a globally dependent feature map; S23, the global dependency feature map is flattened and mapped through a fully connected layer to generate a low-dimensional latent feature vector, which serves as the shared input for the supervised classification head, the self-supervised reconstruction head, and the contrastive learning head.

4. The method of pathogenicity prediction for novel variants of claim 1, wherein, In step S3, the joint update of the feature extractor and the network parameters of each head based on the weighted summation total loss includes: S31, calculate the classification loss of labeled samples in the supervised classification head according to the Focal Loss function, and reduce the weight of simple negative samples by focusing parameters; S32, calculate the reconstruction loss of all samples in the self-supervised reconstruction head according to the mean square error function, and calculate the contrast loss of all sample augmented views in the contrast learning head according to the normalized temperature scale cross-entropy function. S33, the classification loss, reconstruction loss and contrast loss are added together according to the preset hyperparameter weights to obtain the objective function, and the network parameters are updated synchronously along the gradient direction of the objective function according to the optimizer.

5. A novel variant-oriented semi-supervised hybrid convolutional network pathogenicity prediction apparatus, characterized by, include: The generation module is used to acquire new variant site data to be analyzed, perform multi-source functional annotation and quantify clinical evidence, and generate a numerical feature vector containing evolutionary conservation, functional prediction score and quantified clinical evidence. The module is used to construct a feature extractor through the numerical feature vector. The feature extractor is connected to a parallel supervised classification head, a self-supervised reconstruction head, and a contrastive learning head. The feature extractor is internally connected to a one-dimensional convolutional layer for extracting local features and a dilated convolutional layer for extracting global dependent features. The computation module is used to input the training dataset containing labeled samples and unlabeled samples into the feature extractor, calculate the classification loss of the supervised classification head, the reconstruction loss of the self-supervised reconstruction head, and the contrastive loss of the contrastive learning head, and jointly update the network parameters of the feature extractor and each head based on the weighted summation of the total loss. The prediction module is used to predict the target variant by applying the updated feature extractor and the supervised classification head to obtain the pathogenicity probability, and to calculate the contribution value of each feature in the numerical feature vector to the pathogenicity probability. The features with high contribution values ​​are back-mapped to the corresponding clinical evidence items to generate an explanatory report containing pathogenicity level and evidence source. The generation module, the construction module, the calculation module, and the prediction module are controlled to execute the pathogenicity prediction method for emerging variants as described in any one of claims 1 to 4.

6. An electronic device, characterized in that, include: Communication interface, processor, memory; The memory is used to store program instructions, which, when executed by the processor connected to the memory via the communication interface, cause the electronic device to implement the pathogenicity prediction method for emerging variants as described in any one of claims 1 to 4.

7. A computer-readable storage medium having program instructions stored thereon, characterized in that, When the program instructions are executed by a computer, the computer enables the computer to implement the pathogenicity prediction method for any of claims 1 to 4.

Images