New variation-oriented pathogenicity prediction method

Abstract

The invention discloses a new mutation-oriented pathogenicity prediction method, which comprises the following steps of: acquiring new mutation site data to be analyzed, and generating a numerical feature vector containing evolution conservative property, function prediction score and quantitative clinical evidence; constructing a feature extractor through the numerical feature vector; inputting a training data set containing a label sample and a label-free sample into the feature extractor, and updating the network parameters of the feature extractor and each header; and predicting target variation by using the feature extractor after parameter updating and the supervised classification head to obtain a pathogenicity probability, and calculating a contribution value of each feature in the numerical feature vector to the pathogenicity probability. According to the technical scheme provided by the invention, the auxiliary decision-making information conforming to the diagnosis habit of a doctor is output, the workload of manually interpreting the variation with unclear meaning is greatly reduced, and the end-to-end intelligent processing from the original sequencing data to the clinical auxiliary diagnosis is realized.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics data processing, and more particularly to a method for predicting the pathogenicity of emerging variants. Background Technology

[0002] De novo mutations (DNMs) are DNA sequence alterations that first appear in an individual's genome during the formation of parental germ cells or early embryonic development. These mutations are not present in parental somatic cells, thus exhibiting high randomness and uniqueness. With the widespread adoption of genome sequencing technology, research has confirmed that de novo mutations play a crucial pathogenic role in the development of various human genetic diseases and complex illnesses, particularly in the field of early-onset neurological disorders, such as autism spectrum disorder, intellectual disability, early-onset Parkinson's disease, and early-onset Alzheimer's disease. Because de novo mutations often disrupt the function of key genes, accurate identification and interpretation of these mutations are of significant clinical importance for early disease diagnosis, genetic counseling, and precision medicine.

[0003] However, existing predictive technologies for the pathogenicity of gene variants suffer from technical bottlenecks when applied to emerging variant scenarios, including poor generalization performance, low prediction accuracy, and insufficient coverage. Mainstream prediction methods such as CADD and DANN generally employ a purely supervised learning paradigm, which heavily relies on a large amount of high-quality, evenly distributed labeled data for model training. However, in the context of emerging variant research, data exhibits inherent label scarcity and extreme class imbalance, meaning the number of known pathogenic variants is far less than that of benign or unknown variants. With insufficient training data and a severely imbalanced distribution, existing models cannot fully learn the complex, non-linear discriminative boundary between pathogenic and benign variants, often strongly favoring benign patterns. This leads to a sharp decline in the sensitivity to rare pathogenic variants, resulting in numerous false negatives. Furthermore, for variants with severely missing features or not included in knowledge bases, especially a large number of variants with unclear meaning, existing tools, lacking the ability to infer from the underlying data structure, often rely excessively on explicit, complete features, resulting in ineffective predictions and severely limiting their coverage and automation efficiency in real-world clinical testing. On the other hand, existing technologies suffer from serious problems of black-box decision-making and disconnect from clinical standards in clinical translational applications. Although some deep learning methods have improved numerical performance, their internal decision-making processes are highly opaque, only outputting general scores or disease severity levels, failing to provide feature-level contribution analysis. This lack of interpretability makes it difficult for clinical experts to trace the biological evidence behind the model's conclusions, hindering the establishment of trust in high-risk decisions. More critically, the abstract scoring systems of traditional functional prediction tools are disconnected from the structured evidence items in authoritative clinical interpretation guidelines such as ACMG / AMP, lacking a direct quantitative mapping relationship. This results in model predictions not being seamlessly embedded into standardized clinical diagnostic reports, still requiring cumbersome manual review and secondary interpretation, making it difficult to integrate advanced algorithms as truly reliable decision support tools into clinical workflows. How to properly address these issues has become an urgent problem for the industry. Summary of the Invention

[0004] This invention addresses the aforementioned problems of existing technologies by providing a method for predicting the pathogenicity of emerging variants. It utilizes a multi-task joint training framework combining supervised, self-supervised, and contrastive learning to effectively leverage massive amounts of unlabeled data. This overcomes the generalization performance bottleneck caused by the scarcity of emerging variant data and class imbalance, thereby improving the predictive sensitivity for rare pathogenic variants. Furthermore, by fusing local and global feature extraction and establishing a reverse mapping mechanism from feature contribution to clinical evidence items, it breaks down the black-box barrier of traditional algorithms, generating interpretable reports that conform to clinical guidelines and achieving high-precision prediction of the pathogenicity of emerging variants.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for predicting the pathogenicity of emerging variants is provided, including: 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, after the parameter update is completed, 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. The features with high contribution values ​​are back-mapped to the corresponding clinical evidence entries to generate an explanatory report containing pathogenicity level and evidence source.

[0006] Furthermore, in step S1, a numerical feature vector is generated that includes evolutionary conservation, functional prediction score, and quantified clinical evidence, including: 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 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.

[0007] Furthermore, in step S1, acquiring the newly discovered variant site data to be analyzed and constructing a hierarchical labeling system includes: S14. 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 labels. S15, Search the Gene4Denovo database, mark variants with all control phenotypes as benign, and mark variants with all pathogenic phenotypes that meet the frequency threshold as pathogenic, as the second-level labels; S16, merge the first-level pathogenic or benign labels with the second-level labels to form a labeled sample set, and divide the remaining unlabeled variant sites into an unlabeled sample set.

[0008] Furthermore, in step S2, constructing a feature extractor using the numerical feature vector includes: 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.

[0009] Furthermore, 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.

[0010] Furthermore, 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.

[0011] According to a second aspect of the present invention, a semi-supervised hybrid convolutional network pathogenicity prediction device for novel variants is provided, comprising: 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, obtain the pathogenicity probability, calculate the contribution value of each feature in the numerical feature vector to the pathogenicity probability, and back-map the features with high contribution values ​​to the corresponding clinical evidence entries to generate an explanatory report containing pathogenicity level and evidence source.

[0012] The present invention also provides a pathogenicity prediction system for emerging variants, comprising a processor and a computer-readable storage medium interconnected thereto, wherein the computer-readable storage medium stores a computer program, which is executed by the processor to implement the steps of the above-described pathogenicity prediction method for emerging variants.

[0013] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for predicting the pathogenicity of emerging variants.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for predicting pathogenicity of emerging variants.

[0015] Compared with the prior art, the advantages of the present invention are as follows: This invention utilizes a multi-task joint training framework that combines supervised, self-supervised, and contrastive learning to effectively leverage massive amounts of unlabeled data. This overcomes the generalization performance bottleneck caused by the scarcity of emerging variant data and class imbalance, thereby improving the predictive sensitivity for rare pathogenic variants. Simultaneously, by fusing local and global feature extraction and establishing a reverse mapping mechanism from feature contribution to clinical evidence items, it breaks down the black-box barrier of traditional algorithms, generating interpretable reports that conform to clinical guidelines and achieving high-precision prediction of the pathogenicity of emerging variants. Attached Figure Description

[0016] Figure 1 A schematic diagram of the method flow provided for an embodiment of the present invention; Figure 2 A structural diagram of a semi-supervised hybrid convolutional network pathogenicity prediction device for emerging variants is provided for embodiments of the present invention; Figure 3 A schematic diagram of coding region variation and feature selection results provided for embodiments of the present invention; Figure 4 A schematic diagram of the label determination logic provided for an embodiment of the present invention; Figure 5 A schematic diagram of the training phase operation flow provided for an embodiment of the present invention; Figure 6 A schematic diagram illustrating the operational logic of the supervised classification module provided in an embodiment of the present invention; Figure 7 A schematic diagram illustrating the operational logic of the reconstruction module provided in an embodiment of the present invention; Figure 8 A schematic diagram illustrating the operational logic of the comparison learning module provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the operation flow of the inference stage provided for an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0018] The technical solution adopted in this embodiment is as follows: Figure 1 As shown, it includes the following steps: 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, after the parameter update is completed, 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. The features with high contribution values ​​are back-mapped to the corresponding clinical evidence entries to generate an explanatory report containing pathogenicity level and evidence source.

[0019] In one embodiment, a deep semi-supervised learning and multi-task joint optimization framework can address the problems of low prediction accuracy and poor clinical interpretability in existing technologies when genomic data is scarce. The method in this embodiment can run on a bioinformatics analysis server configured with a high-performance graphics processor.

[0020] Step S1 constructs a high-dimensional feature space that combines biological significance with computational operability. The system receives newly discovered variant site data in VCF format to be analyzed and performs multi-source functional annotation on the sites using tools such as Annovar or VEP. The annotation content covers population frequency databases, evolutionary conservation scores, protein structure predictions, and biochemical property predictions. The clinical evidence quantification module pre-configures mapping rules between standard evidence items and numerical vectors from the ACMG / AMP guidelines. For example, for sites that satisfy the criteria of loss of gene function due to truncation variations, the system not only records their sequence changes but also encodes the confidence level of the corresponding PVS1 item as a numerical feature. Through normalization and missing value imputation, the above heterogeneous information is integrated into a standardized numerical feature vector, which retains the underlying sequence physical features while embedding high-level clinical prior knowledge.

[0021] Step S2 is executed to construct a deep neural network model with a specific topology. The main body of this model consists of a feature extractor, which employs a cascaded architecture. The front end is a one-dimensional convolutional layer (1D-CNN) with a small convolutional kernel and a ReLU activation function, specifically designed to capture local mutation patterns in gene sequences and short-range association features between adjacent sites. The back end connects to dilated convolutional layers, which, by setting the dilation rate, exponentially expand the receptive field without increasing the number of parameters, thereby effectively extracting global dependencies between distant regulatory elements in the genome sequence. The output of the feature extractor is connected in parallel to three functionally heterogeneous task heads: a supervised classification head, a self-supervised reconstruction head, and a contrastive learning head. The supervised classification head, composed of fully connected layers and a Softmax activation function, outputs the probability distribution of pathogenicity / benignity. The self-supervised reconstruction head uses a decoder structure 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 non-linear projection layer to map features to a hyperspherical space to calculate the similarity distance between samples.

[0022] Step S3, the joint training phase based on a multi-task loss function, addresses the issue of a small number of labeled samples and a large number of unlabeled samples in newly emerging variant data by employing a hybrid data stream input strategy. The training set is input into the network, consisting of a small number of clinically confirmed pathogenic / benign samples (labeled) and a large number of unclassified population variant samples (unlabeled). During training, the loss function is composed of three weighted parts: First, the cross-entropy loss of the supervised classification head is calculated only using labeled samples, guiding the model to establish a clear discrimination boundary. Second, regardless of whether the samples are labeled, the mean squared error loss (MSE loss) of the self-supervised reconstruction head is calculated using all samples, forcing the feature extractor to capture the underlying distribution patterns and noise-resistant features of the data itself through the reconstruction task. Third, the contrastive loss of the contrastive learning head (such as NT-Xent Loss) is used to narrow the distance between similar variants in the feature space and widen the distance between dissimilar variants. The system dynamically adjusts the weight coefficients; for example, it assigns higher weights to reconstruction loss and contrastive loss in the early stages of training, and emphasizes classification loss in the later stages. The backpropagation algorithm is used to jointly update the network parameters of the feature extractor and each head. This allows the model to learn robust feature representations from unlabeled data even when labels are extremely scarce.

[0023] Step S4, the predictive inference and interpretability report generation stage, is executed. The trained model is used for forward inference on the target novel variant, and the supervised classification head outputs the probability value that the variant is pathogenic, likely pathogenic, or benign. Simultaneously, the system initiates the interpretability analysis module, calculating the contribution value (Shapley Value) of each dimension in the input feature vector to the final predicted probability using the SHAP (Shapley Additive Explanations) algorithm or Integrated Gradients. The system automatically selects the top K key features with the highest contribution values ​​and triggers a back-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: Calculated prediction supports that this variant has a harmful effect and that this site is highly conserved."

[0024] In a preferred embodiment, this invention proposes a semi-supervised hybrid convolutional neural network (SH-CNN) system for predicting the pathogenicity of novel variants in coding regions. This system integrates multi-source functional annotation features with clinical evidence features, combined with an innovative network architecture and a multi-task learning mechanism, to achieve high-accuracy and high-coverage pathogenicity prediction even under label-scarce conditions.

[0025] The specific process is as follows: data preprocessing → label construction → feature system construction → semi-supervised hybrid convolutional network training → interpretable clinical output → prediction result output.

[0026] 1. Data Preprocessing 1.1 Dataset The dataset was obtained from the Gene4Denovo database, which integrates 741,866 newly discovered variants (DNMs). During data preprocessing, entries lacking annotation information and duplicate variants from different literatures were removed, leaving a total of 70,415 coding region variants.

[0027] 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.

[0028] 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.

[0029] 1.2 Quality Control Process 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: (1) Mutation level filtering: Remove variant samples with a feature missing rate of more than 40%. (2) Feature-level filtering: Remove feature dimensions with a missing rate of more than 70% in variant samples. 1.3 Handling Missing Values The K-nearest neighbor imputation algorithm is used to fill in the remaining missing values: Nearest neighbor mutations are selected based on Euclidean distance.

[0030] The selection criterion is that variants with similar functions tend to cluster in the feature space, and KNN can effectively utilize local structural information to restore missing data.

[0031] (1) There is a strong correlation among pathogenicity-related features, and K-nearest neighbor imputation based on proximity variation can provide a more reliable estimate.

[0032] (2) K-nearest neighbor interpolation can better preserve the local structural information of the data and avoid the distribution shift that may be introduced by simple mean or median interpolation. The interpolation formula is as follows:

[0033] The explanations of the symbols in the above formulas are shown in Table 1.

[0034] Table 1

[0035] 1.4 Feature Standardization Processing Min-Max normalization is applied to all numerical features. A linear transformation of the original data maps the resulting values ​​to the range [0–1], eliminating dimensional differences between different features and improving model convergence stability. The transformation function is shown below:

[0036] The explanations of the symbols in the above formulas are shown in Table 2.

[0037] Table 2

[0038] 1.5 Data Partitioning Strategy Training set / validation set / test set ratio: 6:2:2.

[0039] Stratified sampling: Ensures that the label distribution is consistent across all datasets.

[0040] Oversampling: SMOTE oversampling technique is used for the minority class (pathogenic variants) in the training set to alleviate class imbalance.

[0041] 2. Construction of the tag system 2.1 Hierarchical Labeling Strategy A two-level annotation principle is adopted: "ClinVar high confidence level first, Gene4Denovo phenotype supplementation". Level 1: ClinVar high-confidence annotations (highest priority), selected based on ClinVar annotations of criteria_provided\multiple_submitters\no_conflicts (exactly multiple conflict-free submitters) or reviewed_by_expert_panel (reviewed by an expert panel). The pathogenic / benign annotations provided after screening are directly used as the final labels.

[0042] Level 2: Supplementary annotation of Gene4Denovo phenotype (phenotype represents the pathogenicity status of the variant), the label determination logic is as follows. Figure 4 As shown.

[0043] Benign label (label=0): Phenotype is all control (meaning not pathogenic).

[0044] Pathogenicity label (label=1): All phenotypes are pathogenic and reported in ≥2 independent articles or ≥2 study subjects.

[0045] Unknown label (label=-1): Mutations that do not meet the above conditions.

[0046] 2.2 Tag Conflict Resolution Mechanism Basic principle: ClinVar annotations have higher priority than Gene4Denovo.

[0047] Conflict detection: When the same variant has opposite annotations in ClinVar and Gene4Denovo.

[0048] Solution: Use ClinVar high-confidence annotations to ensure the authority of the labels.

[0049] Final label distribution: (1) Gene4Denovo dataset: Of the 70,415 coding region variants, there are 3,956 pathogenic labels (813 from ClinVar and 3,143 from Gene4Denovo), 2,965 benign labels (1,704 from ClinVar and 1,261 from Gene4Denovo), and 63,494 unknown labels.

[0050] (2) Spark independent test set: Among the 33,948 coding region variants, there were 517 pathogenic tags (264 from ClinVar and 253 from Gene4Denovo), 8,614 benign tags (260 from ClinVar and 8,354 from Gene4Denovo), and 24,817 unknown tags.

[0051] 3. Feature System Construction Candidate DNMs were first annotated with basic functions using ANNOVAR (version 2024.09.24), and then integrated with multiple functional prediction tools based on the VarCards2 framework. MutPred2 was used to quantitatively assess the potential disruptive effect of amino acid substitutions on protein function; ReVe combined the results of REVEL and VEST3 to improve the discrimination performance of missense variants; and SpliceAI was used to predict aberrant splicing events that variants might trigger. 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 structural constraints, functional impairment, and splicing regulation.

[0052] Clinical evidence features are derived from ClinVar and the automated ACMG / AMP scoring tool provided by GeneBe. Annotations in ClinVar are submitted manually by experts, while GeneBe is an online tool that automatically extracts the strength of evidence for various evidence clauses (such as PVS1, PM1, PP3, BS1, etc.) based on ACMG / AMP variant interpretation guidelines and outputs quantitative scores to reflect the degree of support for each variant in clinical interpretation. Incorporating these ACMG scores into the feature system effectively bridges deep learning models and clinical interpretation standards, enhancing the interpretability and clinical value of the prediction results.

[0053] To facilitate downstream modeling, all features are divided into six categories: (1) evolutionary conservation; (2) coding region function prediction; (3) cross-regional function prediction features; (4) splicing effect; (5) population allele frequency; and (6) clinical evidence score.

[0054] 3.1 Evolutionary conservation characteristics (12 dimensions) The evolutionary constraints of sequences across different species were measured using indicators such as GERP++, phyloP, phastCons, SiPhy, and LINSIGHT, reflecting the potential functional importance of variations.

[0055] 3.2 Functional prediction of coding region (52 dimensions) It integrates 52 mainstream pathogenicity prediction scores, covering the assessment of the biological impact of amino acid substitutions, sequence model prediction, and integrated scores (such as SIFT, PolyPhen-2, Mutation Taster, REVEL, MutPred2, Alpha Missense, EVE, etc.), to comprehensively evaluate the potential pathogenicity of variants.

[0056] 3.3 Cross-regional functional prediction features (25 dimensions) While some functional prediction tools (such as Eigen and GenoCanyon) were initially designed to assess the regulatory potential of non-coding region variants, their scoring models are based on genome-wide features and can also output results for coding region variants. This study incorporates these features as complementary characteristics into the model to reflect the potential impact of variants on transcriptional regulation, chromatin accessibility, and epigenetic modifications. These features in coding regions help capture regulatory signals beyond traditional protein function scoring, thus providing a more comprehensive characterization of the functional effects of variants.

[0057] 3.4 SpliceAI splicing effect (9 dimensions) SpliceAI predicts the potential splicing effect at each variant site, including four types of delta scores (DS_AG, DS_AL, DS_DG, DS_DL) and delta positions (DS_AG, DS_AL, DS_DG, DS_DL). (), used to describe the location of splicing events. Variants near chromosome ends or restricted to the reference sequence are considered deletion values.

[0058] 3.5 Population frequency of alleles (63 dimensions) Population allele frequency information from multiple authoritative databases, including gnomADv4.1 (exome and genome), ExAC, and ESP, was incorporated, totaling 63 frequency features covering the overall population, sex, ethnic group, and rare allele filtering frequencies (FAF).

[0059] 3.6 Clinical Evidence Scoring There are two main sources: one is pathogenicity assessments submitted and reviewed manually by experts in the ClinVar database; the other is an automated scoring tool based on the ACMG / AMP guidelines provided by GeneBe. The combination of the two ensures both the authority of expert evidence and provides a systematic and quantitative assignment of ACMG entries.

[0060] Determining PS2 / PM6 requires sufficient literature evidence. Currently, the Gene4Denovo database systematically includes supporting literature, while GeneBe cannot directly provide PS2 / PM6 evidence related to novel variants. Therefore, based on the recommendations of the ClinGen SVI working group, we integrated literature data from Gene4Denovo to construct a PS2 / PM6 scoring algorithm to support more accurate pathogenicity assessment: a scoring system was constructed based on parentage verification, phenotypic and genetic consistency, and the number of observed DNMs for each DNM, and converted into quantitative evidence strength according to ClinGen guidelines (high specificity phenotype and parentage verification: 2 points; phenotypic consistency: 1 point; high genetic heterogeneity: 0.5 points; phenotypic inconsistency: 0 points; no parentage verification: the corresponding score is halved). Based on the total score, it was further mapped to an evidence level (0.5: supportive, 1: moderate, 2: strong, 4: very strong). Finally, the evidence strength of all ACMG items was integrated into model features in one-hot encoding to bridge deep learning predictions and clinical standards.

[0061] 4. Semi-supervised hybrid convolutional network training SH-CNN combines the advantages of Convolutional Neural Networks (CNNs) and Dilated Convolutional Networks (IDCNNs). Shallow CNNs can capture local patterns between neighboring features in feature vectors; for example, adjacent conservation scores and functional prediction scores may jointly indicate the importance of a variant. Dilated convolutions, by expanding the receptive field, model feature dependencies across greater distances without reducing feature resolution, allowing the model to simultaneously integrate global information from different categories of features. For example, it can correlate splicing effect scores with conservation scores to identify the pathogenic signal they jointly represent. By fusing CNNs and IDCNNs, SH-CNN can both finely extract local features and synthesize global information, providing a solid structural foundation for semi-supervised feature learning.

[0062] The model consists of three functional modules: feature extractor, supervised classification head, and self-supervised task module (including reconstruction head and contrastive learning projection head).

[0063] 4.1 Feature Extractor Input: One-dimensional annotation feature vector of candidate DNM (1×n, where n is the feature dimension).

[0064] Method: Convolution + pooling modeling local patterns → multi-layer one-dimensional dilated convolution (dilation rate 1, 2, 4, 8) for cross-scale modeling.

[0065] Output: A 128-dimensional latent representation, serving as a unified feature base for three types of tasks (supervised classification, reconstruction, and contrastive learning).

[0066] 4.2 Multi-task head parallel architecture (1) Supervised classification head---learning pathogenic patterns Input: A potential representation of labeled mutations.

[0067] Method: The latent features of the mutation are subjected to Dropout regularization and linear transformation.

[0068] Objective: To learn a discrimination pattern that can distinguish between pathogenic and benign mutations.

[0069] Output: Binary classification probability distribution (used during the training phase to calculate supervised classification loss and update model parameters).

[0070] 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.

[0071] (2) Reconstruction module --- capturing the internal feature structure of the variant Input: A potential representation of unlabeled mutations.

[0072] 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).

[0073] Objective: To improve the fidelity of latent representations to original features and preserve collaborative features within individual variants.

[0074] Output: Reconstructed feature vectors with the same dimensions as the original input, used to calculate the reconstruction loss and update the model parameters.

[0075] 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.

[0076] (3) Comparative learning module---consistency of the same variant and distinguishability of different variants Input: A potential representation of unlabeled mutations.

[0077] Methods: The latent representation is mapped to a 64-dimensional low-dimensional projection space. A positive-negative variant sample comparison mechanism based on the SimCLR framework is used: First, enhancement operations such as random noise, scaling, or feature masking are applied to generate two enhanced views, aiming to encourage the model to focus on the core biological features of the variants rather than random perturbations. Subsequently, the feature space distribution of positive and negative sample pairs is optimized using normalized temperature-scale cross-entropy (NT-Xent) loss: enhanced views of the same variant (positive sample pairs) are close in the feature space, achieving representation stability; different variant samples (negative sample pairs) maintain relative distance even if they are similar in some features, thereby distinguishing the overall biological feature combinations of different variants.

[0078] Objective: To eliminate the influence of accidental noise or input order, maintain the stability of the same variant, and the distinguishability of different variants.

[0079] Output: The projected low-dimensional vector, used to calculate the contrastive loss and update the model parameters.

[0080] The operational logic of the comparative learning module is as follows: Figure 8 As shown.

[0081] The outputs of these three tasks flow to different objectives: the classification head outputs a binary probability distribution, used to calculate the supervised classification loss; the reconstruction head outputs a reconstructed feature vector with the same dimension as the original input, used to calculate the reconstruction error; and the contrastive learning head outputs a low-dimensional embedding vector, used to measure the similarity between augmented views. The three outputs do not interact directly; instead, each calculates its own loss function, which is then weighted and combined to form the total loss. This loss is then used to jointly update the parameters of the feature extractor and the task head through backpropagation. This achieves collaborative optimization of supervised and self-supervised signals during the training phase. The training process is as follows: Figure 5 As shown.

[0082] Training process: Data loading: Load labeled and unlabeled variation data in batches; Forward propagation: The three task heads are computed in parallel; Loss calculation: The total loss is obtained by weighted summation; Backpropagation: Calculates gradients and updates network parameters; Iterative optimization: Repeat the above process until convergence.

[0083] The flowchart of the inference phase after training is completed is shown in the attached figure. Figure 9As shown, all inputs (regardless of whether they are labeled) are processed by a feature extractor and a supervised classification head. The convolutional kernel weights and dilated convolution parameters of the feature extractor have been jointly updated during the training phase using three different tasks, outputting a binary classification probability distribution to predict the pathogenicity of the variant. At this stage, the self-supervised module no longer participates in independent computation, but the latent structure it learns during training is embedded in the feature representation, thereby enhancing the reliability and generalizability of the classification prediction.

[0084] 4.3 Activation Function The classification output layer uses the Softmax activation function to normalize the model output to binary classification probabilities, directly reflecting the pathogenicity risk of the variant. The hidden layers use the Rectified Linear Unit (ReLU) function to mitigate gradient vanishing and maintain computational efficiency; the calculation formula is shown below:

[0085]

[0086] 4.4 Loss Function The loss function follows the multi-task learning approach and consists of three parts: FocalLoss is used for supervised classification to mitigate training bias caused by class imbalance and improve sensitivity to rare pathogenic mutations in minority classes; mean squared error (MSE) is used for self-supervised reconstruction to constrain the stability of potential representations by minimizing the difference between input and reconstructed features; and normalized temperature-scale cross-entropy (NT-Xent) is used for contrastive learning to optimize the consistency of representations among mutations by bringing the representations of the same mutation closer together and distinguishing between different mutations. The three types of losses are weighted and combined, with weights determined experimentally to be λ_recon=1.0 and λ_contrast=0.5 to achieve a balance between supervised and self-supervised tasks. The FocalLoss function is shown below:

[0087]

[0088] The FocalLoss function is used for binary classification, where y is the true label and p is the predicted probability. α is the balancing factor (usually 0.25), and γ is the focusing parameter (usually 2).

[0089] The formula for mean square error is as follows:

[0090] In the mean square error formula It is a primitive feature. N represents the reconstructed features, and N is the number of features.

[0091] 4.5 Training Optimization Strategies Optimizer: Adam (Adaptive Moment Estimator).

[0092] Parameter settings: β1=0.9, β2=0.999, initial learning rate 0.001.

[0093] Learning rate adjustment: Reduce LROn Plateau, dynamically adjusted based on validation set performance.

[0094] Gradient clipping: threshold = 1.0, to prevent gradient explosion.

[0095] Finally, the training-test process was independently repeated 10 times under the same hyperparameters, and the mean and standard deviation were reported to ensure the robustness and reproducibility of the results. Through this design, the supervised task provides explicit discriminative signals, while the self-supervised task provides potential structural constraints. Together, they significantly improve the model's generalization performance under limited label conditions.

[0096] 4.6 Interpretable Clinical Output To enhance the transparency of medical applications, this study introduces SHAP (SHapley Additive ex Planations) into SH-CNN to quantify the contribution of each input feature to the prediction results.

[0097] Based on the Shapley value principle of cooperative game theory, the formula for calculating feature contribution is as follows:

[0098] The symbols in the above formulas are explained in Table 3 below.

[0099] Table 3

[0100] First, a baseline is established, typically the feature mean vector or background variation, representing the model output under conditions of "no effective feature information." Then, the SHAP method sequentially replaces the input features with the baseline and inputs the new feature vectors into the SH-CNN, recalculating the output pathogenicity probability distribution. By comparing the change in predicted probability before and after the replacement, the marginal contribution of each feature is obtained. Furthermore, based on the Shapley value principle, contributions are accumulated across all possible feature combinations to obtain the global importance of each feature to the model's predictions. For example, when the SpliceAI score is removed, the predicted probability drops from 0.85 to 0.65, indicating that the feature has a positive contribution. By accumulating the marginal contributions of different features, SHAP can simultaneously achieve local explanation (single variation) and global explanation (overall feature importance ranking, such as clinical evidence one-hot > SpliceAI > population frequency).

[0101] Clinical evidence features are encoded using one-hot encoding, directly corresponding to ACMG entries. SHAP output contributions can be precisely mapped to specific features and evidence, realizing the transformation from "black box prediction" to "traceable clinical interpretation," enhancing model transparency and verifiability, and providing methodological support for molecular diagnostics and genetic counseling.

[0102] 4.7 Prediction Results Output All performance metrics are the mean ± standard deviation of 10 randomized repeated experiments to reflect the stability of the model results.

[0103] On the Gene4Denovo dataset test set, SH-CNN outperforms existing methods. SH-CNN's area under the ROC curve (AUC) reaches 0.879±0.012, and its precision (AP) is 0.887±0.015, which is significantly higher than 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).

[0104] On both the ROC and PR curves, SH-CNN exhibits a higher overall curve position and a steeper initial segment, indicating that it maintains high sensitivity and accuracy even with a low false positive rate. It is worth noting that 'n' in the figure represents the number of effective labeled variants that the tool can assess in the test set. Unlike other methods, SH-CNN can predict all known labeled variants with a 0% missing rate, while other tools exhibit varying degrees of missing data due to limitations in their applicability or variant type. This characteristic gives SH-CNN both accuracy and coverage advantages in large-scale variant assessment.

[0105] To verify the model's generalization ability, it was evaluated on an independent Spark cohort (Autism Spectrum Disorder, ASD). The results show that SH-CNN maintains good overall discriminative performance (AUC=0.763) on this external dataset, outperforming all control tools except CADD, demonstrating good cross-cohort robustness. However, its mean precision (AP=0.430) is relatively low. Further analysis indicates that this result is mainly due to the high imbalance in label distribution and the extreme proportion of benign variants.

[0106] The Spark test set had a significantly higher proportion of benign variants (label=0) than the Gene4Denovo test set (8614 vs 3956), while the number of pathogenic variants (label=1) was only 517, accounting for less than 6%. Furthermore, this dataset contained a large number of unlabeled variants (label=-1, 24817 in total), further diluting the proportion of valid positive samples. In contrast, the Gene4Denovo dataset had a more balanced number of pathogenic and benign labels (2965 vs 3956), therefore its AP (0.887±0.015) was significantly higher than that of the Spark set.

[0107] In the Spark dataset, the benign labels primarily come from Gene4Denovo phenotypic annotations (97.0%), rather than ClinVar's high-confidence labels. This suggests that these samples may contain label noise or phenotypic heterogeneity, reducing the model's confidence in distinguishing between positive and negative samples. Correspondingly, SH-CNN maintains a high AUC, indicating its overall ability to rank positive and negative samples is stable. However, due to the extremely high proportion of benign samples and the compressed threshold discrimination space, the average precision (AP) decreases.

[0108] The extremely imbalanced class distribution in the Spark queue (significantly too many benign samples and a scarcity of pathogenic samples) and the low confidence of some labels are the main reasons for the relatively low AP. This result demonstrates the robustness of SH-CNN on real-world clinical external data. Even under conditions of extremely imbalanced class distribution and a much larger number of benign samples than pathogenic samples, the model still outperforms most existing tools in overall discriminant performance (AUC) and precision ranking (AP), showing strong generalization ability and adaptability to the complexity of real-world data.

[0109] Based on the test set results of the Gene4Denovo dataset, the analysis results show that: Positive SHAP values: drive predictions toward pathogenicity; Negative SHAP values: drive forecasts toward a more benign trend; Eigen-raw_coding exhibited the highest feature importance in the model, indicating that the integrated score, which combines sequence and evolutionary conservation information, plays a central role in pathogenicity determination. This result is consistent with the biological significance of this feature reflecting the evolutionary selection pressure on functional sites. Following closely behind were HUVEC_confidence_value (functional confidence derived from endothelial cell lines) and ACMG evidence items ps4_{4} and ps4_{10}, suggesting that the model not only captured cell type-specific functional signals but also fully utilized clinical epidemiological evidence related to case frequency. Furthermore, functional prediction and conservation indices such as VARITY_ER_LOO_score, Eigen-phred_coding, FATHMM_MKL_coding, and CADD_coding made significant contributions to model decision-making, further validating the model's ability to comprehensively utilize multi-level biological information. It is worth noting that some cross-regional features (such as RegBase_Can, Funseq2, HUVEC_fitCons_score, etc.) although they come from non-coding or regulatory annotation frameworks, still show significant contributions in coding region prediction. This indicates that the regulatory potential and chromatin openness signals captured by these features can serve as auxiliary discrimination criteria, providing additional contextual information for the model.

[0110] The directionality of feature contributions aligns with biological expectations; high functional scores and high confidence values ​​drive the model's predictions toward pathogenicity, while high allele frequencies (such as the gnomAD_AF feature) are associated with benign variants. More importantly, the high contribution of ACMG evidence items (especially the PS4 series) indicates that the model not only captures variant effects at the functional level but also effectively learns the weighting structure of clinical evidence in pathogenicity assessment, bridging the gap between functional prediction and clinical interpretation.

[0111] The technical solution in this embodiment can achieve the following technical effects.

[0112] (1) A semi-supervised hybrid deep learning framework is proposed to solve the model generalization problem under label scarcity and class imbalance.

[0113] By introducing self-supervised reconstruction and contrastive learning tasks, and jointly optimizing them with supervised classification tasks, the model is forced to autonomously learn the intrinsic structure and contextual relationships between annotated features from massive unlabeled variants, thereby extracting more robust and discriminative feature representations, improving generalization performance under limited label conditions and sensitivity to rare pathogenic variants.

[0114] (2) Design an interpretability deep fusion mechanism to achieve traceability of prediction results at the feature and evidence levels.

[0115] By integrating SHAP interpretability analysis, the model's predicted probabilities are attributed to each input feature. In particular, for quantified ACMG clinical evidence features, a reverse mapping from model output to specific clinical evidence items is achieved, providing transparent and verifiable evidence support for the prediction of each variant, thereby enhancing clinical trust and usability.

[0116] (3) Construct a unified feature engineering and modeling system that integrates functional annotation and clinical evidence, aiming to bridge algorithm prediction and clinical interpretation.

[0117] By systematically quantifying and characterizing key evidence items from the ACMG / AMP guidelines, and jointly training the model with multi-source functional annotation features, the model learns not only the functional impact of biology but also the weighting logic of clinical evidence. This makes its output more closely resemble the decision-making process of clinical experts, achieving a smooth transition from "functional prediction scores" to "interpretation of clinical significance."

[0118] (4) A 100% coverage end-to-end prediction system designed to eliminate prediction blind spots and improve the efficiency of clinical workflow.

[0119] Through the enhancement of feature representation by semi-supervised learning and the unified feature processing process, the model can generate reliable prediction results for all input variants (regardless of whether their labels or evidence are known), achieving a prediction coverage of nearly 100%, thereby supporting the complete automation of large-scale variant screening and reducing manual intervention.

[0120] The beneficial effects of this embodiment are as follows: (1) In terms of performance, through semi-supervised multi-task learning, the structural information of unlabeled variants is fully utilized under limited label conditions. On the Gene4Denovo test set, AUC=0.879±0.012 and AP=0.887±0.015 are achieved, which are significantly better than existing tools.

[0121] (2) In terms of practicality, it achieves 100% variation coverage, solves the problem of prediction loss caused by the limited scope of application of existing tools, and supports the complete workflow of large-scale genome analysis.

[0122] (3) In terms of clinical value, through SHAP interpretability analysis, the model prediction is mapped to specific ACMG evidence items, realizing the transformation from "black box prediction" to "traceable clinical interpretation", providing verifiable methodological support for molecular diagnostics.

[0123] The technical solution in this embodiment can automatically generate structured interpretive reports. Unlike traditional tools that only output a single score, it presents the data through a hierarchical structure. The first level displays the comprehensive prediction conclusion, such as a pathogenicity probability of 0.98, indicating pathogenicity. The second level displays the core evidence chain, listing the high-contribution features supporting this judgment and their corresponding ACMG evidence items in natural language, such as the main evidence: PVS1 (loss of function), PM2 (low frequency in the population), and PP3 (bioinformatics prediction of harmfulness). The third level displays a feature radar chart, intuitively comparing the similarity between the target variant and known pathogenic variants across various dimensions. This embodiment not only solves the problem of black-box models being untrustworthy in clinical practice but also directly outputs auxiliary decision-making information that aligns with doctors' diagnostic habits, greatly reducing the workload of manually interpreting VUS (variables of undetermined significance) and achieving end-to-end intelligent processing from raw sequencing data to clinical auxiliary diagnosis.

[0124] Preferably, in step S1, a numerical feature vector containing evolutionary conservation, functional prediction score, and quantitative clinical evidence is generated, including the following steps S11-S13: 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 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.

[0125] The process of generating numerical feature vectors described in step S1 is not a simple data splicing; it includes data quality control and completion mechanisms, which can solve the sparsity and heterogeneity problems commonly found in genomic data. In step S11, the system initiates a missing value cleaning procedure, performing a bidirectional scan of the newly discovered variant site data matrix to be analyzed. Due to the complex origins of newly discovered variants, some sites may lack specific functional annotations due to insufficient sequencing depth or their location in non-coding regions of the genome, such as the lack of conservation scores for certain species. The system presets filtering criteria, setting a first preset threshold of 50%, i.e., removing low-quality sample rows with a feature missing rate exceeding 50%; and a second preset threshold of 20%, i.e., removing low signal-to-noise ratio feature columns with a missing rate exceeding 20% ​​in all samples. This effectively removes invalid noise from the data, retaining the core feature set with statistical significance, thereby preventing a large number of null values ​​from interfering with the subsequent model convergence process.

[0126] For the few discrete missing values ​​still existing in the feature matrix after filtering, the K-nearest neighbor (KNN) imputation operation described in step S12 is performed. Unlike traditional mean imputation or zero imputation, this embodiment employs an imputation algorithm based on the topological structure of the feature space. Specifically, the system calculates the Euclidean distance between the sample to be imputed and other complete samples in the feature space, selecting the K nearest neighbor samples, where K is preferably set to an integer between 10 and 20. The system assigns weights based on the reciprocal of the Euclidean distance; the closer the neighbor, the greater its contribution to the imputation value. The imputation value is calculated by weighted averaging. This approach leverages the principle of local similarity in biological data—that functionally similar variants often have similar feature distributions—to maximally restore the true distribution of the data.

[0127] Step S13 involves standardizing and vectorizing the complete original feature matrix. The evolutionary conservation score ranges from -10 to 10; the functional prediction score ranges from 0 to 1; and the quantified clinical evidence score is encoded as discrete integers. These three scores have completely different dimensions and physical meanings, and directly inputting them into the neural network would lead to unbalanced gradient updates. Therefore, the system independently performs Min-Max normalization on each dimension of the matrix, linearly mapping all feature values ​​to the closed interval [0, 1]. After processing, the system concatenates the evolutionary features, functional features, and clinical evidence features in a fixed order according to a predetermined channel order strategy, constructing a structurally unified numerical feature vector.

[0128] Preferably, in step S1, acquiring the newly discovered variant site data to be analyzed and constructing a hierarchical labeling system includes the following steps S14-S16: S14. 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 labels. S15, Search the Gene4Denovo database, mark variants with all control phenotypes as benign, and mark variants with all pathogenic phenotypes that meet the frequency threshold as pathogenic, as the second-level labels; S16, merge the first-level pathogenic or benign labels with the second-level labels to form a labeled sample set, and divide the remaining unlabeled variant sites into an unlabeled sample set.

[0129] To address the widespread problem of scarce and inconsistent quality of labeled data in the field of emerging variants, step S1 employs a hierarchical labeling system construction strategy. Step S14 establishes highly reliable Level 1 pathogenic or benign labels. The system retrieves authoritative ClinVar databases via API or local mirror, which aggregates variant interpretation records submitted by clinical laboratories worldwide. To prevent mislabeling from introducing noise into the model, the system employs a filtering method, selecting only loci with conflict-free multiple submitter records or expert panel review records. The system identifies entries with 2 stars or higher in the ClinVar rating system. If their clinical significance is clearly determined to be pathogenic / potentially pathogenic or benign / potentially benign, they are directly labeled as Level 1 pathogenic or benign samples. This ensures that the anchor data used as the basis for model training has extremely high clinical validity and gold standard attributes, reducing the negative impact of label noise on the supervised learning branch.

[0130] Given the limited number of high-confidence variants in ClinVar, insufficient to support adequate training of deep neural networks, the system executes step S15 to expand the second-level labels based on phenotypic evidence using the Gene4Denovo database. Gene4Denovo focuses on collecting newly emerging variant data, containing a large number of sporadic case and control samples. The system infers the presence of sites not included in ClinVar. Variants appearing only in healthy controls and never in disease groups are labeled as benign. For variants appearing only in disease groups, a frequency threshold is introduced to constrain sequencing errors or chance factors. For example, a variant is required to appear independently at least twice in the disease group; only variants meeting this condition are labeled as pathogenic. Through inferences based on statistical regularity and phenotypic correlation, a larger-scale second-level labeling system, with slightly lower confidence than the first-level pathogenic or benign labels, is constructed, alleviating the training bottleneck caused by the long-tailed distribution of data.

[0131] Step S16 involves integrating and partitioning the dataset. The system physically merges the first-level pathogenic or benign labeled samples with the second-level labeled samples, performing deduplication to form a labeled sample set. This labeled sample set will be used for calculating the cross-entropy loss of the supervised classification head in subsequent step S3. Simultaneously, for the remaining variant sites in the database that do not meet the ClinVar high-star rating criteria or the Gene4Denovo phenotypic inference rules—for example, those with conflicting interpretations, insufficient evidence, or rare, singular variants—the system uniformly classifies them into an unlabeled sample set. This invention does not discard this seemingly useless data but uses it as input for self-supervised reconstruction and contrastive learning. Its implicit genomic distribution characteristics assist the model in learning the underlying structure of variants, thereby achieving effective utilization of massive amounts of unsigned variant data.

[0132] Preferably, in step S2, constructing a feature extractor using the numerical feature vector includes the following steps S21-S23: 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.

[0133] The feature extractor described in step S2 is the core connecting the original data with the downstream multi-task head. Its construction process follows a deep learning paradigm of abstracting from local micro-features to global macro-features layer by layer. Step S21 is executed to configure a basic one-dimensional convolutional layer (1D-CNN) to extract local associations. Considering that the input data is a concatenated heterogeneous numerical feature vector, its adjacent elements may represent attributes with biological intrinsic connections; for example, several consecutive dimensions may represent different evolutionary conservation scores for the same site. The system sets the kernel size to a small value (e.g., 3 or 5) and the stride to 1, performing a sliding convolution operation on the feature vector. Through the processing of a non-linear activation function, this one-dimensional convolutional layer can keenly capture short-range coupling patterns between adjacent features, generating a first-level feature map reflecting local feature associations, thus enabling the search for specific short sequence motifs on the genome sequence.

[0134] To overcome the limited receptive field of traditional convolutional layers, step S22 is performed to construct a dilated convolutional layer cascaded after the local feature maps. This dilated convolutional layer consists of multiple cascaded residual blocks, each containing convolutional units with different dilation rates. The system sets the dilation rate parameters to grow exponentially, allowing the network to expand its receptive field exponentially without pooling and thus without loss of resolution. Pathogenicity assessment often relies on nonlinear interactions between distant elements in the feature vector, such as the long-range dependency between sequence features at the head of the vector and clinical evidence quantification features at the tail. The model can generate a global dependency feature map, effectively integrating cross-dimensional information flow.

[0135] Step S23 completes the mapping from the high-dimensional feature map to the low-dimensional latent space. Since the aforementioned globally dependent feature map still retains its spatial structure (number of channels × feature length), it is unsuitable for direct classification or contrastive computation. Therefore, a flattening operation or global average pooling is performed to transform it into a one-dimensional vector. Subsequently, a fully connected layer is used for linear projection, compressing the high-dimensional information into a compact low-dimensional latent feature vector, for example, with dimensions of 128 or 256. This latent feature vector not only removes redundant information but also serves as a shared representation, simultaneously inputting into the subsequent parallel supervised classification head, self-supervised reconstruction head, and contrastive learning head. This ensures that when the model updates parameters during backpropagation, it is simultaneously subject to multiple constraints of classification discriminative power, data reconstruction ability, and sample discrimination power, thereby learning a more robust and essential pathogenic feature expression.

[0136] Preferably, in step S3, the joint update of the network parameters of the feature extractor and each head based on the total loss of weighted summation includes the following steps S31-S33: 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.

[0137] The joint update process described in step S3 is crucial for the model to acquire high generalization ability, and it constructs a composite objective function that balances the supervisory and self-supervised signals. In step S31, addressing the extreme class imbalance problem where the number of benign variants (negative samples) far exceeds the number of pathogenic variants (positive samples) in newly emerging variant data, the supervised classification head abandons the traditional cross-entropy loss function and instead adopts the Focal Loss function. This function introduces a focusing parameter. (For example, set to 2.0) and balance parameters (For example, set to 0.25). During the calculation, the Focal Loss function reduces the weight of easy-to-classify simple negative samples (i.e., benign mutations with extremely high prediction probability), forcing the model to focus the gradient update on difficult samples (i.e., pathogenic mutations at the classification boundary). This prevents the model from getting stuck in local optima due to the dominance of a large number of simple negative samples, thereby improving the recall rate for rare pathogenic mutations.

[0138] Meanwhile, to fully explore the potential value of unlabeled data, step S32 is executed. The system calculates the losses for two auxiliary tasks using all input samples (including labeled and unlabeled data). On one hand, the self-supervised reconstruction head restores the latent features to the original input through the decoder, and calculates the difference between the reconstructed output and the original input using the mean squared error function (MSE Loss), i.e., the reconstruction loss. This loss constrains the feature extractor to retain relevant information of the input data, playing a role in regularization and noise reduction. On the other hand, the contrastive learning head uses data augmentation strategies to generate two augmented views for each sample, and calculates the contrastive loss based on the normalized temperature scale cross-entropy function. Data augmentation strategies include random masking or Gaussian noise injection. This loss can bring the different views of the same sample closer in the feature space, while widening the distance between different samples, thus enabling the model to learn the essential structural features of the samples that are rotation- and translation-invariant.

[0139] Execute step S33 to perform multi-task joint backpropagation. The system constructs the overall objective function based on a preset hyperparameter strategy, expressed as:

[0140] in, , and These are the weight coefficients for classification, reconstruction, and contrastive losses, respectively. During the initial training phase, the system assigns... and Larger weights are used to help the model initialize a good feature space through unsupervised learning in the absence of labels; these weights are gradually increased as the number of training iterations increases. The weights are adjusted to optimize discrimination performance. Based on the calculated weighted total loss, the system uses the AdamW or SGD optimizer to compute gradients and synchronously updates the network parameters of the feature extractor (Backbone) and each task head (Heads) along the gradient descent direction. Through an end-to-end joint training mechanism, the model possesses both the discriminative power of supervised learning and the generalization power of self-supervised learning.

[0141] Preferred, such as Figure 6 As shown, in step S4, the step of back-mapping the features with high contribution values ​​to the corresponding clinical evidence items includes the following steps S41-S43: 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.

[0142] To break down the black-box barrier of deep learning models in clinical applications, step S4 introduces a post-hoc explanation mechanism based on cooperative game theory. In step S41, the system calls the interpreter module (e.g., integrating the SHAP algorithm kernel) to perform attribution analysis on each specific prediction result. Each input numerical feature (e.g., evolutionary conservation score, specific clinical evidence encoding) is considered a participant in the game, while the pathogenicity probability output by the model is considered the total payoff. The system estimates the contribution of each feature dimension by calculating the marginal change in the model's output probability under different feature combinations. Specifically, it simulates the difference in predicted probability when a feature exists and does not exist (or is replaced by a benchmark value), and quantifies the independent contribution of the feature to pushing the result towards pathogenicity or benignity by weighted averaging of all possible feature permutations and combinations. This transforms the complex nonlinear neural network weights into feature importance scores with linear additive properties.

[0143] Step S42 involves filtering and grading the feature contributions. Since the input feature vectors are high-dimensional, typically containing hundreds of features, directly displaying the contribution values ​​of all dimensions to the user would cause information overload. Therefore, the system sorts all features in descending order based on the absolute value of the calculated contribution. The larger the absolute value, the stronger the influence of the feature on the model's final decision; positive values ​​indicate a tendency towards disease, while negative values ​​indicate a tendency towards benign disease. The system sets a filtering threshold N and automatically extracts the top N key features, similar to the differential diagnostic thinking of clinical experts—sifting through complex biological indicators to pinpoint the core pathogenic factors causing the disease and eliminating background noise and redundant information.

[0144] Step S43 executes to achieve reverse mapping from mathematical features to clinical semantics. The system has a pre-set feature index table, which records in detail the original biological meaning and data source corresponding to each dimension index in the feature vector. For selected key features, the system queries the index table. If the key feature is identified as a quantified clinical evidence feature, for example, originating from the encoding of ACMG guideline entries in S1, the system will trigger 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 populates this structured information into the interpretive report template, generating a natural language description such as "Main pathogenicity evidence: PVS1 (detected loss-of-function variant, contribution +0.45)". This not only informs the physician of the predicted result but also clearly elucidates the source of evidence, thus generating a complete diagnostic report that conforms to clinical guidelines.

[0145] In one embodiment, Figure 2 This is a block diagram illustrating a semi-supervised hybrid convolutional network pathogenicity prediction device for emerging variants, according to an exemplary embodiment. Figure 2 As shown, the semi-supervised hybrid convolutional network pathogenicity prediction device for emerging variants includes a generation module 71, a construction module 72, a computation module 73, and a prediction module 74.

[0146] The generation module 71 is used to acquire new variant site data to be analyzed, perform multi-source functional annotation and clinical evidence quantification on it, and generate a numerical feature vector containing evolutionary conservation, functional prediction score and quantified clinical evidence. The building module 72 is used to build 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 computing module 73 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 respectively, 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 74 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. Features with high contribution values ​​are back-mapped to the corresponding clinical evidence entries to generate an explanatory report containing pathogenicity level and evidence source.

[0147] The block diagram of the pathogenicity prediction device for a semi-supervised hybrid convolutional network for emerging variants includes a generation module 71, a construction module 72, a calculation module 73, and a prediction module 74, which are controlled to execute the pathogenicity prediction method for emerging variants described in any of the above embodiments.

[0148] This embodiment also provides a pathogenicity prediction system for emerging variants, including a processor and a computer-readable storage medium interconnected with each other. The computer-readable storage medium stores a computer program, which is executed by the processor to implement the steps of the above-described pathogenicity prediction method for emerging variants.

[0149] This embodiment also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for predicting the pathogenicity of emerging variants.

[0150] This embodiment also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for predicting pathogenicity of emerging variants.

[0151] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may 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, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

Claims

1. A method for predicting the pathogenicity of emerging variants, characterized in that, The method comprises the following steps: S1, obtaining new 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; S2, constructing 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 convolution layer for extracting local features and a dilated convolution layer for extracting global dependency features; 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 contrast loss of the contrast learning head, and jointly updating the network parameters of the feature extractor and each head based on the weighted sum of the total loss; S4, applying the feature extractor and the supervised classification head after the parameter update 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 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.

2. The method of pathogenicity prediction for novel variants of claim 1, wherein, In step S1, the numerical feature vector containing evolutionary conservation, functional prediction score and quantified clinical evidence is generated, comprising: S11, detecting the missing rate of each feature in the new 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; S12, performing K-nearest neighbor imputation on the remaining missing values after removal, calculating the filling value according to the weighted Euclidean distance of the sample in the feature space, and forming a complete original feature matrix; 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.

3. The method of pathogenicity prediction for novel variants of claim 1, wherein, In step S1, the new variant site data to be analyzed is obtained, and a hierarchical label system is constructed, comprising: S14, searching the ClinVar database, and directly marking the variant sites with non-conflicting multi-submitter records or expert group review records as first-level pathogenic or benign labels; S15, searching 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; 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.

4. The method of pathogenicity prediction for novel variants of claim 1, wherein, In step S2, the feature extractor is constructed through the numerical feature vector, comprising: S21, configuring the one-dimensional convolution layer, performing sliding convolution operation on the numerical feature vector according to the small size convolution kernel, and generating local feature maps reflecting the correlation of adjacent features; S22, configuring the dilated convolution layer, cascading multiple dilated convolution units after the local feature maps, setting the expansion rate parameter in an exponential growth mode to expand the receptive field, and generating global dependency feature maps; S23, flattening the global dependency feature map and mapping 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.

5. The method of pathogenicity prediction for novel variants of claim 1, wherein, In step S3, the total loss based on weighted summation jointly updates the network parameters of the feature extractor and each head, including: 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 through a focusing parameter; 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; S33, adding the classification loss, the reconstruction loss and the contrastive loss according to the 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.

6. The method of pathogenicity prediction for novel variants of claim 1, wherein, In step S4, the feature with a high contribution value is reversely mapped to the corresponding clinical evidence item, including: 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; S42, sorting 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; S43, searching the original description corresponding to the key feature according to the feature index table, and outputting the evidence code and pathogenicity support level corresponding to the key feature if the key feature belongs to the quantitative clinical evidence feature.

7. A novel variant-oriented semi-supervised hybrid convolutional network pathogenicity prediction apparatus, characterized by, including: The generating module is configured to obtain new variant site data to be analyzed, perform multi-source functional annotation and clinical evidence quantification on the new variant site data, and generate a numerical feature vector containing evolutionary conservation, functional prediction score and quantitative clinical evidence; 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 contrastive learning head, and the feature extractor is internally connected with a one-dimensional convolution layer for extracting local features and a dilated convolution layer for extracting global dependency features; The calculating 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 contrastive loss of the contrastive learning head, and jointly update network parameters of the feature extractor and each head based on a total loss based on weighted summation; The prediction module is configured to apply the feature extractor and the supervised classification head after parameter update 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 reversely map a feature with a high contribution value to a corresponding clinical evidence item to generate an explanatory report containing a pathogenicity level and evidence traceability.

8. The novel variant-oriented semi-supervised hybrid convolutional network pathogenicity prediction apparatus of claim 7, wherein: The generating module, the constructing module, the calculating module and the prediction module are controlled to perform the pathogenicity prediction method for new variants according to any one of claims 2 to 6.

9. An electronic device, comprising: including: a communication interface, a processor, and a memory; The memory is configured to store program instructions, and the program instructions, when executed by the processor in communication connection with the memory through the communication interface, cause the electronic device to implement the pathogenicity prediction method for de novo mutation according to any one of claims 1 to 6.

10. A computer-readable storage medium having stored thereon program instructions, wherein, The program instructions, when executed by the computer, cause the computer to implement the pathogenicity prediction method for de novo mutation according to any one of claims 1 to 6.

Images