Gene variation category prediction method and device based on convex hull geometric constraint

Abstract

The invention relates to a genetic variation category prediction method and device based on convex hull geometric constraint, and relates to the fields of bioinformatics, artificial intelligence, applied mathematics and the like, and the method comprises the steps: obtaining multi-modal data corresponding to target genetic variation, including a target variation DNA sequence, a target reference DNA sequence and target semantic text information; performing feature extraction on the multi-modal data through the trained gene variation category prediction model, and predicting the category of target gene variation; wherein loss functions adopted in the model training process comprise a classification loss function used for indicating the difference between a prediction category and a real category, the convex hull geometric constraint loss function is used for indicating the concentration degree of the same type of gene variation corresponding to the distribution range of the feature space and the separation degree of different types of gene variation corresponding to the distribution range of the feature space. According to the invention, based on multi-modal feature fusion and convex hull geometric constraint, efficient and accurate prediction of the gene variation category is realized.

Description

Technical Field

[0001] This disclosure relates to fields such as bioinformatics, artificial intelligence, and applied mathematics, and in particular to a method and apparatus for predicting the category of gene variation based on convex hull geometric constraints. Background Technology

[0002] With the rapid development of high-throughput sequencing technology, the cost of human genome sequencing has been decreasing year by year, and the amount of data generated by whole-exome or whole-genome sequencing of a single individual is enormous. Each sequencing sample may contain tens of thousands to hundreds of thousands of single nucleotide variants, small fragment insertions and deletions, etc., most of which are neutral or benign variants, with only a few closely related to the occurrence of diseases. Therefore, how to interpret massive amounts of gene variants with high accuracy and efficiency, especially distinguishing between pathogenic and benign variants, has become a core issue in the fields of precision medicine and clinical genetics. Summary of the Invention

[0003] In view of this, this disclosure proposes a method, apparatus, electronic device, storage medium, and computer program product for class prediction of gene variation based on convex hull geometric constraints.

[0004] According to one aspect of this disclosure, a method for class prediction of gene variants based on convex hull geometric constraints is provided, comprising:

[0005] Obtain multimodal data corresponding to the target gene mutation, wherein the multimodal data includes the target variant DNA sequence, the target reference DNA sequence, and the target semantic text information; wherein the target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description;

[0006] The multimodal data is feature extracted using a trained gene variant category prediction model, and the category of the target gene variant is predicted based on the extracted features. The loss functions used in training the gene variant category prediction model include a convex hull geometric constraint loss function and a classification loss function. The classification loss function indicates the difference between the predicted category and the true category, while the convex hull geometric constraint loss function indicates the degree of concentration of the distribution range of gene variants of the same category in the feature space and the degree of separation of the distribution ranges of gene variants of different categories in the feature space.

[0007] In one possible implementation, the loss function used during the training of the gene variant category prediction model further includes a contrastive learning loss function, which is used to indicate the degree of consistency of features corresponding to the same gene variant and the degree of difference between features corresponding to different gene variants.

[0008] In one possible implementation, feature extraction of the multimodal data includes:

[0009] Feature extraction is performed on the target reference DNA sequence and the target variant DNA sequence to obtain the local mutation features of the target gene variant; wherein, the local mutation features represent the influence of the gene variant on the local structure of the DNA sequence;

[0010] Feature extraction is performed on the target reference DNA sequence and the target variant DNA sequence to obtain the global statistical mutation features of the target gene variant; wherein, the global statistical mutation features represent the impact of the gene variant on the overall structure of the DNA sequence;

[0011] The target semantic text information is processed to obtain the textual semantic features of the target gene mutation.

[0012] In one possible implementation, the step of extracting features from the target reference DNA sequence and the target variant DNA sequence to obtain the local mutation features of the target gene variant includes:

[0013] Local pattern features of the target reference DNA sequence and local pattern features of the target variant DNA sequence are extracted respectively; wherein, the local pattern features represent the local structure of the DNA sequence;

[0014] Based on the local pattern features of the target reference DNA sequence and the local pattern features of the target variant DNA sequence, the local mutation features of the target gene variant are determined.

[0015] In one possible implementation, the step of extracting features from the target reference DNA sequence and the target variant DNA sequence to obtain the global statistical mutation features of the target gene variant includes:

[0016] Based on a preset subsequence length, global statistical features of the target reference DNA sequence and global statistical features of the target variant DNA sequence are extracted respectively; wherein, the global statistical features represent the overall compositional pattern of the DNA sequence;

[0017] Based on the global statistical characteristics of the target reference DNA sequence and the global statistical characteristics of the target variant DNA sequence, the global statistical mutation characteristics of the target gene variant are determined.

[0018] In one possible implementation, the prediction of the category of the target gene variant based on the extracted features includes:

[0019] The local mutation features, global statistical mutation features, and textual semantic features of the target gene variation are fused to generate multimodal fusion features;

[0020] Based on the multimodal fusion features, or based on the multimodal fusion features and the local mutation features, global statistical mutation features, and textual semantic features of the target gene variant, the category of the target gene variant is predicted.

[0021] In one possible implementation, the method further includes:

[0022] During the training process of the gene mutation category prediction model, for any batch of training samples, the gene mutation category prediction model extracts the local mutation features, global statistical mutation features, and textual semantic features of each training sample in the batch of training samples; wherein, each training sample includes: a reference DNA sequence sample, a variant DNA sequence sample, and a semantic text information sample.

[0023] The local mutation features, global statistical mutation features, and text semantic features of each training sample are concatenated to generate corresponding first fusion features;

[0024] Random features are added to the text semantic features of each training sample, and the local mutation features, global statistical mutation features, and text semantic features with added random features of each training sample are concatenated to generate corresponding second fusion features.

[0025] Determine the positive and negative samples corresponding to each fusion feature in the fusion feature set, wherein the fusion feature set includes each first fusion feature and each second fusion feature; wherein, for any fusion feature, the positive sample corresponding to the fusion feature includes the fusion feature and the fusion feature corresponding to the same training sample as the fusion feature, and the negative sample corresponding to the fusion feature includes the remaining fusion features in the fusion feature set;

[0026] Based on the positive and negative samples corresponding to each fusion feature, the function value of the contrastive learning loss function is calculated.

[0027] According to another aspect of this disclosure, a gene variation category prediction device based on convex hull geometric constraints is provided, comprising:

[0028] The acquisition module is used to acquire multimodal data corresponding to the target gene mutation. The multimodal data includes the target mutation DNA sequence, the target reference DNA sequence, and the target semantic text information. The target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description.

[0029] The prediction module is used to extract features from the multimodal data using a trained gene variant category prediction model, and to predict the category of the target gene variant based on the extracted features. The loss functions used in the training of the gene variant category prediction model include a convex hull geometric constraint loss function and a classification loss function. The classification loss function indicates the difference between the predicted category and the true category, while the convex hull geometric constraint loss function indicates the degree of concentration of the distribution range of gene variants of the same category in the feature space and the degree of separation of the distribution ranges of gene variants of different categories in the feature space.

[0030] According to another aspect of this disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method.

[0031] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described method.

[0032] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0033] According to various aspects of this disclosure, multimodal data corresponding to a target gene variant is obtained, the multimodal data including the target variant DNA sequence, the target reference DNA sequence, and target semantic text information; wherein, the target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description; through a trained gene variant category prediction model, features are extracted from the multimodal data, and the category of the target gene variant is predicted based on the extracted features; wherein, the loss function used in the training process of the gene variant category prediction model includes: a convex hull geometric constraint loss function and a classification loss function; the classification loss function is used to indicate the difference between the predicted category and the true category, and the convex hull geometric constraint loss function is used to indicate the degree of concentration of the distribution range of gene variants of the same category in the feature space and the degree of separation of the distribution range of gene variants of different categories in the feature space. This approach extracts features from multimodal data, including target variant DNA sequences, target reference DNA sequences, and target semantic text information, fully exploring the complementary relationships between heterogeneous information from multiple sources. This improves the completeness of feature expression while effectively overcoming the limitations of single-modal information representation. Furthermore, a convex hull geometric constraint strategy is introduced during the training phase to geometrically regularize the distribution structure of different types of gene variants in the feature space. This enables automated, efficient, and high-precision prediction of gene variant categories, such as single nucleotide mutations, insertions / deletions, and other genetic variations, providing auxiliary decision support for clinical genetics and precision medicine. In some examples, a contrastive learning strategy is also introduced during training to further enhance the ability to discriminate gene variant categories, predictive stability, and generalization performance.

[0034] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0035] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0036] Figure 1 A flowchart illustrating a gene variation category prediction method based on convex hull geometric constraints according to an embodiment of the present disclosure is shown.

[0037] Figure 2 A flowchart illustrating feature extraction from multimodal data according to an embodiment of the present disclosure is shown;

[0038] Figure 3 A flowchart illustrating the training process of a gene variation category prediction model according to an embodiment of the present disclosure is shown.

[0039] Figure 4A flowchart is shown illustrating a method for predicting gene variation categories according to an embodiment of the present disclosure;

[0040] Figure 5 A schematic diagram of the structure of a gene variation category prediction system according to an embodiment of this disclosure is shown;

[0041] Figure 6 A structural diagram of a gene variation category prediction device based on convex hull geometric constraints according to an embodiment of the present disclosure is shown.

[0042] Figure 7 A block diagram of an electronic device 1900 according to an embodiment of the present disclosure is shown. Detailed Implementation

[0043] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0044] As used herein, the terms “comprising,” “including,” “having,” or variations thereof are open-ended and include one or more of the stated features, integrals, elements, steps, components, or functions, but do not exclude the presence or addition of one or more other features, integrals, elements, steps, components, functions, or groups thereof.

[0045] When an element is referred to as “connected,” “coupled,” “responding,” or a variation thereof relative to another element, it may be directly connected, coupled, or responding to another element, or there may be an intermediate element present.

[0046] Although the terms first, second, third, etc., may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are only used to distinguish one element / operation from another. Therefore, without departing from the teachings of the inventive concept, a first element / operation in some embodiments may be referred to as a second element / operation in other embodiments.

[0047] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0048] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0049] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant regions.

[0050] In related technologies, methods for predicting gene variation categories mainly rely on manually designed features and rules. Examples include rule-based scoring methods based on evolutionary conservation scores, changes in amino acid physicochemical properties, protein structural influences, or functional domain annotations. While these methods can reveal the biological significance of gene variations to some extent, they have the following limitations: First, they rely on prior knowledge and assumptions, making it difficult to capture complex nonlinear relationships; second, their generalization ability is limited across different genes or disease contexts; and third, they cannot fully utilize rich textual annotation information or multimodal features, resulting in insufficient information utilization.

[0051] In related technologies, machine learning and deep learning techniques are employed in gene variation analysis; for example, support vector machines, random forests, and deep neural networks have been used for multi-feature joint modeling. However, these methods still have the following problems: On the one hand, most only focus on sequence information, usually only using local windows or simple k-mer statistics, making it difficult to take into account both local modal features and global statistical features; on the other hand, a large amount of semantic textual information related to gene variations is not fully utilized; in addition, multimodal features have problems such as large dimensional differences, inconsistent scales, and unstable distributions during joint modeling, leading to unstable model training or an increased risk of overfitting. On the other hand, deep learning methods mostly adopt pure supervised learning, which is highly dependent on high-quality labeled data; however, in real clinical scenarios, the number of clearly labeled pathogenic or benign variations is limited, and the data distribution has class imbalance and sample heterogeneity, making it difficult to guarantee the generalization performance of the model; at the same time, most models only focus on classification performance indicators and rarely constrain the model from the perspective of geometric structure or feature space distribution, resulting in the mixing of different class samples in the feature space and unstable decision boundaries.

[0052] It is evident that the efficiency and accuracy of related technologies in determining the category of gene variants need improvement. To address these technical issues, this disclosure provides a gene variant category prediction method based on convex hull geometric constraints (detailed description below). This method integrates multi-source heterogeneous information (variant DNA sequences, reference DNA sequences, and semantic text information) and introduces a convex hull geometric constraint strategy during training. This achieves high-efficiency and high-accuracy prediction of gene variant categories based on multimodal feature fusion and convex hull geometric constraints, thereby effectively improving the overall model performance and clinical applicability. In some examples, a contrastive learning strategy is also introduced during training, further enhancing the ability to discriminate gene variant categories, prediction stability, and generalization performance.

[0053] For example, the gene variation category prediction method based on convex hull geometric constraints provided in this disclosure can be executed by an electronic device or a part of an electronic device (such as a processor), such as a terminal device or a server. The terminal device can be a desktop terminal or a mobile terminal, for example, a laptop, tablet, desktop computer, smartphone, smart speaker, smartwatch, smart TV, in-vehicle terminal, or other types of electronic devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data and artificial intelligence platforms.

[0054] The following is a detailed description of a gene variation category prediction method based on convex hull geometric constraints provided by embodiments of this disclosure.

[0055] Figure 1 A flowchart illustrating a gene variation category prediction method based on convex hull geometric constraints according to an embodiment of the present disclosure is shown, as follows: Figure 1 As shown, the following steps may be included:

[0056] Step 101: Obtain multimodal data corresponding to the target gene mutation. The multimodal data includes the target variant DNA sequence, the target reference DNA sequence, and the target semantic text information. The target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description.

[0057] In this context, the target gene variant refers to a specific variant, such as a single nucleotide mutation, a small fragment insertion, or a deletion, that is to be predicted. The target gene variant has one or more mutation sites. For example, the target gene variant could be a mutation in exon 17 of the BRCA1 gene, or the R248Q mutation in the TP53 gene, and so on.

[0058] The target variant DNA sequence represents the DNA sequence in which the target gene mutation occurs. This DNA sequence is a segment of a specific length containing the mutation site of the target gene mutation and the upstream and downstream sequences of that mutation site. The length of the target variant DNA sequence can be set according to requirements and is not limited thereto.

[0059] The target reference DNA sequence represents a standard, unmutated DNA sequence corresponding to the target variant DNA sequence. The target reference DNA sequence is identical to the target variant DNA sequence except for the mutation site. For example, the target reference DNA sequence corresponding to the target gene variant can be obtained from a human reference genome.

[0060] The target semantic text information is text information related to the target gene variation. As an example, the target semantic text information may include: chromosome number, gene name, and gene function description; where the gene name is the name of the gene containing the target gene variation, the chromosome number is the chromosome number containing the gene, and the gene function description is a functional annotation of the gene. Furthermore, the target semantic text information may also include other structured or unstructured biological knowledge text, without limitation. For example, the target semantic text information can be obtained from bioinformatics databases (such as NCBI Gene, Ensembl, etc.).

[0061] For example, the target variant DNA sequence and the target reference DNA sequence can be single-stranded DNA sequences or double-stranded DNA sequences, without limitation.

[0062] For example, the target gene mutation is a single nucleotide mutation in the CFTR gene; the target reference DNA sequence is ACGTACGTA...; the target variant DNA sequence is ACGTTCGTA..., where the 5th position is the mutation point, and the base A is mutated to T; the target semantic text information may include: chromosome number chr7, gene name CFTR, and gene function as a chloride channel.

[0063] Step 102: Using the trained gene variant category prediction model, feature extraction is performed on the multimodal data, and the category of the target gene variant is predicted based on the extracted features; wherein, the loss function used in the training process of the gene variant category prediction model includes: convex hull geometric constraint loss function and classification loss function; the classification loss function is used to indicate the difference between the predicted category and the true category, and the convex hull geometric constraint loss function is used to indicate the degree of concentration of the distribution range of gene variants of the same category in the feature space and the degree of separation of the distribution range of gene variants of different categories in the feature space.

[0064] For example, the category of the target gene variant can include pathogenic and benign (or neutral).

[0065] Here, the predicted category represents the category of the gene variant predicted by the gene variant category prediction model for any given sample; the true category represents the actual category of the gene variant corresponding to that sample. By employing a classification loss function during training, the difference between the model's predicted category and the true category of any gene variant can be continuously reduced, thereby enabling the trained gene variant category prediction model to achieve high classification accuracy.

[0066] In this model, for any category of gene mutation, the distribution range of that category of gene mutation in the feature space represents the distribution range of the features of that category of gene mutation in the feature space. A higher degree of concentration in the distribution range of gene mutations of the same category in the feature space indicates a smaller distance between features of the same category in the feature space, and higher intra-class compactness. Conversely, a higher degree of separation in the distribution range of gene mutations of different categories in the feature space indicates a larger distance between features of different categories in the feature space, and higher inter-class separability. Thus, by introducing a convex hull geometric constraint strategy and using a convex hull geometric constraint loss function during the training process of the gene mutation category prediction model, the distribution range of gene mutations of the same category in the feature space can be constrained during training, enhancing the compactness of intra-class feature distribution and expanding the geometric interval between the distributions of gene mutations of different categories in the feature space. This allows features of different categories of gene mutations to form structurally clear and geometrically separable convex hull regions in the feature space, thereby improving the stability and robustness of the classification decisions of the trained gene mutation category prediction model.

[0067] In one possible implementation, the loss function used during the training of the gene variant category prediction model further includes a contrastive learning loss function, which is used to indicate the degree of consistency of features corresponding to the same gene variant and the degree of difference between features corresponding to different gene variants.

[0068] The consistency of features corresponding to the same gene variant indicates the closeness of features of the same gene variant in the feature space across different views. Different views may include different modal combinations or different feature perturbation conditions. For example, features of the same gene variant across different views may include multimodal fusion features obtained by concatenating local mutation features, global statistical mutation features, and textual semantic features of any gene variant, or multimodal fusion features obtained by concatenating local mutation features, global statistical mutation features, and textual semantic features with added random features. The difference between features corresponding to different gene variants indicates the degree of separation of features of different gene variants in the feature space. For example, features of different gene variants may include multimodal fusion features of different gene variants. Thus, a contrastive learning strategy is introduced during training, using a contrastive loss function to optimize the feature space, thereby enhancing the consistency of multimodal fusion features of the same gene variant under different modal combinations or different feature perturbation conditions, while simultaneously widening the distance between multimodal fusion features of different gene variants in the feature space.

[0069] In one possible implementation, predicting the category of the target gene mutation based on the extracted features includes: fusing local mutation features, global statistical mutation features, and textual semantic features of the target gene mutation to generate multimodal fusion features; and predicting the category of the target gene mutation based on the multimodal fusion features, or based on the multimodal fusion features and the local mutation features, global statistical mutation features, and textual semantic features of the target gene mutation.

[0070] For example, a trained gene variant category prediction model can be used to extract features from the multimodal data corresponding to the target gene variant, generating multimodal features such as local mutation features, global statistical mutation features, and textual semantic features of the target gene variant. These multimodal features are then fused to generate multimodal fused features. The specific process of extracting local mutation features, global statistical mutation features, and textual semantic features of the target gene variant is described in detail below.

[0071] For example, the gene variation category prediction model is configured with a multimodal feature fusion module, which can fuse local mutation features, global statistical mutation features, and textual semantic features. The specific type and structure of this multimodal feature fusion module can be configured according to requirements and is not limited thereto. For example, local mutation features, global statistical mutation features, and textual semantic features can be mapped to a feature space and then concatenated to construct a multimodal fused feature, which can be represented as follows:

[0072]

[0073] in, Indicates multimodal fusion features, Local mutation characteristics representing target gene variations. Global statistical mutation characteristics representing target gene variations. Textual semantic features representing target gene variations ( () represents the concatenation function. Thus, when fusing multimodal features, a feature-level fusion method is adopted. First, the features of different modalities are mapped to the feature space respectively, and then concatenated to obtain multimodal fused features.

[0074] For example, the multimodal feature fusion module can also be configured with a projection network, which may include a multi-layer fully connected structure and a nonlinear activation function. The projection network can be used to perform nonlinear mapping and normalization on the fused features after mapping multimodal features to the feature space and concatenating them, generating the final multimodal fused features. Thus, when fusing multimodal features, a feature-level fusion approach is adopted. First, the features of different modalities are mapped to the feature space separately and then concatenated. Then, a nonlinear mapping is achieved through the projection network, thereby obtaining the multimodal fused features.

[0075] Furthermore, the category of target gene mutation can be predicted based on multimodal fusion features; or, the category of target gene mutation can be predicted based on multimodal fusion features, local mutation features of target gene mutation, global statistical mutation features of target gene mutation, and textual semantic features of target gene mutation.

[0076] As an example, the gene variant category prediction model is configured with a stacked classifier that processes multimodal fusion features, local mutation features, global statistical mutation features, and textual semantic features to predict the category of a target gene variant. Exemplarily, the stacked classifier processes these features and can output the predicted category of the target gene variant, such as pathogenic or benign; or, it can output the predicted probability that the target gene variant is pathogenic or benign.

[0077] For example, a stacked classifier may include multiple base classifiers and a meta-classifier. The multiple base classifiers include: a classifier for processing multimodal fusion features, a classifier for processing local mutation features, a classifier for processing global statistical mutation features, and a classifier for processing textual semantic features. Each base classifier processes the features of the corresponding target gene variant and outputs a predicted probability that the target gene variant is pathogenic or benign. The meta-classifier takes the predicted probabilities output by each base classifier as input, performs joint judgment, and outputs a final predicted probability that the target gene variant is pathogenic or benign. In this way, by using a stacked classifier to comprehensively discriminate between multimodal fusion features and individual monomodal features, multi-level feature information is fused, resulting in a more accurate prediction of the target gene variant's category, i.e., whether the target gene variant is pathogenic or benign.

[0078] As another example, a gene variant category prediction model is configured with a multimodal classifier to predict the category of a target gene variant by processing multimodal fusion features. Exemplarily, the multimodal classifier processes the multimodal fusion features and can output the predicted category of the target gene variant, such as pathogenic or benign; or, it can output the predicted probability that the target gene variant is pathogenic or benign.

[0079] The specific types and structures of the base classifiers, meta classifiers, and multimodal classifiers in the stacked classifiers can be flexibly set according to requirements. For example, they can be MLP (Multilayer Perceptron), logistic regression models, fully connected neural network models, other linear or nonlinear classification models, etc.

[0080] In this embodiment, multimodal data corresponding to a target gene mutation is acquired. The multimodal data includes the target mutation DNA sequence, the target reference DNA sequence, and target semantic text information. The target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description. A trained gene mutation category prediction model is used to extract features from the multimodal data, and the category of the target gene mutation is predicted based on the extracted features. The loss functions used in training the gene mutation category prediction model include a convex hull geometric constraint loss function and a classification loss function. The classification loss function indicates the difference between the predicted category and the true category, and the convex hull geometric constraint loss function indicates the degree of concentration of the distribution range of gene mutations of the same category in the feature space and the degree of separation of the distribution ranges of gene mutations of different categories in the feature space. This approach extracts features from multimodal data, including target variant DNA sequences, target reference DNA sequences, and target semantic text information, fully exploring the complementary relationships between heterogeneous information from multiple sources. This improves the completeness of feature expression while effectively overcoming the limitations of single-modal information representation. Furthermore, a convex hull geometric constraint strategy is introduced during the training phase to geometrically regularize the distribution structure of different types of gene variants in the feature space. This enables automated, efficient, and high-precision prediction of gene variant categories, such as single nucleotide mutations, insertions / deletions, and other genetic variations, providing auxiliary decision support for clinical genetics and precision medicine. In some examples, a contrastive learning strategy is also introduced during training to further enhance the ability to discriminate gene variant categories, predictive stability, and generalization performance.

[0081] The following provides an exemplary description of possible implementation methods for feature extraction of the multimodal data in step 102 above.

[0082] Figure 2 A flowchart illustrating feature extraction from multimodal data according to an embodiment of the present disclosure is shown, such as... Figure 2 As shown, it includes the following steps:

[0083] Step 201: Extract features from the target reference DNA sequence and the target variant DNA sequence to obtain the local mutation features of the target gene variant; wherein, the local mutation features represent the influence of the gene variant on the local structure of the DNA sequence.

[0084] Among them, local mutation features can retain rich contextual information and can effectively characterize the impact of gene variations on the local structure of DNA sequences. For example, local mutation features can characterize the local structural changes in DNA sequences caused by mutations such as base substitution, insertion, or deletion.

[0085] In this step, by extracting features from the target reference DNA sequence and the target variant DNA sequence respectively, the local mutation features of the target gene variant can be obtained. The local mutation features of the target gene variant can represent the influence of the target gene variant on the local structure of the DNA sequence.

[0086] In one possible implementation, the step of extracting features from the target reference DNA sequence and the target variant DNA sequence to obtain the local mutation features of the target gene variant includes: extracting local pattern features of the target reference DNA sequence and local pattern features of the target variant DNA sequence, respectively; wherein the local pattern features represent the local structure of the DNA sequence; and determining the local mutation features of the target gene variant based on the local pattern features of the target reference DNA sequence and the local pattern features of the target variant DNA sequence.

[0087] For example, the gene variation category prediction model can be configured with a sequence local coding module, which can extract local pattern features from the DNA sequence. The specific type and structure of this sequence local coding module can be configured according to requirements and is not limited thereto. For example, it can be a sequence local encoder based on a one-dimensional convolutional neural network. As an example, the sequence local coding module can be a one-dimensional convolutional neural network encoder with shared parameters. The encoder performs feature encoding on the target reference DNA sequence and the target variant DNA sequence respectively, extracting local pattern features of the DNA sequence. The feature encoding process of this encoder is shown in the following formula:

[0088]

[0089] in, Indicates the target reference DNA sequence; Indicates the target mutated DNA sequence; Indicates local pattern features of the target reference DNA sequence; This represents the local pattern features of the target variant DNA sequence.

[0090] Furthermore, to highlight the impact of target gene mutations on the local sequence structure of DNA sequences, an embedding difference strategy can be employed. Based on the local pattern features of the target reference DNA sequence and the local pattern features of the target variant DNA sequence, local mutation features of the target gene mutation can be constructed, as shown in the following equation:

[0091]

[0092] in, Local mutation characteristics representing target gene variations. Indicates local pattern features of the target reference DNA sequence; This represents the local pattern features of the target variant DNA sequence.

[0093] In this way, the sequence local encoder based on a one-dimensional convolutional neural network in the gene variation category prediction model encodes the features of the target reference DNA sequence and the target variant DNA sequence respectively, obtaining the local pattern features of the target reference DNA sequence and the local pattern features of the target variant DNA sequence. The local structural changes in the DNA sequence caused by the target gene variation are then characterized by embedding difference between the two. This approach effectively captures the impact of mutations such as single nucleotide substitutions, insertions, or deletions on the upstream and downstream sequence structures of the variant site, avoiding the loss of contextual information caused by modeling only a single variant site, thereby improving the expressive power of the local mutation features of the target gene variation.

[0094] Step 202: Extract features from the target reference DNA sequence and the target variant DNA sequence to obtain the global statistical mutation features of the target gene variant; wherein, the global statistical mutation features represent the influence of the gene variant on the overall structure of the DNA sequence.

[0095] Global statistical mutation features can characterize the changes in the compositional patterns of DNA sequences before and after gene mutations at a global level, thus effectively reflecting the impact of gene mutations on the overall structure of DNA sequences. Compared with local mutation features, global statistical mutation features focus more on describing macroscopic structural changes in DNA sequences, and the two complement each other at the feature level.

[0096] In this step, by extracting features from the target reference DNA sequence and the target variant DNA sequence respectively, the global statistical mutation features of the target gene variant can be obtained. The global statistical mutation features of the target gene variant can represent the impact of the target gene variant on the overall structure of the DNA sequence.

[0097] In one possible implementation, the step of extracting features from the target reference DNA sequence and the target variant DNA sequence to obtain the global statistical mutation features of the target gene variant includes: extracting the global statistical features of the target reference DNA sequence and the global statistical features of the target variant DNA sequence based on a preset subsequence length; wherein the global statistical features represent the overall compositional pattern of the DNA sequence; and determining the global statistical mutation features of the target gene variant based on the global statistical features of the target reference DNA sequence and the global statistical features of the target variant DNA sequence.

[0098] Among them, global statistical features can describe the compositional rules of DNA sequences from an overall perspective, complementing local pattern features at the information level, thereby helping to enhance the comprehensive modeling ability of gene variation category prediction models for complex sequences.

[0099] For example, the gene variation category prediction model can be configured with a global statistical feature extraction module, which can perform global statistical feature extraction on the DNA sequence. The specific type and structure of this global statistical feature extraction module can be configured according to needs and are not limited thereto. As an example, it can be a global statistical feature extraction module that introduces the k-mer natural vector representation method; where k-mer natural vector representation method is a mathematical method to convert a DNA sequence into a numerical vector of a specific length, k-mer represents a nucleotide string (i.e., subsequence) of length k, and k is the preset subsequence length, the specific value of k can be flexibly set according to needs. Using the k-mer natural vector representation method, based on the preset subsequence length k, by performing a sliding window scan on the DNA sequence, all possible k-mer types are counted to obtain a type set. For any k-mer type in the type set The corresponding natural vector components are shown in the following equation:

[0100]

[0101] in, k-mer The natural vector components, express The number of times k-mers appear in the DNA sequence. express The average value of k-mers across all occurrences in the DNA sequence. express The normalized variance (i.e., distribution variance) of the k-mer occurrence positions relative to the mean position is used to characterize The degree of dispersion of k-mer distribution in DNA sequence.

[0102] For example, extracting the global statistical features of the target reference DNA sequence and the target variant DNA sequence based on a preset subsequence length can include: using a k-mer natural vector representation method, extracting the global statistical features of the target reference DNA sequence and the target variant DNA sequence based on a preset subsequence length k. Specifically, based on the preset subsequence length k, a sliding window scan can be performed on the target variant DNA sequence and the target reference DNA sequence to obtain all subsequences (i.e., k-mers) of length k in the target variant DNA sequence and all subsequences (i.e., k-mers) of length k in the target reference DNA sequence; for all k-mers in the target variant DNA sequence, the possible k-mer types are counted, and the natural vector components of each k-mer type corresponding to the target variant DNA sequence are obtained; correspondingly, for all k-mers in the reference variant DNA sequence, the possible k-mer types are counted, and the natural vector components of each k-mer type corresponding to the reference variant DNA sequence are obtained. Furthermore, for the target variant DNA sequence, the corresponding k-mer natural vector components are spliced ​​together in a predetermined order (such as the order of a sliding window scan) to obtain the global statistical features of the target variant DNA sequence; for the target reference DNA sequence, the corresponding k-mer natural vector components are spliced ​​together in a predetermined order to obtain the global statistical features of the target reference DNA sequence, as shown in the following formula:

[0103]

[0104] in, Indicates the target reference DNA sequence; Indicates the target mutated DNA sequence; Represents the global statistical characteristics of the target variant DNA sequence; This represents the global statistical characteristics of the target reference DNA sequence. express Concatenate all corresponding k-mer type natural vector components. express Concatenate all the corresponding k-mer type natural vector components.

[0105] Furthermore, to highlight the impact of the target gene variation on the overall statistical structure of the DNA sequence, a differential modeling strategy can be adopted to construct the global statistical mutation features of the target gene variation based on the global statistical features of the target reference DNA sequence and the global statistical features of the target variant DNA sequence, as shown in the following formula:

[0106]

[0107] in, Global statistical mutation characteristics representing the variation in the target gene; Represents the global statistical characteristics of the target variant DNA sequence; This represents the global statistical characteristics of the target reference DNA sequence.

[0108] Thus, by employing the k-mer natural vector representation method, the statistical characteristics such as the frequency, mean position, and variance of different types of k-mers in the target variant DNA sequence are characterized, forming a global statistical feature reflecting the overall statistical properties of the target variant DNA sequence. Similarly, the statistical characteristics such as the frequency, mean position, and variance of different types of k-mers in the target reference DNA sequence are also characterized, forming a global statistical feature reflecting the overall statistical properties of the target reference DNA sequence. Furthermore, by differencing these two methods, the global statistical mutation characteristics of the target gene mutation are obtained. These global statistical mutation characteristics can characterize the changes in the k-mer statistical structure of the DNA sequence before and after the target gene mutation at a global level, reflecting the impact of the target gene mutation on the overall structure of the DNA sequence.

[0109] Step 203: Process the target semantic text information to obtain the text semantic features of the target gene mutation.

[0110] The target semantic text information may include biological information such as gene function descriptions, chromosome numbers, and gene names related to the target gene variation. In this step, the target semantic text information is processed to obtain the textual semantic features of the target gene variation, thereby enhancing the functional understanding ability of the gene variation category prediction model.

[0111] For example, the gene variation category prediction model can be configured with a semantic text encoding module to encode semantic text information. The specific type and structure of the semantic text encoding module can be configured according to requirements and is not limited thereto. For example, it can be a semantic text encoding module based on a pre-trained language model (such as the Qwen / Qwen3-Embedding-0.6B model), wherein the pre-trained language model can be a general language model or a domain model fine-tuned by biomedical corpus, etc.

[0112] As an example, the semantic text encoding module is for language models. The target semantic text information is embedded and encoded to obtain text semantic features. The encoding process is shown in the following formula:

[0113]

[0114] in, Represents the target semantic text information. Textual semantic features representing target gene mutations.

[0115] In this way, language models are used to encode target semantic text information related to the target gene variation. By using embedding encoding, functional annotations and biological knowledge related to gene variation are mapped into textual semantic features, introducing domain knowledge and biological semantic constraints, effectively compensating for the shortcomings of relying solely on DNA sequence features in functional interpretation and expression of pathogenic mechanisms.

[0116] In this embodiment, feature extraction is performed on multimodal data such as target variant DNA sequence, target reference DNA sequence, and target semantic text information. The complementary relationship between multi-source heterogeneous information is fully explored, which improves the completeness of feature expression and effectively overcomes the problem of limited expression ability of single-modal information. Multimodal features such as local mutation features, global statistical mutation features, and text semantic features that can reflect the changes in DNA sequence before and after the target gene mutation are generated. These features can then be used to effectively improve the accuracy of prediction results in complex real data environments.

[0117] The following is a summary of the above. Figure 1 The training process of the gene variation category prediction model is illustrated by example.

[0118] Figure 3 This diagram illustrates a training flowchart for a gene variation category prediction model according to an embodiment of the present disclosure, such as... Figure 3 As shown, the following steps may be included:

[0119] Step 301: Obtain training data, which includes training samples and category labels for multiple gene variants. Each training sample includes multimodal data such as reference DNA sequence samples, variant DNA sequence samples, and semantic text information samples. The category label for each training sample is pathogenic or benign.

[0120] Each training sample corresponds to a specific gene variant; training samples with the category label of pathogenicity can be called pathogenic variant samples, and training samples with the category label of benign can be called benign variant samples.

[0121] For example, the training data can be divided into multiple batches, where each batch contains B training samples. The specific value of B and the number of batches can be set according to requirements.

[0122] For example, training data can be obtained from public or private databases.

[0123] Step 302: Use the training data to train the gene variation category prediction model to obtain a trained gene variation category prediction model.

[0124] For example, a batch training method can be used to iteratively train the gene variant category prediction model until the training termination condition is met, thereby obtaining a trained gene variant category prediction model. The training termination condition may include: the loss function reaching a preset threshold, the loss function converging, the training reaching a preset number of iterations, etc.

[0125] In one possible implementation, the method further includes: during the training process of the gene variation category prediction model, for any batch of training samples, extracting local mutation features, global statistical mutation features, and textual semantic features of each training sample in the batch of training samples through the gene variation category prediction model; wherein each training sample includes: a reference DNA sequence sample, a variant DNA sequence sample, and a semantic text information sample.

[0126] The specific implementation process for extracting local mutation features, global statistical mutation features, and text semantic features from training samples can be referred to the above. Figure 2 The relevant statements are as described above; among them, the extraction of local mutation features can be referred to the above. Figure 2 Step 201, extraction of global statistical mutation features, can be referred to the above. Figure 2 Step 202, the extraction of text semantic features, can be referred to the above. Figure 2 Step 203.

[0127] For example, a sequence local encoder based on a one-dimensional convolutional neural network can be constructed in a gene variation category prediction model. For any training sample, feature encoding is performed on the reference DNA sequence sample and the variant DNA sequence sample respectively to generate local pattern features of the reference DNA sequence sample and the variant DNA sequence sample. Then, the local structural changes of the DNA sequence caused by the mutation corresponding to the training sample are characterized by embedding difference between the two. This can effectively capture the impact of mutations such as single nucleotide substitution, insertion or deletion on the upstream and downstream sequence structures of the variant site, avoid the problem of loss of contextual information caused by modeling only a single variant site, and thus improve the expression ability of local mutation features.

[0128] For example, a global statistical feature extraction module incorporating the k-mer natural vector representation method can be constructed within a gene variation category prediction model. For any training sample, statistical features such as the frequency, mean position, and variance of different types of k-mers in the variant DNA sequence sample are characterized, forming a global statistical feature reflecting the overall statistical characteristics of the variant DNA sequence sample. Similarly, statistical features such as the frequency, mean position, and variance of different types of k-mers in the reference DNA sample are characterized, forming a global statistical feature reflecting the overall statistical characteristics of the reference DNA sequence sample. Furthermore, by differencing these two features, the global statistical mutation features of the gene variation corresponding to the training sample are obtained. These global statistical mutation features can characterize the changes in the k-mer statistical structure of the DNA sequence before and after the gene variation at a global level, reflecting the impact of the gene variation on the overall structure of the DNA sequence.

[0129] For example, a language model can be constructed in the gene variation category prediction model; for any training sample, the semantic text information sample is encoded using the language model to obtain text semantic features; thereby, through the embedding encoding method, the functional annotations and biological knowledge related to the gene variation corresponding to the training sample are mapped into text semantic features, introducing domain knowledge and biological semantic constraints, effectively making up for the shortcomings of simply relying on DNA sequence features in functional explanation and expression of pathogenic mechanisms.

[0130] In one possible implementation, the method further includes: during the training process of the gene variation category prediction model, for any batch of training samples, the function value of the loss function can be calculated based on the local mutation features, global statistical mutation features, and text semantic features of each training sample in the batch of training samples, and then the gradient descent method can be used to perform backpropagation and update the parameters of the gene variation category prediction model.

[0131] The loss functions used in the training process of the gene variant category prediction model include: convex hull geometric constraint loss function and classification loss function; the classification loss function is used to indicate the difference between the predicted category and the true category, and the convex hull geometric constraint loss function is used to indicate the degree of concentration of the distribution range of gene variants of the same category in the feature space and the degree of separation of the distribution range of gene variants of different categories in the feature space.

[0132] In one possible implementation, for any batch of training samples, the local mutation features, global statistical mutation features, and textual semantic features of each training sample in the batch are fused to generate corresponding fused features. Then, based on the corresponding fused features of each training sample, the function values ​​of the convex hull geometric constraint loss function and the classification loss function can be generated. Here, the fused features are multimodal fused features.

[0133] For example, a multimodal feature fusion module can be constructed in a gene variation category prediction model. For any training sample, the local mutation features, global statistical mutation features, and textual semantic features of the training sample can be fused to generate the fused features of the training sample. For instance, a feature-level fusion approach can be used, first mapping the features of different modalities of the training sample (local mutation features, global statistical mutation features, and textual semantic features) to the feature space respectively, and then concatenating them to obtain the fused features of the training sample.

[0134] As an example, for any training sample, the category of the corresponding gene mutation can be predicted based on the fusion features of the training sample; or, the category of the corresponding gene mutation can be predicted based on the fusion features, local mutation features, global statistical mutation features, and textual semantic features of the training sample. Furthermore, the classification loss function can be calculated based on the predicted category and category label of each training sample in the batch of training samples.

[0135] For example, a stacked classifier can be configured in a gene variation category prediction model to output the predicted category of each training sample. For any given training sample, its fusion features, local mutation features, global statistical mutation features, and textual semantic features can be input into the stacked classifier to output the prediction result.

[0136]

[0137] in, Represents a stacked classifier; Features representing the input training samples; This represents the predicted probability that the gene mutation corresponding to the training sample is pathogenic or benign, and its value can be between 0 and 1.

[0138] The classification loss function used is the cross-entropy loss function, as shown in the following formula:

[0139]

[0140] in, This represents the classification loss function, where i represents the number of categories, which is set to 2 here, i.e., pathogenic or benign. This represents the predicted probability that the gene variant corresponding to the training sample is pathogenic or benign. This represents the class label of the training samples; for example, the class label for pathogenic variant samples is 1, and the class label for benign variant samples is 0. For any batch of training samples, the classification loss function value for that batch of training samples can be obtained based on this formula.

[0141] In this way, a multimodal feature fusion strategy is adopted during training. The category of gene variation is predicted by the multimodal features of the training samples, and the function value of the classification loss function is calculated to update the model parameters. As a result, the difference between the predicted category and the true category is continuously reduced during training, thereby improving the classification accuracy, feature discriminativeness and stability of the model.

[0142] As another example, for any batch of training samples, the function value of the convex hull geometric constraint loss function is calculated based on the fusion features of each training sample in that batch. The result is shown in the following equation:

[0143]

[0144] in, Represents the loss function for convex hull geometric constraints; Indicates intra-class compactness loss. This represents the inter-class separability loss. This represents a smoothing factor or a division-to-zero constant, and can usually be a very small positive number. This represents the set of benign variant samples in this batch of training samples. This indicates the number of training samples in the set of benign variant samples; This represents the set of pathogenic variant samples in the batch of training samples. This indicates the number of training samples in the pathogenic variant sample set; This represents the fusion features of the training samples. express It is a fusion feature of a benign variant sample; express It is a fusion feature of a pathogenic variant sample. The centroid represents the center of the feature space of the fusion feature of the set of benign variant samples. The centroid represents the center of the feature space of the fusion feature of the pathogenic variant sample set. , . The squared Euclidean distance from the fusion features of benign variant samples to the fusion features of the set of benign variant samples at the center of the feature space; The squared Euclidean distance between the fusion features of pathogenic variant samples and the fusion features of the set of pathogenic variant samples at the center of the feature space; This represents the square of the Euclidean distance between the centers of the fusion features of the benign variant sample set and the centers of the fusion features of the pathogenic variant sample set in the feature space. For any batch of training samples, the value of the convex hull geometric constraint loss function can be obtained based on this formula.

[0145] To further alleviate the problems of sample class imbalance and feature space aliasing, a convex hull geometric constraint mechanism is introduced during training to geometrically regularize the distribution structure of the fusion features of benign and pathogenic variant sample sets in the feature space. The convex hull geometric constraint enhances the compactness of intra-class feature distribution and expands the geometric interval between the fusion feature distributions of different classes of training samples by constraining the distribution range of fusion features of samples of the same class in the feature space. This allows the fusion features of different classes of training samples to form clearly structured and geometrically separable convex hull regions in the feature space. During training, the function value of the convex hull geometric constraint loss function is calculated to update the model parameters, constraining the feature space. This gradually increases the concentration of the distribution range of gene variants of the same class in the feature space and the separation of the distribution ranges of gene variants of different classes in the feature space, i.e., continuously improving intra-class compactness and inter-class separation. This enhances the stability and robustness of the model's classification decisions.

[0146] In one possible implementation, the loss function used during the training of the gene variant category prediction model further includes a contrastive learning loss function, which indicates the degree of consistency of features corresponding to the same gene variant and the degree of difference between features corresponding to different gene variants. Thus, introducing a contrastive learning strategy during training further enhances the model's ability to model the feature consistency of the same gene variant under different views.

[0147] In one possible implementation, for any batch of training samples, the local mutation features, global statistical mutation features, and textual semantic features of each training sample in the batch are concatenated to generate corresponding first fusion features; random features are added to the textual semantic features of each training sample, and the local mutation features, global statistical mutation features, and textual semantic features with added random features are concatenated to generate corresponding second fusion features; positive and negative samples corresponding to each fusion feature in the fusion feature set are determined, the fusion feature set including the first fusion features and the second fusion features; wherein, for any fusion feature, the positive sample corresponding to the fusion feature includes the fusion feature and the fusion features corresponding to the same training sample, and the negative sample corresponding to the fusion feature includes the remaining fusion features in the fusion feature set; based on the positive and negative samples corresponding to each fusion feature, the function value of the contrastive learning loss function is calculated.

[0148] The methods for generating the first and second fusion features can be referenced from the previous descriptions of generating multimodal fusion features. For example, a projection network can be configured in the multimodal feature fusion module. For any training sample, a feature-level fusion method can be used. First, the features of different modalities of the training sample are mapped to the feature space and then concatenated. Then, the projection network is used to achieve nonlinear mapping and normalization processing, thereby obtaining the fusion features of the training sample. As shown in the following equation:

[0149]

[0150] in, This represents the fusion features of the training samples. This represents the concatenated features of different modalities in the training samples. This represents the local mutation characteristics of the training samples. This represents the global statistical mutation characteristics of the training samples. The textual semantic features representing the training samples. ( ) represents the concatenation function. This represents a projection network used to improve the discriminativeness of feature representations in the contrast space.

[0151] The above formula can be used to concatenate the local mutation features, global statistical mutation features, and textual semantic features of any training sample to generate the first fusion feature of that training sample; and concatenate the local mutation features, global statistical mutation features, and textual semantic features with added random features to generate the second fusion feature of that training sample. In this way, to enhance the consistency of fusion features under different modal combinations or different feature perturbation conditions for the same gene variant, and to increase the feature distance between different gene variants, two different views of fusion features are constructed based on the multimodal features of the same training sample. The first view is obtained by concatenating the local mutation features, global statistical mutation features, and textual semantic features (i.e., the first fusion feature); the second view, while keeping the local mutation features and global statistical mutation features unchanged, applies random feature masking perturbation to the textual semantic features before feature concatenation, thereby simulating the fusion feature under conditions of missing modal information or noise interference (i.e., the second fusion feature).

[0152] After obtaining the fused features of two different views, positive and negative samples can be further identified for comparative learning. For example, in a batch of training samples, let... Each training sample generates a fusion feature of the two different views described above, denoted as _____. and Therefore, a total of [number] samples can be generated throughout the entire batch of training samples. Let the fusion features be denoted as the fusion feature set. For any fusion feature The fusion features of the other view (i.e., the fusion features corresponding to the same training sample) constitute a positive sample pair, denoted as . ; while in the fusion feature set except and its positive samples All fused features other than those specified are used as negative samples in the contrastive learning process (no explicit construction of negative sample pairs is required). Under temperature parameter adjustment, the contrastive learning loss function is as follows:

[0153]

[0154] in, This represents the contrastive learning loss function; This represents the temperature parameter, and its value can be set according to requirements. Represents the set of fused features; This represents the i-th fusion feature in the fusion feature set. Indicating the fusion feature set External fusion features (negative samples). Indicates and This corresponds to the fused features (positive samples) of the same training sample. For any batch of training samples, the function value of the contrastive learning loss function for that batch of training samples can be obtained based on this formula.

[0155] To further enhance the model's ability to model the consistency of features of the same gene variant across different views, a contrastive learning strategy is introduced during training. This involves constructing fused features from multiple views of the same training sample as positive sample pairs and using fused features from different training samples as negative samples. A contrastive loss function is then used to optimize the feature space, making the fused features of the same training sample more similar across multiple views, while increasing the separation between fused features from different training samples. This effectively enhances the consistency of fused features of the same gene variant across different views, while also improving the separability of fused features of different gene variants in the feature space.

[0156] Furthermore, the total loss function can be obtained based on the aforementioned convex hull geometric constraint loss function and classification loss function, or based on the convex hull geometric constraint loss function, classification loss function, and contrastive learning loss function.

[0157] As an example, the total loss function It can be represented as:

[0158]

[0159] in, Represents the classification loss function. Represents the loss function for convex hull geometric constraints. This represents the contrastive learning loss function, where α and β are the weight coefficients of the convex hull geometric constraint loss function and the contrastive learning loss function, respectively. Their specific values ​​can be configured according to requirements. For any batch of training samples, the total loss function value for that batch can be obtained based on this formula. This value can then be used for backpropagation to adjust the model parameters.

[0160] In this embodiment, during model training, local pattern features, global statistical features, and semantic text features are simultaneously introduced to fully explore the complementary relationships between multi-source heterogeneous information. Relevant features (local mutation features, global statistical mutation features, and textual semantic features) characterizing DNA sequence changes before and after gene mutations are generated, enhancing the model's sensitivity to the impact of gene mutations. Furthermore, convex hull geometric constraints are introduced to geometrically regulate and constrain the distribution structure of samples of different gene mutation categories in the feature space, enabling pathogenic and benign mutations to form clearly structured and stable separable regions in the feature space. This effectively alleviates class imbalance and feature aliasing problems, improving the model's generalization performance. In some examples, a contrastive learning strategy is also introduced during training, using a contrastive loss function to optimize the feature space. This further enhances the consistency of multimodal fusion features of the same gene mutation under different modal combinations or different feature perturbations, while simultaneously widening the distance between multimodal fusion features of different gene mutations in the feature space. Thus, by introducing multimodal contrastive learning and convex hull geometric constraints into the feature space, the model's ability to discriminate the pathogenicity of gene mutations, its prediction stability, and its generalization performance are improved.

[0161] The following section uses a gene variation dataset to verify the performance of the gene variation category prediction model in the above embodiments.

[0162] For example, a gene variant dataset containing 196 pathogenic variant samples and 104 benign variant samples can be obtained from a public database for performance validation. A global statistical feature extraction module incorporating 3-mer natural vector representation is used to extract global statistical features from the DNA sequence. The Qwen / Qwen3-Embedding-0.6B model is used to encode textual semantic information related to gene variants, generating textual semantic features. When predicting the category of gene variants based on the extracted features, multimodal and unimodal feature-based MLP classifiers are used, and joint discrimination is performed through a stacked ensemble approach.

[0163] To evaluate the generalization performance of the gene variant category prediction model, this embodiment employs a five-fold cross-validation strategy for training and testing the multimodal fusion features. In the experiments, the predictive performance of multimodal fusion (f1+f2) using local mutation features (f1) and global statistical mutation features (f2) versus multimodal fusion (f1+f2+f3) using local mutation features (f1), global statistical mutation features (f2), and textual semantic features (f3) is compared; as shown in Table 1:

[0164] Table 1: Results of gene variation category prediction test under different feature combinations

[0165] Experimental results show that when using multimodal features f1+f2+f3 for gene mutation category prediction, the average accuracy is 0.8500±0.0901, precision is 0.8332±0.0805, recall is 0.9744±0.0513, and F1 score is 0.8964±0.0598. However, when using only f1+f2 features for gene mutation category prediction, the average accuracy is 0.6000±0.0298, precision is 0.6773±0.0270, recall is 0.7447±0.0466, and F1 score is 0.7083±0.0240. These experimental results demonstrate that introducing textual semantic features can significantly improve the model's ability to discriminate gene mutation categories, fully reflecting the effectiveness of the multimodal feature fusion strategy adopted in the above embodiments of this disclosure in improving classification accuracy, enhancing feature discriminative power, and improving model stability.

[0166] This disclosure also provides a method for predicting gene variation categories based on multimodal feature fusion and convex hull geometric constraints. Figure 4 A flowchart illustrating a gene variation category prediction method according to an embodiment of the present disclosure is shown, as follows: Figure 4The process includes stages such as data acquisition, multimodal feature extraction, multimodal feature fusion, convex hull geometric constraint-supervised training, and discriminative output. First, in the data acquisition stage, for the gene variant to be analyzed, the corresponding reference DNA sequence and variant DNA sequence are acquired, and semantic text information related to the gene variant is acquired simultaneously. Second, in the multimodal feature extraction stage, deep sequence coding models, sequence statistical modeling methods (k-mer natural vector representation), and pre-trained language models are used to encode the features of the above different modalities, obtaining local mutation features (f1), global statistical mutation features (f2), and textual semantic features (f3). Subsequently, in the modal feature fusion stage, the above multimodal features are uniformly mapped and fused to obtain multimodal fused features. Further, a convex hull geometric constraint mechanism is introduced during training to geometrically regularize the distribution structure of different gene variant categories in the feature space. Finally, in the discriminative output stage, a stacked classifier is used to discriminate the multimodal fused features and output the predicted category (pathogenic or benign) of the corresponding gene variant.

[0167] This disclosure also provides a gene variation category prediction system based on multimodal feature fusion and convex hull geometric constraints. Figure 5 This diagram illustrates the structure of a gene variation category prediction system according to an embodiment of this disclosure, as shown below. Figure 5 As shown, it includes: a data acquisition module, a multimodal feature extraction module (local feature representation module, global feature representation module, semantic representation module), a feature fusion module, a convex hull geometric constraint module, and a classification prediction module; the implementation process of each module's function can be referred to the relevant descriptions above, and will not be repeated here.

[0168] Based on the same inventive concept in the above method embodiments, this disclosure also provides a gene variation category prediction device based on convex hull geometric constraints, which can be used to execute the technical solutions described in the above method embodiments.

[0169] Figure 6 This diagram illustrates a structural representation of a gene variation category prediction device based on convex hull geometric constraints according to an embodiment of the present disclosure. Figure 6As shown, the device may include: an acquisition module 601, used to acquire multimodal data corresponding to a target gene variant, the multimodal data including a target variant DNA sequence, a target reference DNA sequence, and target semantic text information; wherein the target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description; and a prediction module 602, used to extract features from the multimodal data using a trained gene variant category prediction model, and predict the category of the target gene variant based on the extracted features; wherein the loss function used in the training process of the gene variant category prediction model includes: a convex hull geometric constraint loss function and a classification loss function; the classification loss function is used to indicate the difference between the predicted category and the true category, and the convex hull geometric constraint loss function is used to indicate the degree of concentration of the distribution range of gene variants of the same category in the feature space and the degree of separation of the distribution range of gene variants of different categories in the feature space.

[0170] In this embodiment, multimodal data corresponding to a target gene mutation is acquired. The multimodal data includes the target mutation DNA sequence, the target reference DNA sequence, and target semantic text information. The target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description. A trained gene mutation category prediction model is used to extract features from the multimodal data, and the category of the target gene mutation is predicted based on the extracted features. The loss functions used in training the gene mutation category prediction model include a convex hull geometric constraint loss function and a classification loss function. The classification loss function indicates the difference between the predicted category and the true category, and the convex hull geometric constraint loss function indicates the degree of concentration of the distribution range of gene mutations of the same category in the feature space and the degree of separation of the distribution ranges of gene mutations of different categories in the feature space. In this way, feature extraction is performed on multimodal data such as target variant DNA sequences, target reference DNA sequences, and target semantic text information, fully exploring the complementary relationships between multi-source heterogeneous information. This not only improves the completeness of feature expression but also effectively overcomes the problem of limited expression capacity of single-modal information. Furthermore, a convex hull geometric constraint strategy is introduced during the training phase to geometrically regularize the distribution structure of different types of gene variants in the feature space. This enables automated, efficient, and high-precision prediction of gene variants such as single nucleotide mutations, insertions, deletions, and other genetic variations, providing auxiliary decision support for clinical genetics and precision medicine.

[0171] In one possible implementation, the loss function used during the training of the gene variant category prediction model further includes a contrastive learning loss function, which is used to indicate the degree of consistency of features corresponding to the same gene variant and the degree of difference between features corresponding to different gene variants.

[0172] In one possible implementation, the prediction module 602 is further configured to: extract features from the target reference DNA sequence and the target variant DNA sequence respectively to obtain local mutation features of the target gene variant; wherein the local mutation features represent the influence of the gene variant on the local structure of the DNA sequence; extract features from the target reference DNA sequence and the target variant DNA sequence respectively to obtain global statistical mutation features of the target gene variant; wherein the global statistical mutation features represent the influence of the gene variant on the overall structure of the DNA sequence; and process the target semantic text information to obtain the textual semantic features of the target gene variant.

[0173] In one possible implementation, the prediction module 602 is further configured to: extract local pattern features of the target reference DNA sequence and local pattern features of the target variant DNA sequence, respectively; wherein the local pattern features represent the local structure of the DNA sequence; and determine the local mutation features of the target gene variant based on the local pattern features of the target reference DNA sequence and the local pattern features of the target variant DNA sequence.

[0174] In one possible implementation, the prediction module 602 is further configured to: extract global statistical features of the target reference DNA sequence and global statistical features of the target variant DNA sequence based on a preset subsequence length; wherein the global statistical features represent the overall compositional pattern of the DNA sequence; and determine the global statistical mutation features of the target gene variant based on the global statistical features of the target reference DNA sequence and the global statistical features of the target variant DNA sequence.

[0175] In one possible implementation, the prediction module 602 is further configured to: fuse the local mutation features, global statistical mutation features, and textual semantic features of the target gene variant to generate a multimodal fusion feature; and predict the category of the target gene variant based on the multimodal fusion feature, or based on the multimodal fusion feature and the local mutation features, global statistical mutation features, and textual semantic features of the target gene variant.

[0176] In one possible implementation, the apparatus further includes a training module, configured to: during the training process of the gene variation category prediction model, for any batch of training samples, extract local mutation features, global statistical mutation features, and textual semantic features of each training sample in the batch of training samples using the gene variation category prediction model; wherein each training sample includes: a reference DNA sequence sample, a variant DNA sequence sample, and a semantic text information sample; concatenate the local mutation features, global statistical mutation features, and textual semantic features of each training sample to generate corresponding first fusion features; add random features to the textual semantic features of each training sample, and... The local mutation features, global statistical mutation features, and text semantic features after adding random features of each training sample are concatenated to generate corresponding second fusion features; the positive and negative samples corresponding to each fusion feature in the fusion feature set are determined, the fusion feature set including each first fusion feature and each second fusion feature; wherein, for any fusion feature, the positive sample corresponding to the fusion feature includes the fusion feature and the fusion feature corresponding to the same training sample, and the negative sample corresponding to the fusion feature includes the remaining fusion features in the fusion feature set; based on the positive and negative samples corresponding to each fusion feature, the function value of the contrastive learning loss function is calculated.

[0177] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0178] This disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0179] This disclosure also provides a non-volatile computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0180] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0181] Figure 7 A block diagram of an electronic device 1900 according to an embodiment of the present disclosure is shown. For example, the electronic device 1900 may be provided as a server or a terminal device. (Refer to...) Figure 7The electronic device 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions, such as application programs, that can be executed by the processing component 1922. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 1922 is configured to execute instructions to perform the methods described above.

[0182] Electronic device 1900 may also include a power supply component 1926 configured to perform power management of electronic device 1900, a wired or wireless network interface 1950 configured to connect electronic device 1900 to a network, and an input / output interface 1958 (I / O interface). Electronic device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM Mac OS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0183] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of an electronic device 1900 to perform the above-described method.

[0184] Computer-readable storage media can be tangible devices capable of holding and storing programs / instructions used by instruction execution devices. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0185] The computer program (or computer-readable program instructions) described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage medium in the respective computing / processing device.

[0186] The computer program (or computer program instructions) used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions to implement various aspects of this disclosure.

[0187] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should 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-readable program instructions.

[0188] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0189] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0190] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0191] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for class prediction of gene variation based on convex hull geometric constraints, characterized in that, include: Obtain multimodal data corresponding to the target gene mutation, wherein the multimodal data includes the target variant DNA sequence, the target reference DNA sequence, and the target semantic text information; wherein the target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description; The multimodal data is feature extracted using a trained gene variant category prediction model, and the category of the target gene variant is predicted based on the extracted features. The loss functions used in training the gene variant category prediction model include a convex hull geometric constraint loss function and a classification loss function. The classification loss function indicates the difference between the predicted category and the true category, while the convex hull geometric constraint loss function indicates the degree of concentration of the distribution range of gene variants of the same category in the feature space and the degree of separation of the distribution ranges of gene variants of different categories in the feature space.

2. The method according to claim 1, characterized in that, The loss function used in the training process of the gene variant category prediction model also includes a contrastive learning loss function, which is used to indicate the degree of consistency of features corresponding to the same gene variant and the degree of difference between features corresponding to different gene variants.

3. The method according to claim 1 or 2, characterized in that, The feature extraction of the multimodal data includes: Feature extraction is performed on the target reference DNA sequence and the target variant DNA sequence to obtain the local mutation features of the target gene variant; wherein, the local mutation features represent the influence of the gene variant on the local structure of the DNA sequence; Feature extraction is performed on the target reference DNA sequence and the target variant DNA sequence to obtain the global statistical mutation features of the target gene variant; wherein, the global statistical mutation features represent the impact of the gene variant on the overall structure of the DNA sequence; The target semantic text information is processed to obtain the textual semantic features of the target gene mutation.

4. The method according to claim 3, characterized in that, The step of extracting features from the target reference DNA sequence and the target variant DNA sequence to obtain the local mutation features of the target gene variant includes: Local pattern features of the target reference DNA sequence and local pattern features of the target variant DNA sequence are extracted respectively; wherein, the local pattern features represent the local structure of the DNA sequence; Based on the local pattern features of the target reference DNA sequence and the local pattern features of the target variant DNA sequence, the local mutation features of the target gene variant are determined.

5. The method according to claim 3, characterized in that, The step of extracting features from the target reference DNA sequence and the target variant DNA sequence to obtain the global statistical mutation features of the target gene variant includes: Based on a preset subsequence length, global statistical features of the target reference DNA sequence and global statistical features of the target variant DNA sequence are extracted respectively; wherein, the global statistical features represent the overall compositional pattern of the DNA sequence; Based on the global statistical characteristics of the target reference DNA sequence and the global statistical characteristics of the target variant DNA sequence, the global statistical mutation characteristics of the target gene variant are determined.

6. The method according to claim 3, characterized in that, The prediction of the target gene variant category based on the extracted features includes: The local mutation features, global statistical mutation features, and textual semantic features of the target gene variation are fused to generate multimodal fusion features; Based on the multimodal fusion features, or based on the multimodal fusion features and the local mutation features, global statistical mutation features, and textual semantic features of the target gene variant, the category of the target gene variant is predicted.

7. The method according to claim 2, characterized in that, The method further includes: During the training process of the gene mutation category prediction model, for any batch of training samples, the gene mutation category prediction model extracts the local mutation features, global statistical mutation features, and textual semantic features of each training sample in the batch of training samples; wherein, each training sample includes: a reference DNA sequence sample, a variant DNA sequence sample, and a semantic text information sample. The local mutation features, global statistical mutation features, and text semantic features of each training sample are concatenated to generate corresponding first fusion features; Random features are added to the text semantic features of each training sample, and the local mutation features, global statistical mutation features, and text semantic features with added random features of each training sample are concatenated to generate corresponding second fusion features. Determine the positive and negative samples corresponding to each fusion feature in the fusion feature set, wherein the fusion feature set includes each first fusion feature and each second fusion feature; wherein, for any fusion feature, the positive sample corresponding to the fusion feature includes the fusion feature and the fusion feature corresponding to the same training sample as the fusion feature, and the negative sample corresponding to the fusion feature includes the remaining fusion features in the fusion feature set; Based on the positive and negative samples corresponding to each fusion feature, the function value of the contrastive learning loss function is calculated.

8. A gene variation category prediction device based on convex hull geometric constraints, characterized in that, include: The acquisition module is used to acquire multimodal data corresponding to the target gene mutation. The multimodal data includes the target mutation DNA sequence, the target reference DNA sequence, and the target semantic text information. The target semantic text information includes at least one of the following: chromosome number, gene name, and gene function description. The prediction module is used to extract features from the multimodal data using a trained gene variant category prediction model, and to predict the category of the target gene variant based on the extracted features. The loss functions used in the training of the gene variant category prediction model include a convex hull geometric constraint loss function and a classification loss function. The classification loss function indicates the difference between the predicted category and the true category, while the convex hull geometric constraint loss function indicates the degree of concentration of the distribution range of gene variants of the same category in the feature space and the degree of separation of the distribution ranges of gene variants of different categories in the feature space.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images