A method and apparatus for training gene data models based on self-supervision and multimodality.

Abstract

This invention provides a method and apparatus for training a gene data model based on self-supervised and multimodal approaches, belonging to the field of model training technology. The method involves obtaining the sequence vector representation of a gene sequence in a sample DNA fragment; obtaining the region vector representation of the regulatory region in the sample DNA fragment; and training based on the matching relationship between the sequence vector representation and the region vector representation to obtain a gene data model. Because the region vector representation of the regulatory region in the DNA fragment is obtained, it can be applied to different cell types, thus improving the applicability of the gene data model.

Description

Technical Field

[0001] This invention relates to the field of model training technology, and in particular to a method and apparatus for training gene data models based on self-supervision and multimodality. Background Technology

[0002] Mining information from gene sequence data is of great significance for many practical health issues, such as disease risk prediction. Current prediction models mainly include the Enformer model and the DNABERT model. The Enformer model primarily uses a widened convolutional neural network and a transformer structure to model gene data, while the DNABERT model treats DNA as a nucleotide sequence and uses BERT to model the sequence. Both models are designed for the same type of cell.

[0003] Therefore, how to establish gene data models for different cell types has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention provides a method and apparatus for training gene data models based on self-supervision and multimodality, which addresses the shortcomings of existing technologies that cannot target different cell types. It enables the establishment of gene data models through training by adding regulatory regions, making them applicable to different cell types.

[0005] This invention provides a method for training gene data models based on self-supervision and multimodality, comprising:

[0006] Obtain the sequence vector representation of gene sequences in a sample DNA fragment;

[0007] Obtain the region vector representation of the regulatory region in the sample DNA fragment;

[0008] Gene data models are obtained by training based on the matching relationship between the sequence vector representation and the region vector representation.

[0009] According to the present invention, a self-supervised and multimodal gene data model training method is provided, wherein training based on the matching relationship between the sequence vector representation and the region vector representation to obtain the gene data model includes:

[0010] Determine the cosine similarity between the sequence vector representation and the region vector representation;

[0011] Based on the cosine similarity, a gene data model representing the matching relationship between the entire gene sequence and each regulatory region is obtained through training.

[0012] According to the self-supervised and multimodal gene data model training method provided by the present invention, before training based on the cosine similarity, the method further includes:

[0013] Maximize the cosine similarity between the correct sequence vector representation and the region vector representation;

[0014] Minimize the cosine similarity between the erroneous sequence vector representation and the region vector representation.

[0015] According to a self-supervised and multimodal gene data model training method provided by the present invention, maximizing the cosine similarity between the correct sequence vector representation and the region vector representation, and minimizing the cosine similarity between the incorrect sequence vector representation and the region vector representation, includes:

[0016] Based on the objective loss function, the cosine similarity between the correct sequence vector representation and the region vector representation is maximized; the cosine similarity between the incorrect sequence vector representation and the region vector representation is minimized.

[0017] According to the self-supervised and multimodal gene data model training method provided by the present invention, the step of obtaining the sequence vector representation of gene sequences in sample DNA fragments includes:

[0018] The sequence vector representation of gene sequences in sample DNA fragments is obtained through BERT Transformer, which includes a masked language model and / or a next-sentence prediction model.

[0019] According to the self-supervised and multimodal gene data model training method provided by the present invention, the sequence vector representation of the gene sequence includes:

[0020] Vector representation of each substring in the gene sequence, vector representation of the target gene segment corresponding to the substring in the gene sequence, and vector representation of the position of the substring;

[0021] Correspondingly, the step of obtaining the sequence vector representation of the gene sequence in the sample DNA fragment through BERT includes:

[0022] Input the vector representation corresponding to each substring, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring into the BERT Transformer. Summate the vector representation corresponding to each substring, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring to form a context vector representation, which is then used as the sequence vector representation.

[0023] According to a self-supervised and multimodal gene data model training method provided by the present invention, the step of obtaining the region vector representation of the regulatory region in the sample DNA fragment includes:

[0024] The region vector representation of the regulatory region in the sample DNA fragment is obtained using the Swing Transformer.

[0025] The present invention also provides a gene data model training device based on self-supervision and multimodality, comprising:

[0026] The first acquisition module is used to acquire the sequence vector representation of the gene sequence in the sample DNA fragment;

[0027] The second acquisition module is used to acquire the region vector representation of the regulatory region in the sample DNA fragment;

[0028] The model training module is used to train the gene data model based on the matching relationship between the sequence vector representation and the region vector representation.

[0029] The present invention also provides a disease prediction method, comprising:

[0030] Obtain the DNA fragment to be tested;

[0031] The DNA fragment to be detected is input into the gene data model, and the disease prediction result is output. The gene data model is trained using the gene data model training method based on self-supervision and multimodality as described above.

[0032] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the self-supervised and multimodal gene data model training method described above.

[0033] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the gene data model training method based on self-supervised and multimodal methods described above.

[0034] The present invention also provides a computer program product, comprising a computer program that, when executed by a processor, implements the steps of the gene data model training method based on self-supervised and multimodal methods described above.

[0035] This invention provides a method and apparatus for training a gene data model based on self-supervised and multimodal approaches. The method involves obtaining the sequence vector representation of a gene sequence in a sample DNA fragment; obtaining the region vector representation of a regulatory region in the sample DNA fragment; and training based on the matching relationship between the sequence vector representation and the region vector representation to obtain a gene data model. Because the region vector representation of the regulatory region in the DNA fragment is obtained, it can be applied to different cell types, thus improving the applicability of the gene data model. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating the gene data model training method based on self-supervision and multimodality provided in an embodiment of the present invention.

[0038] Figure 2 This is a schematic diagram of the structure of a gene data model training device based on self-supervision and multimodality provided in an embodiment of the present invention;

[0039] Figure 3 This is a flowchart illustrating the disease prediction method provided in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0042] The following is combined with Figures 1-4 This invention describes a method and apparatus for training a gene data model based on self-supervision and multimodality.

[0043] BERT: An open-source representation learning model from Google. This model can transform sequential data, such as text sentences, into numerical vectors, which can then be used as input for downstream tasks, such as text classification, achieving excellent performance in these tasks. Self-supervised learning: A machine learning algorithm that does not require data labels. It designs an objective function to train a neural network model by defining a task relevant to the data itself, such as predicting intermediate words given the context of a word in a sentence. Transformer: A deep learning model that utilizes attention mechanisms to enhance model training. It has wide applications in natural language processing, computer vision, and other fields.

[0044] Figure 1 This is a flowchart illustrating the gene data model training method based on self-supervision and multimodality provided in this embodiment of the invention.

[0045] like Figure 1 As shown in the figure, an embodiment of the present invention provides a gene data model training method based on self-supervision and multimodality, comprising the following steps:

[0046] 101. Obtain the sequence vector representation of the gene sequence in the sample DNA fragment.

[0047] Specifically, obtaining the sequence vector representation of the gene sequence in the sample DNA fragment, which is the pre-training of the gene sequence, can be achieved using BERT Transformer, which includes a masked language model and / or a next-sentence prediction model. For this step, which is the gene sequence embedding in pre-training, the input contains three types of vector representations: the vector representation corresponding to each substring (k-mer, a substring of length k) in the gene sequence; the vector representation indicating the target gene fragment corresponding to the substring (k-mer), i.e., the segment vector representation indicating which part of the sequence the k-mer comes from; and the vector representation of the position of the substring (k-mer).

[0048] Specifically, for the three types of vector representations, obtaining the sequence vector representation of the gene sequence in the sample DNA fragment using BERT can be achieved by inputting the vector representation corresponding to each substring, the vector representation of the target gene fragment corresponding to the substring in the gene sequence, and the vector representation of the substring's position into the BERT Transformer. The vector representations of each substring, the target gene fragment corresponding to the substring in the gene sequence, and the vector representation of the substring's position are summed to form a context vector representation, which is then used as the sequence vector representation. After inputting into the BERT-based Transformer model, all three types of vector representations are summed into a single context vector representation. These context vector representations then become the sequence vector representation E corresponding to the gene sequence. seqThe learning process involves using a masked language model and a next-sentence prediction model. For the masked language model, special symbols are used to randomly mask the k-mer part of the input, and then the model is trained to predict the masked part. For the next-sentence prediction model, the model is trained to distinguish whether a given pair of sequences is continuous in the genome based on the input vector representation, thus ultimately learning the sequence vector representation E of the gene sequence. seq .

[0049] 102. Obtain the region vector representation of the regulatory region in the sample DNA fragment.

[0050] Specifically, obtaining the region vector representation of the regulatory region in the sample DNA fragment, i.e., the pre-training of the regulatory region, can be achieved using the Swing Transformer. For the region features in the pre-training, a powerful backbone model (Swin) can be used to extract the region vector representation E. reg The encoder can apply a Swin transformer model pre-trained on the ImageNet dataset to the input regions. During pre-training, the model parameters are not fixed, allowing for continuous updates to learn better region vector representations. In the pre-training setup, each input regulatory region has a corresponding gene sequence, enabling the capture of multimodal data—that is, the matching relationship between regulatory regions and gene sequences. Of course, the representation learning of regulatory regions can also be accomplished using other similar transformer structures.

[0051] 103. Gene data models are obtained by training based on the matching relationship between sequence vector representation and region vector representation.

[0052] Specifically, in order for a gene data model to learn the correspondence between gene sequences and regulatory regions, a mechanism of matching gene sequences and regulatory regions can be used. This means training based on the matching relationship between sequence vector representations and region vector representations to obtain a gene data model. Specifically, this can involve determining the cosine similarity between sequence vector representations and region vector representations; training based on the cosine similarity yields a gene data model representing the matching relationship between the entire gene sequence and each regulatory region. In other words, these cosine similarities are integrated together for training to obtain the matching relationship between the entire gene sequence and each regulatory region.

[0053] To improve the accuracy of the gene data model, before training based on cosine similarity, the following steps are included: optimizing the matching relationship. This can be achieved by maximizing the cosine similarity between the correct sequence vector representation and the region vector representation, and minimizing the cosine similarity between the incorrect sequence vector representation and the region vector representation. Specifically, this can be done by maximizing the cosine similarity between the correct sequence vector representation and the region vector representation, and minimizing the cosine similarity between the incorrect sequence vector representation and the region vector representation, based on a target loss function. The optimization process uses similarity information to measure the degree of matching between gene sequences and regulatory regions using a target loss function. Through this target loss function, the similarity between matching gene sequences and regulatory regions can be maximized, while the similarity between incorrect matches can be minimized. The corresponding information of the entire gene sequence and each regulatory region can be learned simultaneously during pre-training. Therefore, the overall goal of pre-training includes two sequence pre-training tasks and a sequence-region matching task, thereby obtaining the gene data model. Through this training method, the cosine similarity between the correct gene sequence and the regulatory region embedding is maximized, while the cosine similarity between the incorrect gene sequence and the regulatory region embedding is minimized. Intuitively, the entire alignment based on the sequence and each regulatory region is achieved through learning the gene data during pre-training. The expression for the objective loss function is as follows (1):

[0054]

[0055] Where b is the batch size, E seq E is a sequence vector representation. reg Let i be a region vector, and j be natural numbers.

[0056] This invention provides a gene data model training method based on self-supervised and multimodal approaches. It obtains the sequence vector representation of gene sequences in a sample DNA fragment; obtains the region vector representation of regulatory regions within the sample DNA fragment; and trains the model based on the matching relationship between the sequence vector representation and the region vector representation. Because it obtains the region vector representation of regulatory regions in the DNA fragment, it is applicable to different cell types. This invention combines the advanced deep learning model BERT, and for the first time models the interaction between regulatory regions and gene sequences in a multimodal manner, learning better representations through self-supervised learning, thus improving the applicability of the gene data model. It develops a multimodal deep gene data model with good generalization performance and transferability, expanding the modeling scope to functional non-coding regions; it incorporates the interactions between various non-coding regulatory regions; and it provides a new paradigm for analyzing gene mutations in non-coding regions, expanding the interpretation of non-coding region gene detection results for the industry.

[0057] Based on the same general inventive concept, the present invention also provides a gene data model training device based on self-supervision and multimodality. The gene data model training device based on self-supervision and multimodality provided by the present invention will be described below. The gene data model training device based on self-supervision and multimodality described below can be referred to in correspondence with the gene data model training method based on self-supervision and multimodality described above.

[0058] Figure 2 This is a schematic diagram of the structure of a gene data model training device based on self-supervision and multimodality provided in an embodiment of the present invention.

[0059] like Figure 2 As shown in the figure, an embodiment of the present invention provides a gene data model training device based on self-supervision and multimodality, comprising:

[0060] The first acquisition module 21 is used to acquire the sequence vector representation of the gene sequence in the sample DNA fragment;

[0061] The second acquisition module 22 is used to acquire the region vector representation of the regulatory region in the sample DNA fragment;

[0062] The model training module 23 is used to train the gene data model based on the matching relationship between the sequence vector representation and the region vector representation.

[0063] This embodiment provides a gene data model training device based on self-supervised and multimodal approaches. It obtains the sequence vector representation of gene sequences in a sample DNA fragment; obtains the region vector representation of regulatory regions in the sample DNA fragment; and trains the model based on the matching relationship between the sequence vector representation and the region vector representation. Because the region vector representation of the regulatory regions in the DNA fragment is obtained, the model can be applied to different cell types, thus improving its applicability.

[0064] Furthermore, in this embodiment, the model training module 23 is specifically used for:

[0065] Determine the cosine similarity between the sequence vector representation and the region vector representation;

[0066] Based on the cosine similarity, a gene data model representing the matching relationship between the entire gene sequence and each regulatory region is obtained through training.

[0067] Furthermore, the model training module 23 in this embodiment is specifically used for:

[0068] Maximize the cosine similarity between the correct sequence vector representation and the region vector representation;

[0069] Minimize the cosine similarity between the erroneous sequence vector representation and the region vector representation.

[0070] Furthermore, the model training module 23 in this embodiment is specifically used for:

[0071] Based on the objective loss function, the cosine similarity between the correct sequence vector representation and the region vector representation is maximized; the cosine similarity between the incorrect sequence vector representation and the region vector representation is minimized.

[0072] Furthermore, in this embodiment, the first acquisition module 21 is specifically used for:

[0073] The sequence vector representation of gene sequences in sample DNA fragments is obtained through BERT Transformer, which includes a masked language model and / or a next-sentence prediction model.

[0074] Furthermore, the sequence vector representation of the gene sequence in this embodiment includes:

[0075] Vector representation of each substring in the gene sequence, vector representation of the target gene segment corresponding to the substring in the gene sequence, and vector representation of the position of the substring;

[0076] Correspondingly, the first acquisition module 21 is specifically used for:

[0077] Input the vector representation corresponding to each substring, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring into the BERT Transformer. Summate the vector representation corresponding to each substring, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring to form a context vector representation, which is then used as the sequence vector representation.

[0078] Furthermore, in this embodiment, the second acquisition module 22 is specifically used for:

[0079] The region vector representation of the regulatory region in the sample DNA fragment is obtained using the Swing Transformer.

[0080] Based on the same general inventive concept, this application also protects a method for predicting diseases.

[0081] Figure 3 This is a flowchart illustrating the disease prediction method provided in an embodiment of the present invention.

[0082] like Figure 3 As shown in the figure, an embodiment of the present invention provides a disease prediction method, which includes the following steps:

[0083] 301. Obtain the DNA fragment to be tested.

[0084] 302. Input the DNA fragment to be detected into the gene data model and output the disease prediction result, wherein the gene data model is trained using the gene data model training method based on self-supervised and multimodal methods of any of the above embodiments.

[0085] By establishing self-supervised and multimodal gene data models, multimodal mining of gene information can be achieved, leading to a better understanding of disease causes and breakthroughs in drug target discovery and disease risk control. This can be applied to new drug development and basic medical pathology research in research institutions; it can serve sequencing institutions / pharmaceutical companies / hospitals in developing new algorithms (pre-trained models), identifying potential drug targets, and optimizing treatment plans; and it can serve consumers in newborn disease screening, tumor diagnosis and treatment, and genetic diseases. Through testing and validation on multiple downstream tasks, it has achieved excellent performance and demonstrated significant value in disease prediction.

[0086] Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention.

[0087] like Figure 4 As shown, the electronic device may include a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a gene data model training method based on self-supervised and multimodal approaches. This method includes: obtaining sequence vector representations of gene sequences in a sample DNA fragment; obtaining region vector representations of regulatory regions in the sample DNA fragment; and training based on the matching relationship between the sequence vector representations and the region vector representations to obtain a gene data model.

[0088] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the gene data model training method based on self-supervised and multimodal methods provided above. The method includes: obtaining sequence vector representations of gene sequences in a sample DNA fragment; obtaining region vector representations of regulatory regions in the sample DNA fragment; and training based on the matching relationship between the sequence vector representations and the region vector representations to obtain a gene data model.

[0090] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for training a gene data model based on self-supervised and multimodal approaches provided by the methods described above. The method includes: obtaining sequence vector representations of gene sequences in a sample DNA fragment; obtaining region vector representations of regulatory regions in the sample DNA fragment; and training based on the matching relationship between the sequence vector representations and the region vector representations to obtain a gene data model.

[0091] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0092] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training gene data models based on self-supervision and multimodality, characterized in that, include: Obtain the sequence vector representation of gene sequences in a sample DNA fragment; Obtain the region vector representation of the regulatory region in the sample DNA fragment; A gene data model is obtained by training based on the matching relationship between the sequence vector representation and the region vector representation; The acquisition of sequence vector representations of gene sequences in sample DNA fragments includes: The sequence vector representation of gene sequences in sample DNA fragments is obtained through BERT Transformer, wherein BERT includes a masked language model and / or a next-sentence prediction model; The sequence vector representation of the gene sequence includes: the vector representation of each substring in the gene sequence, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring. Correspondingly, the step of obtaining the sequence vector representation of the gene sequence in the sample DNA fragment through BERT includes: Input the vector representation corresponding to each substring, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring into the BERT Transformer. Summate the vector representation corresponding to each substring, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring to form a context vector representation, which is then used as the sequence vector representation.

2. The self-supervised and multimodal gene data model training method according to claim 1, characterized in that, The process of training based on the matching relationship between the sequence vector representation and the region vector representation to obtain a gene data model includes: Determine the cosine similarity between the sequence vector representation and the region vector representation; Based on the cosine similarity, a gene data model representing the matching relationship between the entire gene sequence and each regulatory region is obtained through training.

3. The self-supervised and multimodal gene data model training method according to claim 2, characterized in that, Before training based on the cosine similarity, the method further includes: Maximize the cosine similarity between the correct sequence vector representation and the region vector representation; Minimize the cosine similarity between the erroneous sequence vector representation and the region vector representation.

4. The self-supervised and multimodal gene data model training method according to claim 3, characterized in that, Maximizing the cosine similarity between the correct sequence vector representation and the region vector representation; and minimizing the cosine similarity between the incorrect sequence vector representation and the region vector representation, includes: Based on the objective loss function, the cosine similarity between the correct sequence vector representation and the region vector representation is maximized; the cosine similarity between the incorrect sequence vector representation and the region vector representation is minimized.

5. The self-supervised and multimodal gene data model training method according to claim 1, characterized in that, The process of obtaining the region vector representation of the regulatory region in the sample DNA fragment includes: The region vector representation of the regulatory region in the sample DNA fragment is obtained using the Swing Transformer.

6. A gene data model training device based on self-supervised and multimodal methods, characterized in that, include: The first acquisition module is used to acquire the sequence vector representation of the gene sequence in the sample DNA fragment; The second acquisition module is used to acquire the region vector representation of the regulatory region in the sample DNA fragment; The model training module is used to train the gene data model based on the matching relationship between the sequence vector representation and the region vector representation. The first acquisition module is specifically used to: acquire the sequence vector representation of the gene sequence in the sample DNA fragment through BERT Transformer, wherein BERT includes a masked language model and / or a next sentence prediction model; The sequence vector representation of the gene sequence includes: the vector representation of each substring in the gene sequence, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring. The first acquisition module is specifically used to: input the vector representation corresponding to each substring, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring into the BERT Transformer, and sum the vector representation corresponding to each substring, the vector representation of the target gene segment corresponding to the substring in the gene sequence, and the vector representation of the position of the substring as a context vector representation as a sequence vector representation.

7. A disease prediction method, characterized in that, include: Obtain the DNA fragment to be tested; The DNA fragment to be detected is input into the gene data model, and the disease prediction result is output. The gene data model is trained using the gene data model training method based on self-supervision and multimodality as described in any one of claims 1-5.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the gene data model training method based on self-supervision and multimodality as described in any one of claims 1 to 5.

Images