A method and device for determining the authenticity of a variant, a computer device and a storage medium

Abstract

The application provides a method for judging the authenticity of a variation, comprising: obtaining sequencing data of a target sample; determining a candidate variation set and corresponding variation text features according to the sequencing data; for any candidate variation in the candidate variation set, generating a sequencing read coverage map according to the corresponding sequencing data; generating a first prompt according to the candidate variation, the corresponding variation text features, the sequencing read coverage map and a prompt template; and inputting the first prompt into a target large model to obtain a result of judging the authenticity of the candidate variation. In this scheme, the sequencing information related to the judgment of the variation is integrated and input into the large model, which can be provided to the model for prediction without complex manual processing. The large language model can fully and comprehensively capture the signals of the authenticity of the variation from the sequencing information, thereby improving the accuracy and efficiency of the judgment of the authenticity of the variation, providing reliable variation judgment basis for the fields of biological research, medical diagnosis and the like, reducing false positive results, and assisting the development of related research and application.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer device, and storage medium for determining the authenticity of mutations. Background Technology

[0002] In the field of genetic disease diagnosis, accurately identifying the authenticity of genomic variations is a crucial prerequisite for clinical decision-making. While current next-generation sequencing technologies can detect variations, systematic errors introduced during the sequencing process lead to frequent false positives. Currently, verification primarily relies on Sanger sequencing or qPCR. However, these methods require additional experimental procedures, have long detection cycles, and are costly, severely limiting clinical diagnostic efficiency. Existing bioinformatics alternatives largely depend on manually setting static screening thresholds; these empirical rules suffer from low accuracy and limited versatility. Therefore, there is an urgent need for a method that can efficiently distinguish between true and false variations. Summary of the Invention

[0003] The purpose of this application is to address at least one of the aforementioned technical deficiencies, particularly the problems of low efficiency, high dependence on manual labor, and strong subjectivity in traditional technologies.

[0004] Firstly, this application provides a method for determining the authenticity of a mutation, including:

[0005] Obtain sequencing data of the target sample;

[0006] Based on sequencing data, candidate variant sets and their corresponding variant text features are determined;

[0007] For any candidate variant in the candidate variant set, generate the corresponding sequencing read coverage map based on the sequencing data;

[0008] The first prompt word is generated based on the candidate variants and their corresponding variant text features, sequencing read coverage map, and prompt word template.

[0009] Input the first prompt word into the target large model to obtain the true or false judgment result of the candidate mutation.

[0010] In one embodiment, generating a corresponding sequencing read coverage map based on sequencing data includes:

[0011] Identify the variant regions corresponding to candidate variants from sequencing data;

[0012] A sequencing read coverage map is generated based on the variant regions and sequencing alignment data; the sequencing alignment data is obtained by aligning the sequencing data with a reference genome.

[0013] In one embodiment, the prompt word template includes a role definition domain, a text feature domain, and an image feature domain; the role definition domain is used to specify the role played by the target large model, the text feature domain is used to insert variant text features, and the image feature domain is used to insert sequencing read coverage maps.

[0014] In one embodiment, the variant text features include the location information, type information, homozygosity, naming information, sequencing depth, population frequency, corresponding gene and transcript and whether it is located in a genomic repetitive region tag, the proportion of reads at the mutation site, the number of false positives detected by the variant in the historical database, and various other factors in the quality control software prediction results.

[0015] In one embodiment, the target large model is obtained by fine-tuning a pre-trained large language model, and the fine-tuning process includes:

[0016] Obtain the first training set; the first training set includes multiple first training data, which include training variants and their corresponding training text features, training sequencing read coverage map and discrimination labels;

[0017] The pre-trained large language model is fine-tuned using the first training set to match the discrimination results of the pre-trained large language model with the discrimination labels.

[0018] In one embodiment, the pre-trained large language model is fine-tuned using a first training set to match the discrimination results of the pre-trained large language model with the discrimination labels, including:

[0019] The weights of the image processing network of the pre-trained large language model are frozen, and the training variants and their corresponding training text features in the first training data are used as input to fine-tune the pre-trained large language model according to the first loss function.

[0020] If the pre-trained large language model satisfies the first termination condition, the weights of the image processing network are unfrozen. The training variants and their corresponding training text features and training sequencing read coverage map in the first training data are used as inputs. The pre-trained large language model is then fine-tuned according to the second loss function to obtain the target large model.

[0021] In one embodiment, fine-tuning the pre-trained large language model according to a second loss function to obtain the target large model further includes:

[0022] The first training data with multiple correct discriminant labels are perturbed to obtain a second training set; the second training set includes multiple second training data, which are obtained by perturbing the training text features and / or training coverage sequencing maps of the first training data with errors and modifying the discriminant labels to be incorrect;

[0023] Fine-tuning of the pre-trained large language model based on the second loss function;

[0024] If the pre-trained large language model satisfies the second termination condition, the training variants and their corresponding training text features and training sequencing read coverage map in the second training data are used as inputs, and the pre-trained large language model is fine-tuned according to the third loss function to obtain the target large model.

[0025] Secondly, this application provides a device for determining the authenticity of a mutation, comprising:

[0026] The data acquisition module is used to acquire sequencing data of the target sample;

[0027] The text feature acquisition module is used to determine the candidate variant set and its corresponding variant text features based on sequencing data.

[0028] The image feature acquisition module is used to generate a corresponding sequencing read coverage map based on sequencing data for any candidate variant in the candidate variant set.

[0029] The prompt word generation module is used to generate the first prompt word based on the candidate variant and its corresponding variant text features, sequencing read coverage map and prompt word template;

[0030] The judgment module is used to input the first prompt word into the target large model to obtain the true or false judgment result of the candidate mutation.

[0031] Thirdly, this application provides a computer device, including one or more processors and a memory, in which computer-readable instructions are stored. When the computer-readable instructions are executed by one or more processors, the mutation truth-determination method in any of the above embodiments is executed.

[0032] Fourthly, this application provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the mutation truth-determination method in any of the above embodiments.

[0033] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0034] Based on the mutation authenticity determination method in this embodiment, the sequencing data of the target sample is first acquired. Then, a candidate mutation set and its variant text features are determined based on the sequencing data to screen potential mutation sites and extract relevant feature information. Subsequently, for each candidate mutation in the candidate mutation set, a sequencing read coverage map is generated based on the sequencing data. Next, a first prompt word is generated based on the candidate mutation, variant text features, sequencing read coverage map, and prompt word template. Finally, the first prompt word is input into the target large-scale model to obtain the candidate mutation authenticity determination result. This scheme integrates sequencing information related to mutation determination into the large-scale model, providing it for model prediction without complex manual processing. The large-scale language model can fully and comprehensively capture the signals of mutation authenticity from the sequencing information, improving the accuracy and efficiency of mutation authenticity determination. This provides a reliable basis for mutation determination in fields such as biological research and medical diagnosis, reduces false positive results, and contributes to the development of related research and applications. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart illustrating a method for determining the authenticity of a variant provided in one embodiment of this application;

[0037] Figure 2 This is a schematic diagram of the sequencing read coverage map in one embodiment of this application;

[0038] Figure 3 This is a schematic diagram of the process for generating a sequencing read coverage map in one embodiment of this application;

[0039] Figure 4 This is a schematic diagram of the process of fine-tuning the target large model in one embodiment of this application;

[0040] Figure 5 This is a flowchart illustrating the process of fine-tuning to obtain the target large model in another embodiment of this application;

[0041] Figure 6 This is a flowchart illustrating the process of fine-tuning the target large model in another embodiment of this application;

[0042] Figure 7 A flowchart illustrating the method for determining the authenticity of variations provided in one embodiment of this application;

[0043] Figure 8 This is an internal structural diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0044] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0045] This application provides a method for determining the authenticity of mutations. Please refer to [link / reference]. Figure 1 This includes steps S102 to S110.

[0046] S102, Obtain sequencing data of the target sample.

[0047] As can be understood, the target sample refers to the biological sample used as the research object in this method for determining the authenticity of variants. The sample contains genetic information and forms the basis for subsequent analysis. Sequencing data is data obtained by measuring the nucleic acids (such as DNA or RNA) in the target sample using specific sequencing technologies (also known as sequencing data). It may also include annotation results obtained by bioinformatics annotation of the sequencing results. Here, sequencing can be a diagnostic method based on next-generation sequencing (NGS) technology, such as whole-exome sequencing (WES) and whole-genome sequencing (WGS), and other high-throughput sequencing technologies. The sequences obtained from these sequencing technologies are often compared with a reference genome to identify variants at the nucleic acid level.

[0048] S104, determine the candidate variant set and its corresponding variant text features based on sequencing data.

[0049] Candidate variant sets are understood to be collections of potentially variable sites selected from whole-genome sequencing data. These sites may contain genuine variants or false positives due to sequencing errors or other factors. Sequencing data undergoes alignment with a reference genome, sorting and deduplication, and variant detection steps to identify candidate variants. Specifically, bioinformatics software such as BWA (Burrows-Wheeler Aligner) can be used for reference genome alignment to obtain sequencing alignment data. GATK (Genome Analysis Toolkit) is then used to sort, deduplicate, and detect variants in the sequencing alignment data. GATK uses a series of algorithms and workflows to perform base quality calibration, variant detection, and other operations on the sequencing alignment data, ultimately outputting a list of candidate variants.

[0050] Variant text features are textual descriptions of information related to candidate variants in sequencing alignment data. After obtaining sequencing alignment data, the volume is typically very large, containing a significant amount of normal sequence information and potential variant information. Analyzing and processing this data to identify candidate variant sets narrows the analysis scope to potentially mutated sites, improving the efficiency of subsequent analysis. Variant text features provide specific descriptions of candidate variants, reflecting their characteristics and providing necessary information for generating sequencing read coverage maps and cue words. For variant text feature extraction, scripts can be written to extract relevant information from sequencing alignment data files and variant detection result files, such as the chromosomal location, type, homozygosity, naming information, sequencing depth, population frequency, corresponding gene and transcript, and whether it is located in a genomic repetitive region, the read ratio of the mutation site, the number of false positives detected in the historical database, and various predictions from quality control software. This information is then compiled into text format. Location information reflects a wealth of information; therefore, to achieve better results, chromosomal location information of the mutation site can be preferentially included in the variant text features.

[0051] S106: For any candidate variant in the candidate variant set, generate the corresponding sequencing read coverage map based on the sequencing data.

[0052] As can be understood, a sequencing read coverage map is a visual graph used to show the coverage of sequencing data at candidate variant sites and their surrounding regions. For example... Figure 2 As shown, it can intuitively reflect information such as sequencing depth and base distribution. Sequencing read coverage maps graphically present the coverage depth distribution, base distribution, and alignment quality of sequencing reads near candidate variant sites on the reference genome, allowing researchers to more intuitively observe the sequencing status of candidate variant sites. By observing sequencing read coverage maps, some abnormal sequencing patterns can be detected, such as sudden changes in sequencing depth and uneven base distribution; these anomalies may be related to false mutations. Simultaneously, sequencing read coverage maps can also provide another modality of information for models to determine the authenticity of mutations, increasing the amount of reference information during the judgment process.

[0053] S108, generate the first prompt word based on the candidate variant and its corresponding variant text features, sequencing read coverage map and prompt word template.

[0054] It's understandable that a cue word template is a predefined text format used to standardize the content and structure of cue words. It contains placeholders; during cue word generation, candidate variants and their corresponding variant text features, sequencing read coverage maps, and other information are filled into these placeholders to form specific cue words. The first cue word, generated based on this information, is input into the target large-scale model to describe the candidate variant in detail, enabling the model to determine whether it's true or false. The cue word template's role is to standardize and structure the various information about candidate variants, ensuring consistency and comparability of the information input into the target large-scale model. By integrating variant text features and sequencing read coverage maps into the cue words, more comprehensive and detailed information can be provided to the target large-scale model, helping it better understand the characteristics of candidate variants and thus improving the accuracy of its true / false cue word determination.

[0055] S110, input the first prompt word into the target large model to obtain the true or false judgment result of the candidate mutation.

[0056] In essence, a target large-scale model refers to an AI model with powerful language understanding and reasoning capabilities, such as the GPT series models and the BERT model. These models are trained on large-scale corpora and are able to understand natural language text and perform related reasoning and judgments. The true / false judgment result is the target large-scale model's judgment on whether a candidate variant is a true variant based on the first prompt word in the input, and the result is usually "true variant" or "false variant". Here, the target large-scale model is a large language model that can process at least two modalities of information: image and text. It can be an ordinary large language model without fine-tuning, or a large language model that has been fine-tuned to improve the judgment performance.

[0057] Based on the mutation authenticity determination method in this embodiment, the sequencing data of the target sample is first acquired. Then, a candidate mutation set and its variant text features are determined based on the sequencing data to screen potential mutation sites and extract relevant feature information. Subsequently, for each candidate mutation in the candidate mutation set, a sequencing read coverage map is generated based on the sequencing data. Next, a first prompt word is generated based on the candidate mutation, variant text features, sequencing read coverage map, and prompt word template. Finally, the first prompt word is input into the target large-scale model to obtain the candidate mutation authenticity determination result. This scheme integrates sequencing information related to mutation determination into the large-scale model, providing it for model prediction without complex manual processing. The large-scale language model can fully and comprehensively capture the signals of mutation authenticity from the sequencing information, improving the accuracy and efficiency of mutation authenticity determination. This provides a reliable basis for mutation determination in fields such as biological research and medical diagnosis, reduces false positive results, and contributes to the development of related research and applications.

[0058] In one embodiment, a corresponding sequencing read coverage map is generated based on the sequencing data. Please refer to [link to relevant documentation]. Figure 3 This includes steps S302 to S304.

[0059] S302, determine the corresponding variant segment based on the candidate variant set.

[0060] It can be understood that a variant region refers to a genomic region of a predetermined length flanking the candidate variant's genomic location. The authenticity of a candidate variant may be influenced by its surrounding sequence environment. By extending the candidate variant's location to both sides by a predetermined length to determine the variant region, richer information can be obtained, such as the presence of adjacent variants and changes in sequencing depth. This information helps to more accurately determine the authenticity of candidate variants. Different types of variants can be identified using different methods to determine the variant region. For SNVs / INDELs, a segment of a predetermined length is truncated to both sides of the candidate variant's location as the variant region. The predetermined length is generally chosen to be between 100bp and 200bp. For LargeCNVs (Large Copy Number Variations), the start and end regions of the LargeCNV can be used as the variant region, or the entire chromosome region plus the start and end regions of the LargeCNV can be used as the variant region. For exonCNVs (exon Copy Number Variations), a certain range of segments can be selected as the variant region based on its location.

[0061] S304 generates a sequencing read coverage map based on the variant regions of candidate variants and sequencing alignment data. The sequencing alignment data is obtained by aligning the sequencing data with a reference genome.

[0062] Sequencing alignment, as we understand it, involves comparing short sequence fragments (sequencing reads) obtained from sequencing with a known reference genome to determine the optimal matching position of each sequencing read on the reference genome. This is a fundamental step in subsequent analysis of sequencing data, helping to understand the differences between the sample's genome and the reference genome. Sequencing alignment data is information obtained after sequencing alignment, typically stored in SAM (Sequence Alignment / Map) or BAM (Binary Alignment / Map) formats. These results record information such as the alignment position of each sequencing read on the reference genome, alignment quality, and the presence of variations. The short sequence fragments obtained from sequencing are isolated; by aligning them with the reference genome, these short sequence fragments can be located on the genome, thus constructing the overall information of the sample's genome. Sequencing alignment algorithms calculate the optimal alignment position of each sequencing read on the reference genome based on factors such as the degree of base matching, mismatch rate, deletions, and insertions between the sequencing read and the reference genome. Through this comparison, we can discover possible variations in the sample genome, such as single nucleotide polymorphisms (SNPs) and insertions / deletions (Indels), providing a basis for subsequent analysis.

[0063] The generation of sequencing read coverage maps is based on variant regions and sequencing alignment data. The location of each sequencing read on the genome is determined using sequencing alignment data. Then, the sequencing depth and read distribution within the variant regions are statistically analyzed, and this information is finally displayed graphically. Sequencing read coverage maps help us visually observe the sequencing quality of regions containing candidate variants, such as whether there are regions with excessively low sequencing depth, and whether the reads are evenly distributed, thereby determining the reliability and biological significance of candidate variants. Different generation methods can be used for different variant types. For SNVs / INDELs, open-source tools such as IGV, Samtools, plotReads, or custom programs can be used to generate sequencing read coverage maps of specific lengths on both sides of the variant location. For LargeCNVs, the open-source tool cnvkit.pyscatter command or a custom script can be used. For exonCNVs, open-source tools such as ExomeDepth or a custom script can be used.

[0064] In one embodiment, the prompt word template includes a role definition domain, a text feature domain, and an image feature domain. The role definition domain is used to specify the role played by the target large model, the text feature domain is used to insert variant text features, and the image feature domain is used to insert sequencing read coverage maps.

[0065] It is understandable that the role definition domain in this area serves to clarify the role played by the target large model within the prompt word template. By defining roles, the target large model is guided to process and analyze input information from a specific perspective and in a specific manner, enhancing the model's understanding of the task and its targeted execution. The text feature domain is specifically designed to accommodate variant text features. Variant text features include key information related to sequencing data, described in textual form, such as the chromosomal location of the variant site, type information, homozygosity, naming information, sequencing depth, population frequency, corresponding genes and transcripts, and whether it is located in a genomic repetitive region. Accurately inserting this information into this domain provides detailed textual data support for the target large model. The function of the image feature domain is to insert sequencing read coverage maps. Sequencing read coverage maps are visual graphics that show the coverage of sequencing data at candidate variant sites and their surrounding areas. Inserting them into this domain allows the target large model to obtain direct image information, aiding in the judgment of variant status.

[0066] The prompt word template provides a multi-dimensional and structured information input method for the target large model by dividing it into a role definition domain, a text feature domain, and an image feature domain. Different types of information are crucial for the model's accurate judgment in the process of determining the true or false nature of variants. The role definition domain enables the target large model to analyze subsequent information from the perspective of a specific professional role (such as a senior geneticist or bioinformatics analyst), enhancing the model's depth of understanding of the task. The variant text features inserted in the text feature domain provide the model with specific details about candidate variants through precise textual descriptions; these details are the foundational data for the model's logical reasoning and judgment. The sequencing read coverage map inserted in the image feature domain utilizes the intuitiveness of images, allowing the model to extract information from graphics that is difficult to convey directly through text, such as the overall distribution characteristics of sequencing data and local anomaly patterns. These three domains work together to provide information to the target large model from different angles, helping the model to comprehensively and deeply understand the situation of candidate variants, thereby improving the accuracy of variant true / false judgment.

[0067] In one embodiment, the target large model is obtained by fine-tuning a pre-trained large language model; see [link to relevant documentation]. Figure 4 The fine-tuning process includes steps S402 to S404.

[0068] S402, Obtain the first training set. The first training set includes multiple first training data sets, which include training variants and their corresponding training text features, training sequencing read coverage maps, and discrimination labels.

[0069] The first training set is understood to be a dataset used for fine-tuning the pre-trained large language model, containing multiple first training data sets. Each first training data set is the basic unit within it, containing training variants, training text features, a training sequencing read coverage map, and a discriminant label. Training variants are known true / false variants detected from actual sequencing samples and validated by Sanger or qPCR, used to simulate candidate variants that need to be judged as true / false in real-world applications. Training text features are textual descriptions of the features associated with the training variants, similar to the variant text features mentioned earlier. The training sequencing read coverage map is a visualization generated from the sequencing data corresponding to the training variant, showing the sequencing coverage of the variant site and its surrounding region. The discriminant label is an identifier that clearly marks whether a training variant is a true or false variant, usually represented by binary (e.g., 0 for false variant, 1 for true variant) or other classification identifiers.

[0070] Obtaining the initial training set can be done through several methods. On one hand, existing sequencing data and variant information can be collected from publicly available bioinformatics databases. This data is usually validated and labeled, allowing for the selection of suitable training variants, their associated textual features, and sequencing read coverage maps. Discriminant labels can then be added based on the validation results. For example, variants with clear true / false determinations can be selected from publicly available medical databases as training samples. On the other hand, data can be obtained through in-house sequencing experiments. The sequenced variants can be manually validated or other reliable methods can be used to determine their true / false status before compiling this data into the initial training set. After data collection, preprocessing is necessary, including data cleaning, feature extraction, and standardization. For example, for training textual features, noise can be removed and the data format standardized; for training sequencing read coverage maps, image enhancement and normalization can be performed to improve data quality and consistency.

[0071] S404 uses the first training set to fine-tune the pre-trained large language model so that the discrimination results of the pre-trained large language model match the discrimination labels.

[0072] Fine-tuning, as we understand it, refers to further training the model using domain-specific training data (i.e., the first training set) based on pre-training. This involves adjusting some or all of the model's parameters to better adapt it to the needs of a specific task. The discrimination result refers to the judgment given by the pre-trained large language model after analyzing the training variants and related features of the input, indicating whether the variant is a true variant or a false variant. The discrimination label is the pre-labeled correct judgment result in the first training set. The goal of fine-tuning is to make the model's discrimination result as consistent as possible with the discrimination label.

[0073] While pre-trained large language models possess powerful language understanding and processing capabilities, the generality of their training data may prevent them from accurately capturing the features and patterns of specific domains when faced with domain-specific tasks. By fine-tuning the model using an initial training set, the model can learn the intrinsic relationships between training variants, training text features, and training sequencing read coverage maps in the variant true / false judgment task. During training, the model calculates a loss function based on the discriminant label, which measures the difference between the model's judgment result and the label. By continuously adjusting the model's parameters using optimization algorithms (such as stochastic gradient descent and Adam), the loss function is gradually reduced, allowing the model's judgment result to get closer and closer to the label, thus improving the model's performance on the variant true / false judgment task.

[0074] In one embodiment, please refer to Figure 5 The pre-trained large language model is fine-tuned using the first training set so that the discrimination result of the pre-trained large language model matches the discrimination label, including steps S502 and S504.

[0075] S502, freeze the weights of the image processing network of the pre-trained large language model, take the training variants in the first training data and their corresponding training text features as input, and fine-tune the pre-trained large language model according to the first loss function.

[0076] Image processing networks are the network structures within large language models responsible for processing image data. They typically consist of convolutional layers, pooling layers, etc., used to extract image features. Weights in the model are the strength values ​​of the connections between neurons in the neural network, determining how the input signal is transmitted and processed within the network. Freezing weights means not updating these weight values ​​during training. The first loss function measures the difference between the model's discrimination result and the discrimination label. In this step, the discrimination result is calculated based on the training variance and its corresponding training text features, and this function is used to evaluate the model's performance. The first loss function can be the cross-entropy loss function, used to optimize the model's ability to process text modal data.

[0077] Freezing the weights of the image processing network is an effective strategy in the initial stage of fine-tuning. Pre-trained large language models have already learned rich linguistic knowledge and patterns on general data, and the training text features are important textual information related to mutations. By fine-tuning with training mutations and their corresponding training text features as input, the model can focus on learning the relationship between text features and the truth value of mutations. This avoids introducing too much interference from the initial weight updates of the image processing network, allowing the model to more stably adjust text-processing-related parameters, thereby optimizing text information processing.

[0078] S504, if the pre-trained large language model satisfies the first termination condition, unfreeze the weights of the image processing network, take the training variants and their corresponding training text features and training sequencing read coverage map in the first training data as input, and fine-tune the pre-trained large language model according to the second loss function to obtain the target large model.

[0079] The first termination condition is understood to be one or more pre-defined conditions used to determine whether the model has reached a suitable state when fine-tuned with training variants and training text features, such as achieving a certain accuracy or the loss value decreasing to a certain threshold. Unfreezing, on the other hand, is the opposite of freezing, meaning that the weights of the image processing network are allowed to be updated during training. The second loss function, in this step, measures the difference between the model's discrimination result and the discrimination label when simultaneously considering training variants, training text features, and training sequencing read coverage maps. It differs from the first loss function. Since it only considers text information, the first loss function is relatively simple in form and can typically use common classification loss functions, such as the cross-entropy loss function. This simple loss function can effectively measure the model's performance in text classification tasks and is used to optimize the model's ability to process text modal data. To better utilize text and image information, the second loss function can be a joint loss function combining text and image modalities, such as a combination of weighted cross-entropy loss and mean squared error loss, to optimize the model's overall ability to process multimodal data.

[0080] When the model meets the first termination condition (e.g., the number of training epochs reaches a preset value or the model's accuracy on the validation set reaches a certain threshold), it indicates that the model has achieved good results in processing training text features. At this point, the weights of the image processing network are unfrozen, and the training sequencing read coverage map is introduced as input, allowing the model to further learn the relationship between image information and the true / false nature of mutations. The training sequencing read coverage map contains information that text features cannot fully express, such as the distribution pattern of sequencing data and local anomalies. By combining text and image information, the model can gain a more comprehensive understanding of mutations, thereby improving the accuracy of the discrimination results. Therefore, using the training mutations in the first training data, the corresponding training text features, and the training sequencing read coverage map as input, the pre-trained large language model is further fine-tuned according to the second loss function. Through the above two-stage fine-tuning process, the target large model is obtained. To improve model performance, the target large model obtained in this stage can be further fine-tuned and optimized.

[0081] Furthermore, the unfreezing process here can be gradual, dividing the image processing network into multiple layers. Different layering methods can be used for different image processing network architectures. Unfreezing occurs sequentially based on the proximity to the output layer, with layers closer to the output layer unfrozen first. After each layer completes a phase of training, the next layer is unfrozen. This is because the parameters of higher-level structures closer to the output layer have the strongest correlation with the specific task. In the pre-trained model, although some general knowledge has been learned, when facing the specific task of mutation true / false judgment, higher-level networks (i.e., networks closer to the output layer) need to be significantly adjusted to adapt to the new task requirements. Unfreezing higher-level networks first allows the model to quickly learn for the new task, utilizing the general features extracted from the lower and middle layers to initially adapt to the mutation true / false judgment task in a shorter time, improving the model's initial performance on the new task and enhancing training effectiveness.

[0082] In one embodiment, the pre-trained large language model is fine-tuned according to the second loss function to obtain the target large model, and the method further includes steps S602 to S606.

[0083] S602, perturb the first training data with multiple correct discriminant labels to obtain the second training set.

[0084] It is understandable that perturbation modifies or damages the data to a certain extent, causing it to deviate from its original state. In this step, errors are injected into the training text features and / or training sequencing read coverage map in the first training data. The second training set consists of the perturbated first training data and contains multiple second training data sets. The second training data is new data obtained after perturbing the first training data, in which the training text features and / or training sequencing read coverage map are modified, and the discriminant labels are modified to errors, in order to simulate the situation of false mutations and enhance the model's ability to identify false mutations.

[0085] In the task of distinguishing true and false mutations, training solely with real samples may lead to overfitting. By perturbing the first training data with correct labels to generate a second training set containing false mutation information, the model is exposed to morphologies of false mutations that are similar to the true correct samples. This allows the model to better learn the key features that distinguish true and false mutations during subsequent training, improving its ability to distinguish false mutations in practical applications and enhancing its robustness and generalization ability. Perturbations to training text features can be achieved by randomly modifying base information or changing sequencing depth values. For example, randomly selecting base positions in the training text features and replacing them with other bases; or randomly increasing or decreasing the sequencing depth value within a certain range. Perturbations to the training sequencing read coverage map can be achieved using image processing techniques, such as adding noise, changing image brightness and contrast, or randomly cropping parts of the image.

[0086] S604, fine-tunes the pre-trained large language model based on the second loss function.

[0087] This step can be referred to in the description of step S504.

[0088] S606: If the pre-trained large language model satisfies the second termination condition, the pre-trained large language model is fine-tuned according to the third loss function using the training variants and their corresponding training text features and training sequencing read coverage map in the second training data as input to obtain the target large model.

[0089] The second termination condition is a pre-defined condition for judging whether the model has reached a suitable state when fine-tuned using the second training set. Examples include the model's accuracy on the validation set reaching a certain threshold or the loss value converging to a certain range. The third loss function is used to fine-tune the model in the final stage, further optimizing its performance. The design of the third loss function can refer to the design of the second loss function. When the model meets the second termination condition, it indicates that the model has achieved good results when processing the second training set. However, to further improve the model's performance and stability, fine-tuning with the second training data as input has a significant effect on improving the judgment effect, allowing the model to better adapt to various situations in practical applications and improving the accuracy of mutation true / false judgments.

[0090] In one embodiment, another method for determining the truth value of a mutation is also provided; please refer to [link to relevant documentation]. Figure 7 This includes steps S702 to S710.

[0091] S702, acquire sequencing data of the target sample.

[0092] As can be understood, the target sample refers to the biological sample used as the research object in this method for determining the authenticity of variants. The sample contains genetic information and forms the basis for subsequent analysis. Sequencing data is obtained by measuring the nucleic acids (such as DNA or RNA) in the target sample using specific sequencing technologies. It may also include annotation results obtained by bioinformatics annotation of the sequencing results. Here, sequencing can be a diagnostic method based on next-generation sequencing (NGS) technology, such as high-throughput sequencing technologies like whole-exome sequencing (WES) and whole-genome sequencing (WGS). The sequences obtained from these sequencing technologies are often compared with a reference genome to identify variants at the nucleic acid level.

[0093] S704, determine the candidate variant set based on sequencing data.

[0094] Candidate variant sets are understood to be collections of potentially variable sites selected from whole-genome sequencing data. These sites may contain genuine variants or false positives due to sequencing errors or other factors. Sequencing data undergoes genome alignment, deduplication, and variant detection steps to identify candidate variants. Specifically, bioinformatics software such as BWA (Burrows-Wheeler Aligner) can be used for genome alignment to obtain sequencing alignment data. GATK (Genome Analysis Toolkit) is then used to sort, deduplicate, and detect variants in the sequencing alignment data. GATK employs a series of algorithms and workflows to perform base quality calibration, variant detection, and other operations on the sequencing alignment data, ultimately outputting a list of candidate variants.

[0095] S706: For any candidate variant in the candidate variant set, generate the corresponding sequencing read coverage map based on the sequencing data.

[0096] As can be understood, a sequencing read coverage map is a visual graph used to show the coverage of sequencing data at candidate variant sites and their surrounding regions. For example... Figure 2 As shown, it can intuitively reflect information such as sequencing depth and base distribution. Sequencing read coverage maps graphically present the coverage depth distribution, base distribution, and alignment quality of sequencing reads near candidate variant sites on the reference genome, allowing researchers to more intuitively observe the sequencing status of candidate variant sites. By observing sequencing read coverage maps, some abnormal sequencing patterns can be detected, such as sudden changes in sequencing depth and uneven base distribution; these anomalies may be related to false mutations. Simultaneously, sequencing read coverage maps can also provide another modality of information for models to determine the authenticity of mutations, increasing the amount of reference information during the judgment process.

[0097] S708 generates judgment prompts based on candidate variants, sequencing read coverage maps, and prompt word templates.

[0098] It's understandable that a cue word template is a predefined text format used to standardize the content and structure of cue words. It contains placeholders; during cue word generation, candidate variants, their corresponding variant text features, sequencing read coverage maps, and other information are filled into these placeholders to form specific cue words. Judgment cue words are generated based on the above information and are input into the target large-scale model to describe the candidate variant in detail, enabling the model to determine its authenticity. A candidate variant itself is only a preliminary discovery; its authenticity requires comprehensive information from multiple sources. Sequencing read coverage maps provide sequencing quality and read distribution information for the region containing the candidate variant, which is crucial for determining the actual existence of the variant. Cue word templates provide a standardized way to integrate this information. By organizing key information from candidate variants and sequencing read coverage maps into judgment cue words according to the template, the target large-scale model can more accurately understand and analyze this information, thereby improving the accuracy of its judgment on the authenticity of candidate variants.

[0099] S710 inputs the judgment prompt words into the target large model to obtain the true or false judgment results of the candidate mutations.

[0100] In essence, a target large-scale model refers to an AI model with powerful language understanding and reasoning capabilities, such as the GPT series models and the BERT model. These models are trained on large-scale corpora and are able to understand natural language text and perform related reasoning and judgments. The true / false judgment result is the target large-scale model's judgment on whether a candidate variant is a true variant based on the input judgment prompts, and the result is usually "true variant" or "false variant". Here, the target large-scale model is a large language model that can process at least two modalities of information: image and text. It can be an ordinary large language model without fine-tuning, or a large language model that has been fine-tuned to improve the judgment performance.

[0101] This application provides a device for determining the authenticity of variants, including a data acquisition module, a text feature acquisition module, an image feature acquisition module, a prompt word generation module, and a judgment module.

[0102] The data acquisition module acquires sequencing data from the target sample. The text feature acquisition module determines the candidate variant set and its corresponding variant text features based on the sequencing data. The image feature acquisition module generates a sequencing read coverage map for any candidate variant in the candidate variant set based on its corresponding variant text features. The prompt word generation module generates a first prompt word based on the candidate variant, its corresponding variant text features, the sequencing read coverage map, and the prompt word template. The judgment module inputs the first prompt word into the target large model to obtain the true or false result of the candidate variant.

[0103] Specific limitations regarding the mutation authenticity determination device can be found in the limitations of the mutation authenticity determination method above, and will not be repeated here. Each module in the aforementioned mutation authenticity determination device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. It should be noted that the module division in this embodiment is illustrative and only represents a logical functional division; other division methods may be used in actual implementation.

[0104] This application provides a computer device, including one or more processors and a memory, in which computer-readable instructions are stored. When the computer-readable instructions are executed by one or more processors, the mutation truth-determination method in any of the above embodiments is performed.

[0105] Indicatively, such as Figure 8 As shown, Figure 8 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. (Refer to...) Figure 8 The computer device 800 includes a processing component 802, which further includes one or more processors, and memory resources represented by memory 801 for storing instructions executable by the processing component 802, such as application programs. The application programs stored in memory 801 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 802 is configured to execute instructions to perform the variation truth / false determination method of any of the above embodiments.

[0106] The computer device 800 may also include a power supply component 803 configured to perform power management of the computer device 800, a wired or wireless network interface 804 configured to connect the computer device 800 to a network, and an input / output (I / O) interface 805.

[0107] This application provides a storage medium storing computer-readable instructions. When executed by one or more processors, the computer-readable instructions cause the one or more processors to perform the mutation truth-determination method in any of the above embodiments.

[0108] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0109] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0110] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for determining the authenticity of a mutation, characterized in that, include: Obtain sequencing data of the target sample; Based on the sequencing data, a candidate variant set and its corresponding variant text features are determined; For any candidate variant in the candidate variant set, a corresponding sequencing read coverage map is generated based on the sequencing data; A first prompt word is generated based on the candidate variant and its corresponding variant text features, the sequencing read coverage map, and the prompt word template. The first prompt word is input into the target large model to obtain the true or false judgment result of the candidate mutation.

2. The method for determining the authenticity of mutations according to claim 1, characterized in that, The step of generating a corresponding sequencing read coverage map based on the sequencing data includes: Based on the genomic location of the candidate variants, the variant segments of the candidate variants are determined; Based on the variant regions and sequencing alignment data, a sequencing read coverage map is generated; the sequencing alignment data is obtained by aligning the sequencing data with a reference genome.

3. The method for determining the authenticity of mutations according to claim 1, characterized in that, The prompt word template includes a role definition domain, a text feature domain, and an image feature domain; the role definition domain is used to specify the role played by the target large model, the text feature domain is used to insert the variant text features, and the image feature domain is used to insert the sequencing read coverage map.

4. The method for determining the authenticity of mutations according to claim 1, characterized in that, The variant text features include the location information, type information, homozygosity, naming information, sequencing depth, population frequency, corresponding gene and transcript, whether it is located in a genomic repetitive region, the proportion of reads at the mutation site, the number of false positives detected by the variant in the historical database, and various other factors from the quality control software prediction results.

5. The method for determining the authenticity of mutations according to claim 1, characterized in that, The target large model is obtained by fine-tuning a pre-trained large language model. The fine-tuning process includes: Obtain a first training set; the first training set includes multiple first training data, the first training data including training variants and their corresponding training text features, training sequencing read coverage map and discrimination labels; The pre-trained large language model is fine-tuned using the first training set so that the discrimination result of the pre-trained large language model matches the discrimination label.

6. The method for determining the authenticity of mutations according to claim 5, characterized in that, The step of fine-tuning the pre-trained large language model using the first training set to match the discrimination result of the pre-trained large language model with the discrimination label includes: The weights of the image processing network of the pre-trained large language model are frozen, and the training variants and their corresponding training text features in the first training data are used as input to fine-tune the pre-trained large language model according to the first loss function. If the pre-trained large language model satisfies the first termination condition, the weights of the image processing network are unfrozen. The training variants and their corresponding training text features and the training sequencing read coverage map in the first training data are used as inputs. The pre-trained large language model is then fine-tuned according to the second loss function to obtain the target large model.

7. The method for determining the authenticity of mutations according to claim 6, characterized in that, The step of fine-tuning the pre-trained large language model according to the second loss function to obtain the target large model further includes: The first training data with multiple correct discrimination labels are perturbed to obtain a second training set; the second training set includes multiple second training data, which are obtained by perturbing the training text features and / or the training coverage sequencing map of the first training data with errors, and by modifying the discrimination labels to be incorrect; The pre-trained large language model is fine-tuned based on the second loss function; If the pre-trained large language model satisfies the second termination condition, the training variants and their corresponding training text features and the training sequencing read coverage map in the second training data are used as inputs, and the pre-trained large language model is fine-tuned according to the third loss function to obtain the target large model.

8. A device for determining the authenticity of a mutation, characterized in that, include: The data acquisition module is used to acquire sequencing data of the target sample; The text feature acquisition module is used to determine the candidate variant set and its corresponding variant text features based on the sequencing data. The image feature acquisition module is used to generate a corresponding sequencing read coverage map based on the sequencing data for any candidate variant in the candidate variant set. The prompt word generation module is used to generate a first prompt word based on the candidate variant and its corresponding variant text features, the sequencing read coverage map, and the prompt word template. The judgment module is used to input the first prompt word into the target large model to obtain the true or false judgment result of the candidate mutation.

9. A computer device, characterized in that, It includes one or more processors and a memory storing computer-readable instructions, which, when executed by the one or more processors, perform the mutation truth-determination method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the mutation truth / falseness determination method as described in any one of claims 1-7.

Images