A tumor mutation detection model training method, device and equipment

Abstract

The application provides a tumor mutation detection model training method, device and equipment, the method comprising: obtaining a plurality of preset data for tumor mutation detection model training; wherein the preset data comprises target mutation information and its annotation information, patient clinical information, sample other mutation information and its annotation information, and database comparison information; sorting the preset data according to a preset thinking paradigm, establishing the association between the target mutation and the sample other mutation, converting the preset data into structured data to obtain a structured data set; determining a question prompt word according to the structured data set, and determining an answer prompt word according to an expert interpretation result; grouping the question prompt word and the answer prompt word into a training question and answer prompt word to determine a fine-tuning training set; fine-tuning a pre-training large language model according to the fine-tuning training set based on a preset prompt word structure and according to a fine-tuning strategy to obtain a fine-tuning large model.

Description

Technical Field

[0001] This application relates to mutation site detection technology, and more particularly to a method, apparatus, device, and storage medium for training a tumor mutation detection model. Background Technology

[0002] In the precision oncology system, tumor mutation detection is a crucial prerequisite for developing targeted therapy plans and assessing patient prognosis. The development and progression of tumors is essentially a malignant transformation process driven by the accumulation of mutations in genetic material. These mutations alter key biological pathways such as cell proliferation, apoptosis, and migration, leading to abnormal cell proliferation and tumor formation. Tumor mutation detection, as a core technology for elucidating the molecular mechanisms of tumors and guiding precision treatment, has evolved from early single-gene detection to a comprehensive, genome-wide analysis, constructing a personalized "mutation atlas" for each cancer patient. Based on the stage of mutation occurrence, tumor mutations can be divided into germline mutations and somatic mutations. Germline mutations are hereditary mutations carried in germ cells, while somatic mutations are acquired mutations accumulated in somatic cells during tumorigenesis. Somatic mutations are the main targets of precision oncology treatment and a current focus of research in detection technologies.

[0003] Systemic mutations can be classified into three categories according to their molecular mechanisms and structural characteristics, and their detection faces different technical challenges: (1) Systemic single nucleotide variants (SNVs): refer to the substitution, insertion or deletion of a single base, which is the most common type of mutation in tumors. Systemic SNVs can directly lead to changes in protein structure and function, and are an important target for targeted drug design. However, the detection of systemic SNVs is easily affected by sequencing errors and normal cell contamination, especially in samples with low tumor purity. Traditional filtering methods based on sequencing quality are difficult to effectively distinguish between real mutations and technical errors. (2) Systemic copy number variants (CNVs): refer to the amplification or deletion of genomic fragments, ranging in length from several thousand bases to the entire chromosome arm. Systemic CNVs affect the balance of cell signaling pathways by changing the copy number dose of oncogenes or tumor suppressor genes. However, CNV detection depends on the uniformity of genome coverage. Sequencing biases in aneuploid and repetitive sequence regions of chromosomes in tumor samples, as well as signal dilution caused by normal cell contamination, often lead to errors in the judgment of CNV boundaries. (3) Somatic Structural Variants (SVs): These include chromosomal translocations, inversions, and large-fragment insertions / deletions. Although their incidence is lower than that of SNVs and CNVs, they have a significant impact on the malignant phenotype of tumors. The detection of somatic SVs faces two major challenges: First, the breakpoints are often located in repetitive sequences or low-complexity regions, making it difficult to align sequencing reads and leading to missed detections; second, some SVs are subclonal mutations (existing only in some tumor cells), with low signal intensity, and traditional detection methods based on breakpoint depth have low recall rates for these SVs.

[0004] The current mainstream methods for judging the authenticity of tumor mutations mainly rely on traditional bioinformatics tools and manual review, including: (1) Filtering methods based on sequencing quality: such as GATK, Mutect2, etc., which filter low-quality variant sites and reduce false positive mutations by detecting sequencing depth, base quality value, alignment quality and other indicators. However, such methods only focus on errors at the sequencing technology level and cannot identify false mutations caused by sample biological characteristics or data analysis logic defects, resulting in a high false negative rate. (2) Phenotypic or clinical information-assisted verification: combining clinical information such as patient tumor type, pathological stage, and previous treatment history, manual judgment is made on whether the mutation conforms to the pathogenesis logic of tumors. This method relies on expert experience and has the defects of being knowledge-intensive and labor-intensive. Moreover, the consistency of judgments on "rare tumor mutations" among different experts is low, which makes it difficult to meet the efficiency requirements of large-scale clinical sample testing. (3) Database comparison method: the detected mutations are compared with tumor mutation databases such as NSCLC, COSMIC, TCGA, etc. If the mutation appears frequently in the database, it is judged as a "true mutation". However, such methods are limited by the database coverage, and cannot effectively identify newly discovered tumor mutations (mutations not included in the database), nor can they distinguish between "polymorphic variations in the normal population included in the database" and "tumor-specific mutations." Traditional machine learning-based methods, while able to integrate multi-dimensional information, essentially rely on manually defined rules or feature engineering, making it difficult to effectively model unstructured and semantically complex clinical text information, let alone understand its inherent medical logical connections. This results in a superficial use of clinical background information, thus affecting the overall accuracy of interpretation.

[0005] In recent years, large language models (LLMs) have shown advantages in semantic understanding and multi-dimensional information integration. LLMs are deep learning models based on the Transformer architecture and pre-trained with massive amounts of text data. Their core capability lies in capturing semantic associations and logical rules in language sequences, enabling full-chain task processing from text understanding to complex reasoning. In the biomedical field, the application of LLMs has extended from basic text processing (such as document abstract generation and case structuring) to solving professional tasks. However, LLMs have not made substantial breakthroughs in tumor mutation analysis and still have many limitations, such as (1) insufficient adaptation of domain knowledge: the pre-training data is mainly general text (such as books and web pages), with a small proportion of biomedical data, resulting in biases in the semantic understanding of tumor-specific concepts. (2) prominent "illusion" problem: due to data conflicts in tumor mutation databases (such as COSMIC) (the pathogenicity labeling of the same mutation is contradictory in different studies), LLMs are prone to generating seemingly reasonable but inconsistent results when lacking external verification mechanisms. (3) The ability to handle complex mutation types is weak, and the sensitivity to identify "low-frequency true tumor mutations" is insufficient, which cannot meet the clinical demand for "high accuracy and high sensitivity" in tumor mutation judgment. These limitations make it impossible for unoptimized LLMs to be directly applied to the scenario of judging the authenticity of tumor mutations. Summary of the Invention

[0006] This application provides a method, apparatus, device, and storage medium for training a tumor mutation detection model, in order to construct a tumor mutation detection model adapted to tumor scenarios, overcoming the technical limitations of at least one of the aforementioned traditional methods.

[0007] In a first aspect, embodiments of this application provide a method for training a tumor mutation detection model, including: Acquire multiple preset data for training the tumor mutation detection model; the preset data includes target mutation information and its annotation information, patient clinical information, other mutation information of the sample and its annotation information, and database comparison information; The preset data is sorted according to a preset thinking paradigm, and the correlation between the target mutation and other mutations in the sample is established. The preset data is then converted into structured data to obtain a structured dataset. Question hints are determined based on the structured dataset, and answer hints are determined based on expert interpretation results; The question prompts and answer prompts are combined to form training question-and-answer prompts, and the fine-tuning training set is determined. Based on the fine-tuning training set, and using a preset prompt word structure, the pre-trained large language model is fine-tuned according to the fine-tuning strategy to obtain the fine-tuned large model.

[0008] Secondly, embodiments of this application also provide a tumor mutation detection model training device, comprising: The data acquisition module is used to acquire multiple preset data for training the tumor mutation detection model; the preset data includes target mutation information and its annotation information, patient clinical information, other mutation information of the sample and its annotation information, and database comparison information; The structured dataset determination module is used to sort the preset data according to a preset thinking paradigm, establish the association between the target mutation and other mutations in the sample, and convert the preset data into structured data to obtain a structured dataset. The prompt word determination module is used to determine question prompt words based on the structured dataset and answer prompt words based on expert interpretation results; The fine-tuning training set determination module is used to combine the question prompts and the answer prompts to form training question-and-answer prompts, and determine the fine-tuning training set; The model fine-tuning training module is used to fine-tune the pre-trained large language model according to the fine-tuning training set, based on the preset prompt word structure and the fine-tuning strategy, to obtain the fine-tuned large model.

[0009] Thirdly, embodiments of this application also provide an electronic device, the electronic device comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the tumor mutation detection model training method provided in any embodiment of this application.

[0010] Fourthly, embodiments of this application also provide a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the tumor mutation detection model training method provided in any embodiment of this application.

[0011] The tumor mutation detection model training method, apparatus, device, and storage medium provided in this application, through the design of targeted fine-tuning strategies for all types of tumor mutations, and targeted fine-tuning for the multi-type characteristics and clinical relevance of tumor mutations, enable the pre-trained large language model to learn the value of the fusion positions of different genes, preset prompt word structures to guide the model's reasoning logic, and combine domain knowledge engineering and external database verification to solve the problem of semantic understanding bias and the easy generation of seemingly reasonable but inconsistent results, thus providing an efficient and reliable diagnostic tool for tumor mutation detection. Attached Figure Description

[0012] The above and other features, advantages and aspects of the embodiments of this application will become more apparent when taken in conjunction with the accompanying drawings and the following detailed description.

[0013] Figure 1 This is a flowchart of a tumor mutation detection model training method according to an embodiment of the present invention; Figure 2 This is a flowchart of the fine-tuning large model inference process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a tumor mutation detection model training device according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0014] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0015] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0016] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0017] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0018] In the following embodiments, each embodiment provides optional features and examples. The features described in the embodiments can be combined to form multiple optional solutions. Each numbered embodiment should not be regarded as only one technical solution.

[0019] Figure 1This is a flowchart illustrating a tumor mutation detection model training method provided in an embodiment of this application. This embodiment is applicable to the fine-tuning training of a neural network model for detecting the true and false nature of tumor mutations. This method can be implemented by a tumor mutation detection model training device, which can be implemented in hardware and / or software and is generally integrated into an electronic device, such as a server or personal computer. Figure 1 As shown, the method specifically includes: Step 110: Obtain multiple preset data for training the tumor mutation detection model.

[0020] The pre-set data is used to construct a fine-tuning training set for training the tumor mutation detection model. This pre-set data includes target mutation information and its annotations, patient clinical information, other mutation information and its annotations, and database comparison information. The pre-set data primarily concerns information about the target mutation and other mutations in the sample, and includes sequencing quality and patient clinical background information. Mutation information includes the target mutation and other mutations in the sample, and includes multiple types of tumor mutation features, such as systemic SNVs, systemic CNVs, and systemic SVs. Furthermore, the dataset composed of the pre-set data can include correlations in clinical information (such as the relationship between mutations and drug responses). Combined with domain knowledge engineering and external database validation, an adaptation to the tumor scenario is constructed. This allows for targeted fine-tuning of the pre-trained large language model during training. The tumor mutation detection model is a fine-tuned large language model obtained by fine-tuning the pre-trained large language model, used as the model for detecting tumor mutations. The pre-set data can be obtained by collecting tumor mutation data from patient samples (e.g., target mutation sites, sequencing quality information, corresponding tumor types, patient clinical information, and true / false mutations verified by pathological biopsy). A joint annotation approach involving professional tumor pathologists and bioinformatics engineers can be used to clarify the "true" (clinically validated tumor-related mutations) or "false" (sequencing errors, polymorphisms in normal populations, etc.) attribute of each mutation, and supplement the annotation with clinically relevant information such as tumor stage and family history corresponding to the mutation.

[0021] Step 120: Sort the preset data according to the preset thinking paradigm, establish the correlation between the target mutation and other mutations in the sample, and convert the preset data into structured data to obtain a structured dataset.

[0022] The pre-set data may be multi-sourced and heterogeneous clinical sequencing data. To fine-tune the training of the large language model, a shift from data stacking to logical clue compilation is needed, transforming it into structured input adapted to the large language model's reasoning. Therefore, simply concatenating all data fields is insufficient for logical clue compilation. Thus, a pre-set thinking paradigm can be determined based on the expert thinking paradigm of tumor mutation review, forcibly ranking the input features. It can also proactively establish associations between the target mutation information and other mutation information in the sample. For example, when judging a fusion inverse mutation, it simultaneously extracts whether there is a positive fusion mutation of the same gene in the sample, and uses this information as context input to the large language model. The structured dataset's data structure can be: first, the target mutation information; then, the patient's clinical information; and finally, the remaining mutation information in samples other than the target mutation.

[0023] Step 130: Determine question prompts based on the structured dataset, and determine answer prompts based on expert interpretation results.

[0024] The prompt words guide the large language model in reasoning. The composition and order of the prompt words can be set in the prompt word design. The selectable reasoning logic generally guides the model to reason in the order of sequencing quality verification, target mutation quality, quality and information of other mutations in the sample, and clinical information corroboration. Information from databases such as NSCLC is used as matching vectors and incorporated into each step of the reasoning process, forming a three-in-one reasoning loop of technical indicators, clinical background, and molecular atlas. This ensures that the model's judgment is both consistent with the authenticity of the sequencing data and aligned with the clinical diagnosis and treatment of tumors. The "thought chain" (COT) is constructed to address the "illusion" problem that occurs in existing large language models in tumor mutation detection. After the prompt words are determined, experts interpret them to obtain the answer prompt words.

[0025] Step 140: Combine the question prompts and answer prompts to form training question-and-answer prompts, and determine the fine-tuning training set.

[0026] After obtaining the question prompts and answer prompts, the two are matched together to form the training question prompts. Combined with the structured dataset obtained from the preset data, the fine-tuning training set can be determined.

[0027] Step 150: Based on the fine-tuning training set and the preset prompt word structure, fine-tune the pre-trained large language model according to the fine-tuning strategy to obtain the fine-tuned large model.

[0028] In this process, a pre-trained large language model with moderate parameter size and strong semantic understanding capability is selected. Within a pre-defined training framework, the model is fine-tuned using a fine-tuning training set. The pre-defined prompt word structure is a specialized prompt word designed for tumor mutations. After fine-tuning the pre-trained large language model, a fine-tuned large model is obtained, which can be used as a tumor mutation detection model.

[0029] The data in the sample set can be divided into a training set and a validation set according to a certain ratio. It's understood that when dividing the sample set, a portion of the data can be selected and allocated to the training and validation sets; that is, in addition to the training and validation sets, there is independent data. This data can be used as the test set. The validation and test sets do not need to include the answer hints from the fine-tuning training. For example, when fine-tuning a pre-trained large model using dataset partitioning, the fine-tuning training set has 22500 true CNV mutations and 22500 false CNV mutations, 18000 true SNV mutations and 18000 false SNV mutations, and 1575 true and false SV mutations. The validation set has 2500 true CNV mutations and 2500 false CNV mutations, 2000 true and 2000 false SNV mutations, and 175 true and false SV mutations.

[0030] Optionally, after fine-tuning the pre-trained large language model according to the fine-tuning strategy based on the fine-tuning training set to obtain the fine-tuned large model, the following steps are also included: Based on the test set and validation set determined by the question prompts, the fine-tuned large model is tested, and hyperparameters are fine-tuned using the validation set.

[0031] The performance of the fine-tuned large model was evaluated using a test set. The model was trained using the fine-tuning training and validation sets, and hyperparameters were fine-tuned using the validation set to determine the final fine-tuned large model for tumor mutation detection.

[0032] Optionally, the preset data includes sequencing quality information, patient clinical information, basic information on the target mutation, basic information on mutations in other systems of the sample, basic information on mutations in other germline systems of the sample, database comparison results, true / false attribute annotations for each mutation in the sample, and association annotations between mutations and patient clinical information.

[0033] Each feature requires corresponding data content. Sequencing quality information includes sample type, sequencing depth, sample contamination rate, details of quality control anomalies, DNA input, and cell tumor content. Patient clinical information includes tumor type, stage, and family history. System CNVs include gene, state, copy number, total length percentage, and auto tag. System SNVs include mutation name, mutation frequency, mutation depth, auto tag, GSR, function, and exon. System SVs include fusion gene, fusion information, mutation frequency, auto tag, exon, splice reads, and Discordant Pair. Germline SNVs include mutation name, variant significance, and auto tag. Germline CNVs include mutation name, variant significance, and detection range. Database comparison information includes mutation significance in databases such as NSCLC, mutation detection in self-developed databases, and mutation significance.

[0034] Optionally, the preset data is sorted according to a preset thinking paradigm, and the association between the target mutation and other mutations in the sample is established. The preset data is then converted into structured data to obtain a structured dataset, including: The pre-defined data are sorted in the following order: target mutation information, patient clinical information, sequencing quality and sample information, mutation spectrum of other systems in the sample, and / or mutation spectrum of other germline systems in the sample. This structure aims to guide the model simulation experts to first focus on the core mutation, and then make a comprehensive judgment based on the clinical scenario, technical artifacts, and overall genomic background.

[0035] Establish a correlation between the target mutation information and other mutation information in the sample, and use the correlation information as context. The sequencing data in the pre-designed algorithm is transformed into structured data adapted to the pre-trained large language model. The following is an example of the data required for the SNV target mutation in the inference system: 1) Target system SNV mutation: {Mutation name: CTNNB1: c.110C>T (p.S37F), Mutation frequency: 0.016, Mutation depth: 31, auto label: HMY, GSR: 19; 10, Germ: N, Function: missense, Exon: EX3, label: Mutation frequency greater than true mutation reporting threshold} 2) Patient Clinical Information: {Gender: Male, Clinical Diagnosis: Lung Adenocarcinoma; Family History: None, Cancer Type: Non-Small Cell Lung Cancer, Product Focus: Focus on previous tests, EGFR sensitive mutations 19del / 858, etc., Sample Type: Plasma, Contamination Rate: 0.0092%, Quality Control Abnormalities Details: {}, MSI Status: MSS, HRD Status: Negative, DNA Input: 14.3, Tumor Cell Content (%): 50} 3) System-wide mutation spectrum: {System SNV} |Mutation Name|Mutation Frequency|Mutation Depth|auto Tag|GSR|Germ|Function|Exon| |---|---|---|---|---|---|---|---| |EGFR:c.2238_2248delATTAAGAGAAGinsGC(p.L747_A750delinsP)|0.025|58|HHY|30;22|H|cds-indel|EX19| |CTNNB1:c.110C>T(p.S37F)|0.016|31|HMY|19;10|N|missense|EX3| System CNV |Gene|State|Ratio|GCN|Total_LP|auto tag| |---|---|---|---|---|---| |MSH2|loss|0.8249|-1.502|0.0275|GR-Part| |BRCA2|loss|0.873|-0.54|0.3814|GR-Part| |PALB2|loss|0.8835|-0.33|0.2974|GR-Part| }, 4) Germline complete mutation spectrum: {germline SNV} |Mutation Name|Mutation Significance|auto Tag| |---|---|---| |NM_005378.4(MYCN):c.141G>A(p.(E47=))| } Optionally, the preset prompt word structure is set as follows: The model consists of structured prompts for role setting, input features, thought chain guidance, and output format. The thought chain guidance model is set to logically deduce step by step based on sequencing quality verification, target mutation quality, quality and information of other mutations in the sample, and patient clinical information.

[0036] Among them, the prompt word structure of the tumor mutation-specific prompt word engineering design can be designed to include structured prompt words that include role setting, input features, thought chain guidance, and output format. The role is clearly defined as an expert in interpreting and analyzing tumor NGS gene detection. This expert possesses expertise in understanding tumor mutation interpretation standards, mastering the spectral characteristics of mutations in various cancer types, and integrating multi-dimensional evidence for logical reasoning. For example, as an expert in interpreting and analyzing tumor NGS gene detection, the model should comprehensively judge the authenticity of SNV mutations in the target system based on the following factors: 1. Mutation name, mutation frequency, mutation depth, auto tag, etc. in the target system's SNV mutations; 2. Patient's clinical information; 3. System's complete mutation spectrum (SNV / CNV / SV); 4. Germline complete mutation spectrum (SNV / CNV). Input feature structuring involves constructing a standardized dataset according to a pre-defined data structure. The thought process guidance guides the model through a step-by-step logical reasoning process: sequencing quality verification – target mutation quality – quality and information of other mutations in the sample – clinical information corroboration. The output format is strictly limited to binary results, i.e., "Y" (true mutation) or "N" (false mutation), without any additional explanatory text, ensuring direct parsing during subsequent system integration.

[0037] Optionally, based on the fine-tuning training set and a preset prompt word structure, the pre-trained large language model is fine-tuned according to the fine-tuning strategy to obtain the fine-tuned large model, including: Based on the MS-Swift training framework, the low-rank adaptation module is applied to the linear layer of the pre-trained large language model; Training rounds are conducted based on a preset initial learning rate, preset learning rate warm-up ratio, and preset batch size, using a preset prompt word structure that includes role setting, input features, thought chain guidance, and output format.

[0038] Among them, the pre-trained large language model can be a large language model with a moderate parameter size and strong semantic understanding ability. Optional pre-trained large language models include Qwen3-4B, Qwen3-8B, and Qwen3-14B, balancing model performance and deployment efficiency.

[0039] The fine-tuning strategy employs a parameter-efficient fine-tuning mode, trained on a GPU cluster with eight NVIDIA A100 processors (80GB VRAM) based on the MS-Swift training framework. Core optimizations include LoRA adaptation design, applying the low-rank adaptation (LoRA) module to all linear layers of the model (including attention and feedforward layers), setting the rank to 4 and the scaling factor (alpha) to 16. Training parameter optimizations include setting the training epochs to 4, the initial learning rate to 1e-4, the learning rate warm-up ratio to 0.05, and the batch size to 1.

[0040] Optionally, the fine-tuned large model can be integrated into the tumor mutation judgment system. An analysis system integrating the fine-tuned large language model can be developed, supporting users to upload tumor sample files, such as xlsx format files (containing mutation information and sequencing quality) and clinical information tables (containing tumor type and stage). The system automatically extracts input features, calls the model to judge the authenticity of mutations, and outputs a report containing "mutation site - judgment result" that can be directly downloaded.

[0041] In one implementation, to evaluate model performance, a self-developed software was created to simulate human decision-making. This software primarily uses machine learning methods to construct a decision tree script that simulates human decision-making based on information such as mutation frequency, sequencing depth, and mutation tags. Using human decision-making results as the standard, the performance of the fine-tuned large model and the self-developed software were comprehensively evaluated. The results showed that the fine-tuned large model's results were closer to human decision-making results, with a consistency rate as high as 99%. The self-developed software had a consistency rate of 97%. Compared with the self-developed software built using traditional machine learning methods, the fine-tuned large model significantly improved mutation identification accuracy by 2%, greatly enhancing the reliability of mutation detection reports.

[0042] Validated on a set of 100,000 independent test cases, the accuracy of the fine-tuned large model was significantly improved. In the test set, the fine-tuned large model (taking the SNV system as an example) achieved a recall rate of over 99% and a false mutation recognition rate of over 99%, with an accuracy rate of over 99%, which is significantly better than the self-developed software (accuracy rate of around 97%). In particular, for low-frequency true mutations (database frequency < 0.1%), the recall rate reached 99.7%, solving the problem of insufficient sensitivity of traditional methods for identifying rare tumor mutations.

[0043] The efficiency of fine-tuning the large model is significantly improved. While self-developed software and manual review of a single tumor sample (containing 200-300 candidate variants) takes 1-2 hours, the system integrating the fine-tuned large model automatically analyzes the sample in just 1-8 minutes. Furthermore, if the fine-tuned large model assists expert review, it can significantly reduce the workload of experts. It has strong clinical adaptability, supporting mutation detection for multiple types of tumors (lung cancer, breast cancer, colorectal cancer, and many other common tumors). It can integrate multi-dimensional information such as sequencing quality, tumor clinical information, and database data, eliminating the need for manual integration of scattered data and adapting to over 90% of clinical tumor mutation detection scenarios. The results are highly reliable; 1000 repeated inferences on 10 randomly selected tumor samples showed 100% consistency, meeting the stringent clinical requirements for reproducibility.

[0044] In some specific embodiments, Example 1: Performance verification of a fine-tuned large model based on Qwen3-4B on the system SNV test set. Experimental materials include the Qwen3-4B fine-tuned model, trained on a tumor mutation-specific dataset. Comparison method: Model performance is mainly compared between the accuracy of the self-developed software and the accuracy of the fine-tuned model. The self-developed software was jointly developed by multiple experts and bioinformatics engineers to simulate human decision-making. Test set: Mutations cover various tumors such as lung cancer, breast cancer, gastric cancer, and colorectal cancer, including 27,488 true mutations and 16,047 false mutations. Experimental steps, model inference: as follows. Figure 2 As shown in the flowchart, the "basic information + sequencing quality + clinical information of tumor patients" of the test set mutations are input into the fine-tuning model to obtain the "true / false mutation" judgment results. Performance evaluation: The true mutation recall rate, false mutation identification rate, and overall accuracy of each group are calculated. Experimental results are shown in Table 1: Table 1

[0045] Experiments show that the fine-tuned large model is significantly more accurate than the self-developed software in judging the true and false SNVs in the system, with obvious advantages in recall and specificity, and can effectively reduce false negatives and false positives.

[0046] Example 2: Performance comparison of different pre-trained large language models after fine-tuning. Experimental materials included Qwen3-4B, Qwen3-8B, and Qwen3-14B (all fine-tuned using the same tumor mutation-specific dataset and the same prompt words). Experimental steps: The three models performed true / false detection on the test set, and the true mutation recall rate, false mutation detection rate, and overall accuracy were calculated.

[0047] The experimental results are shown in Table 2: Table 2

[0048] Experiments show that the finely tuned Qwen3-4B model has the best overall performance and the smallest model size, making it the optimal choice for clinical deployment; Qwen3-8B and Qwen3-14B have similar performance and can be considered as alternatives.

[0049] Example 3: Performance Comparison After Fine-tuning Different Dataset Structures. Experimental materials included a Qwen3-4B fine-tuned model trained on a tumor mutation-specific dataset. Comparison Method: Dataset performance was primarily compared between the accuracy of the model trained on the single-mutation dataset and the accuracy of the model combining single mutations with background genomic information. Test Set: Mutations covered various tumors including lung cancer, breast cancer, gastric cancer, and colorectal cancer, with 27,488 true mutations and 16,047 false mutations. Experimental Procedure: The accuracy of the model trained on the single-mutation dataset and the model combining single mutations with background genomic information were used to determine true and false mutations on the test set, calculating the true mutation recall, false mutation identification rate, and overall accuracy.

[0050] The experimental results are shown in Table 3: Table 3

[0051] Experiments show that the dataset structure combining single mutations with background mutation information is significantly better than the single mutation dataset structure.

[0052] Figure 3 This application provides a tumor mutation detection model training device, such as... Figure 3 As shown, the device specifically includes: The data acquisition module 310 is used to acquire multiple preset data for training the tumor mutation detection model; the preset data includes target mutation information and its annotation information, patient clinical information, other mutation information of the sample and its annotation information, and database comparison information; The structured dataset determination module 320 is used to sort the preset data according to the preset thinking paradigm, establish the association between the target mutation and other mutations in the sample, and convert the preset data into structured data to obtain a structured dataset. The prompt word determination module 330 is used to determine question prompt words based on the structured dataset and answer prompt words based on expert interpretation results; The fine-tuning training set determination module 340 is used to combine question prompts and answer prompts to form training question-answer prompts and determine the fine-tuning training set. The model fine-tuning training module 350 is used to fine-tune the pre-trained large language model according to the fine-tuning training set, based on the preset prompt word structure and the fine-tuning strategy, to obtain the fine-tuned large model.

[0053] Optionally, the tumor mutation detection model training device also includes: The model validation module is used to fine-tune the pre-trained large language model according to the fine-tuning strategy based on the fine-tuning training set. After obtaining the fine-tuned large model, it is tested based on the test set and validation set determined by the question prompt words, and the hyperparameters are fine-tuned through the validation set.

[0054] Optionally, the preset data includes sequencing quality information, patient clinical information, basic information on the target mutation, basic information on mutations in other systems of the sample, basic information on mutations in other germline systems of the sample, database comparison results, true / false attribute annotations for each mutation in the sample, and association annotations between mutations and patient clinical information.

[0055] Optionally, the structured dataset determination module 320 is specifically used for: The pre-defined data are sorted in the following order: target mutation information, patient clinical information, sequencing quality and sample information, mutation spectrum of other systems in the sample, and / or mutation spectrum of other germline systems in the sample. Establish a correlation between the target mutation information and other mutation information in the sample, and use the correlation information as context. The sequencing data in the preset design is transformed into structured data adapted to the pre-trained large language model.

[0056] Optionally, the preset prompt word structure is set as follows: The model consists of structured prompts for role setting, input features, thought chain guidance, and output format. The thought chain guidance model is set to logically deduce step by step based on sequencing quality verification, target mutation quality, quality and information of other mutations in the sample, and patient clinical information.

[0057] Optional, model fine-tuning training module 350, specifically used for: Based on the MS-Swift training framework, the low-rank adaptation module is applied to the linear layer of the pre-trained large language model; Training rounds are conducted based on a preset initial learning rate, preset learning rate warm-up ratio, and preset batch size, using a preset prompt word structure that includes role setting, input features, thought chain guidance, and output format.

[0058] Optional pre-trained large language models include Qwen3-4B, Qwen3-8B, and Qwen3-14B.

[0059] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 A schematic diagram of the structure of a server 400 suitable for implementing embodiments of this application is shown. Figure 4 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0060] like Figure 4As shown, electronic device 400 may include a processing device (e.g., a central processing unit, a graphics processor, etc.) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from storage device 406 into random access memory (RAM) 403. RAM 403 also stores various programs and data required for the operation of electronic device 400. Processing device 401, ROM 402, and RAM 403 are interconnected via bus 404. Input / output (I / O) interface 405 is also connected to bus 404.

[0061] Typically, the following devices can be connected to I / O interface 405: input devices 406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 406 including, for example, magnetic tapes, hard disks, etc.; and communication devices 409. Communication device 409 allows electronic device 400 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 An electronic device 400 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0062] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 409, or installed from storage device 406, or installed from ROM 402. When the computer program is executed by processing device 401, it performs the functions defined in the methods of embodiments of this application.

[0063] It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0064] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0065] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0066] The aforementioned computer-readable medium carries one or more programs. When the aforementioned one or more programs are executed by the electronic device, the electronic device causes the following: it acquires multiple preset data sets for training a tumor mutation detection model; the preset data sets include target mutation information and its annotation information, patient clinical information, other mutation information and its annotation information, and database comparison information; it sorts the preset data sets according to a preset thinking paradigm, establishes the association between the target mutation and other mutations in the sample, and converts the preset data sets into structured data to obtain a structured dataset; it determines question prompts based on the structured dataset and determines answer prompts based on expert interpretation results; it combines the question prompts and answer prompts to form training question-and-answer prompts, and determines a fine-tuning training set; based on the fine-tuning training set and the preset prompt structure, it fine-tunes the pre-trained large language model according to a fine-tuning strategy to obtain a fine-tuned large model.

[0067] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including but not limited to object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0068] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0069] The units described in the embodiments of this application can be implemented in software or hardware. The names of the units are not, in some cases, limiting the scope of the unit itself.

[0070] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0071] In the context of this application, a computer-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0072] According to one or more embodiments of this application, a storage medium containing computer-executable instructions is provided, which, when executed by a computer processor, are used to perform a tumor mutation detection model training method as provided in any embodiment of this application.

[0073] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

[0074] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. Multitasking and parallel processing may be advantageous in certain environments. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this application. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0075] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for training a tumor mutation detection model, characterized in that, include: Acquire multiple preset data for training the tumor mutation detection model; wherein, the preset data includes target mutation information and its annotation information, patient clinical information, other mutation information of the sample and its annotation information, and database comparison information; The preset data is sorted according to a preset thinking paradigm, and the correlation between the target mutation and other mutations in the sample is established. The preset data is then converted into structured data to obtain a structured dataset. Question hints are determined based on the structured dataset, and answer hints are determined based on expert interpretation results; The question prompts and answer prompts are combined to form training question-and-answer prompts, and the fine-tuning training set is determined. Based on the fine-tuning training set, and using a preset prompt word structure, the pre-trained large language model is fine-tuned according to the fine-tuning strategy to obtain the fine-tuned large model.

2. The tumor mutation detection model training method according to claim 1, characterized in that, After fine-tuning the pre-trained large language model according to the fine-tuning training set, based on the preset prompt word structure and following the fine-tuning strategy to obtain the fine-tuned large model, the process further includes: The fine-tuned large model is tested based on the test set and validation set determined by the question prompts, and hyperparameters are tuned using the validation set.

3. The tumor mutation detection model training method according to claim 1, characterized in that, The preset data includes sequencing quality information, patient clinical information, basic information on the target mutation, basic information on mutations in other systems of the sample, basic information on mutations in other germline systems of the sample, database comparison results, true / false attribute annotations for each mutation in the sample, and association annotation information between the mutation and the patient's clinical information.

4. The tumor mutation detection model training method according to claim 3, characterized in that, The preset data is sorted according to a preset thinking paradigm, and the correlation between the target mutation and other mutations in the sample is established. The preset data is then converted into structured data to obtain a structured dataset, including: The preset data are sorted in the following order: target mutation information, patient clinical information, sequencing quality and sample information, mutation spectrum of other systems in the sample, and / or mutation spectrum of other germline systems in the sample. Establish a correlation between the target mutation information and other mutation information in the sample, and use the correlation information as context. The sequencing data in the preset design is transformed into structured data adapted to the pre-trained large language model.

5. The tumor mutation detection model training method according to claim 4, characterized in that, The preset prompt word structure is set as follows: The model consists of structured prompts for role setting, input features, thought chain guidance, and output format. The thought chain guidance model is set to logically deduce step by step based on sequencing quality verification, target mutation quality, quality and information of other mutations in the sample, and patient clinical information.

6. The tumor mutation detection model training method according to claim 5, characterized in that, Based on the fine-tuning training set, and according to the preset prompt word structure, the pre-trained large language model is fine-tuned according to the fine-tuning strategy to obtain the fine-tuned large model, including: Based on the MS-Swift training framework, the low-rank adaptation module is applied to the linear layer of the pre-trained large language model; Training rounds are conducted based on the preset initial learning rate, preset learning rate warm-up ratio, and preset batch size, using the preset prompt word structure of role setting, input features, thought chain guidance, and output format.

7. The tumor mutation detection model training method according to claim 6, characterized in that, The pre-trained large language models include Qwen3-4B, Qwen3-8B, and Qwen3-14B.

8. A tumor mutation detection model training device, characterized in that, include: The data acquisition module is used to acquire multiple preset data for training the tumor mutation detection model; wherein, the preset data includes target mutation information and its annotation information, patient clinical information, other mutation information of the sample and its annotation information, and database comparison information; The structured dataset determination module is used to sort the preset data according to a preset thinking paradigm, establish the association between the target mutation and other mutations in the sample, and convert the preset data into structured data to obtain a structured dataset. The prompt word determination module is used to determine question prompt words based on the structured dataset and answer prompt words based on expert interpretation results; The fine-tuning training set determination module is used to combine the question prompts and the answer prompts to form training question-and-answer prompts, and determine the fine-tuning training set; The model fine-tuning training module is used to fine-tune the pre-trained large language model according to the fine-tuning training set, based on the preset prompt word structure and the fine-tuning strategy, to obtain the fine-tuned large model.

9. An electronic device, characterized in that, The electronic device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the tumor mutation detection model training method as described in any one of claims 1-7.

10. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the tumor mutation detection model training method as described in any one of claims 1-7.

Images