Method for realizing automatic interpretation of genome copy number variation

Abstract

The invention discloses a method for realizing automatic interpretation of genome copy number variation. The method comprises the following steps: acquiring a ChatGLM-6B large language model; establishing a genome copy number variation detection result interpretation data set, desensitizing, correcting and normalizing the data, and converting the data into training data; performing adjustment learning on the large language model by using a low-rank adaptation algorithm, so that the large language model obtains the interpretation capability for a copy number variation result report; designing a plurality of cue words to obtain corresponding input-output pairs in the large language model for learning adjustment, manually aligning the input-output pairs, and performing secondary adjustment on the model by using a reinforcement learning algorithm; and generating a new copy number variation detection result interpretation data set. According to the method, the genome copy number variation detection result can be accurately, reliably and intelligently interpreted, so that the working efficiency of CNV data analysis is greatly improved, the diagnosis of clinical genetic cases is more efficient and quicker, and the method has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for realizing automatic interpretation of genome copy number variation based on a large language model. Background Art

[0002] Copy number variation (CNV), as an important form of genetic variation, plays a key role in the research and diagnosis of various genetic diseases. Genome copy number variation (CNV) is associated with a variety of human diseases. In order to find the genetic causes of children with "developmental delay, mental retardation, multiple malformations", patients with recurrent spontaneous abortions, and fetuses with structural malformations indicated by prenatal ultrasound, whole genome copy number variation detection has become a frontline clinical examination method.

[0003] The interpretation of copy number variation test results involves a large amount of bioinformatics analysis, genomic principles, and clinical pathology, making the writing of CNV test results interpretation particularly arduous. There are many types of CNVs, including deletions, duplications, inversions, and other forms of variation, each of which may correspond to different biological effects and clinical significance. Accurately understanding and distinguishing the role of these variations in the occurrence and development of diseases is a very challenging task. Although CNV detection technology continues to develop, when it is translated into clinical practice, how to integrate complex genetic variation information into the existing diagnosis and treatment process and convert it into clear and practical reporting language is currently a major difficulty. CNV detection and its report writing are also affected by differences in medical resources and technical strength between regions. Especially in underdeveloped or remote areas, limited by technology and talent resources, it has become more difficult to effectively implement CNV detection and accurately interpret reports. In recent years, although a variety of CNV detection technologies have been developed, such as chromosome microarray analysis and genomic copy number variation detection technology based on high-throughput sequencing, the large number of copy number variations detected by these technologies have brought great challenges to data analysis and genetic counseling.

[0004] In the current field of medical technology, with the rapid development of big data and artificial intelligence technology, especially the progress in natural language processing (NLP), the use of large models to assist in the interpretation of medical test reports has become an important direction of cutting-edge research. The traditional process of interpreting medical test reports is highly dependent on the professional knowledge and experience of doctors, which may be affected by subjective factors and has limited efficiency when faced with massive amounts of data. The original model that has not been professionally adapted has not been specifically optimized for professional terminology, pathophysiological expressions, and the logical relationship of complex medical test indicators in the medical context.

[0005] Genetic reports are different from most imaging or test numerical reports. Human genetic material is more than 6Gb, and its copy number variation detection results are composed of three elements: position, size, and copy number variation. The interpretation of its results requires consideration of factors such as genes and regions for relevant pathogenicity rating and analysis. Based on this, it is necessary to develop a method that can accurately and reliably interpret copy number variation (CNV) case detection reports. Summary of the invention

[0006] The present invention aims to solve the technical problem that various copy number variation (CNV) detection results cannot be accurately interpreted in clinical practice, and provides a method for realizing automatic interpretation of genomic copy number variation. This method can accurately and reliably interpret the genomic copy number variation detection results intelligently, thereby greatly improving the work efficiency of CNV data analysis.

[0007] In order to solve the above technical problems, the present invention is achieved through the following technical solutions:

[0008] In one aspect of the present invention, a method for realizing automated interpretation of genomic copy number variation is provided, comprising the following steps:

[0009] Get the large language model of ChatGLM-6B;

[0010] Establish a data set for interpreting the results of genome copy number variation detection, and desensitize, correct, normalize, and convert the data into training data;

[0011] Using a low-rank adaptation algorithm to adjust and learn the large language model, so that the large language model acquires the ability to interpret copy number variation result reports;

[0012] In the large language model for adjustment learning, multiple prompts are designed to obtain corresponding input-output pairs, the input-output pairs are manually aligned, and the model is adjusted twice using a reinforcement learning algorithm;

[0013] Generate new copy number variation detection result interpretation dataset.

[0014] Preferably, after the new copy number variation detection result interpretation data set is generated, it is learned online and continuously fine-tuned through an online learning module.

[0015] Preferably, the data desensitization is for desensitizing user privacy-sensitive information, and a hash coding algorithm is used to complete ciphertext coding conversion of user information;

[0016] And / or, the data correction is to manually correct the erroneous data and statements appearing in the historical data to obtain correct information for model training;

[0017] And / or, the data normalization is to process the result interpretation data with different formats into a general format to adjust the model training;

[0018] And / or, the conversion into training data is to convert the result interpretation data into a training data format of natural language question-natural language answer.

[0019] Preferably, the use of a low-rank adaptation algorithm to adjust and learn the large language model comprises the following steps:

[0020] In the initialization stage, for the weight matrix of the self-attention layer to be fine-tuned in the pre-trained model, it is regarded as a fixed original parameter matrix W0; two low-rank matrices A and B are defined for each matrix to be adjusted, where A∈Rr×k, B∈Rd×r, r is a rank much smaller than d and k, d and k represent the number of rows and columns of the original weight matrix respectively, and matrix A is initialized with Gaussian distribution, and B is initialized as a zero matrix;

[0021] In the training phase, the low-rank matrices A and B are trained alternately using the stochastic gradient descent scheme, so that the model update is only achieved through ΔW=BA, where ΔW is the adjustment change relative to the original parameters; the model is trained using the dataset of the target task, and the loss function is optimized to obtain the optimal A and B;

[0022] In the application stage, W=W0+ΔW forms a new parameter matrix, and then W is used to adjust the copy number variation results of the large model for interpretation and inference.

[0023] In some embodiments of the present invention, the step of adjusting the large language model using a low-rank adaptation algorithm uses 16 RTX GPUs with a total computing cost of more than 72 hours.

[0024] Preferably, manually aligning the input-output pair and performing secondary adjustment on the model using a reinforcement learning module comprises the following steps:

[0025] Manually review the input-output pairs. If the answers do not conform to human values ​​or are unreasonable, then add the answer scores to construct a negative sample set score. The added answer scores include: "The following answer is unreasonable", "The following answer is incorrect", or "The following answer is not in line with values";

[0026] Through PPO (Proximal Policy Optimization) in reinforcement learning, the score is used to update the above low-rank matrix B / A to obtain ΔW'=B'A';

[0027] Apply W = W0 + ΔW', and then use W to adjust the copy number variation results of the large model for interpretation and inference.

[0028] Preferably, the online learning module includes:

[0029] The new data cleaning function is used to clean and organize the new result interpretation data into a unified model requirement format, and convert the result interpretation data into a pair of natural language questions and natural language answers;

[0030] The training mode of the learning module is to use the new data as the adjustment data set and train W” using the above method. The setting of the rank parameter r is consistent with the rank parameter of the current fine-tuning model.

[0031] Application module, apply W = W0 + ΔW", and then use W to interpret the copy number variation results of the fine-tuned large model.

[0032] In another aspect of the present invention, a device for realizing automatic interpretation of genome copy number variation is provided, comprising:

[0033] Memory, used to store programs;

[0034] A processor is used to execute the program stored in the memory to implement an automated interpretation method for genome copy number variation based on a large language model.

[0035] In another aspect of the present invention, a system for automatically analyzing genome copy number variation is provided, comprising:

[0036] A device for realizing automatic interpretation of genome copy number variation, which is used to train and adjust the large language model so that it can obtain the ability to interpret the copy number variation result report;

[0037] An information input module, used to input copy number variation data to be analyzed;

[0038] The data analysis statement generation module enables the trained large language model to automatically interpret the input copy number variation data and output the corresponding results.

[0039] In another aspect of the present invention, a computer-readable storage medium is provided, which stores a program, and the program can be executed by a processor to implement the above-mentioned method for automated interpretation of genomic copy number variation.

[0040] The present invention realizes a method for automatically interpreting genomic copy number variation, and through targeted training and performance optimization of a large language model, enables it to understand and accurately interpret various types of professional information in a CNV test report, so that the large language model is most effectively applied to the interpretation of CNV fragments. Moreover, the large language model of the present invention has the ability of uninterrupted learning and self-updating, and can well cope with the situation where the CNV interpretation results are frequently updated. In addition, the system of the present invention is very convenient to use. Only the CNV information needs to be input to obtain a test report that automatically interprets the CNV fragment, thereby effectively assisting the diagnosis of clinical genetic cases, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0042] Figure 1 is a flow chart of a method for realizing automatic interpretation of genome copy number variation of the present invention;

[0043] Figure 2 It is a diagram of the steps of using the automated analysis system for genome copy number variation of the present invention;

[0044] Figure 3 It is a schematic diagram of input content of the present invention;

[0045] Figure 4 Schematic diagram of input and output contents of Embodiment 1 of the present invention;

[0046] Figure 5 Schematic diagram of input and output contents of Embodiment 2 of the present invention;

[0047] Figure 6 Schematic diagram of input and output contents of Embodiment 3 of the present invention;

[0048] Figure 7 is a schematic diagram of input and output contents of Embodiment 4 of the present invention;

[0049] Figure 8 is a schematic diagram of input and output contents of Embodiment 5 of the present invention;

[0050] Fig. 9 Schematic diagram of input and output contents of Embodiment 6 of the present invention;

[0051] Fig.10 It is a schematic diagram of the input and output contents of Example 7 of the present invention. DETAILED DESCRIPTION

[0052] The method for realizing automatic interpretation of genome copy number variation provided by the present invention performs targeted training and optimization on the pre-trained large language model, so that the model can understand and accurately interpret various professional information in the CNV detection report.

[0053] The method of the present invention for realizing automatic interpretation of genome copy number variation, such as Figure 1 As shown, the specific steps include:

[0054] Get the large language model of ChatGLM-6B from the Internet, including all language model parameters;

[0055] Establish a data set for interpreting the results of genome copy number variation detection, and desensitize, correct, normalize, and convert the data into training data;

[0056] After the data preparation is completed, the low-rank adaptation algorithm (LoRA) is used to adjust and learn the large language model so that the large language model can acquire the ability to interpret the copy number variation result report;

[0057] In the large language model for adjustment learning, multiple prompts are designed to obtain corresponding input-output pairs, the input-output pairs are manually aligned, and the model is adjusted twice using a reinforcement learning algorithm;

[0058] After generating a new copy number variation detection result interpretation data set, the new data is learned online and continuously fine-tuned through the online learning module.

[0059] In some embodiments of the present invention, a genome copy number variation detection result interpretation data set is established through the operation process of obtaining, cleaning, and organizing the copy number variation detection result interpretation data set. Among them, data desensitization is to desensitize user privacy sensitive information, and use a hash coding algorithm to complete the ciphertext encoding conversion of user information; data correction is to manually correct the erroneous data and statements in historical data to obtain correct information for model training; data normalization is to unify the format of historical result interpretation data and other issues into a general format to adjust model training; conversion to training data is to convert the form of result interpretation data into a training data format of natural language question-natural language answer.

[0060] In some embodiments of the present invention, a low-rank adaptation algorithm is used to adjust and learn a large language model, including the following steps:

[0061] In the initialization stage, for the weight matrix of the self-attention layer to be fine-tuned in the pre-trained model, it is regarded as a fixed original parameter matrix W0; two low-rank matrices A and B are defined for each matrix to be adjusted, where A∈Rr×k, B∈Rd×r, r is a rank much smaller than d and k, d and k represent the number of rows and columns of the original weight matrix respectively, and matrix A is initialized with Gaussian distribution, and B is initialized as a zero matrix;

[0062] In the training phase, the low-rank matrices A and B are trained alternately using the Stochastic Gradient Descent scheme, so that the model update is only achieved through ΔW=BA, where ΔW is the adjustment change relative to the original parameters; the model is trained using the dataset of the target task, and the loss function is optimized to obtain the optimal A and B;

[0063] In the application stage, W=W0+ΔW forms a new parameter matrix, and then W is used to adjust the copy number variation results of the large model for interpretation and inference.

[0064] This step used 16 RTX GPUs and had a total computational cost of over 72 hours.

[0065] In some embodiments of the present invention, manually aligning input-output pairs and re-adjusting the model using a reinforcement learning algorithm includes the following steps:

[0066] Manually review the input-output pairs, and add answer scores for answers that do not conform to human values ​​or are unreasonable, such as "the following answer is unreasonable / the following answer is incorrect / the following answer is inconsistent with values", etc., to construct a negative sample set score;

[0067] Use the PPO algorithm to update the low-rank matrix B / A using the score to obtain ΔW'=B'A';

[0068] Apply W = W0 + ΔW', and then use W to adjust the copy number variation results of the large model for interpretation and inference.

[0069] In some embodiments of the present invention, after a new copy number variation detection result interpretation data set is generated, the new data is online learned and continuously fine-tuned through an online learning module, wherein the online learning module includes:

[0070] The new data cleaning function is used to clean and organize the new result interpretation data into a unified model requirement format, and convert the result interpretation data into a pair of natural language questions and natural language answers;

[0071] The training mode of the learning module is to use the new data as the adjustment data set and train W” using the above method. The setting of the rank parameter r is consistent with the rank parameter of the current fine-tuning model.

[0072] Application module, apply W = W0 + ΔW", and then use W to interpret the copy number variation results of the fine-tuned large model.

[0073] The present invention enables the large language model to have uninterrupted learning and self-updating capabilities through an online learning module, which can well cope with the situation where CNV interpretation results are frequently updated.

[0074] In some embodiments of the present invention, a device for realizing automated interpretation of genomic copy number variation includes: a memory for storing programs; and a processor for executing the programs stored in the memory to realize an automated interpretation method of genomic copy number variation based on a large language model.

[0075] The device includes a computer device, such as a desktop computer, a handheld computer, a cloud server, etc. The computer device and the user can interact with each other through a keyboard, a mouse, a remote control, a touch pad or a voice control device.

[0076] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (such as SD or D interface display memory, etc.), random access memory, static random access memory, read-only memory, electrically erasable programmable read-only memory, programmable read-only memory, magnetic memory, disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of a computer, such as a hard disk or memory of a computer. In other embodiments, the memory can also be an external storage device of a computer, such as a plug-in hard disk, a smart memory card, a secure digital card, a flash memory card, etc. The memory can also include both an internal storage unit of a computer device and an external storage device thereof. In an embodiment of the present invention, the memory is commonly used to store an operating system and various application software installed on a computer device, such as a program code for automatic interpretation of genome copy number variation based on a large language model, etc. In addition, the memory can also be used to temporarily store various types of data that have been output or are to be output.

[0077] The processor may be a central processing unit, a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor is generally used to control the overall operation of a computer device. In an embodiment of the present invention, the processor is used to run a program code stored in a memory or process data, such as running a program code of an automated interpretation method for genome copy number variation based on a large language model.

[0078] In some embodiments of the present invention, a computer-readable storage medium stores a program, which can be executed by a processor to implement an automated interpretation method for genomic copy number variation based on a large language model.

[0079] The computer-readable storage medium stores an interface display program, which can be executed by at least one processor to perform the steps of the method for automated interpretation of genomic copy number variation based on a large language model.

[0080] In some embodiments of the present invention, a system for automated analysis of genomic copy number variation comprises:

[0081] A device for realizing automatic interpretation of genome copy number variation, which is used to train and adjust the large language model so that it can obtain the ability to interpret the copy number variation result report;

[0082] An information input module, used to input copy number variation data to be analyzed;

[0083] The data analysis statement generation module enables the trained large language model to automatically interpret the input copy number variation data and output the corresponding results.

[0084] In some embodiments of the present invention, Figure 2 As shown, the method for using the automated genome copy number variation analysis system of the present invention comprises the following steps:

[0085] The user submits the CNV data to be analyzed. The CNV data to be analyzed can be input according to the ISCN standard format. The information categories of the CNV data include chromosome, upstream position, downstream position and copy number;

[0086] When the user submits the CNV data to be analyzed, the large language model trained and optimized by the present invention will parse the number of genes contained in the CNV data, and automatically generate a CNV interpretation result with fluent language according to the missing or repeated type information of the CNV data.

[0087] In some embodiments of the present invention, the information input module also includes an ISCN standard format input data submission method, which is used for the user's existing CNV data written in the ISCN standard. The user inputs the CNV data in the ISCN standard format into the automated analysis system.

[0088] In some embodiments of the present invention, the CNV data interpretation content returned and output by the automatic analysis system for genome copy number variation of the present invention is designed to include relevant records in the DGV database, Decipher database, ClinGene database, OMIM phenotype database and Genereview database. Among them, the DGV database records the CNV data of the normal population, the Decipher database records the CNV data in the disease population, the ClinGene database records whether the segment to be retrieved involves haploinsufficient or triple-dose sensitive genes, and the OMIM phenotype database records whether the segment to be retrieved contains haploinsufficient or triple-dose sensitive genes in the ClinGene database. The relevant results are all displayed in the analysis results.

[0089] In some embodiments of the present invention, the input content is as follows: Figure 3As shown, "Please enter the command:" is a fixed text of the automatic analysis system for genome copy number variation of the present invention, and the input content is entered after the ":", and the input format is: "Explain the genetic test results: XXXXX". XXXXX is the genetic test result in the ISCN standard format.

[0090] Example 1

[0091] like Figure 4 As shown, the following format is input in the automated genome copy number variation analysis system of the present invention:

[0092] Please enter the command: "Explain the genetic test results: arr(21)×3"

[0093] The answer is:

[0094] "1. The result shows that the fetus is a carrier of Down syndrome (21-trisomy syndrome), which is also known as congenital idiocy or Down syndrome. The main clinical manifestations include growth retardation, mild to moderate intellectual disability, characteristic facial features (wide eyes, small palpebral fissures, upward slanting outer corners of the eyes, inner canthus skin, low nasal root, small ears, fat tongue, frequent tongue extension, drooling, etc.), abnormal bone development (the thumb is close to the palm when making a fist, and the little finger is close to the other four fingers), etc. 2. It is recommended that the parents of the fetus undergo genetic testing."

[0095] Example 2

[0096] like Figure 5 As shown, the following format is input in the automated genome copy number variation analysis system of the present invention:

[0097] Please enter the command: "Explain the genetic test results: arr[GRCh38]2q11.2(96017630-97589819)×3"

[0098] The answer is:

[0099] "1. The sample was found to have a duplication of approximately 1.57Mb in the q11.2 region of chromosome 2. (1) This region contains one protein-coding gene. (2) The ClinGen database does not include triple dosage sensitive genes / regions in this region. (3) The DECIPHER database contains a small number of individuals with copy number duplications in this region, and clinical manifestations include developmental delay, autism, etc. [Patient ID: 289813, 312634, etc.]. (4) The DGV database does not include this variation in the general population. In summary, according to the ACMG CNV interpretation and reporting technical standards, its score ranges from -0.89 to 0.89, which is a CNV of unknown clinical significance."

[0100] Example 3

[0101] like Figure 6 As shown, the following format is input in the automated genome copy number variation analysis system of the present invention:

[0102] Please enter the command: "Explain the genetic test results: arr[GRCh38]8p22p21.3(18331312-19532064)×3"

[0103] The answer is:

[0104] "1. The sample was found to have a duplication of approximately 1207.56Kb in the p22p21.3 region of chromosome 8. (1) This region contains one protein-coding gene. (2) The ClinGen database does not include triple dosage sensitive genes / regions in this region. (3) The DECIPHER database includes a small number of individuals with copy number duplications in this region, and clinical manifestations include developmental delay, etc. [Patient ID: 466437, etc.]. (4) The DGV database does not include this variation in the general population. In summary, according to the ACMG CNV interpretation and reporting technical standards, its score ranges from -0.89 to 0.89, which is a CNV of unknown clinical significance."

[0105] Example 4

[0106] like Figure 7 As shown, the following format is input in the automated genome copy number variation analysis system of the present invention:

[0107] Please enter the command: "Explain the genetic test results: arr[GRCh38]15q11.2(22784734-23102074)×0"

[0108] The answer is:

[0109] "1. This sample was found to have a deletion of approximately 317.34Kb in the 15q11.2 region of chromosome 15. (1) This region contains 4 protein-coding genes. (2) The ClinGen database contains a triple dosage-sensitive region in this region: 15q11.2 recurrent region (BP1-BP2) (includes NIPA1) [ISCA-37414]. (3) The excess dosage pathogenicity score of the 15q11.2 recurrent region (BP1-BP2) (includes NIPA1) region is 3, with a penetrance of approximately 8%-10%, which can cause the disease 15q11q12 microduplication syndrome. 15q11q12 microduplication syndrome has clinical heterogeneity, and patients mainly present with developmental delay, intellectual disability, epilepsy, autism spectrum disorder, schizophrenia, congenital heart disease, endocrine abnormalities, etc. (4) The DGV database does not include this variant in the general population. In summary, according to the ACMG CNV interpretation and reporting technical standards, its score is higher than 0.99 points, which is a pathogenic CNV. "

[0110] Example 5

[0111] like Figure 8 As shown, the following format is input in the automated genome copy number variation analysis system of the present invention:

[0112] Please enter the command: "Explain the genetic test results: arr[GRCh38]12p11.21p11.1(32978474-34022859)×1"

[0113] The answer is:

[0114] "1. This sample was found to have a deletion of approximately 1048.86Kb in the p11.21p11.1 region of chromosome 12. (1) This region contains 5 protein-coding genes. (2) The ClinGen database does not include haploinsufficiency-sensitive genes / regions in this region. (3) The DECIPHER database does not include individuals with copy number deletions in this region. (4) The DGV database does not include this variant in the general population. In summary, according to the ACMG CNV interpretation and reporting technical standards, its score ranges from -0.89 to 0.89, which is a CNV of unknown clinical significance."

[0115] Example 6

[0116] like Fig. 9 As shown, the following format is input in the automated genome copy number variation analysis system of the present invention:

[0117] Please enter the command: "Explain the genetic test results: arr(X)×2,(Y)×1"

[0118] The answer is:

[0119] "The result showed that the fetus had 47,XXY syndrome, also known as Klinefelter syndrome. This abnormal chromosome can cause the fetus to have clinical manifestations such as testicular dysplasia, tall stature, infertility, breast development, fine and white skin, sparse beard, and a thin and high-pitched voice."

[0120] Example 7

[0121] like Fig.10 As shown, the following format is input in the automated genome copy number variation analysis system of the present invention:

[0122] Please enter the command: "Explain the genetic test results: seq[hg19]dup(2)(q34)chr2:g.212600001_213806000dup"

[0123] The answer is:

[0124] "1. The sample was found to have a duplication of about 1.2Mb in the q34 region of chromosome 2. (1) This region contains three protein-coding genes. (2) The ClinGen database does not include triple dosage sensitive genes / regions in this region. (3) The DECIPHER database contains a small number of individuals with copy number duplications in this region, and clinical manifestations include autism, developmental delay, etc. [Patient ID: 506842, 314718, etc.]. (4) The DGV database does not include this variation in the general population. In summary, according to the ACMG CNV interpretation and reporting technical standards, its score ranges from -0.89 to 0.89, which is a CNV of unknown clinical significance. 2. Genetic counseling is recommended."

[0125] From the output results of Examples 1-7, it can be seen that the automated interpretation method and analysis system of genome copy number variation based on a large language model of the present invention can automatically make accurate and reliable interpretations of CNV fragment detection results and issue interpretation reports, making the diagnosis of clinical genetic cases more efficient.

[0126] Those skilled in the art will appreciate that all or part of the functions of the various methods in the above-mentioned embodiments can be implemented by hardware or by computer program. When all or part of the functions in the above-mentioned embodiments are implemented by computer program, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, disk, CD, hard disk, etc., and the program is executed by a computer to implement the above-mentioned functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above-mentioned functions can be implemented. In addition, when all or part of the functions in the above-mentioned embodiments are implemented by computer program, the program can also be stored in a storage medium such as a server, another computer, disk, CD, flash disk or mobile hard disk, and saved in the memory of the local device by downloading or copying, or the system of the local device is updated, and when the program in the memory is executed by the processor, all or part of the functions in the above-mentioned embodiments can be implemented.

[0127] The above-mentioned embodiments only express the implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A method for realizing automated interpretation of genome copy number variation, characterized in that: The following steps are involved: Get the large language model of ChatGLM-6B; Establish a data set for interpreting the results of genome copy number variation detection, and desensitize, correct, normalize, and convert the data into training data; Using a low-rank adaptation algorithm to adjust and learn the large language model, so that the large language model acquires the ability to interpret copy number variation result reports; In the large language model for adjustment learning, multiple prompt words are designed to obtain corresponding input-output pairs, the input-output pairs are manually aligned, and the model is adjusted twice using a reinforcement learning algorithm; Generate new copy number variation detection result interpretation dataset.

2. The method according to claim 1, characterized in that After the new copy number variation detection result interpretation data set is generated, it is learned online and continuously fine-tuned through an online learning module.

3. The method according to claim 1, characterized in that The data desensitization is to desensitize the user's privacy-sensitive information, and the user information is converted into ciphertext using a hash coding algorithm; And / or, the data correction is to manually correct the erroneous data and statements appearing in the historical data to obtain correct information for model training; And / or, the data normalization is to process the result interpretation data with different formats into a general format to adjust the model training; And / or, the conversion into training data is to convert the result interpretation data into a training data format of natural language question-natural language answer.

4. The method according to claim 1, characterized in that: The method of using a low-rank adaptation algorithm to adjust and learn a large language model includes the following steps: In the initialization stage, for the weight matrix of the self-attention layer to be fine-tuned in the pre-trained model, it is regarded as a fixed original parameter matrix W0; two low-rank matrices A and B are defined for each matrix to be adjusted, where A∈Rr×k, B∈Rd×r, r is a rank much smaller than d and k, d and k represent the number of rows and columns of the original weight matrix respectively, and matrix A is initialized with Gaussian distribution, and B is initialized as a zero matrix; In the training phase, the low-rank matrices A and B are trained alternately using the stochastic gradient descent scheme, so that the model update is only achieved through ΔW=BA, where ΔW is the adjustment change relative to the original parameters; the model is trained using the dataset of the target task, and the loss function is optimized to obtain the optimal A and B; In the application stage, W=W0+ΔW forms a new parameter matrix, and then W is used to adjust the copy number variation results of the large model for interpretation and inference.

5. The method according to claim 1, characterized in that: The step of manually aligning the input-output pair and performing secondary adjustment on the model using a reinforcement learning algorithm comprises the following steps: Manually review the input-output pairs. If the answers do not conform to human values ​​or are unreasonable, then add the answer scores to construct a negative sample set score. The added answer scores include: "The following answer is unreasonable", "The following answer is incorrect", or "The following answer is not in line with values"; By using the PPO algorithm in reinforcement learning, the low-rank matrix B / A described in claim 4 is updated using the score to obtain ΔW'=B'A'; Apply W = W0 + ΔW', and then use W to adjust the copy number variation results of the large model for interpretation and inference.

6. The method according to claim 2, characterized in that The online learning modules include: The new data cleaning function is used to clean and organize the new result interpretation data into a unified model requirement format, and convert the result interpretation data into a pair of natural language questions and natural language answers; A learning module, wherein the training mode is to use the new data as an adjustment data set and train W” using the method of claim 4, wherein the rank parameter r is set to be consistent with the rank parameter of the current fine-tuning model; Application module, apply W = W0 + ΔW", and then use W to interpret the copy number variation results of the fine-tuned large model.

7. A device for realizing automatic interpretation of genome copy number variation, characterized in that: include: Memory, used to store programs; A processor, configured to execute the program stored in the memory to implement the method according to any one of claims 1 to 6.

8. An automated analysis system for genome copy number variation, characterized in that: include: The device for realizing automatic interpretation of genome copy number variation according to claim 7 is used to train, adjust and learn the large language model so that it can obtain the ability to interpret the copy number variation result report; An information input module, used to input copy number variation data to be analyzed; The data analysis statement generation module enables the trained large language model to automatically interpret the input copy number variation data and output the corresponding results.

9. The automated genome copy number variation analysis system according to claim 8, characterized in that: The corresponding results of the output are presented in written language.

10. A computer-readable storage medium, characterized in that: The device stores a program, which can be executed by a processor to implement the method according to any one of claims 1 to 6.

Images