Thyroid cancer NGS conjoint analysis method based on fusion gene dynamic monitoring

Abstract

The invention relates to the technical field of nucleic acid molecule detection, and discloses a thyroid cancer NGS conjoint analysis method based on fusion gene dynamic monitoring. According to the method, high-throughput sequencing and a DNABERT deep learning algorithm are combined, so that highly specific single nucleotide variation (SNV) and fusion gene markers can be identified from massive sequencing data and are used for judging whether residual tumors or recurrence signs exist in the body of a patient or not, and whether the patient needs to pass a further detection method or not is judged; and diagnosis is carried out, or effect evaluation on the drug reaction is realized. According to the method, a deep learning algorithm is introduced into dynamic monitoring of the thyroid cancer ctDNA, so that the detection rate and the accuracy of the micro marker are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of nucleic acid molecule detection, and in particular to a combined NGS analysis method for thyroid cancer for dynamic monitoring of fusion genes. Background Art

[0002] Thyroid cancer patients remain at risk of recurrence and metastasis after surgical treatment, necessitating long-term follow-up and dynamic monitoring. Traditional follow-up relies primarily on imaging studies (such as ultrasound) and serum markers (such as thyroglobulin Tg), but these studies have limited detection rates for early recurrences of small lesions. The emergence of circulating tumor DNA (ctDNA) offers a new approach for dynamic tumor monitoring: ctDNA is sampled and analyzed for fusion gene content. This intermediate result is used to determine whether further testing is warranted to detect abnormal cell proliferation, thereby reflecting molecular changes in the tumor and overcoming tumor heterogeneity.

[0003] A fusion gene is a process in which the partial or complete sequences of two or more genes form a new chimeric gene. Fusion genes can lead to abnormalities in protein expression levels, function, and site of action, which can cause abnormal cell proliferation and promote the occurrence or development of tumors.

[0004] To improve the reliability of thyroid cancer marker detection, existing MRD (minimal residual disease) testing protocols typically identify specific mutations in a patient's initial tumor and then perform ultra-deep sequencing of these mutations in follow-up blood samples. However, the massive amount of data generated by high-throughput sequencing (NGS) requires efficient analysis methods to distinguish true low-frequency mutation signals from sequencing errors. Existing research has shown that training machine learning models to analyze sequencing data from patient blood can significantly improve the signal-to-noise ratio, making marker detection more reliable.

[0005] In view of this, it is necessary to propose a new method for detecting thyroid cancer markers to improve its detection accuracy and sensitivity, and use it as intermediate result information to determine whether further testing is needed to diagnose thyroid cancer. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a combined NGS analysis method for thyroid cancer for dynamic monitoring of fusion genes, which combines high-throughput sequencing (NGS) technology with a deep learning algorithm to monitor postoperative recurrence of thyroid cancer patients and evaluate the efficacy of targeted drugs.

[0007] The specific technical solutions of the present invention are: In one aspect, the present invention provides a combined NGS analysis method for thyroid cancer for dynamic monitoring of fusion genes, comprising the following steps: (1) Obtaining gene sequencing data of a target biological sample, wherein the sequencing data includes candidate mutation information; (2) Inputting the sequence fragments corresponding to the candidate mutation information into the pre-trained DNABERT deep learning model, the DNABERT model calculates the confidence score of each candidate mutation sequence based on the k-mer pattern features of the sequence; (3) Screening out high-confidence variant sequences based on the confidence score and preset rules; (4) Output the identification information of the high-confidence variant sequence and its abundance information in the sample.

[0008] The method of the present invention combines high-throughput sequencing with the DNABERT deep learning algorithm, making it possible to identify highly specific single nucleotide variants (SNVs) and fusion gene markers from massive sequencing data. This can be used to determine whether there are residual tumors or signs of recurrence in the patient's body, thereby determining whether the patient needs further testing methods for diagnosis, or to evaluate the effectiveness of drug response.

[0009] The DNABERT deep learning model is used to perform specific analysis on the identified mutation sequences, screening for a highly specific marker set. High specificity refers to the fact that these mutations are extremely rare in normal individuals and clearly originate from tumor clones, thus avoiding misinterpretation of benign polymorphisms or somatic background mutations as signals of recurrence. The method of the present invention organically combines high-sensitivity NGS detection with intelligent deep learning analysis. On the one hand, NGS is used to capture multiple markers, achieving a ctDNA LoD (limit of detection) of less than 0.05%. On the other hand, the DNABERT model is used to improve the accuracy of signal interpretation and eliminate most false positive noise. This method can timely and accurately capture molecular changes in thyroid cancer patients after surgery, providing early warning and personalized treatment decision support for clinicians. The method of the present invention is used for long-term follow-up and dynamic monitoring of patients after surgical treatment.

[0010] Deep learning models can be used to assess the degree of abnormality or significance of each mutation sequence. For example, the DNABERT model extracts features from mutant sequence fragments and, compared to a large training set of normal sequences, assigns high weights to variants that significantly deviate from normal sequence patterns, helping to select the most representative tumor markers.

[0011] As a preferred embodiment of the above method, in step (1), the gene sequencing data is obtained by: (a) performing targeted capture and enrichment of nucleic acids in the biological sample based on a preset personalized probe combination; (b) constructing and sequencing the enriched nucleic acid library, wherein the library construction process introduces a unique molecular identifier (UMI) tag; (c) Performing preliminary alignment and variant screening on the sequencing data to obtain the candidate variant information.

[0012] Further preferably, the design of the personalized probe combination is based on tumor-specific mutation markers screened out after gene sequencing of tumor tissue samples from the same source, and the markers include single nucleotide variations (SNVs) and / or fusion gene breakpoint regions. Further preferably, the personalized probe combination covers a predetermined length sequence before and after the SNV site and / or a predetermined length sequence before and after the fusion gene breakpoint. As a preferred embodiment of the above method, in step (1), the sequencing depth reaches or exceeds 20,000x.

[0013] As a preferred embodiment of the above method, in step (2), the calculation of the DNABERT model includes: extracting contextual k-mer features of the candidate variant sequence fragment, and judging the degree of its deviation from the normal sequence pattern based on the features, and outputting a confidence score.

[0014] As a preferred embodiment of the above method, in step (3), the preset rules include: the confidence score is higher than a preset threshold, and the variant sequence is supported by reads of at least two independent UMI labels.

[0015] As a preferred embodiment of the above method, in step (4), the abundance information outputted is the allele frequency of the high-confidence variant sequence or the ratio of the number of supporting reads to all valid reads. As a preferred embodiment of the above method, the biological sample is a body fluid sample or an isolate thereof derived from a human body.

[0016] Based on the above analysis method, on the other hand, the present invention provides a data processing system, comprising: A data input interface for receiving gene sequencing data or candidate mutation information of biological samples; A processor configured to execute: (i) Call the pre-trained DNABERT deep learning model to calculate the input candidate variant sequence fragments and output the confidence score of each candidate variant sequence; (ii) screening out high-confidence variant sequences based on the confidence score and preset rules; (iii) outputting identification information of the high-confidence variant sequence and its abundance information in the sample; An output interface is used to output the identification information and abundance information of the high-confidence variant sequence.

[0017] Compared with the prior art, the present invention has the following technical effects: 1. The method of the present invention combines high-throughput sequencing with the DNABERT deep learning algorithm to identify highly specific single nucleotide variants (SNVs) and fusion gene markers from massive sequencing data. This can be used to determine whether there are residual tumors or signs of recurrence in the patient's body, thereby determining whether the patient needs further testing methods for diagnosis, or to evaluate the effect of drug response.

[0018] 2. The method of the present invention introduces deep learning algorithms into the dynamic monitoring of ctDNA in thyroid cancer, thereby improving the detection rate and accuracy of micromarkers. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the following embodiments. Those skilled in the art will be able to implement the present invention based on these descriptions. Furthermore, the embodiments of the present invention described below are generally only a portion of the embodiments of the present invention, rather than all of the embodiments. Therefore, all other embodiments derived by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0020] Example 1 This embodiment provides a combined NGS analysis method for thyroid cancer for dynamic monitoring of fusion genes, which includes the following steps: (1) Obtaining gene sequencing data of a target biological sample, wherein the sequencing data includes candidate mutation information; (2) Inputting the sequence fragments corresponding to the candidate mutation information into the pre-trained DNABERT deep learning model, the DNABERT model calculates the confidence score of each candidate mutation sequence based on the k-mer pattern features of the sequence; (3) Screening out high-confidence variant sequences based on the confidence score and preset rules; (4) Output the identification information of the high-confidence variant sequence and its abundance information in the sample.

[0021] This method combines high-throughput sequencing with the DNABERT deep learning algorithm to identify highly specific single nucleotide variants (SNVs) and fusion gene markers from massive sequencing data. This can be used to determine whether there are residual tumors or signs of recurrence in the patient's body, thereby determining whether the patient needs further testing for diagnosis or to evaluate the effectiveness of drug response.

[0022] The DNABERT deep learning model is used to perform specific analysis on the identified mutation sequences to screen for a highly specific marker set. High specificity refers to the fact that these mutations are extremely rare in normal individuals and clearly originate from tumor clones, thus avoiding misinterpretation of benign polymorphisms or somatic background mutations as recurrence signals. This method organically combines high-sensitivity NGS detection with intelligent deep learning analysis. On the one hand, it uses NGS to capture multiple markers, achieving a ctDNA LoD (limit of detection) of less than 0.05%. On the other hand, it utilizes the DNABERT model to improve the accuracy of signal interpretation and eliminate most false-positive noise. This method can timely and accurately capture molecular changes in patients with thyroid cancer after surgery, providing early warning and personalized treatment decision support for clinicians. The method of the present invention is used for long-term follow-up and dynamic monitoring of patients after surgical treatment.

[0023] Deep learning models can be used to assess the degree of abnormality or significance of each mutation sequence. For example, the DNABERT model extracts features from mutant sequence fragments and, compared to a large training set of normal sequences, assigns high weights to variants that significantly deviate from normal sequence patterns, helping to select the most representative tumor markers.

[0024] In this embodiment, as a preferred embodiment of the above method, in step (1), the gene sequencing data is obtained by: (a) performing targeted capture and enrichment of nucleic acids in the biological sample based on a preset personalized probe combination; (b) constructing and sequencing the enriched nucleic acid library, wherein the library construction process introduces a unique molecular identifier (UMI) tag; (c) Performing preliminary alignment and variant screening on the sequencing data to obtain the candidate variant information.

[0025] Further preferably, the design of the personalized probe combination is based on tumor-specific mutation markers screened out after gene sequencing of tumor tissue samples from the same source, and the markers include single nucleotide variations (SNVs) and / or fusion gene breakpoint regions. Further preferably, the personalized probe combination covers a predetermined length sequence before and after the SNV site and / or a predetermined length sequence before and after the fusion gene breakpoint. In this embodiment, as a preferred embodiment of the above method, in step (1), the sequencing depth reaches or exceeds 20,000x.

[0026] In this embodiment, as a preferred embodiment of the above method, in step (2), the calculation of the DNABERT model includes: extracting contextual k-mer features of the candidate variant sequence fragment, and judging the degree of its deviation from the normal sequence pattern based on the features, and outputting a confidence score.

[0027] In this embodiment, as a preferred embodiment of the above method, in step (3), the preset rule includes: the confidence score is higher than a preset threshold, and the variant sequence is supported by reads of at least two independent UMI labels.

[0028] In this embodiment, as a preferred embodiment of the above method, in step (4), the abundance information output is the allele frequency of the high-confidence variant sequence or the ratio of the number of supporting reads to all valid reads. Example 2 Based on the method provided in Example 1, this embodiment provides another embodiment mode, specifically: This embodiment provides a combined NGS analysis method for thyroid cancer for dynamic monitoring of fusion genes, which includes the following steps: (1) Sample preparation and panel design: Tumor tissues from 10 patients with advanced thyroid cancer and corresponding follow-up plasma samples within six months after surgery were selected. First, whole-exome sequencing was performed on the tumor tissues of each patient, and it was found that each case contained an average of 15 high-frequency somatic SNV mutations (such as BRAF V600E, NRAS Q61R, etc.) and 1-2 gene fusion events (such as RET-CCDC6, EML4-ALK fusion, etc.). Based on the results of each patient, a custom capture probe panel was designed, with a size of approximately 0.5Mb, covering the exon regions and fusion breakpoint regions of all identified mutant genes. Ten patients with benign thyroid nodules were selected as negative controls.

[0029] (2) Sequencing and preliminary analysis: cfDNA (approximately 20 ng per sample) was extracted from the 20 plasma samples mentioned above, and the capture library was constructed using the corresponding panel and sequenced (NovaSeq 6000, PE150 mode). After the sequencing data were generated, the standard process was used for sequence alignment and variant screening. The results showed that several low-frequency mutation candidates were detected in the plasma of actual thyroid cancer patients. For example, the BRAF V600E mutation allele frequency was 0.1%, the TERT promoter C228T mutation was 0.05%, and 5 RET / PTC1 fusion support reads were initially screened in the plasma of patient A. However, only a small number of background mutations and no clear fusion reads were detected in the control group patients. At this time, it is difficult to determine which signals are of true tumor origin by conventional variant detection alone.

[0030] (3) Deep learning screening: All candidate mutation information was input into the pre-trained DNABERT model for evaluation. The model gave a score to each sequence, with a threshold set at 0.8 (out of 1.0). In the data of patient A, the BRAF V600E mutation fragment received a high score of 0.95 and was supported by ≥10 UMIs, which was determined to be a true positive; the TERT mutation fragment scored 0.90 and had a support number of 8, which was also a positive signal; the RET / PTC1 fusion fragment was not included in the high confidence list because it had 5 supporting reads but the model score was only 0.6 (possibly due to the poor quality of some reads). In another patient B, DNABERT identified the NRAS Q61R mutation (frequency 0.08%) as a true positive signal, while excluding several scattered variants with low dispersion in different genes (determined to be sequencing noise). All candidate variants in the control group were judged by the model to be of low confidence (score <0.5), which was consistent with the expected situation of no tumor signal.

[0031] (4) Report generation: Based on the model screening results, the system automatically generates an analysis report. For example, the report for patient A contains the following summary: "Braf V600E (allele frequency 0.10%) and TERT promoter mutations (0.05%) related to thyroid cancer were detected, suggesting the possibility of molecular residual lesions. It is recommended to strengthen imaging follow-up. No new drug-resistant mutations were detected." The report for patient B shows: "NRAS Q61R mutation (0.08%) was detected, and no abnormalities were found in other markers. Compared with the previous results, the NRAS mutation copy number has increased, suggesting a possible tendency for minor recurrence. It is recommended to repeat the neck ultrasound." All reports contain a list of mutations, a time-dependent dynamic change graph, and a corresponding risk level description.

[0032] (5) Model verification: Among the 10 patients mentioned above, 3 were clinically confirmed to have radiographic recurrence or metastasis within the next 6 months. It is worth noting that these 3 patients showed a clear trend of increasing ctDNA markers in the detection method of the present invention, with an average warning of 4 months earlier than the radiographic diagnosis. In the 7 patients who did not relapse, this method did not detect a continuous increase in tumor markers, and no signs of recurrence were found during their follow-up, which proved the specificity of this method. In contrast, no tumor-related mutations were detected in patients with benign nodules in the control group.

[0033] It can be seen that the method of the present invention can monitor molecular changes in thyroid cancer patients with high sensitivity, detect signs of recurrence in advance, and effectively distinguish real tumor signals from noise (specificity of approximately 100%), which has important clinical application value.

[0034] Example 3 Based on the method provided in Example 1, this embodiment provides a data processing system, which includes: A data input interface for receiving gene sequencing data or candidate mutation information of biological samples; A processor configured to execute: (i) Call the pre-trained DNABERT deep learning model to calculate the input candidate variant sequence fragments and output the confidence score of each candidate variant sequence; (ii) screening out high-confidence variant sequences based on the confidence score and preset rules; (iii) outputting identification information of the high-confidence variant sequence and its abundance information in the sample; An output interface is used to output the identification information and abundance information of the high-confidence variant sequence.

[0035] Unless otherwise specified, the raw materials and equipment used in the present invention are commonly used in the art; the methods used in the present invention are conventional methods in the art unless otherwise specified.

[0036] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent transformation made to the above embodiment based on the technical essence of the present invention still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A combined NGS analysis method for thyroid cancer for dynamic monitoring of fusion genes, characterized by: The following steps are involved: (1) Obtaining gene sequencing data of a target biological sample, wherein the sequencing data includes candidate mutation information; (2) Inputting the sequence fragments corresponding to the candidate mutation information into the pre-trained DNABERT deep learning model, the DNABERT model calculates the confidence score of each candidate mutation sequence based on the k-mer pattern features of the sequence; (3) Screening out high-confidence variant sequences based on the confidence score and preset rules; (4) Output the identification information of the high-confidence variant sequence and its abundance information in the sample.

2. The analysis method according to claim 1, wherein: In step (1), gene sequencing data is obtained by: (a) performing targeted capture and enrichment of nucleic acids in the biological sample based on a preset personalized probe combination; (b) constructing and sequencing the enriched nucleic acid library, wherein the library construction process introduces UMI tags; (c) Performing preliminary alignment and variant screening on the sequencing data to obtain the candidate variant information.

3. The analysis method according to claim 2, wherein: The design of the personalized probe combination is based on tumor-specific mutation markers screened after gene sequencing of tumor tissue samples from the same source, and the markers include SNV and / or fusion gene breakpoint regions.

4. The analysis method according to claim 3, wherein: The personalized probe combination covers a predetermined length sequence before and after the SNV site and / or a predetermined length sequence before and after the fusion gene breakpoint.

5. The analysis method according to claim 2, wherein: In step (1), the sequencing depth is at least 20,000x.

6. The analysis method according to claim 1, wherein: In step (2), the calculation of the DNABERT model includes: extracting context k-mer features of the candidate variant sequence fragment, and judging the degree of its deviation from the normal sequence pattern based on the features, and outputting a confidence score.

7. The analysis method according to claim 1, wherein: In step (3), the preset rules include: the confidence score is higher than a preset threshold, and the variant sequence is supported by reads of at least two independent UMI labels.

8. The analysis method according to claim 1, wherein: In step (4), the output abundance information is the allele frequency of the high-confidence variant sequence or the ratio of the number of supporting reads to all valid reads.

9. The analysis method according to claim 1, wherein: The biological sample is a body fluid sample or a separate substance thereof derived from a human body.

10. A data processing system, characterized in that: include: A data input interface for receiving gene sequencing data or candidate mutation information of biological samples; A processor configured to execute: (i) Call the pre-trained DNABERT deep learning model to calculate the input candidate variant sequence fragments and output the confidence score of each candidate variant sequence; (ii) screening out high-confidence variant sequences based on the confidence score and preset rules; (iii) outputting identification information of the high-confidence variant sequence and its abundance information in the sample; An output interface is used to output the identification information and abundance information of the high-confidence variant sequence.