A reagent kit for differentiating benign and malignant pulmonary nodules and its application.

The kit, which integrates whole-genome methylation profiles and fragment omics features of cfDNA to construct an XGBoost algorithm model, solves the problems of high false positive rate and insufficient detection sensitivity in lung nodule screening, and achieves high-precision, non-invasive lung nodule identification, reducing the risks of invasive examinations and overdiagnosis.

CN123108148APending Publication Date: 2026-11-17SUN YAT SEN UNIVERSITY CANCER CENTER (CANCER HOSPITAL AFFILIATED TO SUN YAT SEN UNIVERSITY CANCER RESEARCH INSTITUTE OF SUN YAT SEN UNIVERSITY)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611292484.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing lung nodule screening methods suffer from high false positive rates, significant risks associated with invasive biopsies, and insufficient sensitivity of traditional molecular testing, making it difficult to accurately distinguish between early-stage malignant and benign lung nodules.

Method used

We integrated and analyzed the whole-genome methylation profile and fragment omics characteristics of cfDNA to construct a diagnostic model for non-small cell lung cancer based on the XGBoost algorithm. The model was evaluated using a kit to differentiate between benign and malignant lung nodules, including the application of nucleic acid extraction, library construction, and machine learning models.

Benefits of technology

It achieves high-precision, non-invasive differentiation between benign and malignant lung nodules, reduces the false positive rate, reduces unnecessary invasive examinations, improves the ability to identify and diagnose early lung cancer, and reduces the risk of overtreatment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123108148A_ABST
    Figure CN123108148A_ABST
Patent Text Reader

Abstract

This invention discloses a kit for differentiating benign and malignant pulmonary nodules and its application. The kit includes nucleic acid extraction reagents for extracting cell-free DNA from plasma, library construction reagents for constructing GM-seq libraries, and a non-small cell lung cancer diagnostic model based on the XGBoost algorithm. The GM-seq library construction reagents include DNA fragmentation reagents, end repair reagents, A / T tail addition reagents, adapter ligation reagents, TET enzyme, pyridine borane, PCR amplification reagents, and library cyclization reagents. The kit enables lung cancer risk assessment based on the following 12 features: nine differentially methylated regions (DMR27734, DMR55120, DMR114989, DMR66665, DMR122523, DMR3325, DMR14627, DMR3582, and DMR35337), and three fragment group features: 90-150 bp frequency, 151-220 bp frequency, and ratio.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of biomedical detection technology, specifically relating to a reagent kit for differentiating benign and malignant pulmonary nodules and its application. Background Technology

[0002] Lung cancer is the leading cause of cancer-related morbidity and mortality in my country, and early diagnosis and treatment are crucial for improving the survival of lung cancer patients. However, due to atypical clinical manifestations, the current diagnosis rate for early-stage lung cancer is only 19%. As a major early manifestation of lung cancer, the scientific assessment of the nature of pulmonary nodules is of decisive significance for early diagnosis and treatment. Low-dose spiral CT (LDCT) is the main method for early lung cancer screening, but it suffers from a high false-positive rate (20-27%). This high false-positive rate can lead to overdiagnosis and increased patient burden, among other clinical problems. Currently, the assessment of pulmonary nodules detected through screening is mainly based on their imaging size, shape, location, and dynamic comparison. Finding a balance between avoiding missed diagnoses of lung cancer and unnecessary treatment of benign nodules has become a significant challenge for clinicians.

[0003] In recent years, liquid biopsy technology, with cell-free DNA (cfDNA) detection as its core, has shown great potential in the field of early cancer diagnosis. DNA methylation, as an important epigenetic marker, changes in the early stages of lung cancer development and has strong tissue specificity and stability, making it an ideal biomarker for early lung cancer diagnosis.

[0004] However, existing cfDNA methylation detection technologies still have significant limitations. Traditional bisulfite sequencing methods cause severe chemical degradation of DNA during processing, leading to reduced library complexity and limited sequence capture efficiency. Meanwhile, existing commercially available detection chips have low coverage of genome-wide methylation sites (only about 2%-3%), making it difficult to provide sufficient sensitivity to accurately distinguish between early-stage malignant and benign pulmonary nodules. Therefore, there is an urgent clinical need for a highly sensitive, non-invasive molecular detection tool that can accurately differentiate between benign and malignant pulmonary nodules to compensate for the shortcomings of imaging screening and guide the development of personalized treatment plans. Summary of the Invention

[0005] To address the technical challenges in current lung nodule screening, such as high false positive rates, significant risks associated with invasive biopsies, and insufficient sensitivity of traditional molecular testing, this invention aims to construct a high-precision prediction model by integrating and analyzing the whole-genome methylation profile and fragment omics characteristics of cfDNA. This will provide a non-invasive, highly sensitive, and highly specific auxiliary diagnostic kit and application for benign and malignant lung nodules in clinical practice.

[0006] To achieve the above objectives, the technical solution provided by this invention is as follows: The kit for differentiating benign and malignant pulmonary nodules includes nucleic acid extraction reagents for extracting cell-free DNA from plasma and library construction reagents for constructing GM-seq libraries. The GM-seq library construction reagents include DNA fragmentation reagents, end repair reagents, A / T tail addition reagents, adapter ligation reagents, TET enzyme, pyridine borane, PCR amplification reagents, and library cyclization reagents. The kit uses the XGBoost algorithm-based non-small cell lung cancer diagnostic model to assess lung cancer risk based on the following 12 features: 9 differentially methylated regions DMR27734 (chr1:1289004), DMR55120... (chr1:2097679), DMR114989 (chr1:4715576), DMR66665 (chr1:2433914), DMR122523 (chr1:5366740), DMR3325 (chr1:437583), DMR14627 (chr1:980109), DMR3582 (chr1:448987) and DMR35337 (chr1:1488625), and three fragment group characteristics: 90-150bp frequency, 151-220bp frequency and ratio.

[0007] Preferably, the DNA fragmentation reagent is used to randomly fragment genomic DNA to 200-300 bp.

[0008] Preferably, the TET enzyme is used to oxidize 5-methylcytosine and 5-hydroxymethylcytosine, and the pyridineborane is used to reduce the oxidation product to dihydrouracil.

[0009] Preferably, the non-small cell lung cancer diagnostic model is a machine learning model.

[0010] Preferably, the kit for identifying benign and malignant pulmonary nodules further includes a positive control and a negative control; the positive control is pUC19 DNA and the negative control is lambda DNA.

[0011] Preferably, in the fragment group features, the proportion of fragments with a frequency of 90-150bp and a length of 100-150bp in the total fragments, and the proportion of fragments with a frequency of 151-220bp and a length of 151-220bp in the total fragments, where ratio is the ratio of short fragments to long fragments.

[0012] Preferably, the area under the curve (AUC) of the non-small cell lung cancer diagnostic model on the validation set is not less than 0.9.

[0013] Preferably, the non-small cell lung cancer diagnostic model has a sensitivity of not less than 0.9 and a specificity of not less than 0.8.

[0014] The kit for differentiating benign and malignant pulmonary nodules described in this invention can be used to prepare early diagnostic products for lung cancer.

[0015] The present invention will be further described below: The kit for differentiating benign and malignant pulmonary nodules described in this invention includes: a nucleic acid extraction reagent for extracting cell-free DNA from plasma, a library construction reagent for constructing a GM-seq library, and a non-small cell lung cancer diagnostic model based on the XGBoost algorithm. The GM-seq library construction reagent includes DNA fragmentation reagents, end repair reagents, A / T tail addition reagents, adapter ligation reagents, TET enzyme, pyridine borane, PCR amplification reagents, and library cyclization reagents. The non-small cell lung cancer diagnostic model assesses lung cancer risk based on the following 12 features: (1) 9 differentially methylated regions: DMR27734 (chr1:1289004), DMR55120 (chr1:2097679), DMR114989 (chr1:4715576), DMR66665 (chr1:2433914), DMR122523 (chr1:5366740), DMR3325 (chr1:437583), DMR14627 (chr1:980109), DMR3582 (chr1:448987) and DMR35337 (chr1:1488625); (2) 3 fragment omics features: 90-150bp frequency, 151-220bp frequency and ratio. The non-small cell lung cancer diagnostic model based on the XGBoost algorithm is a machine learning model. XGBoost (eXtreme Gradient Boosting) is an efficient gradient boosting decision tree algorithm with advantages such as high accuracy, high efficiency, and easy parallelization, making it suitable for large-scale data analysis.

[0016] The process of differentiating benign from malignant pulmonary nodules using the kit described in this invention is as follows: (1) Extract cell-free DNA from peripheral blood samples of the subjects; (2) The plasma cell-free DNA was used to construct a GM-seq library and sequence it to obtain raw sequencing data; (3) The original sequencing data were filtered, aligned and methylation sites were detected, and the methylation levels of 9 differentially methylated regions were extracted; (4) Perform fragment omics analysis on the original sequencing data and extract three fragment omics features: 90-150bp frequency, 151-220bp frequency, and ratio; (5) Input the methylation levels of the nine differentially methylated regions and the omics features of the three fragments into the non-small cell lung cancer diagnostic model to obtain lung cancer risk assessment results.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) High precision and high sensitivity: This invention overcomes the shortcomings of insufficient signal from a single biomarker in early lung cancer detection by integrating dual-dimensional information from methylation (epigenetics) and fragmentomics (structural features). Experimental data show that the area under the curve (AUC) of this invention on the validation set is approximately 0.9, the F1 score exceeds 0.8, the sensitivity is close to 0.93, and the negative predictive value (NPV) is approximately 0.91, significantly improving the ability to identify early lung cancer.

[0018] (2) High specificity: The validation set specificity reached 0.83 or higher, indicating that the technology of the present invention also performed well in accurately identifying benign nodules and could effectively reduce unnecessary invasive examinations and surgeries.

[0019] (3) Non-invasive and high compliance: This invention only requires the collection of a small amount of peripheral blood from the subject, avoiding the risk of complications such as bleeding and pneumothorax caused by puncture biopsy, making it suitable for continuous dynamic monitoring of high-risk groups and improving patient compliance.

[0020] (4) Reduce overtreatment: This invention can effectively identify benign nodules (such as inflammation, granulomas, etc.) by accurately reclassifying lung nodules detected by LDCT, reducing unnecessary surgical resection, protecting the patient's lung function, and has important clinical value and socio-economic benefits.

[0021] (5) Feasibility of clinical translation: The nine core differential methylation sites screened in this invention have high consistency in tissues and plasma, ensuring the stability of the detection results and the feasibility of clinical translation.

[0022] (6) High positive predictive value: The positive predictive value (PPV) of the present invention exceeds 75%, which is ideal in clinical screening models. It can significantly narrow down the population that needs further puncture, biopsy or high-resolution CT examination, and greatly improve diagnostic efficiency. Attached Figure Description

[0023] Figure 1 Summary of results from 9 models and 6 indicators. Detailed Implementation

[0024] The kit for differentiating benign and malignant lung nodules includes nucleic acid extraction reagents for extracting cell-free DNA from plasma, library construction reagents for constructing GM-seq libraries, and a non-small cell lung cancer diagnostic model based on the XGBoost algorithm. The GM-seq library construction reagents include DNA fragmentation reagents, end-repair reagents, A / T tail addition reagents, adapter ligation reagents, TET enzyme, pyridine borane, PCR amplification reagents, and library cyclization reagents. The kit enables lung cancer risk assessment based on the following 12 features: 9 differentially methylated regions DMR27734 (chr1:1289004), DMR55120 (chr1... The sequences at these sites (2097679), DMR114989 (chr1:4715576), DMR66665 (chr1:2433914), DMR122523 (chr1:5366740), DMR3325 (chr1:437583), DMR14627 (chr1:980109), DMR3582 (chr1:448987), and DMR35337 (chr1:1488625) are all available online, and three fragment group characteristics are: 90-150 bp frequency, 151-220 bp frequency, and ratio.

[0025] The DNA fragmentation reagent is used to randomly fragment genomic DNA to 200-300 bp. The TET enzyme is used to oxidize 5-methylcytosine and 5-hydroxymethylcytosine, and the pyridineborane is used to reduce the oxidation product to dihydrouracil. The non-small cell lung cancer diagnostic model is a machine learning model. The kit for differentiating benign and malignant lung nodules also includes a positive control and a negative control; the positive control is pUC19 DNA, and the negative control is lambda DNA. In the fragment group characteristics, the 90-150 bp frequency is the proportion of fragments with a length of 100-150 bp in the total fragments, the 151-220 bp frequency is the proportion of fragments with a length of 151-220 bp in the total fragments, and the ratio is the ratio of short fragments to long fragments. The area under the curve (AUC) of the non-small cell lung cancer diagnostic model on the validation set is not less than 0.9. The sensitivity of the non-small cell lung cancer diagnostic model is not less than 0.9, and the specificity is not less than 0.8.

[0026] The usage process of the reagent kit described in this embodiment is as follows: 1. DNA Extraction and Quality Control Genomic DNA was isolated using a TANBead nucleic acid extractor. DNA purity and concentration were assessed using a NanoDrop 8000 and Qubit® 4.0. Samples with an N / Q ratio < 6, DNA concentration > 100 ng / μL, and acceptable DNA degradation levels and no RNA or protein contamination as determined by electrophoresis were considered suitable for further analysis.

[0027] 2. GM-seq library construction and sequencing 300 ng of DNA was taken from each sample, with a certain proportion of negative control (lambda DNA) and a certain proportion of positive control (pUC19 DNA) added. Genomic DNA was first randomly fragmented to 200-300 bp using a Covaris LE220 and then processed using the Hieff NGS Ultima Pro DNA library construction kit. After end repair, A-tailing, and adapter ligation, 5-methylcytosine (5mC) and 5-hydroxymethylcytosine (5hmC) were oxidized using TET enzymes, and then reduced to dihydrouracil (DHU) using pyridineborane. Transformation efficiency was monitored by incorporating controls (0.5% methylated and 1% unmethylated synthetic DNA). DHU can be used as a polymerase chain reaction (PCR) template and is recognized by DNA polymerases that recognize U. PCR enrichment yielded a PCR product converting C to T. The products were quantified using a Qubit real-time fluorescence / ELISA reader, and the length distribution range of the products was also quantified, ensuring the total amount met the requirements and that there was no adapter or large fragment contamination. Finally, the double-stranded target region library was denatured, circularized, and digested to obtain single-stranded circular DNA. The single-stranded circular DNA was then amplified using rolling circle amplification (RCA) technology to obtain DNA nanospheres (DNB). After library construction, Qubit was used for quantitative quality control. Once the library quality was deemed acceptable, sequencing was performed on the DNBSEQ-T7 platform. Specifically, the prepared DNB was loaded onto a patterned array, and sequencing was performed using combined probe anchoring polymerization (cPAS). After the sequencing primers and fluorescent probes polymerized on the DNA nanospheres, a high-resolution imaging system was used to acquire, read, and identify the light signals to obtain single-base sequence information. The process was repeated to obtain the next base sequence information, ultimately generating 150 bp paired-end sequencing data.

[0028] 3. GM-seq methylation data analysis Analysis of GM-seq methylation sequencing data: After obtaining the raw sequencing data, the data was filtered and quality controlled, compared and quality controlled, methylated C was detected, and methylated C was statistically evaluated.

[0029] The raw image data files obtained from high-throughput sequencing are converted into sequenced reads through base calling analysis. The results are stored in FASTQ (fq) file format, which we call raw reads. A FASTQ file contains the name of each sequenced read, its base sequence, and its corresponding sequencing quality information. In the FASTQ file, each base corresponds to a base quality character. The ASCII code value of each base quality character minus 33 is the sequencing quality score (Phred Quality Score). Different sequencing quality scores represent different sequencing error rates. For example, Phred Quality Score values ​​of 20 and 30 represent sequencing error rates of 1.0% and 0.1%, respectively. A higher quality score indicates a lower sequencing error rate and higher accuracy. A small portion of raw reads contain adapter information, low-quality bases, or undetected bases, which can interfere with subsequent data analysis. The raw reads were filtered and quality control statistics were performed using the FastP software (https: / / github.com / OpenGene / fastp1), according to the following criteria: (1) Remove the residual connector sequences at both ends of the reads; (2) Reads with a length of less than 35 bp after removing the linker sequence and low-quality bases at both ends; (3) Remove reads with more than 5% N (N indicates that the base information cannot be determined) base content; (4) Remove reads from single-end reads where the number of low-quality bases (quality value Qff 15) accounts for more than 40% of the total read length.

[0030] Since GM-seq technology only converts mC to T, and mC only accounts for about 5% of the C bases in the entire genome, it has virtually no impact on genome alignment. Here, we use Sentieon software (https: / / www.sentieon.com / ) to align the filtered cleansreads to the reference genome, then sort and label repetitive sequences to obtain the alignment result BAM file. Then, we use Samtools software (https: / / github.com / samtools / samtools) to split the BAM into three parts: human, lambda, and pUC19, and analyze the alignment information for each. After obtaining the alignment results, we use asTair software (https: / / bitbucket.org / bsblabludwig / astair / src / master / ) to detect methylation sites. asTair is a tool specifically designed for analyzing data generated by TAPS and other "mCtoT" technologies that do not rely on pyrophosphate conversion. For the identified methylation sites, the proportions of methylated C, CpG, CHG, and CHH in human, lambda, and PUC19 samples were statistically analyzed. After DNA fragmentation during library construction, end repair occurs, which may introduce unmethylated C sites at the DNA ends. We used an M-biasplot to assess its impact. Theoretically, under random fragmentation without bias, the methylated C sites at each position in the reads should form a uniformly distributed straight line.

[0031] For identified CpG methylation sites, their methylation levels (mC / C) were calculated, and the distribution of different methylated CpG levels and the depth distribution of methylated CpG were analyzed. The methylation levels of CpG sites in genomic functional regions (promoter, 5'UTR, exon, intron, 3'UTR) were statistically analyzed. The methylation site motif represents the base distribution characteristics of the upstream and downstream 9 bp including the mC site, which can characterize sequence conservation and may play a role in gene expression regulation. Motif analysis is of great significance for the discovery of DNA-protein binding sites, etc. For each sequence context (CG, CHG, CHH), we used the sequence information of identified methylated sites (taking 9 bp including the site) to create Logo Plots to study the sequence characteristics upstream and downstream of methylated cytosine in different contexts.

[0032] Methylation status can vary across genomes from various samples (tissues, cells, individuals, etc.), and differentially methylated regions (DMRs) are considered to potentially participate in transcriptional regulation of genes. Identifying differential methylation allows for a comprehensive examination of epigenetic differences across tissues. DNA methylation is well-known to be associated with cell differentiation and proliferation, and many DMRs have been identified during embryonic reprogramming and developmental stages. Cancer samples, compared to normal samples, also exhibit overall methylation loss and hypermethylation in localized regions such as CpG islands in tumor suppressor gene promoter regions. Here, we perform DMR analysis across comparative combinations. We use three software programs—methylkit, rnBeads, and metilene—to identify DMRs, then annotate significantly different DMR regions to obtain differentially expressed gene sets. We filter the raw data using FastP and align it to the hg19 reference genome using Sentieon. We remove duplicate reads and assess alignment quality using SAMtools. Finally, the methylation sites DMR27734, DMR55120, DMR114989, DMR66665, DMR122523, DMR3325, DMR14627, DMR3582, and DMR35337 were identified.

[0033] 4. Fragment group feature analysis and screening High-quality reads (mapping quality ≥ 30) were extracted, sorted, and indexed. The autosomal genome was divided into 504 non-overlapping 5 Mb windows (bins). The number of short (100–150 bp) and long (151–220 bp) reads within each window was counted, and the short / long ratio was calculated for each window. Regional bias was corrected using the Locally Weighted Linear Regression Smoothing (LOESS) method. Three indicators were selected: 90–150 bp frequency, 151–220 bp frequency, and ratio.

[0034] 5. Identification of diagnostic models for non-small cell lung cancer (NSCLC) The 12 features detected above—DMR27734, DMR55120, DMR114989, DMR66665, DMR122523, DMR3325, DMR14627, DMR3582, DMR35337, 90-150bp frequency, 151-220bp frequency, and ratio—are used to identify the risk of lung cancer in the examinee using a product-specific XGBoost-based non-small cell lung cancer (NSCLC) diagnostic model. The model output is a lung cancer risk score, which can be judged as positive or negative based on a preset threshold.

[0035] The XGBoost non-small cell lung cancer (NSCLC) diagnostic model construction process is as follows: (1) Data preparation and sample grouping: Read the sample feature data, divide it into three groups: train / val / test, and construct binary classification labels (response0 / response1). (2) Feature pre-screening: Candidate features are initially screened in parallel based on three strategies: Random Forest (Ranger) feature importance, XGBoost feature importance and stability selection (stabsel), and the union is taken; (3) LASSO feature selection: Based on the cross-validation AUC criterion, logistic regression with L1 regularization (cv.glmnet) is performed on candidate features, and non-zero coefficient features are selected according to lambda.1se; (4) Univariate AUC filtering: Eliminate features with weak univariate discrimination ability (AUC ≤ 0.65); (5) Final feature set determination: Take the intersection of the pre-screened features, LASSO features, and univariate AUC features. If there are more than 50 features, select the top 50 features according to their importance. (6) Data preprocessing pipeline: Perform constant term removal, median imputation, numerical selection and standardization (scale) on features; (7) Candidate model and hyperparameter tuning: For the nine candidate algorithms, 5-fold cross-validation and model-independent optimization (MBO) were used for hyperparameter search; (8) Optimal model evaluation: Select and evaluate the final model by comprehensively considering the AUC, sensitivity, specificity, PPV, NPV and other indicators of the three sets of data: Train / Val / Test. (9) Stage sensitivity analysis: On the training set, the classification threshold is determined according to the target specificity, and the sensitivity of the model to each disease stage (AIS / MIA, I, II, III) is evaluated.

[0036] The optimal hyperparameters of the XGBoost non-small cell lung cancer (NSCLC) diagnostic model are shown in Table 1.

[0037] Table 1

[0038] The performance metrics of the model are as follows (see summary of 6 metrics for model 9). Figure 1 ): AUC (Area Under the Curve): Measures the overall ability of a model to distinguish between patients and non-patients. The AUC of this model is about 0.9, indicating that the overall discrimination accuracy is very high. F1 score: the harmonic mean of precision and recall. This model has an F1 score of over 0.8. Sensitivity: The sensitivity of this model is close to 0.93, indicating that the model has a very strong ability to capture lung cancer patients with very few missed diagnoses, making it very suitable for early screening; Specificity: The validation set specificity reached above 0.83, demonstrating the high quality and robustness of the model in feature extraction; PPV (Positive Predictive Value): Over 75%, which is ideal in clinical screening models and can significantly narrow down the population that needs further testing. NPV (Negative Predictive Value): Approximately 0.91, indicating that the model has a high reliability in excluding diagnoses, and doctors can have more confidence in the model's negative results.

[0039] The above description is a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and concept of the present invention, should be covered within the scope of protection of the claims of the present invention.

Claims

1. A kit for differentiating benign and malignant pulmonary nodules, characterized in that, The kit includes nucleic acid extraction reagents for extracting cell-free DNA from plasma and library construction reagents for constructing GM-seq libraries. The GM-seq library construction reagents include DNA fragmentation reagents, end repair reagents, A / T tail addition reagents, adapter ligation reagents, TET enzyme, pyridine borane, PCR amplification reagents, and library cyclization reagents. The kit uses the XGBoost algorithm-based non-small cell lung cancer diagnostic model to assess lung cancer risk based on the following 12 features: 9 differentially methylated regions DMR27734 (chr1:1289004). The fragments identified include DMR55120 (chr1:2097679), DMR114989 (chr1:4715576), DMR66665 (chr1:2433914), DMR122523 (chr1:5366740), DMR3325 (chr1:437583), DMR14627 (chr1:980109), DMR3582 (chr1:448987), and DMR35337 (chr1:1488625), as well as three fragment group characteristics. 90-150bp frequency, 151-220bp frequency and ratio.

2. The kit for differentiating benign and malignant pulmonary nodules as described in claim 1, characterized in that, The DNA fragmentation reagent is used to randomly fragment genomic DNA to 200-300 bp.

3. The kit for differentiating benign and malignant pulmonary nodules as described in claim 1, characterized in that, The TET enzyme is used to oxidize 5-methylcytosine and 5-hydroxymethylcytosine, and the pyridineborane is used to reduce the oxidation product to dihydrouracil.

4. The kit for differentiating benign and malignant pulmonary nodules as described in claim 1, characterized in that, The diagnostic model for non-small cell lung cancer is a machine learning model.

5. The kit for differentiating benign and malignant pulmonary nodules as described in claim 1, characterized in that, The kit for differentiating benign and malignant pulmonary nodules also includes a positive control and a negative control; the positive control is pUC19 DNA and the negative control is lambda DNA.

6. The kit for differentiating benign and malignant pulmonary nodules as described in claim 1, characterized in that, In the segment group features, the 90-150bp frequency represents the proportion of segments with a length of 100-150bp in the total segments, the 151-220bp frequency represents the proportion of segments with a length of 151-220bp in the total segments, and the ratio represents the ratio of short segments to long segments.

7. The kit for differentiating benign and malignant pulmonary nodules as described in claim 1, characterized in that, The area under the curve (AUC) of the non-small cell lung cancer diagnostic model in the validation set is not less than 0.

9.

8. The kit for differentiating benign and malignant pulmonary nodules as described in claim 1, characterized in that, The diagnostic model for non-small cell lung cancer has a sensitivity of not less than 0.9 and a specificity of not less than 0.

8.

9. The use of the kit for differentiating benign and malignant pulmonary nodules as described in any one of claims 1 to 8 in the preparation of early lung cancer diagnostic products.