Early-stage lung cancer prediction method based on multi-mode eccDNA marker

Abstract

The invention discloses an early-stage lung cancer prediction method based on a multi-modal eccDNA marker, and relates to the technical field of liquid biopsy, and the method comprises the following steps: collecting a peripheral blood sample, carrying out differential centrifugal separation on plasma, constructing a cfDNA library, and carrying out double-end sequencing to obtain original sequencing data; extracting structural features through an eccDNA model, and inputting the structural features into a first machine learning model to generate a circular DNA score; extracting variation features through an SNV model, and inputting the variation features into a second machine learning model to generate a spectrum feature score; extracting copy number features through a CNV model, and inputting the copy number features into a third machine learning model to generate copy number scores; carrying out probability distribution calibration on the circular DNA score, the spectrum feature score and the copy number score; inputting the three types of molecular features into a deep neural network to generate a first fusion score; inputting the calibrated score into a logistic regression model to generate a second fusion score; and generating a final lung cancer risk probability according to the first fusion score and the second fusion score.

Description

Technical Field

[0001] This invention relates to the field of liquid biopsy technology, and more particularly to a method for predicting early lung cancer based on multimodal eccDNA biomarkers. Background Technology

[0002] Lung cancer is the leading cause of cancer-related deaths worldwide. According to statistics from the World Health Organization, lung cancer accounts for approximately 18% of all cancer deaths globally. More than 2.2 million new cases of lung cancer are diagnosed each year, and more than 1.8 million people die from it annually. Lung cancer can be divided into non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC), with NSCLC accounting for more than 85%. Early diagnosis of lung cancer is crucial for patient prognosis; patients diagnosed early have a five-year survival rate of over 70%, while the survival rate for late-stage patients drops significantly to below 15%.

[0003] Although existing chest imaging methods, such as X-rays, CT scans, and tissue biopsies, are widely used in the early diagnosis of lung cancer, they suffer from insufficient sensitivity, low specificity, and low patient acceptance. Tissue biopsies are usually used as a last resort, but their high invasiveness and potential harm to patients limit their widespread application. In recent years, liquid biopsy technology has emerged as a new research direction in the field of early lung cancer detection, especially the detection of cell-free DNA (cfDNA) molecular markers in plasma, which has shown significant potential. Summary of the Invention

[0004] This application provides an early lung cancer prediction method based on multimodal eccDNA biomarkers, addressing the shortcomings of existing technologies such as insufficient sensitivity and radiation risks associated with traditional imaging examinations, the invasiveness and low patient acceptance of tissue biopsies, and the difficulty in achieving both sensitivity and specificity with a single liquid biopsy biomarker in early lung cancer detection. The method achieves significant technical improvements: complementary multimodal features enhance diagnostic performance, specifically by constructing a three-level model through the fusion of eccDNA structural features, SNV variant profiles, and CNV copy number features. The implementation of this technical improvement includes a multimodal fusion mechanism that enhances discriminative power through dual-path fusion; and calibration techniques that optimize the process by utilizing temperature scaling to address probability output biases in sub-models, ensuring consistent score distribution before fusion. This comprehensively improves the accuracy, accessibility, and clinical application value of early lung cancer prediction.

[0005] This application provides a method for predicting early lung cancer based on multimodal eccDNA biomarkers, including:

[0006] S1: Collect peripheral blood samples, separate plasma by differential centrifugation, construct cfDNA libraries and perform paired-end sequencing to obtain raw sequencing data; generate standardized BAM files based on raw sequencing data through quality filtering, adapter removal and genome alignment.

[0007] S2: Perform three types of molecular characterization analyses in parallel based on the standardized BAM file:

[0008] Structural features are extracted using the eccDNA model and input into the first machine learning model to generate a circular DNA score.

[0009] The variation features are extracted using the SNV model and then input into a second machine learning model to generate spectral feature scores.

[0010] Copy number features are extracted using the CNV model and then input into a third machine learning model to generate a copy number score.

[0011] S3: Perform probability distribution calibration on circular DNA scoring, spectral feature scoring, and copy number scoring;

[0012] S4: Input the three types of molecular features into a deep neural network to generate a first fusion score; input the calibrated score into a logistic regression model to generate a second fusion score; generate the final lung cancer risk probability based on the first and second fusion scores.

[0013] Furthermore, the eccDNA recognition model includes the following steps:

[0014] The standardized BAM files were analyzed using Circle-Map, and high-confidence candidate circular sites that met the criteria of split reads and inconsistent reads were retained.

[0015] Extract multidimensional feature vectors of candidate loop sites, including structural features and functional annotation features;

[0016] The structural features include chromosome start position, termination position, and breakpoint conservation score;

[0017] The functional annotation features include oncogene / tumor overlap markers, neighbor enhancer distance, and GC content;

[0018] Each candidate circular site is assigned a weight, a random forest model is trained based on the weighted features, the single-event prediction probability is output, and the circular DNA score is calculated using a summation formula.

[0019] Furthermore, the SNV model includes the following steps:

[0020] Use GATK Mutect2 to identify single nucleotide variants in BAM files as candidate SNVs;

[0021] Three types of labels are assigned to each candidate SNV: strong labels, intermediate labels, and weak labels.

[0022] The strong label is a tissue-verified variant in a lung cancer driver gene;

[0023] The intermediate label is a reliable variant predicted by a resilient network logistic regression model;

[0024] The weak label is a background variation in the cfDNA of the healthy control group;

[0025] Feature vectors of candidate SNVs are extracted, and a resilient network logistic regression model is trained to output spectral feature scores.

[0026] Furthermore, the CNV model includes the following steps:

[0027] The copy number of a standardized BAM file is estimated in a 5Mb window using Control-FreeC, and the copy number ratio for each window is calculated.

[0028] By analyzing copy number data using GISTIC2.0, high-frequency copy number variation hotspots were identified.

[0029] Extract the copy number feature vector and perform Z-score normalization. Input it into the generalized linear model for training and output the copy number score.

[0030] Further, the probability distribution calibration includes: extracting the uncalibrated linear output value of the sub-model output; optimizing the temperature parameter T using a grid search on the validation set, and generating the calibrated probability using the scaling formula σ(Z / T), where Z is the uncalibrated linear output value and T is the optimized temperature parameter;

[0031] The validation set is a subset independently partitioned from the standardized BAM file and is used for hyperparameter optimization.

[0032] Furthermore, the first fusion score is the first lung cancer risk probability value output by the deep neural network after nonlinearly fusing the three original molecular features of eccDNA, SNV, and CNV, with a value range of [0,1].

[0033] Furthermore, the second fusion score is a second lung cancer risk probability value generated by linearly weighting and fusing the temperature-scaled calibrated circular DNA score, spectral feature score, and copy number score using a logistic regression model, with a value range of [0,1].

[0034] Furthermore, the final lung cancer risk probability also includes: collecting peripheral blood cfDNA samples from patients monthly, extracting time-series feature matrices for each time point; performing temperature scaling calibration to generate a calibration probability vector, and fusing it in two ways: short-term fusion inputs the calibration probability into a logistic regression model to output a short-term time-series score, and long-term fusion inputs the original features into a deep neural network to output a long-term time-series score; constructing the proportion of drug-resistant clones based on the short-term and long-term time-series scores, and outputting the recurrence risk probability when the proportion of drug-resistant clones is greater than or equal to 0.1.

[0035] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0036] By analyzing the three-modal synergistic mechanism of eccDNA, SNV, and CNV in parallel, this study addresses the industry pain point of difficulty in achieving both sensitivity and specificity in early lung cancer detection using a single liquid biopsy biomarker. Temperature scaling calibration technology is employed, utilizing grid search to optimize the temperature parameter T and resolve the probability distortion problem of the sub-model. A non-invasive, full-cycle management system combining static screening and dynamic monitoring is constructed, improving patient compliance by 40% and reducing screening costs by 60%, providing a non-invasive, precise, and sustainable closed-loop solution for early lung cancer screening. Attached Figure Description

[0037] Figure 1 This is a flowchart of an early lung cancer prediction method based on multimodal eccDNA biomarkers in an embodiment of the present invention;

[0038] Figure 2 This is a violin diagram for verifying the age, sex, and stage subgroup of eccDNA in this embodiment of the invention.

[0039] Figure 3 This is a violin diagram validating the age, gender, and stage subgroups of SNVs in this embodiment of the invention.

[0040] Figure 4 This is a violin diagram validating the age, gender, and stage subgroup of CNV in this embodiment of the invention. Detailed Implementation

[0041] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0043] Example 1: As Figure 1 As shown, an early lung cancer prediction method based on multimodal eccDNA biomarkers.

[0044] S1: Collect peripheral blood samples, separate plasma by differential centrifugation, construct cfDNA libraries and perform paired-end sequencing to obtain raw sequencing data; generate standardized BAM files based on raw sequencing data through quality filtering, adapter removal and genome alignment.

[0045] Specifically, peripheral blood samples were obtained from the subject population, placed in EDTA vacuum EP tubes, and transported at room temperature (4°C). Within 3 hours of sample collection, peripheral blood samples were transferred to sterile enzyme-free centrifuge tubes using pipette tips in a UV-sterilized environment. Plasma was separated by differential centrifugation using a rotary rotor in a pre-chilled refrigerated centrifuge. After an initial centrifugation at 1900×g and 4°C for 10 minutes, the supernatant was transferred to new centrifuge tubes and centrifuged at 16,000×g and 4°C for 10 minutes using a fixed-angle rotor to remove residual cells and debris. The new supernatant was aliquoted and stored in 2.0–3.5 mL portions. For long-term storage, short-term storage (up to one month) at -20°C or long-term storage at -80°C was permitted.

[0046] use The Cyclic Nucleic Acid Kit isolates ccfDNA, including linear DNA and cf-eccDNA, from thawed plasma aliquots (3.5 mL).

[0047] Add blocked proteinase K to the centrifuge tube containing the plasma sample to be tested to remove any remaining cell debris. Incubate at 56°C with shaking for 10 minutes. After protein precipitation, add 200 μL of ethanol and repeat the centrifugation process. Transfer the mixture to... Add bufferAW1 to the centrifuge column and centrifuge at 8000 rpm for 1 min, then remove the centrifuge tube. Add bufferAW2 and centrifuge at 14000 rpm for 3 min. Finally, add an appropriate amount of elution buffer, centrifuge at 14000 rpm for 1 min, and elute the DNA from the column to a final elution volume of 20 μL–50 μL. The concentration of the separated cfDNA sample can be measured using a Qubit 3.0 fluorometer.

[0048] Linear DNA was removed using the Epicentre Plasma-Safe ATP-Dependent DNase Kit (Biosearch Technologies). 20 μL–50 μL of eluted DNA was then incubated in a water or metal bath at 37°C for 1 hour for complete digestion, followed by 70°C for 5 minutes to remove linear DNase. After removing the linear DNA, the remaining eccDNA was cleaned using the MinElute Reaction ELISA Kit (Qiagen) to remove residual enzymes and other impurities. The final purified eccDNA was then quantified using a Qubit 3.0 fluorometer for quality control.

[0049] The cf-eccDNA obtained was cut using a Covaris S220 ultrasonic fragment cutter to obtain a median DNA fragment size of 400 bp. DNA-seq kits were used to prepare DNA libraries from the cut DNA. The quality of each library was assessed using an Agilent DNF-915 kit. The DNA libraries were sequenced on an Illumina NovaSeq 6000 using a 150bp paired-end pattern. Raw sequencing data were obtained, and cf-ecc DNA libraries were constructed based on the images and base sequences.

[0050] FastQC was used to assess data quality, fastp was used to remove adapters and low-quality bases, BWA-MEM was aligned to the hg19 human reference genome, and Samtools and Picard tools were used for deduplication, sorting, and indexing to obtain the BAM file.

[0051] S2: Perform three types of molecular characterization analyses in parallel based on the standardized BAM file:

[0052] Structural features are extracted using the eccDNA model and input into the first machine learning model to generate a circular DNA score.

[0053] The eccDNA recognition model includes the following steps:

[0054] The standardized BAM files were analyzed using Circle-Map, and high-confidence candidate circular sites that met the criteria of split reads and inconsistent reads were retained.

[0055] Extract multidimensional feature vectors of candidate loop sites, including structural features and functional annotation features;

[0056] The structural features include chromosome start position, termination position, and breakpoint conservation score;

[0057] The functional annotation features include oncogene / tumor overlap markers, neighbor enhancer distance, and GC content;

[0058] Each candidate circular site is assigned a weight, a random forest model is trained based on the weighted features, the single-event prediction probability is output, and the circular DNA score is calculated using a summation formula.

[0059] Specifically, circular locus detection uses a circle-map to analyze the standardized BAM file, retaining high-confidence candidate circular loci that meet the following criteria: ≥2 split reads and ≥3 discordant reads. Candidate loci are then locally re-aligned using the circle-map realign module to eliminate false positives, ultimately outputting a BED file containing information such as chromosome, start, and end sites.

[0060] The 18-dimensional feature vectors of candidate circular sites are extracted. A weighted random forest model is used to standardize the 18-dimensional features using Z-scores. The input to the random forest model (RandomForestClassifier parameters: n_estimators = 100, max_depth = 10, oob_score = True) assigns event weights based on the functional annotation features. The circular DNA score is calculated using the following formula:

[0061]

[0062] Among them, S ecc_raw The original circular DNA score is represented by ω, where ω is the event weight. If the site overlaps with an oncogene / tumor region (Oncogene / Tumor Overlap = 1), the weight ω = 1.5; otherwise, ω = 0.8. k Here, N represents the predicted probability for a single event, and N is the total number of eccDNA events in the sample. The original score is mapped to the [0,1] interval using the sigmoid function to obtain the circular DNA score.

[0063]

[0064] Among them, S ecc Scoring of circular DNA.

[0065] Step S2 also includes: extracting variant features through the SNV model and inputting them into the second machine learning model to generate spectral feature scores;

[0066] The SNV model includes the following steps:

[0067] Use GATK Mutect2 to identify single nucleotide variants in BAM files as candidate SNVs;

[0068] Three types of labels are assigned to each candidate SNV: strong labels, intermediate labels, and weak labels.

[0069] The strong label is a tissue-verified variant in a lung cancer driver gene;

[0070] The intermediate label is a reliable variant predicted by a resilient network logistic regression model;

[0071] The weak label is a background variation in the cfDNA of the healthy control group;

[0072] Feature vectors of candidate SNVs are extracted, and a resilient network logistic regression model is trained to output spectral feature scores.

[0073] Specifically, GATK Mutect2 was used to analyze the standardized BAM files to identify single nucleotide variants (SNVs) as candidate SNVs. Only variants meeting the following criteria were retained: passed quality filtering; Variant allele frequency (VAF) ≥1% to exclude low-frequency noise interference; and gene variants associated with known clonal hematopoietic mutations (CH) (including DNMT3A, TET2, ASXL1, PPM1D, GNB1, CBL, JAK2, STAT3, GNAS, MYD88, SF3B1) were removed to avoid interference from the hematopoietic system background. The retained candidate SNVs were annotated with ANNOVAR to filter variants with a frequency >1% in the gnomAD database, ensuring that only rare or tumor-specific mutations were retained.

[0074] Each candidate SNV is assigned three types of labels: strong labels, intermediate labels, and weak labels.

[0075] Strong labels are derived from tumor mutations in lung cancer driver genes (such as EGFR, KRAS, PIK3CA, BRAF, MET, U2AF1, NFE2L2, TERT, ERBB2, HRAS, NRAS, and RAF1) that have been validated by tissue samples.

[0076] The intermediate labels are the credible variations predicted by the ElasticNet Logistic Regression (ELR) model under the Leave-One-out Cross-Validation (LOOCV) framework.

[0077] Weak labels were derived from background variations in cfDNA from healthy control groups and used to construct a noisy baseline.

[0078] Extract a 26-dimensional feature vector for each candidate SNV, covering location, sequence, read support, and functional annotation information:

[0079] SNV spectral features refer to the variation distribution patterns obtained statistically based on different allele frequency thresholds. After feature extraction, 14 patient-level features (such as total SNVs, proportion of driver gene mutations, and average VAF) are summarized for each sample to form the input matrix X. snv ∈R 14 The feature matrix is ​​Z-score normalized to ensure comparability of features at different scales.

[0080] For a dataset containing N samples, one sample is retained as the validation set in each iteration, and the remaining N-1 samples are used as the training set. On each training set, 5-fold internal cross-validation is performed to optimize the regularization parameter λ. The grid search traverses the range λ∈{0.01,0.1,1,10}, with the objective function being to maximize the validation set AUC (Area Under Curve) or minimize the log loss. Using the optimal λ value, a resilient network logistic regression model is trained on the full training set. Key parameters include: a regularization mixture parameter α = 0.5, balancing L1 (Lasso) and L2 (Ridge) penalties; L1 promotes feature sparsity, and L2 prevents overfitting; the loss function is binary cross-entropy, and the optimizer uses coordinate descent. For the retained validation set samples, the model outputs the sample-level probability S. snv ∈[0,1], representing the spectral feature score. To improve stability, the process is repeated 30 times (Bootstrap sampling), and the final S snv The average of 30 predicted probabilities is used to reduce the impact of random variance.

[0081] Step S2 also includes: extracting copy number features through the CNV model and inputting them into a third machine learning model to generate a copy number score;

[0082] The CNV model includes the following steps:

[0083] The copy number of a standardized BAM file is estimated in a 5Mb window using Control-FreeC, and the copy number ratio for each window is calculated.

[0084] By analyzing copy number data using GISTIC2.0, high-frequency copy number variation hotspots were identified.

[0085] Extract the copy number feature vector and perform Z-score normalization. Input it into the generalized linear model for training and output the copy number score.

[0086] Specifically, Control-FreeC was used to perform copy number analysis on the normalized BAM file based on a fixed 5Mb window, and the copy number ratio (log2) of each window was calculated as the raw CNV signal. The CNV data of cfDNA were input into GISTIC2.0 to identify high-frequency copy number variation hotspots in the genome (such as chromosome amplification / deletion hotspots) and output significantly enriched regions.

[0087] Interference signals were filtered out, retaining only CNV segments with a |Z-score|>2.58 from the background noise; regions with a |Z-score|>2.58 that also appeared in ≥20% of healthy control samples were excluded; regions with a |Z-score|>2.58 detected in matched peripheral blood leukocyte (WBC) samples were also excluded. The genomic background p-value of the CNV segments was required to be <0.05 to ensure signal reliability. All variations on chromosome 19 (Chr19) were deleted to avoid false positives caused by highly repetitive sequences near telomeres.

[0088] Based on the filtered results, a 3D feature X is extracted for each sample. snv ∈R3: filtered_5MB_CNV_count (total number of CNV events retained in the 5Mb window after filtering), filtered_GISTIC_CNV_count (number of significant hotspot regions identified by GISTIC 2.0), FisherLogP (CNV enrichment significance score based on Fisher's exact test). Z-score standardization (mean = 0, standard deviation = 1) is applied to the 3D features to eliminate dimensional differences.

[0089] The CNV model is trained using Logistic Regression (GLM). Core parameters include: regularization with an L2 penalty term; hyperparameter optimization: adjusting the regularization strength parameter C (search range: C∈{0.001,0.01,0.1,1,10}) through grid search cross-validation (GridSearchCV) to maximize the validation set AUC. The model outputs sample-level probability values ​​S. cnv ∈[0,1] represents the copy number score.

[0090] S3: Perform probability distribution calibration on circular DNA scoring, spectral feature scoring, and copy number scoring;

[0091] The probability distribution calibration includes: extracting the uncalibrated linear output value of the sub-model; optimizing the temperature parameter T using a grid search on the validation set; and generating the calibrated probability using the scaling formula σ(Z / T), where Z is the uncalibrated linear output value and T is the optimized temperature parameter.

[0092] The validation set is a subset independently partitioned from the standardized BAM file and is used for hyperparameter optimization.

[0093] Specifically, the raw linear output value Z, untransformed by the Sigmoid function, is extracted from the sub-model output layer instead of directly using the probability value. A subset is independently partitioned from the corresponding samples of the standardized BAM file, specifically for hyperparameter optimization. The expected calibration error (ECE) is minimized, with an objective of ECE ≤ 0.02. The parameter space is traversed through a grid search, with the search range being the optimization temperature parameter T, T∈[0.1,5.0] (step size 0.1). The combined loss of ECE and negative log-likelihood is jointly optimized.

[0094] L = ECE + τ·NLL,

[0095] Where L is the loss function, NLL is the negative log-likelihood, and τ is the weight that balances the calibration error and the likelihood loss, τ = 0.5. The temperature parameter T is corrected and optimized through the loss function L.

[0096] The temperature scaling formula generates the calibration probability-scaled original linear output value Z and applies the Sigmoid function to obtain:

[0097]

[0098] Among them, S cal To calibrate the probability, each sub-model independently optimizes the temperature parameters and outputs a calibrated score S. ecc-cal S snv-cal S cnv-cal .

[0099] S4: Input the three types of molecular features into a deep neural network to generate the first fusion score;

[0100] The first fusion score is the first lung cancer risk probability value output by the deep neural network after nonlinearly fusing the three original molecular features of eccDNA, SNV and CNV, with a value range of [0,1].

[0101] Specifically, the three original molecular features are directly input into a deep neural network. After Z-score standardization, missing values ​​are filled in using KNN-impute (K-nearest neighbor imputation). The first fully connected layer maps 62 dimensions to 128 dimensions and uses ReLU activation, followed by a batch normalization layer and Dropout (regularization) of 0.25. The second fully connected layer maps 128 dimensions to 32 dimensions and uses ReLU activation. The third fully connected layer maps 32 dimensions to 6 dimensions and uses linear activation. The classification head is a fully connected layer that takes a 6-dimensional latent vector as input and outputs a 1-dimensional probability using Sigmoid activation. The encoder layers are initialized with He-normal weights and a weight decay coefficient of 1×10. -4The loss function is the sum of the binary cross-entropy and the mean squared error of reconstruction, with a reconstruction term weight of 0.1; the optimizer is Adam, and the learning rate is 3×10⁻⁶. -4 Batch size 64; during training, if the validation set loss does not improve after 15 rounds, training is terminated to ensure consistent output variance of each layer in the early stages of training and to avoid gradient vanishing or exploding; the final output is the first fusion score S. fusion .

[0102] Step S4 also includes inputting the calibrated score into the logistic regression model to generate a second fusion score;

[0103] The second fusion score is a second lung cancer risk probability value generated by linearly weighting and fusing the temperature-scaled calibrated circular DNA score, spectral feature score, and copy number score using a logistic regression model, with a value range of [0,1].

[0104] Specifically, the probabilities of the temperature-scaled calibrated eccDNA, SNV, and CNV models are used as input to the three-dimensional feature vector X = [S ecc-cal S snv-cal S cnv-cal Z-score standardization was performed to eliminate dimensional differences. Logistic regression was then used for fusion prediction.

[0105] S fusion_2 =σ(β0+β1S) ecc-cal +β2S snv-cal +β3S cnv-cal ),

[0106] Among them, S fusion_2 For the second fusion score, β0 is the baseline risk probability, σ is the Sigmoid function, which converts the linear combination into a probability value to ensure that the output conforms to the probability distribution, and β1, β2, and β3 are the feature weight coefficients.

[0107] Step S4 also includes generating the final lung cancer risk probability based on the first fusion score and the second fusion score.

[0108] Specifically, calculate the final probability of lung cancer risk:

[0109] S final =α·S fusion +(1-α)·S fusion_2 ,

[0110] Among them, S final Let S be the final lung cancer risk probability, and α be a dynamic weighting coefficient optimized through a validation set grid search (α∈[0,1]), with a default value of α=0.6. final A score ≥0.7 indicates a high risk of lung cancer, while a score ≤0.3 indicates a low risk. final A score <0.7 indicates a medium risk of lung cancer.final A score <0.3 indicates a low risk of lung cancer.

[0111] To validate the generalization ability of the models across different clinical subgroups, the models were analyzed by grouping patients according to age, sex, and stage based on their clinical information. Results showed that the CNV and eccDNA models performed significantly better in the age ≥60 years group than in the age <60 years group. Figure 2 , 4 As shown. In particular, the CNV model achieved an AUC of 0.93 in the ≥60-year-old group, significantly higher than the <60-year-old group (AUC = 0.89), indicating that tumor characteristics may be more pronounced in older patients. Within the gender subgroups, the overall model performance was slightly better in the male group than in the female group. Specifically, the CNV model showed an AUC of 0.92 in the male subgroup and 0.89 in the female subgroup, suggesting that CNV features are more prominent in the male group. Figure 4 As shown, the model performs significantly better in patients with advanced lung cancer than in those with early-stage lung cancer. The CNV model has an AUC of 0.94 in advanced patients, compared to 0.88 in early-stage patients. Figure 4 As shown, the eccDNA and SNV models also exhibit similar trends, such as... Figure 2 , 3 As shown, the differences between stages are relatively small. Decision curve analysis shows that all three models, after calibration, achieved positive net benefits within the threshold probability range of 0.10–0.50, significantly outperforming the two extreme strategies of "full intervention" and "no intervention at all." Specifically, the SNV model curve was slightly higher than other models in most threshold ranges, followed closely by the eccDNA curve, while the CNV model, although with a slightly lower net benefit, still maintained its advantage. This means that in actual screening scenarios, if a 10–50% probability of disease is used as the intervention threshold, this scoring system can avoid missed diagnoses and effectively reduce unnecessary treatments that may harm patients, maximizing the overall benefits for both patients and clinicians.

[0112] The final model is deployed on a cloud or local server, and accepts cfDNA sequencing data uploaded by users through an API interface. It automates the entire process from data quality control and comparison to model prediction, and generates an early lung cancer risk probability report.

[0113] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0114] This application achieves a significant breakthrough in non-invasive early lung cancer detection by integrating multimodal molecular features with machine learning technology. It employs a three-modal synergistic mechanism that analyzes eccDNA structural features, SNV variant profiles, and CNV copy number variations in parallel, effectively addressing industry issues such as insufficient sensitivity of traditional imaging examinations (CT screening false negative rate up to 22%) and radiation risks, as well as the highly invasive nature of tissue biopsies (patient acceptance rate <50%). Using Circle-Map, GATK Mutect2, and Control-FreeC tools, 62-dimensional original molecular features are accurately extracted. Temperature scaling calibration technology is combined to address the probability distortion problem of sub-models, reducing the ECE of the calibrated eccDNA, SNV, and CNV sub-models to below 0.015. A dual-path fusion architecture is constructed: a deep neural network performs a 128→32→6-dimensional nonlinear transformation on the original features to capture complex interactions, outputting a first fusion score; logistic regression linearly weights the calibrated probabilities to generate a highly interpretable second fusion score. After fusion, the test set AUC reached 0.94, a 12% improvement compared to the single biomarker model, with sensitivity and specificity reaching 96.2% and 93.8%, respectively. This provides a non-invasive, precise, and sustainable closed-loop solution for early lung cancer screening.

[0115] Example 2: Example 1 only supports static screening at a single time point and cannot capture the dynamic evolution of tumor clones. In order to meet the needs of dynamic monitoring and recurrence early warning for lung cancer patients after treatment, this example further supplements the content of Example 1.

[0116] The final lung cancer risk probability also includes: collecting peripheral blood cfDNA samples from patients monthly, extracting time-series feature matrices for each time point; performing temperature scaling calibration to generate a calibration probability vector, and fusing it in two ways: short-term fusion inputs the calibration probability into a logistic regression model to output a short-term time-series score, and long-term fusion inputs the original features into a deep neural network to output a long-term time-series score; constructing the proportion of drug-resistant clones based on the short-term and long-term time-series scores, and outputting the recurrence risk probability when the proportion of drug-resistant clones is greater than or equal to 0.1.

[0117] Specifically, peripheral blood cfDNA samples were collected from patients monthly (t1, t2, ... t). m The calibration probability vector is repeatedly generated at each time point. And extract the time series feature matrix.

[0118] P t Input a logistic regression model, where each probability value has been calibrated by temperature scaling, and output a short-term time-series score:

[0119]

[0120] in, It provides short-term time-series scoring and is sensitive to capturing instantaneous signals.

[0121] Will Input a deep neural network and output a long-term time series score. Tracking the cumulative trend of genomic instability.

[0122] The percentage of drug-resistant clones was calculated based on short-term and long-term time-series scores.

[0123]

[0124] Where, N y The percentage of drug-resistant clones is represented by m, where m is the total number of detection time points, and t is the time point index. The mean absolute deviation, The deviation between short-term noise and long-term trend is quantified to reflect the extent of drug-resistant clone expansion; a continuous deviation (>0.05) indicates the expansion of drug-resistant clones, and a cumulative mean of ≥0.1 indicates that drug-resistant clones are dominant.

[0125] When N y When the level is ≥0.1, a high-risk warning is activated, a recurrence risk report is generated, and the report is pushed to the clinical intervention system; when the level is ≥0.1 for two consecutive months... And N y When N is ≥0.05, a medium-risk warning is activated, the monitoring period is shortened from 1 month to 2 weeks, and the cfDNA sampling frequency is increased; when N y <0.05 and When necessary, initiate low-risk routine monitoring, maintain routine monthly monitoring, and automatically generate stability reports.

[0126] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0127] This application achieves a closed-loop, full-cycle management system based on static screening through dynamic monitoring and recurrence early warning mechanisms, significantly improving the accuracy and timeliness of postoperative management for lung cancer patients. Peripheral blood cfDNA samples are collected monthly to construct a time-series feature matrix. A calibration probability vector is generated using temperature scaling calibration technology, and an innovative dual-path fusion architecture is employed: the short-term fusion path reuses a static logistic regression model to process the calibrated probability, quickly capturing transient signals such as treatment response and inflammatory interference; the long-term fusion path reuses a deep neural network to analyze original molecular characteristics and track the cumulative trend of genomic instability. A quantitative model of the proportion of drug-resistant clones is constructed based on the absolute deviation between short-term and long-term time-series scores. When the proportion is ≥0.1, it indicates the dominant group of drug-resistant clones, and the system automatically generates a recurrence risk report and triggers targeted therapy adjustments. For intermediate-risk patients, the monitoring cycle is shortened to 2 weeks to capture early drug resistance signals. Clinical validation shows that this approach increases the median early warning period for recurrence to 5.3 months compared to traditional CT scans, reduces screening costs by 60%, and improves patient compliance by 40%, providing a sustainable closed-loop management tool for precision treatment of lung cancer.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting early lung cancer based on multimodal eccDNA biomarkers, characterized in that, include: S1: Peripheral blood samples were collected, plasma was separated by differential centrifugation, a cfDNA library was constructed, and paired-end sequencing was performed to obtain raw sequencing data; Standardized BAM files are generated based on raw sequencing data through quality filtering, adapter removal, and genome alignment. S2: Perform three types of molecular characterization analyses in parallel based on the standardized BAM file: Structural features are extracted using the eccDNA model and input into the first machine learning model to generate a circular DNA score. The variation features are extracted using the SNV model and then input into a second machine learning model to generate spectral feature scores. Copy number features are extracted using the CNV model and then input into a third machine learning model to generate a copy number score. S3: Perform probability distribution calibration on circular DNA scoring, spectral feature scoring, and copy number scoring; S4: Input the three types of molecular features into a deep neural network to generate a first fusion score; input the calibrated score into a logistic regression model to generate a second fusion score; generate the final lung cancer risk probability based on the first and second fusion scores.

2. The method for predicting early lung cancer based on multimodal eccDNA biomarkers as described in claim 1, characterized in that, The eccDNA recognition model includes the following steps: The standardized BAM files were analyzed using Circle-Map, and high-confidence candidate circular sites that met the criteria of split reads and inconsistent reads were retained. Extract multidimensional feature vectors of candidate loop sites, including structural features and functional annotation features; The structural features include chromosome start position, termination position, and breakpoint conservation score; The functional annotation features include oncogene / tumor overlap markers, neighbor enhancer distance, and GC content; Each candidate circular site is assigned a weight, a random forest model is trained based on the weighted features, the single-event prediction probability is output, and the circular DNA score is calculated using a summation formula.

3. The method for predicting early lung cancer based on multimodal eccDNA biomarkers as described in claim 1, characterized in that, The SNV model includes the following steps: Use GATK Mutect2 to identify single nucleotide variants in BAM files as candidate SNVs; Three types of labels are assigned to each candidate SNV: strong labels, intermediate labels, and weak labels. The strong label is a tissue-verified variant in a lung cancer driver gene; The intermediate label is a reliable variant predicted by a resilient network logistic regression model; The weak label is a background variation in the cfDNA of the healthy control group; Feature vectors of candidate SNVs are extracted, and a resilient network logistic regression model is trained to output spectral feature scores.

4. The method for predicting early lung cancer based on multimodal eccDNA biomarkers as described in claim 1, characterized in that, The CNV model includes the following steps: The copy number of a standardized BAM file is estimated in a 5Mb window using Control-FreeC, and the copy number ratio for each window is calculated. By analyzing copy number data using GISTIC2.0, high-frequency copy number variation hotspots were identified. Extract the copy number feature vector and perform Z-score normalization. Input it into the generalized linear model for training and output the copy number score.

5. The method for predicting early lung cancer based on multimodal eccDNA biomarkers as described in claim 1, characterized in that, The probability distribution calibration includes: extracting the uncalibrated linear output value of the sub-model; optimizing the temperature parameter T using a grid search on the validation set; and generating the calibrated probability using the scaling formula σ(Z / T), where Z is the uncalibrated linear output value and T is the optimized temperature parameter. The validation set is a subset independently partitioned from the standardized BAM file and is used for hyperparameter optimization.

6. The method for predicting early lung cancer based on multimodal eccDNA biomarkers as described in claim 1, characterized in that, The first fusion score is the first lung cancer risk probability value output by the deep neural network after nonlinearly fusing the three original molecular features of eccDNA, SNV and CNV, with a value range of [0,1].

7. The method for predicting early lung cancer based on multimodal eccDNA biomarkers as described in claim 1, characterized in that, The second fusion score is a second lung cancer risk probability value generated by linearly weighting and fusing the temperature-scaled calibrated circular DNA score, spectral feature score, and copy number score using a logistic regression model, with a value range of [0,1].

8. The method for predicting early lung cancer based on multimodal eccDNA biomarkers as described in claim 1, characterized in that, The final lung cancer risk probability also includes: collecting peripheral blood cfDNA samples from patients monthly, extracting time-series feature matrices for each time point; performing temperature scaling calibration to generate a calibration probability vector, and fusing it in two ways: short-term fusion inputs the calibration probability into a logistic regression model to output a short-term time-series score, and long-term fusion inputs the original features into a deep neural network to output a long-term time-series score; constructing the proportion of drug-resistant clones based on the short-term and long-term time-series scores, and outputting the recurrence risk probability when the proportion of drug-resistant clones is greater than or equal to 0.1.

Images