Method for dynamically identifying low-abundance mutation of tumor gene based on isolated forest algorithm

Abstract

The invention discloses a method for dynamically identifying low-abundance mutation of a tumor gene based on an isolated forest algorithm. The method comprises the following steps: constructing a background noise frequency database; extracting multi-dimensional characteristics of sequencing depth, mutation support number and base quality of a to-be-detected mutation site of the tumor gene; and calculating the path length and the abnormal score of the mutation site by using an isolated forest algorithm, dynamically adjusting the feature weight through the confidence coefficient, and distinguishing the real mutation from the background noise. According to the method, the calculation efficiency is improved, the method is suitable for large-scale tumor gene next-generation sequencing data analysis, the adopted isolated forest algorithm has linear time complexity, compared with a traditional machine learning method, the method has higher calculation efficiency, and the method is suitable for large-scale analysis of high-throughput sequencing data such as whole exome sequencing WES or whole genome sequencing WGS.

Description

Technical Field

[0001] This invention relates to the field of gene sequencing, specifically to a method for dynamically identifying low-abundance mutations in tumor genes based on the isolated forest algorithm. Background Technology

[0002] Low-abundance somatic mutations in tumors are of great significance for cancer genome research. However, the ability to detect low-abundance mutations in tumors is limited by errors introduced during sample preparation, sequencing, and analysis.

[0003] The main issues are: 1. Severe background contamination interference: Low biomass samples are easily contaminated by the experimental environment. Traditional methods, relying solely on abundance filtering, struggle to distinguish between real signals and contaminants. 2. Artificial mutations introduced during preparation: Tissue collection, formalin fixation, and preservation processes may artificially introduce mutations, such as the deamination of cytosine into uracil. C>T and G>A mutations are the main artificially introduced single-base mutations, directly interfering with the determination of somatic variations in the dataset. 3. Systematic errors: In the tumor next-generation sequencing (NGS) end-to-end detection system, interference from low-abundance mutation signals caused by DNA oxidative damage during extraction, amplification bias, and sequencing errors makes it difficult to accurately distinguish between real mutations and background noise.

[0004] With the widespread adoption of next-generation sequencing (NGS) technology in tumor gene testing, the demand for somatic mutation detection is increasing. While current NGS technology has achieved single-base resolution, it still faces a "false positive dilemma" in clinical applications—a 2023 survey report by the College of American Pathologists (CAP) showed that the false positive recall rate due to background noise in clinical laboratories was as high as 17%.

[0005] The background noise in the "false positive dilemma" mainly stems from: 1. Cross-contamination mechanisms in biological samples: Low-biomass samples are susceptible to aerosol DNA contamination in the experimental environment, and traditional filtering strategies based on a single abundance threshold are insufficient to effectively distinguish between genuine biological signals and contaminating background interference; 2. Mutagenic effects from sample processing: During the collection, formalin fixation, and long-term preservation of tissue samples, deamination reactions caused by nucleic acid damage lead to the systematic introduction of characteristic artificial mutations such as C>T and G>A, severely affecting the accurate determination of somatic cell variations; 3. Cumulative technical deviations throughout the entire process: From sample pretreatment to data generation, systematic deviations in various technical stages exhibit a cumulative effect, including but not limited to oxidative damage during nucleic acid extraction, PCR amplification bias, and error accumulation at key technical nodes such as optical signal crosstalk in sequencing chips. The industry urgently needs a multi-factor, dynamic, and quantifiable noise reduction method to solve the problem of accurate identification of low-frequency mutation signals in high-throughput sequencing data.

[0006] Furthermore, existing background processing methods for next-generation sequencing (NGS) technologies have the following drawbacks: Rigid threshold filtering: Static thresholds lead to a surge in false positives in high-depth regions. When identifying background noise mutations in NGS sequencing data, traditional methods using fixed mutation information thresholds or empirical models cannot dynamically adjust according to sample type, resulting in high false positive rates and an inability to dynamically adapt to different sequencing depths and site heterogeneity, leading to loss of effective data or residual noise. Black-box machine learning models: Rely on large amounts of labeled training data and suffer from cross-platform generalization barriers, failing to effectively distinguish between systematic technical biases and random sequencing errors. Bayesian methods: High computational complexity, increasing clinical sample analysis time by 3-5 times, failing to integrate experimental specific biases and ignoring artificially introduced biases, resulting in posterior probability estimation errors exceeding the 95% confidence interval's impact on noise distribution.

[0007] In traditional NGS mutation detection, feature weights are typically assigned a fixed weight (e.g., equal weight) or adjusted based on simple statistics (e.g., variance), leading to the following problems: (1) Excessive interference from low-quality data; mutation sites with low sequencing quality (e.g., base quality score <20) or strand-specific errors have extremely low confidence. If not weighted, this type of noise can significantly increase the abnormality score, resulting in false positives.

[0008] (2) Weight distortion of high-depth samples; In high-sequencing-depth samples, the absolute value of the mutation support number may mask the true biological signal. For example: If the depth of a certain site is 1000X and Alt Count is 5 (VAF = 0.5%), if the original count is used directly, the isolated forest may be misjudged as abnormal due to the "sparseness" of its value and its low confidence may be ignored.

[0009] (3) Parameter coupling problem of low abundance mutations: There is a strong correlation between features of low frequency mutations (such as VAF<1%). Traditional methods require manual adjustment of the threshold of each feature, which is cumbersome and difficult to standardize.

[0010] Furthermore, traditional mutation detection methods typically rely solely on conventional characteristics such as sequencing quality and mutation frequency, without fully considering unmeasurable or unpredictable systematic interference factors in experiments and bioinformatics analyses. These include: sample preparation-related interference such as common C>T base damage pseudomutations in FFPE (formalin-fixed paraffin-embedded) samples; experimental environment contamination such as false positive signals caused by exogenous DNA contamination or cross-contamination; systematic errors in sequencing platforms such as phasing errors in Illumina sequencing and alignment biases in low-complexity regions (such as microsatellite sequences); and batch effects such as technical noise introduced by different sequencing batches or library preparation methods. Summary of the Invention

[0011] Based on the above background technology, the present invention aims to provide a method for dynamically identifying low-abundance mutations in tumor genes based on the isolated forest algorithm. This method is a second-generation sequencing data processing method used to efficiently identify and filter background noise mutations, thereby improving the accuracy and reliability of somatic mutation detection.

[0012] The technical solution adopted in this invention is as follows: A method for dynamically identifying low-abundance mutations in tumor genes based on the isolated forest algorithm, comprising the following steps: Step S1: Construct a background noise frequency database; Step S2: Extract multi-dimensional features of the expected mutation readings, quality weighting, strand bias index, mutation quality ratio, and systematic errors of the tumor gene mutation sites to be tested. Step S3: Calculate the path length and anomaly score of tumor gene mutation sites using the isolated forest algorithm, and dynamically adjust the feature weights of multi-dimensional features based on confidence level. Step S4: Distinguish between real mutations and background noise based on anomaly scores.

[0013] Furthermore, in step S1, a background noise frequency database is constructed, specifically as follows: Step S11: Select white blood cell samples from tumor patients that do not contain somatic mutations as the baseline dataset; Step S12: Library construction and high-throughput sequencing are performed using amplicon sequencing of the target region from the benchmark dataset to obtain raw sequencing data; Step S13: The raw sequencing data is processed by the standard analysis workflow of the GATK genome analysis toolkit, and the detected mutations are treated as systematic technical noise. Step S14: Analyze the frequency distribution characteristics of systematic technical noise mutations and establish a background noise frequency reference database.

[0014] Furthermore, in step S2, the multi-dimensional feature extraction includes: calculating the expected mutation reads based on background noise frequency and sequencing depth; calculating the quality-weighted mutation allele frequency by combining mutation depth, base quality, and sequencing depth; calculating the strand bias index by the number of mutations in the forward and reverse strands; obtaining the mutation quality ratio by using the ratio of mutation quality score to average base quality; and calculating the site systematic error rate based on the benchmark dataset. Feature extraction of tumor gene mutation sites to be tested: based on the sequencing depth of a specific mutation in the tumor sample to be tested. Background noise frequencies corresponding to abrupt changes in the background noise database Calculate the expected mutation site Read the numerical value; ; It also extracts the expected mutation readings, quality-weighted VAF, strand bias index, mutation quality ratio, and systematic errors of the tumor gene mutation sites to be tested; it artificially introduces bias: C>T / G>A and specific DNA damage types: C>A / G>T and multifactor features (C>T refers to a C base mutation to a T base at a certain location in the genome sequence, and so on).

[0015] Furthermore, in step S3, the path length and anomaly score of the mutation site are calculated using the isolated forest algorithm, and the feature weights of the multi-dimensional features are dynamically adjusted based on the confidence level; specifically: Step S31: Calculate the path length of a single tree by using the segmentation distance and weights of multiple feature dimensions, and normalize the path length based on the sample size. Step S32: Calculate the anomaly score based on the average path length of multiple trees, and calibrate the anomaly score sensitivity using the Sigmoid function; Step S33: Introduce feature values ​​and integrate the ratio of sequencing depth to system error rate, mutation quality ratio, and strand bias index parameters; Step S34: Calculate the feature confidence level using a nonlinear function, and then dynamically correct the Fisher weight matrix.

[0016] Further, in step S31, the path length is calculated, as follows: For each desired mutation site x, the path length in a single tree is calculated as follows: ; in, For path length, w represents the segmentation distance of the expected mutation site x in the i-th feature dimension; i d represents the feature weights; d represents the number of feature dimensions. This represents the maximum segmentation distance along the i-th feature dimension; The normalized path length is expressed as: ; Where c(n) is the normalization factor, n is the sample size in the current subsample, the number of samples randomly drawn when constructing a single tree, H is the harmonic series, H(k) is the harmonic series when the sample size is k, reflecting the average search path length when the sample size is k, used for standardization, converting the original path length into a comparable value that is independent of the sample size. Step S32, calculate the anomaly score s(x) based on the average path length of multiple trees, expressed as: ; in, Represents the path length in all trees The mean; This represents the Sigmoid parameter fitted using negative control samples, used to control scoring sensitivity; Step S33 introduces feature values, specifically: extracting various feature values ​​of the mutation site to be tested: expected mutation reading, quality-weighted VAF, strand bias index, mutation quality ratio, and systematic error; each feature is defined as: (1) The desired mutation site x is represented as ; in, This represents the background noise frequency corresponding to the abrupt change in the background noise database. Indicates the sequencing depth of a specific mutation; (2) Quality-weighted VAF mutation frequency, expressed as: ; Among them, VAF adj This is expressed as the quality-weighted mutation frequency. This indicates the mutation depth of the mutation site; BaseQual indicates the average base quality of the mutation site; depth indicates the sequencing depth of the mutation site. (3) Chain bias index, expressed as: ; Where StrandBias represents the chain bias index, F alt Indicates the number of forward strand mutations; R alt Indicates the number of reverse strand mutations; (4) Mutation quality ratio, expressed as: ; Where VQratio represents the mutation quality ratio, VariantQual represents the mutation quality score, and AvgBaseQual represents the average base quality of the mutation site; (5) Systematic errors, represented as: ; in, i This represents the systematic error at position i, which is the ratio of the frequency of the corresponding base at position i in the benchmark dataset to the sequencing depth at position i; g and m are base g and base m, respectively; #reads with nucleotide m at position i represents the number of sequencing reads for base m at position i; Total#reads at position i represents the total number of sequencing reads at position i. S3.4 Feature weight allocation, specifically as follows: For each feature, the ability to distinguish between noisy abrupt changes and true abrupt changes is calculated, resulting in a weight matrix, which is represented as follows: ; Among them, w i Represents feature weights, This indicates that the mean and variance are calculated using noisy mutations and known positive mutations in the benchmark dataset. i，noise This represents the mean of the i-th feature in the noise abrupt change. i，real This represents the mean of the i-th feature in real mutations. This represents the variance of the i-th feature in the noise abrupt change. Let represent the variance of the i-th feature in the real mutation.

[0017] Further, in step S34, the feature confidence is calculated using a nonlinear function, and then the Fisher weight matrix is ​​dynamically corrected, specifically as follows: Step S341: Dynamically adjust the weight matrix by jointly optimizing the confidence of the features; Set confidence level C i , represented as: ; in, This means compressing the ratio of sequencing depth to systematic errors to the (0,1) range to avoid extreme values ​​dominating; This indicates a non-linear scaling of the mutation quality ratio. The median of the mutation quality ratio data was used to highlight high-quality mutations; The chain bias index is normalized to [-1,1] to the range (0,1) to suppress extreme chain bias; ϵ is a small value to prevent division by zero; Step S342, set the confidence level C i The initial weight matrix applied to step S341 is expressed as: ; in, This represents the corrected weight matrix. represents the initial weight matrix, and ⊙ represents element-wise multiplication; To adjust the parameters.

[0018] Furthermore, the confidence score is calculated by compressing the ratio of sequencing depth to systematic error to the range of (0,1), scaling the mutation quality ratio nonlinearly, normalizing the strand bias index to the range of (0,1), and finally generating a comprehensive confidence factor.

[0019] Furthermore, it also includes a dynamic threshold determination step: comparing the abnormal score of the mutation site with the 99th percentile of the mutation score in the benchmark dataset; if it is lower than the threshold, it is determined to be a true mutation. The input data includes a background noise database constructed in a standard format and a VCF file of the test sample. During output, the identified background noise mutations are marked as background noise mutations (BN) and filtered.

[0020] Furthermore, the method is applicable to the detection of low-abundance somatic mutations in tumors, and suppresses false positive signals caused by background contamination, artificial mutations, and systematic errors through multi-factor joint modeling.

[0021] Furthermore, mutations are divided into reliable and suspicious sets using multi-dimensional threshold rules. Mutations to be tested are pre-screened using background noise database, VCF of the test sample, sequencing depth, and frequency of the population database. Reliable mutations are retained, and suspicious mutations are dynamically identified as low-abundance mutations based on a multi-factor isolated forest algorithm.

[0022] The beneficial effects achieved by this invention are as follows: 1. This invention improves computational efficiency and is suitable for large-scale tumor gene next-generation sequencing (NGS) data analysis: The isolated forest algorithm used in this invention has linear time complexity (O(n)), which has higher computational efficiency than traditional machine learning methods (such as support vector machines or random forests), and is suitable for large-scale analysis of high-throughput sequencing data (such as whole exome sequencing WES or whole genome sequencing WGS).

[0023] 2. Optimize the mutation detection process and reduce the false positive rate: By combining the factor characteristics of mutation sites, this invention can more accurately identify abnormal mutation patterns, reduce false positive results caused by sequencing errors or technical noise, and improve the reliability of mutation detection.

[0024] 3. Dynamic threshold adjustment mechanism to adapt to different sample noise levels: In response to the background noise differences that may exist in different samples, such as different tumor types or sequencing batches, this invention adopts a dynamic threshold optimization strategy to automatically adjust the abnormality scoring threshold, ensuring a stable noise filtering effect under different data conditions. Attached Figure Description

[0025] Figure 1 This is a flowchart of the steps of the method of the present invention. Detailed Implementation

[0026] The following embodiments are only used to illustrate the technical solutions of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0027] Reference Figure 1A method for dynamically identifying low-abundance mutations in tumor genes based on the isolated forest algorithm includes the following steps: Step S1: Construct a background noise frequency database; Step S2: Extract multi-dimensional features of the expected mutation readings, quality weighting, strand bias index, mutation quality ratio, and systematic errors of the tumor gene mutation sites to be tested. Step S3: Calculate the path length and anomaly score of tumor gene mutation sites using the isolated forest algorithm, and dynamically adjust the feature weights of multi-dimensional features based on confidence level. Step S4: Distinguish between real mutations and background noise based on anomaly scores.

[0028] Furthermore, in step S1, a background noise frequency database is constructed, specifically as follows: Step S11: Select white blood cell samples from tumor patients that do not contain somatic mutations as the baseline dataset; Step S12: Library construction and high-throughput sequencing are performed using amplicon sequencing of the target region from the benchmark dataset to obtain raw sequencing data; Step S13: The raw sequencing data is processed by the standard analysis workflow of the GATK genome analysis toolkit, and the detected mutations are treated as systematic technical noise. Step S14: Analyze the frequency distribution characteristics of systematic technical noise mutations and establish a background noise frequency reference database.

[0029] The GATK standard analysis process includes base mass fraction recalibration (BQSR), insertion / deletion realignment (Indel Realigner) for systematic error correction, and the application of conventional filtering parameters for mutation screening, ultimately obtaining a set of high-confidence mutation sites.

[0030] Among them, systematic technical noise is defined as the detection of mutations based on the premise that white blood cell samples theoretically do not carry somatic cell mutations.

[0031] Furthermore, in step S2, the multi-dimensional feature extraction includes: calculating the expected mutation reads based on background noise frequency and sequencing depth; calculating the quality-weighted mutation allele frequency by combining mutation depth, base quality, and sequencing depth; calculating the strand bias index by the number of mutations in the forward and reverse strands; obtaining the mutation quality ratio by using the ratio of mutation quality score to average base quality; and calculating the site systematic error rate based on the benchmark dataset. Feature extraction of tumor gene mutation sites to be tested: based on the sequencing depth of a specific mutation in the tumor sample to be tested. Background noise frequencies corresponding to abrupt changes in the background noise database Calculate the expected mutation site Read the numerical value; ; It also extracts the expected mutation readings, quality-weighted VAF, strand bias index, mutation quality ratio, and systematic errors of the tumor gene mutation sites to be tested; it artificially introduces bias: C>T / G>A and specific DNA damage types: C>A / G>T and multifactor features (C>T refers to a C base mutation to a T base at a certain location in the genome sequence, and so on).

[0032] Furthermore, in step S3, the path length and anomaly score of the mutation site are calculated using the isolated forest algorithm, and the feature weights of the multi-dimensional features are dynamically adjusted based on the confidence level; specifically: Step S31: Calculate the path length of a single tree by using the segmentation distance and weights of multiple feature dimensions, and normalize the path length based on the sample size. Step S32: Calculate the anomaly score based on the average path length of multiple trees, and calibrate the anomaly score sensitivity using the Sigmoid function; Step S33: Introduce feature values ​​and integrate the ratio of sequencing depth to system error rate, mutation quality ratio, and strand bias index parameters; Step S34: Calculate the feature confidence level using a nonlinear function, and then dynamically correct the Fisher weight matrix.

[0033] Further, in step S31, the path length is calculated, as follows: For each desired mutation site x, the path length in a single tree is calculated as follows: ; in, For path length, w represents the segmentation distance of the expected mutation site x in the i-th feature dimension; i d represents the feature weights; d represents the number of feature dimensions. This represents the maximum segmentation distance along the i-th feature dimension; The normalized path length is expressed as: ; Where c(n) is the normalization factor, n is the sample size in the current subsample, the number of samples randomly drawn when constructing a single tree, H is the harmonic series, H(k) is the harmonic series when the sample size is k, reflecting the average search path length when the sample size is k, used for standardization, converting the original path length into a comparable value that is independent of the sample size. Step S32, calculate the anomaly score s(x) based on the average path length of multiple trees, expressed as: ; in, Represents the path length in all trees The mean; This represents the Sigmoid parameter fitted using negative control samples, used to control scoring sensitivity; Step S33 introduces feature values, specifically: extracting various feature values ​​of the mutation site to be tested: expected mutation reading, quality-weighted VAF, strand bias index, mutation quality ratio, and systematic error; each feature is defined as: (1) The desired mutation site x is represented as: ; in, This represents the background noise frequency corresponding to the abrupt change in the background noise database. Indicates the sequencing depth of a specific mutation; (2) Quality-weighted VAF mutation frequency, expressed as: ; Among them, VAF adj This is expressed as the quality-weighted mutation frequency. This indicates the mutation depth of the mutation site; BaseQual indicates the average base quality of the mutation site; depth indicates the sequencing depth of the mutation site. (3) Chain bias index, expressed as: ; Where StrandBias represents the chain bias index, F alt Indicates the number of forward strand mutations; R alt Indicates the number of reverse strand mutations; (4) Mutation quality ratio, expressed as: ; Where VQratio represents the mutation quality ratio, VariantQual represents the mutation quality score, and AvgBaseQual represents the average base quality of the mutation site; (5) Systematic errors, represented as: ; in, i This represents the systematic error at position i, which is the ratio of the frequency of the corresponding base at position i in the benchmark dataset to the sequencing depth at position i; g and m are base g and base m, respectively; #reads with nucleotide m at position i represents the number of sequencing reads for base m at position i; Total#reads at position i represents the total number of sequencing reads at position i. S3.4 Feature weight allocation, specifically as follows: For each feature, the ability to distinguish between noisy abrupt changes and true abrupt changes is calculated, resulting in a weight matrix, which is represented as follows: ; Among them, w i Represents feature weights, This indicates that the mean and variance are calculated using noisy mutations and known positive mutations in the benchmark dataset. i，noise This represents the mean of the i-th feature in the noise abrupt change. i，real This represents the mean of the i-th feature in real mutations. This represents the variance of the i-th feature in the noise abrupt change. Let represent the variance of the i-th feature in the real mutation.

[0034] Further, in step S34, the feature confidence is calculated using a nonlinear function, and then the Fisher weight matrix is ​​dynamically corrected, specifically as follows: Step S341: Dynamically adjust the weight matrix by jointly optimizing the confidence of the features; Set confidence level C i , represented as: ; in, This means compressing the ratio of sequencing depth to systematic errors to the (0,1) range to avoid extreme values ​​dominating; This indicates a non-linear scaling of the mutation quality ratio. The median of the mutation quality ratio data was used to highlight high-quality mutations; The chain bias index is normalized to [-1,1] to the range (0,1) to suppress extreme chain bias; ϵ is a small value to prevent division by zero; Step S342, set the confidence level C i The initial weight matrix applied to step S341 is expressed as: ; in, This represents the corrected weight matrix. represents the initial weight matrix, and ⊙ represents element-wise multiplication; To adjust the parameters.

[0035] Furthermore, the confidence score is calculated by compressing the ratio of sequencing depth to systematic error to the range of (0,1), scaling the mutation quality ratio nonlinearly, normalizing the strand bias index to the range of (0,1), and finally generating a comprehensive confidence factor.

[0036] Furthermore, it also includes a dynamic threshold determination step: comparing the abnormal score of the mutation site with the 99th percentile of the mutation score in the benchmark dataset; if it is lower than the threshold, it is determined to be a true mutation. The input data includes a background noise database constructed in a standard format and a VCF file of the test sample. During output, the identified background noise mutations are marked as background noise mutations (BN) and filtered.

[0037] Furthermore, the method is applicable to the detection of low-abundance somatic mutations in tumors, and suppresses false positive signals caused by background contamination, artificial mutations, and systematic errors through multi-factor joint modeling.

[0038] Furthermore, mutations are divided into reliable and suspicious sets using multi-dimensional threshold rules. Mutations to be tested are pre-screened using background noise database, VCF of the test sample, sequencing depth, and frequency of the population database. Reliable mutations are retained, and suspicious mutations are dynamically identified as low-abundance mutations based on a multi-factor isolated forest algorithm.

[0039] Example The background noise frequency was collected, specifically: First, several clinically validated leukocyte samples without somatic mutations were selected as the baseline dataset. Library construction and high-throughput sequencing were performed using a target region amplicon sequencing panel to obtain the raw sequencing data. The raw sequencing data were processed using the GATK standard analysis workflow, which included systematic error correction steps such as base quality fraction recalibration (BQSR) and insertion / deletion realignment (Indel Realigner). Conventional filtering parameters were then applied to screen for variants, ultimately obtaining a set of high-confidence mutation sites.

[0040] Since white blood cell samples theoretically do not carry somatic cell mutations, the mutations detected above are defined as systematic technical noise.

[0041] By systematically analyzing the frequency distribution characteristics of these noise abrupt changes, a statistically significant background noise frequency baseline reference database is finally constructed.

[0042] Feature extraction of the mutation site to be tested: Based on the sequencing depth of a specific mutation in the sample to be tested (usually referring to the genomic sequence obtained by next-generation sequencing of DNA extracted from tumor tissue cells). Background noise frequencies corresponding to abrupt changes in the background noise database Calculate the expected mutation site Read the numerical value.

[0043] The expected mutation reads, quality-weighted VAF, strand bias index, mutation quality ratio, and systematic errors for this site were extracted. This includes artificially introduced biases: C>T / G>A, and specific DNA damage types: C>A / G>T, among other multi-factor characteristics. (C>T refers to a C base mutation to a T base at a certain location in the genome sequence, and so on). S3. The mutation points obtained in S2 are processed based on the isolated forest algorithm, specifically as follows: S3.1: The path length is normalized after calculation (correcting for the influence of sample size), and is expressed as: Where n is the sample size in the current subsample (i.e., the number of samples randomly drawn when constructing a single iTree); H(k) is the harmonic number, which reflects the average search path length when the sample size is k and is used for standardization; c(n) is the normalization factor, which converts the original path length into a comparable value that is independent of the sample size. S3.2: Calculate the anomaly score to introduce feature confidence, expressed as: ; in, Represents the path length in all trees The mean; This represents the Sigmoid parameter fitted using negative control samples, used to control scoring sensitivity; S3.3; Eigenvalue introduction, specifically: extracting various eigenvalues ​​of the mutation to be tested: expected mutation reading, quality-weighted VAF, chain bias index, mutation quality ratio, and systematic error; The benchmark dataset consists of genomic sequences extracted from a series of white blood cells. Theoretically, it does not contain any systemic mutations. Therefore, any suspected systemic mutations appearing in the benchmark dataset are assumed to be caused by systematic errors. Specifically, the systematic error rate is calculated by dividing the mutation frequency of a base at a certain position (#reads with nucleotide m at position i) by the total number of bases at that position (Total#reads at position i). S3.4; Feature weight allocation; S3.5; This invention designs a dynamic nonlinear weight correction method based on confidence level, specifically as follows: The weight matrix is ​​dynamically adjusted by jointly optimizing the confidence (sequencing reliability) of the features. This method eliminates the need for individual feature adjustments and automatically adapts to the data distribution.

[0044] Set confidence level C i , represented as: ; in, This means compressing the ratio of sequencing depth to systematic errors to the (0,1) range to avoid extreme values ​​dominating; This indicates a non-linear scaling of the mutation quality ratio. The median of the mutation quality ratio data was used to highlight high-quality mutations; This indicates that the chain bias index is normalized to [-1,1] to the range of (0,1) to suppress extreme chain bias; To prevent small quantities from being reduced to zero; Confidence level C i The initial weight matrix applied is represented as follows: ; in, This represents the corrected weight matrix. represents the initial weight matrix, and ⊙ represents element-wise multiplication; To adjust the parameters.

[0045] It should be further noted that, in one example, refer to Table 1: Table 1

[0046] The list of mutations after processing is shown in Table 2: Table 2

[0047] In one example: Referring to Table 3, 75 clinical samples with prior results were collected, and the results of correctly identifying noise using the fixed noise recognition threshold were statistically analyzed.

[0048] Table 3

[0049] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamically identifying low-abundance mutations of tumor genes based on the Isolation Forest algorithm, characterized in that: The method comprises the following steps: Step S1, constructing a background noise frequency database; Step S2, extracting the multi-dimensional features of the expected mutation reads, quality weighting, strand bias index, mutation quality ratio and systematic error of the tumor gene mutation site to be detected; Step S3, calculating the path length and anomaly score of the tumor gene mutation site by using the Isolation Forest algorithm, and dynamically adjusting the feature weight of the multi-dimensional features through confidence; Step S4, distinguishing the true mutation from the background noise based on the anomaly score.

2. The method for dynamically identifying low-abundance mutations of tumor genes based on the Isolation Forest algorithm according to claim 1, characterized in that: In step S1, the background noise frequency database is constructed, specifically: Step S11, selecting a white blood cell sample of a tumor patient without somatic mutation as a benchmark data set; Step S12, performing library construction and high-throughput sequencing by using target region amplicon sequencing of the benchmark data set to obtain raw sequencing data; Step S13, processing the raw sequencing data by using the Genome Analysis Toolkit (GATK) standard analysis process, and detecting mutations as systematic technical noise; Step S14, analyzing the frequency distribution characteristics of the systematic technical noise mutations, and establishing a background noise frequency reference database.

3. The method for dynamically identifying low-abundance mutations of tumor genes based on the Isolation Forest algorithm according to claim 2, characterized in that: In step S2, the multi-dimensional feature extraction includes: calculating the expected mutation reads based on the background noise frequency and sequencing depth; calculating the quality weighted mutant allele frequency by combining the mutation depth, base quality and sequencing depth; calculating the strand bias index by the number of forward and reverse strand mutations; obtaining the mutation quality ratio by using the ratio of the mutation quality score to the average base quality; and calculating the site systematic error rate based on the benchmark data set; Feature extraction of mutation sites to be tested in tumor genes: sequencing depth of a certain mutation in a tumor sample to be tested and the background noise frequency of the corresponding mutation in the background noise database Calculate the expected mutation site read value; ; And extracting the expected mutation reads, quality weighted VAF, strand bias index, mutation quality ratio and systematic error of the tumor gene mutation site to be detected; introducing human bias: C>T / G>A and specific DNA damage type: C>A / G>T and multi-factor features.

4. The method for dynamically identifying low-abundance mutations of tumor genes based on the Isolation Forest algorithm according to claim 3, characterized in that: In step S3, the path length and anomaly score of the mutation site are calculated by using the Isolation Forest algorithm, and the feature weight of the multi-dimensional features is dynamically adjusted through confidence; Specifically: Step S31, calculating the path length of a single tree by the segmentation distance and weight of multiple feature dimensions, and normalizing the path length based on the sample size; Step S32, calculating the anomaly score based on the average path length of multiple trees, and calibrating the anomaly score sensitivity by using the Sigmoid function; Step S33, introducing feature values, integrating the ratio of sequencing depth to systematic error rate, mutation quality ratio, and strand bias index parameters; Step S34, calculating the feature confidence by using a nonlinear function, and then dynamically correcting the Fisher weight matrix.

5. The method for dynamically identifying low-abundance mutations of tumor genes based on the Isolation Forest algorithm according to claim 4, wherein: Step S31, path length calculation, is represented as: For each expected mutation site x, the path length in a single tree is calculated as: ； wherein, is the path length, represents the split distance of the desired mutation site x on the i-th feature dimension; w i represents the feature weight; d represents the number of feature dimensions, represents the split maximum distance on the i-th feature dimension; The normalized path length is represented as: ； Wherein, c(n) is a normalization factor, n is the sample size in the current sub-sample, the number of samples randomly extracted when constructing a single tree, H is a harmonic series, H(k) is a harmonic series when the sample size is k, reflecting the average search path length when the sample size is k, for standardization, converting the original path length into a comparable value independent of the sample size; Step S32, the average path length calculation anomaly score s(x) of the multi-tree is represented as: ； wherein, represents the mean of the path lengths in all trees; represents the Sigmoid parameter fitted by the negative control samples for controlling the scoring sensitivity;​ Step S33, the eigenvalue is introduced, specifically: extracting each eigenvalue of the to-be-tested mutation site: expected mutation reads, quality-weighted VAF, strand bias index, mutation quality ratio and systematic error; each feature is defined as: (1) The expected mutation site x is represented as: ； wherein, represents the background noise frequency in the background noise database corresponding to the mutation, represents the sequencing depth for a certain mutation; (2) The quality-weighted VAF mutation frequency is represented as: ； wherein VAF adj is represented as a mass-weighted mutation frequency, is the mutation depth of the mutation site; BaseQual represents the average base quality of the mutation site; and depth represents the sequencing depth of the mutation site. (3) The strand bias index is represented as: ； wherein StrandBias represents a strand bias index, F alt represents the number of forward strand mutations; R alt represents the number of reverse strand mutations; (4) The mutation quality ratio is represented as: ； Wherein, VQratio represents the mutation quality ratio, VariantQual represents the mutation quality score, and AvgBaseQual represents the average base quality of the mutation site; (5) The systematic error is represented as: ； wherein, i represents the systematic error at the i-th position, i.e. the ratio of the frequency of the i-th position corresponding base in the reference data set to the sequencing depth of the i-th position; g and m are base g and base m, respectively; # reads with nucleotide m at position i represents the number of sequencing reads of base m at position i; Total # reads at position i represents the total number of sequencing reads at position i; S3.4 Feature weight distribution, specifically: For each feature, the ability to distinguish noise mutations from real mutations is calculated to obtain a weight matrix, represented as: ； where w i denotes the feature weight, denotes the mean and variance computed from noise mutations and known positive mutations in the benchmark dataset, i，noise denotes the mean of the ith feature in noise mutations, i，real denotes the mean of the ith feature in true mutations, denotes the variance of the ith feature in noise mutations, denotes the variance of the ith feature in true mutations.

6. The method for dynamically identifying low-abundance mutations of tumor genes based on the Isolation Forest algorithm according to claim 5, characterized in that: Step S34, the feature confidence is calculated through a nonlinear function, and then the Fisher weight matrix is dynamically corrected, specifically: Step S341, the confidence of the feature is optimized by joint optimization, and the weight matrix is dynamically adjusted. Setting the confidence C i is expressed as: ； wherein, represents compressing the ratio of sequencing depth to systematic error to the range (0, 1) to avoid extreme values dominating; represents a non-linear scaling of the mutation quality ratio, takes the median of the mutation quality ratio data to highlight high quality mutations; represents the chain bias index normalized to [-1, 1] converted to the range (0, 1) to suppress extreme chain bias; and e is a small quantity to prevent division by zero. Step S342, the confidence C i The initial weight matrix applied to step S341 is denoted as: ； wherein, denotes the modified weight matrix, denotes the initial weight matrix, and denotes element-wise multiplication; is a tuning parameter.

7. The method for dynamically identifying low-abundance mutations of tumor genes based on the Isolation Forest algorithm according to claim 6, characterized in that: The confidence is calculated by the following method: the ratio of sequencing depth to systematic error is compressed to the range of (0, 1), the mutation quality ratio is nonlinearly scaled, the strand bias index is normalized to the range of (0, 1), and finally a comprehensive confidence factor is generated.

8. The method for dynamically identifying low-abundance mutations of tumor genes based on the Isolation Forest algorithm according to claim 7, characterized in that: It also includes a dynamic threshold determination step: comparing the anomaly score of the mutation site with the 99th percentile of the mutation score in the benchmark data set, and determining it as a real mutation if it is below the threshold; The input data includes a background noise database constructed in a standard format and a file of the to-be-tested sample VCF, and when output, the identified background noise mutations are marked as background noise mutations BN and filtered.

Images