tNGS Multi-Target Copy Number Absolute Quantitative Analysis Calculation Method

By using target-specific primers with unique molecular identifiers and a multi-level quality control correction method in tNGS technology, the problems of imperfect multiplex amplification and quantitative calibration are solved, and the accuracy and robustness of target copy number analysis are improved.

CN123120022APending Publication Date: 2026-11-17探因医学科技(浙江)有限公司 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611613203.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In tNGS multitarget copy number analysis, existing technologies suffer from problems such as multiple amplification competition correction and imperfect global quantitative calibration, leading to inconsistent amplification efficiency and differences in sequencing read quality, which affects the accuracy of quantitative results.

Method used

Target-specific primers with unique molecular identifiers were used to amplify and sequence the nucleic acids in the samples. Quantitative calibration was performed by extracting original read clusters, filtering read quality, merging UMIs, and introducing internal reference targets. Multi-level quality control correction and competitive suppression were combined with graph attention networks and sparse Bayesian learning frameworks to generate absolute copy number quantification values.

Benefits of technology

It improves the accuracy of target molecule count and the robustness of quantification results, reduces the impact of PCR repeat counting and sequencing errors, and achieves accurate and consistent quantification of multi-target copy numbers.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120022A_ABST
    Figure CN123120022A_ABST
Patent Text Reader

Abstract

This invention provides a method for calculating the absolute copy number of multiple targets using tNGS, belonging to the field of molecular biology detection / gene sequencing analysis. It employs target-specific primers with unique molecular identifiers (UMIs) for multiplex amplification and sequencing of sample nucleic acids; an internal reference target with a known copy number is added to the sample; amplification and sequencing are performed to complete quantitative calibration; and read clusters are extracted according to target-specific sequences and UMI families. This invention improves the accuracy and consistency of absolute copy number quantification of multiple targets through target-specific primers with unique molecular identifiers, extraction of original read clusters, two-stage composite quality control correction, and target read retention rate compensation. Under non-negative constraints, it solves the joint posterior of the number of endogenous template molecules and the competitive inhibition coefficient using sparse Bayesian prior and Markov chain Monte Carlo (MCMC), achieving closed-loop quality control from amplification, sequencing to quantitative output; and it retains and outputs low-quality target reference information and provides indicators of successful or failed quantification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of molecular biology detection / gene sequencing analysis, and more specifically, to a method for calculating the absolute quantitative analysis of multi-target copy number in tNGS. Background Technology

[0002] tNGS (targeted next-generation sequencing) can enrich multiple targets in a sample through multiplex amplification and sequence the amplified products, making it valuable in pathogen detection, gene mutation detection, copy number analysis, and other scenarios. Absolute quantification usually requires the introduction of an internal reference target or standard with a known copy number, and calibration in conjunction with sequencing read counts.

[0003] During multiplex amplification, competitive amplification may occur between different target primers, leading to inconsistent amplification efficiencies and differences in sequencing read quality and the ratio of forward to reverse strand reads. Using a unique molecular identifier (UMI) can reduce the impact of PCR repeats on molecular counting, but UMI merging, read quality filtering, inter-target competition correction, and global quantitative calibration still affect the final quantitative results. Therefore, a feasible tNGS method for calculating the absolute copy number of multiple targets is needed to comprehensively address these issues. Summary of the Invention

[0004] To overcome the above deficiencies, this invention provides a method for calculating the absolute quantitative analysis of multi-target copy number in tNGS, aiming to improve the problems of imperfect multiple amplification competitive correction and global quantitative calibration in the prior art.

[0005] This invention is implemented as follows: This invention provides a method for absolute quantitative analysis of multi-target copy number in tNGS. It employs target-specific primers with unique molecular identifiers to amplify and sequence sample nucleic acids multiplex, extracting raw read clusters for each target. After filtering the read quality, the target single-molecule count is obtained by homologous merging based on unique molecular identifiers. i ); Add an internal control target with a known copy number to the sample, amplify and sequence to complete quantitative calibration, including: S1. Target-specific primers with a unique molecular identifier at the 5' end are used for multiplex amplification and sequencing of sample nucleic acids. Original read clusters of each target are extracted according to the primer binding sites. The unique molecular identifier is located at the 5' end of the target-specific primer, is 8-16 nt in length, consists of incompletely random bases, and is limited to no more than 3 consecutive identical bases. The GC content is 40%-60%. A 1-5 base spacer sequence is placed between the unique molecular identifier and the target-specific sequence. During the first round of extension, the unique molecular identifier is incorporated into the amplification product along with the primers. S2. After normalizing the mean base quality of the reads, the window mismatch rate, and the ratio of forward to reverse strand reads, a weighted fusion is performed to generate a first-level composite quality control correction factor. Reads below the lower limit of judgment are removed, and the retained reads are included in the effective read cluster. The target read retention rate is calculated. The mean base quality is linearly scaled to 0-1 according to the quality value range; the window mismatch rate is normalized after taking the upper limit of the number of mismatched bases; the ratio of forward to reverse strand reads is normalized according to the upper limit of the ratio; the first-level composite quality control correction factor is the weighted sum of the three normalized parameters, and the lower limit of judgment is that the first-level composite quality control correction factor is not less than 0.4; the target read retention rate is the ratio of the number of effective read clusters of the target to the number of original read clusters. S3. Construct a read similarity map based on unique molecular identifiers, cluster to obtain single-molecule families, use the number of single-molecule families as the target single-molecule count, divide by the target read retention rate to obtain the target observed molecules, and use effective reads with the same target identifier and identical unique molecular identifiers as nodes; node features include UMI sequence encoding, window mismatch rate, number of consecutive identical UMI bases, chain orientation, and quality value; edge features or edge weights include inter-read base editing distance, UMI correction distance, and window mismatch rate; edge weights are inter-read base editing distances. The algorithm employs a weighted combination of UMI correction distance and window mismatch rate; it aggregates neighborhood features using a graph attention network with UMI error correction awareness, the loss function of which is a weighted sum of contrast loss, UMI correction distance regularization term, primer mismatch regularization term, family size sparsity regularization term, and chain balance regularization term; it divides single-molecule families according to the aggregation feature similarity threshold of 0.80-0.90; it generates consistent sequences for reads within a family using variational inference; and the target observed molecule number is the number of single-molecule families divided by the target read retention rate. S4. Divide the sample nucleic acid into several equal parts, add the internal control target with the same known copy number, add the target synthetic template with a logarithmic interval gradient covering the estimated number of endogenous template molecules of 0.25-4 times, and set up unlabeled equal parts. Amplify and sequence under the same conditions to obtain the number of target molecules observed in each equal part. S5. After normalizing the amplicon coverage uniformity, target site base mismatch ratio, and target read retention rate, a weighted fusion is performed to generate a second-level composite quality control correction factor. Targets with a value below the lower limit are removed as low-quality targets. The amplicon coverage uniformity is normalized using the coefficient of variation; the target site base mismatch ratio is normalized according to the proportion of mismatched reads to the total reads; the target read retention rate is normalized; the lower limit is that the second-level composite quality control correction factor is not less than 0.3, and targets with a value below 0.3 are considered low-quality targets. S6. Set a sliding window and compare the window similarity of the target primer and template binding regions. When the window similarity is higher than the preset threshold and the window mismatch rate is lower than the preset threshold, set the initial value of the competitive inhibition coefficient of the target pair to the product of the window similarity and the preset proportional coefficient; otherwise, set it to zero. Arrange the competitive inhibition coefficients between each target pair to form a competitive network matrix. The off-diagonal elements are the absolute values ​​of the competitive inhibition coefficients of the corresponding target pairs, and the diagonal elements are set to zero. Normalize by row. Group the targets according to the connected components of the competitive network matrix. Establish a competitive saturation response equation within each group: the number of observed target molecules is equal to the product of the intrinsic amplification response coefficient and the number of endogenous template molecules divided by the competitive inhibition term. The competitive inhibition term is 1 plus the sum of the products of the competitive inhibition coefficient and the vector of the number of endogenous template molecules within the group. Under non-negative constraints, the number of observed target molecules in unspiked and spiked equal parts is used as the fitting target. The known amount of target synthetic template added in the spiked equal parts is included as a known term in the competitive saturation response equation. The second-level composite quality control correction factor was used as the residual weighting coefficient. A prior was applied to the competition inhibition coefficient using a sparse Bayesian learning framework. The joint posterior distribution of the endogenous template molecule number vector and the competition inhibition coefficient was solved iteratively using Markov Chain Monte Carlo (MCMC) posterior sampling. A cross-target group inter-group coupling term was introduced for global correction to obtain the competitive correction copy number of each target. The sparse Bayesian learning framework applied a Laplace prior to the competition inhibition coefficient and a Gamma prior to the endogenous template molecule number. The Markov Chain Monte Carlo posterior sampling used the Metropolis-Hastings algorithm or Gibbs sampler, with 10,000-50,000 iterations, a burn-in rate of 20%-50%, a thin rate of 5-10, and an acceptance rate of 20%-40%. The convergence criterion was that the Euclidean distance between the endogenous template molecule number vectors of two adjacent samples was less than 1% of their mean, or the weighted residual norm was less than 0.05. The inter-group coupling coefficient (d im The inter-group coupling term is independently estimated based on inter-group window similarity, batch effect, or shared resource competition; im ×T m sum; S7. The ratio of the competing calibration copy number to the known copy number of the internal control target in each aliquot is used as the response factor for each aliquot. The mean of the response factors for each aliquot is used as the global correction coefficient. The quantitative value of the absolute copy number of each target is output. For each spiked aliquot, the response factor of the internal control target is calculated, and the mean of the response factors for each aliquot is used as the global correction coefficient. If the response factor of a certain aliquot deviates from the mean ±30%, it is removed and recalculated. The absolute copy number of each target is the competing calibration copy number of each target divided by the global correction coefficient. The spike recovery rate and the coefficient of determination (R²) of the linear regression are calculated from the observed number of target molecules in each spiked aliquot and the estimated value obtained according to the competitive saturation response equation. 2When the recovery rate of each spiked aliquot is within the range of 70%-130% and the linear regression determination coefficient is not less than 0.90, the output will indicate successful quantification; otherwise, the output will indicate failed quantification. In S6, the preset proportional coefficient β is obtained by using the gradient-spiked amplification data obtained in S4, performing linear regression with the target pair window similarity as the independent variable and the measured amplification efficiency decrease ratio as the dependent variable, and using the regression slope as the preset proportional coefficient β.

[0006] Preferably, in step S1, a unique molecular identifier is placed at the 5' end of the target-specific primer and incorporated with the primer during the first extension; the machine-readable segment is compared with a preset target primer sequence library, the start and end sites of the primer binding region are recorded, and the read fragment is extracted; reads with the same target identifier and whose fragment lengths conform to the amplicon length range are grouped into the original read cluster, where the fragment length conforms to the amplicon length range means that the expected amplicon length is ±10% or ±20bp.

[0007] By adopting the above technical solution, a unique molecular identifier is set at the 5' end of the target-specific primer and incorporated with the primer during the first extension, so that each original template molecule is distinguishably labeled. After the sequencing reads are compared with the preset target primer sequence library, the start and end sites of the primer binding region are located, and the amplicon length range is screened, they are classified into the original read cluster. This associates the target identity, the unique molecular identifier identity, and the original template molecule, reducing PCR duplicate counting, primer region mismatch, and target assignment errors, and providing a reliable basis for subsequent target single molecule counting and absolute quantification.

[0008] Preferably, in step S2, the average base quality is linearly scaled to 0-1 according to the quality value range; the window mismatch rate is normalized after taking the upper limit of the number of mismatched bases; the ratio of forward and reverse strand read segments is normalized according to the upper limit of the ratio; and the lower limit of judgment is that the first-level composite quality control correction factor is not less than 0.4.

[0009] By adopting the above technical solution, the mean base quality of reads, window mismatch rate, and forward / reverse read ratio are normalized respectively. The weights of each parameter are determined according to their correlation with the target read retention rate and then weighted and fused to generate a first-level composite quality control correction factor. Low-quality reads are removed with a lower limit of not less than 0.4, which can simultaneously control sequencing quality, primer matching quality, and strand balance deviation. The target read retention rate is calculated by the ratio of the number of effective read clusters to the number of original read clusters, which can compensate for the molecular loss caused by quality control removal in subsequent target single molecule counting and improve the accuracy of target molecule observation.

[0010] Preferably, in step S3, effective reads identified by the same target are used as nodes, and effective reads with UMI correction distance and read edit distance as edge weights are used as nodes; node features include UMI sequence encoding, window mismatch rate, number of consecutive identical UMI bases, chain orientation, and quality value; edge features or edge weights include inter-read base edit distance, UMI correction distance, and window mismatch rate; edge weights are a weighted combination of inter-read base edit distance, UMI correction distance, and window mismatch rate; a graph attention network with UMI error correction awareness is used to aggregate neighborhood features, and the loss function is a weighted sum of contrastive loss, UMI correction distance regularization term, primer region mismatch regularization term, family size sparsity regularization term, and chain balance regularization term; consistent sequences are generated for reads within a family using variational inference.

[0011] By adopting the above technical solution, effective reads with the same target identifier and unique molecular identifier are used as nodes. Features such as UMI correction distance, window mismatch rate, inter-read base editing distance, number of consecutive identical UMI bases, strand orientation and quality value are integrated. Clustering is performed using a graph attention network with UMI error correction awareness, contrast loss, UMI correction distance regularization term, primer mismatch regularization term, family size sparsity regularization term and strand balance regularization term. This can accurately classify single-molecule families when UMI errors and sequencing errors exist. The number of single-molecule families is used as the target single-molecule count and divided by the target read retention rate to obtain the target observed molecule number, which can reduce the impact of PCR repetition and quality control loss on quantification.

[0012] Preferably, in step S4, target synthetic templates with known copy numbers are added at logarithmic intervals, covering 0.25-4 times the estimated number of endogenous template molecules for the target; the internal control target is an exogenous synthetic nucleic acid target; each aliquot is amplified, library constructed, and sequenced using the same multiplex amplification system, primer pool, and cycling conditions, in 4-8 aliquots, preferably 6 aliquots; the amount of internal control target added is 10³-10¹⁰ per aliquot. 6 copies, preferably 10 4 copies; the logarithmic interval gradient is 0.25 times, 0.5 times, 1 time, 2 times, 4 times; if the number of equal parts is 8, intermediate gradient points such as 0.35 times, 1.5 times, 3 times can be added.

[0013] By adopting the above technical solution, the nucleic acid of the sample is divided into multiple equal parts, and the same known copy number of exogenous internal reference target is added to each part. The target synthetic template is added at log intervals of 0.25-4 times the estimated number of endogenous template molecules, and unspiked equal parts are set. Library construction and sequencing are carried out under the same multiplex amplification system, primer pool and cycling conditions. The number of observed target molecules can be obtained under the same background, which provides sufficient and comparable data for subsequent competitive network fitting, endogenous template molecule number solution, global correction coefficient calculation and evaluation of spiked recovery rate and linear relationship.

[0014] Preferably, in S5, the amplicon coverage uniformity is calculated by dividing the standard deviation of sequencing depth by the mean; the target site base mismatch ratio is calculated by the proportion of mismatched reads to the total reads; the weight of the second-level composite quality control correction factor is determined by normalizing the positive correlation part of the Spearman correlation coefficient between each parameter and the target read retention rate; only positive correlation coefficients are taken, and the weight of a parameter is obtained by dividing the positive correlation coefficient of a certain parameter by the sum of all positive correlation coefficients; the weight of a correlation coefficient that is zero or negative is set to zero and renormalized; if all correlation coefficients are zero or negative, the weights are evenly distributed or a preset default weight is used, and the sum of the weights is 1; if all Spearman correlation coefficients are negative, the weights are evenly distributed or a default weight is set. The lower limit for judgment is that the second-level composite quality control correction factor is not lower than 0.3; low-quality targets that are rejected are not included in S6 fitting and S7 recalibration.

[0015] By adopting the above technical solution, after normalizing the amplicon coverage uniformity, target site base mismatch ratio, and target read retention rate, the parameters are weighted and fused according to their correlation with the target quantitative repeatability to generate a second-level composite quality control correction factor. Low-quality targets are eliminated with a threshold of not less than 0.3. This avoids interference from targets with uneven coverage, high target site base mismatch ratio, or severe read loss in the competitive network construction and global quantitative calibration. The eliminated targets do not enter the S6 fitting and S7 recalibration, which can improve the repeatability and robustness of multi-target absolute quantification.

[0016] Preferably, in step S6, the sliding window length is 80%-200% of the primer sequence length, preferably 100%-150%; the step size is 1-5 nt; the window similarity is calculated as the ratio of the number of identical bases within the sliding window to the window length; the window mismatch rate is calculated as the ratio of the number of mismatched bases within the window to the window length; the preset similarity threshold is 0.75-0.90, preferably 0.80; the preset window mismatch rate threshold is 0.10-0.25, preferably 0.20; the preset proportional coefficient is obtained by regression fitting of gradient-spiked amplification data through window similarity-measured amplification efficiency; row normalization is performed so that the sum of the elements in each row of the competition network matrix is ​​1.

[0017] By adopting the above technical solution, a sliding window is set with a length of 80%-200% of the primer sequence. The window similarity and window mismatch rate are calculated base by base. When the window similarity is higher than the preset threshold and the window mismatch rate is lower than the preset threshold, the product of the window similarity and the preset proportional coefficient obtained by regression fitting of the gradient-spiked amplification data is used as the initial value of the competition inhibition coefficient. Otherwise, it is set to zero. The competition network matrix is ​​formed according to the target pairs. The absolute value of the competition inhibition coefficient is taken for the off-diagonal elements and the diagonal elements are set to zero and normalized by row. The similarity of the primer binding region can be transformed into a quantifiable competition relationship, highlighting the main competitive edges and suppressing irrelevant edges, providing a reliable input for connected component grouping and competition correction.

[0018] Preferably, in step S6, the Markov chain Monte Carlo posterior sampling iteratively samples the joint posterior distribution of the endogenous template molecule number vector and the competition inhibition coefficient using the Metropolis-Hastings algorithm or the Gibbs sampler; the endogenous template molecule number is updated using Gibbs or the log-normal random walk Metropolis-Hastings; the competition inhibition coefficient is updated using the normal random walk Metropolis-Hastings, with a proposed distribution standard deviation of 0.01-0.1; the number of iterations is 10,000-50,000, the burn-in is 20%-50%, and the thin is 5-10; the acceptance rate is 20%-40%; the convergence criterion is that the Euclidean distance between two adjacent sampling endogenous template molecule number vectors is less than 1% of their mean, or the weighted residual norm is less than 0.05.

[0019] By adopting the above technical solution, the targets are grouped according to the connected components of the competitive network matrix, and a competitive saturated response equation is established within each group. Under non-negative constraints, the number of observed molecules in the unspiked and spiked target portions is used as the fitting target. The known amount of synthetic template added to the spiked portion is included as a known term in the competitive saturated response equation. The second-level composite quality control correction factor is used as the residual weighting coefficient. A Laplace prior is applied to the competition inhibition coefficient, and a Gamma prior is applied to the number of endogenous template molecules. The joint posterior distribution is solved iteratively using the Metropolis-Hastings algorithm or Gibbs sampler. At the same time, an inter-group coupling term is introduced across target groups for global correction, which can effectively suppress overfitting. The number of endogenous template molecules and the competition inhibition coefficient are jointly estimated, and intra-group competition and inter-group resource sharing or batch effects are corrected to obtain a more accurate competitive correction copy number.

[0020] Preferably, in step S7, the ratio of the competing correction copy number to the known copy number of the internal reference target in each fraction is used as the response factor for each fraction, and the mean of the response factors for each fraction is used as the global correction coefficient. If the response factor of a certain fraction deviates from the mean by ±30%, it is removed and recalculated. The quantitative value of the absolute copy number of each target is obtained by dividing the competing correction copy number of each target by the global correction coefficient. The spike recovery rate and the linear regression determination coefficient are calculated from the observed number of target molecules in each spiked fraction and the estimated value obtained according to the competitive saturation response equation.

[0021] By adopting the above technical solution, after removing the fractions that deviate from the mean by ±30%, the calculation is recalculated, and the competitive correction copy number of each target is divided by the global correction coefficient, which can eliminate the overall response differences between different fractions and batches and output the quantitative value of the absolute copy number of each target. Furthermore, the acceptance criteria are that the recovery rate of each spiked fraction is in the range of 70%-130% and the linear regression determination coefficient is ≥0.90.

[0022] Preferably, for the low-quality targets removed in S5, their original read clusters, the number of observed molecules of the target, and the value of the second-level composite quality control correction factor are retained and output along with the quantitative value of the absolute copy number of each target; the values ​​of low-quality targets are not involved in the fitting of S6 and the calculation of the global correction coefficient in S7, but are only stored as reference information in the output results.

[0023] By adopting the above technical solution, the original read clusters, the number of observed molecules of the target, and the value of the second-level composite quality control correction factor of the low-quality targets removed in S5 are retained and output along with the quantitative value of the absolute copy number of each target. At the same time, they are not included in the fitting of S6 and the calculation of the global correction coefficient in S7. This can ensure that the main quantitative model is not affected by low-quality data, while retaining the traceable reference information of low-quality targets, which is convenient for subsequent verification and method optimization.

[0024] The beneficial effects of this invention are: 1. This invention reduces the impact of PCR repeat counting, sequencing errors, primer mismatches, and strand bias on molecule counting by using target-specific primers with unique molecular identifiers, original read cluster extraction, two-stage composite quality control correction, and target read retention rate compensation, thereby improving the accuracy of target molecule count.

[0025] 2. This invention corrects for inter-target competition and inter-group coupling in multiplex amplification by using equal-splitting spiking, internal reference targets, log-interval gradient synthesis templates, competitive network matrices, competitive saturation response equations, sparse Bayesian priors, and MCMC combined posterior sampling, thereby improving the accuracy and consistency of absolute quantification of multi-target copy numbers.

[0026] 3. This invention achieves closed-loop quality control from amplification and sequencing to quantitative output by global recalibration of internal reference targets, removal of abnormal fractions, and determination of spiked recovery rate and linear regression coefficient of determination, thereby improving the robustness, comparability and feasibility of quantitative results.

[0027] 4. This invention balances the robustness of the quantitative model, the traceability of results, and the need for automated analysis by retaining and outputting low-quality target reference information and providing quantitative success or failure indicators. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the tNGS multi-target copy number absolute quantitative analysis calculation method provided by the embodiments of the present invention; Figure 2 This is a schematic diagram of the target-specific primer structure and amplification label with unique molecular identifier for the tNGS multi-target copy number absolute quantitative analysis calculation method provided by the embodiments of the present invention. Figure 3 This is a schematic diagram of the original read cluster extraction process of the tNGS multi-target copy number absolute quantitative analysis calculation method provided by the embodiments of the present invention; Figure 4 This is a schematic diagram of the generation of the first-level composite quality control correction factor and the calculation of target read retention rate in the tNGS multi-target copy number absolute quantitative analysis method provided by the embodiments of the present invention. Figure 5 This is a schematic diagram of UMI single-molecule family clustering and target observation molecule number calculation of the tNGS multi-target copy number absolute quantitative analysis calculation method provided in the embodiments of the present invention; Figure 6 This is a schematic diagram of equal-part spiking and internal reference target addition in the tNGS multi-target copy number absolute quantitative analysis calculation method provided by the embodiments of the present invention; Figure 7 This is a schematic diagram of the generation of the second-level composite quality control correction factor and the removal of low-quality targets in the tNGS multi-target copy number absolute quantitative analysis calculation method provided in the embodiments of the present invention. Figure 8 This is a schematic diagram of the competitive network matrix construction and connected component grouping of the tNGS multi-target copy number absolute quantitative analysis calculation method provided in the embodiments of the present invention; Figure 9 This is a schematic diagram of the connected component grouping and the establishment of the competitive saturation response equation within the group in the tNGS multi-target copy number absolute quantitative analysis calculation method provided by the embodiments of the present invention. Figure 10 This is a schematic diagram of the competitive saturated response equation and MCMC joint solution process of the tNGS multi-target copy number absolute quantitative analysis calculation method provided by the embodiments of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Reference Figures 1-10The tNGS method for absolute quantitative analysis of multi-target copy number uses target-specific primers with unique molecular identifiers to amplify and sequence sample nucleic acids multiplex, extracting raw read clusters for each target. After filtering the read quality, homologous molecule counts of the targets are obtained by merging them according to the unique molecular identifier. An internal reference target with a known copy number is added to the sample, and quantitative calibration is completed by amplification and sequencing. This includes extracting raw read clusters for each target, generating a first-level composite quality control correction factor and calculating the target read retention rate, performing UMI single-molecule family clustering to obtain the number of observed target molecules, obtaining multi-gradient observation data through equal-splitting and internal reference targets, generating a second-level composite quality control correction factor and removing low-quality targets, establishing a competition network matrix and a competition saturation response equation, jointly solving for the number of endogenous template molecules and the competition inhibition coefficient to obtain the competition correction copy number, and performing global recalibration using the internal reference target to output the absolute copy number quantification value and provide a judgment on the quantitative validity. Reference Figure 2 S1. Target-specific primers with a unique molecular identifier at the 5' end are used for multiplex amplification and sequencing of sample nucleic acids. Original read clusters of each target are extracted according to the primer binding sites. A 1-5 base spacer sequence is placed between the unique molecular identifier and the target-specific sequence. The unique molecular identifier is an incompletely random sequence of 8-16 nt, with a limit of no more than 3 consecutive identical bases. The GC content is 40%-60%, and the Hamming distance between different unique molecular identifiers is not less than 3. Specifically, the unique molecular identifier is placed at the 5' end of the target-specific primer. In another embodiment, it can also be introduced by ligation or transposition. During the first round of extension, the unique molecular identifier is incorporated into the amplification product along with the primer.

[0032] Its working mechanism is as follows: In the reverse transcription or first-round extension stage, each original template molecule is randomly labeled with a unique molecular identifier. The repetitive reads generated by subsequent PCR amplification will carry the same unique molecular identifier, thereby distinguishing the original template molecule from the PCR repeats in the data analysis stage and reducing the counting bias caused by amplification repeats.

[0033] Multiplex amplification refers to adding multiple pairs of target-specific primers to the same amplification system, enabling the simultaneous enrichment of multiple targets in the same reaction. The machine-readable reads are first aligned with a pre-defined target primer sequence library, recording the start and end sites of the primer binding regions, and then extracting read fragments. The purpose of this alignment is to determine which target each read belongs to and to remove uncertain sequences outside the primer regions.

[0034] Reference Figure 3The machine-generated reads are compared with a pre-defined target primer sequence library. The start and end sites of the primer binding regions are recorded, and read fragments are extracted. Reads with the same target identifier and whose fragment lengths conform to the amplicon length range are grouped into original read clusters. The amplicon length range can be set according to the expected target fragment length during primer design, for example, by adding or subtracting a certain number of bases from the expected amplicon length. Thus, the original read clusters contain both target identity information and unique molecular identifier information, providing a basis for subsequent homology grouping.

[0035] Specifically, the process is divided into the following steps: First, sample nucleic acid preparation: The sample can be DNA or RNA. If it is RNA, reverse transcription should be performed first, with the 5' end of the reverse transcription primer or target-specific primer carrying UMI. The amount of nucleic acid input can be 1-100 ng, preferably 5-50 ng. The sample can be derived from pathogen nucleic acid, genomic DNA, cfDNA, FFPEDNA / RNA, etc.

[0036] Second, UMI design: The UMI is positioned at the 5′ end of the target-specific primer. The UMI length can be 8-16 nt, and the UMI can be composed of random N bases, with N being randomly incorporated with equal probability of A / T / C / G. To reduce errors and homopolymerization within the UMI itself, it can be designed as a partially random sequence. A 1-5 base spacer sequence, such as TT or AC, can be placed between the UMI and the target-specific sequence to assist in identifying the UMI boundary.

[0037] The unique molecular identifier is preferably 12nt in length, composed of incompletely random bases, with a limit of no more than 3 consecutive identical bases, a GC content of 40%-60%, and a Hamming distance of not less than 3 between different unique molecular identifiers.

[0038] Third, target-specific primer design: The target-specific sequence length can be 18-30 nt, preferably 20-25 nt; the Tm can be 55-65℃, preferably 58-62℃; the amplicon length can be 60-200 bp, preferably 80-150 bp. The concentration of each primer in the multiplex primer pool can be 50-200 nM, preferably 80-120 nM. Primer design should avoid primer dimers, hairpin structures, and highly homologous regions between targets as much as possible.

[0039] Fourth, multiplex amplification: Multiple target-specific primers with a UMI at the 5' end are added to the same reaction system to amplify the sample nucleic acid in multiplexes. During the first extension, the UMI is incorporated into the amplification product along with the primers, ensuring that subsequent PCR repeat reads generated from the same original template molecule carry the same UMI. The number of amplification cycles can be 18-28, preferably 20-24, to avoid amplification saturation and template repeat count bias. After purification, the amplification product is used to construct a library, add sequencing adapters and sample indexes, and sequence using the Sikun 2000 platform in SE50 or SE75 mode.

[0040] Fifth, read alignment and extraction of original read clusters: The data is first split according to the sample index to obtain FASTQ files. The reads are then aligned with a pre-defined target primer sequence library or a reference target sequence. Alignment tools such as BWA-MEM, Bowtie2, and Minimap2 can be used, allowing 0-2 mismatches in the primer region. After alignment, the target identifier, primer binding region start site, and end site corresponding to each read are recorded, and read fragments between primer binding regions are extracted.

[0041] If the original reads are grouped into clusters based on target identifier and fragment length, and their UMI information is recorded, then single-molecule family clustering is performed based on UMI in S3. Further information such as forward and reverse strand information, base quality, and window mismatch rate can be recorded within the original read clusters.

[0042] Reference Figure 4 S2. After normalizing the average base quality of the read segment, the window mismatch rate, and the ratio of forward to reverse strand read segments, the values ​​are weighted and fused to generate the first-level composite quality control correction factor. Read segments below the judgment limit are removed to obtain the effective read segment cluster and the target read segment retention rate.

[0043] In this embodiment, the mean base quality is linearly scaled to 0-1 according to the quality value range. The sequencing quality value is usually the Phred quality value, and its working mechanism is Q=-10·log 10 P, where P is the probability of sequenced errors, and a higher Q indicates a lower error probability. The window mismatch rate is normalized after rounding up to the upper limit of the number of mismatched bases. That is, the fewer the number of mismatched bases, the higher the normalized value; rounding up aims to make higher quality values ​​larger, facilitating uniform weighting. The normalization of the number of mismatched bases in the primer region is as follows: Q norm =clip((Q mean -Q min ) / (Q max -Q min ),0,1), Q min =20, Q max =40.

[0044] Calculate the mean base quality Q for each read segment or read cluster. mean The relationship between the Phred quality value Q and the base error probability P is Q = -10·log 10 P. Q mean Linear scaling to 0-1, for example: Q norm =(Q mean -Q min ) / (Q max -Q min ), And truncate to [0,1].

[0045] Q min It can be 20, Q max It can be 40.

[0046] Alternatively, the following can be used: Q norm =1-10 (-Qmean / 10) , As the quality value after error probability correction.

[0047] The forward and reverse read ratios are normalized to their upper limit. This ratio measures the balance of double-stranded template amplification; deviations from the expected ratio may indicate single-strand bias or amplification anomalies. To normalize the window mismatch rate: M norm =1−min(m,m max ) / m max , Where m is the number of mismatched bases within the window, m max =5.

[0048] If m is the mismatch rate, then m max It should be 0.2 or 0.25.

[0049] The three types of parameters are summed according to preset weights, with the weights determined by the absolute value of the Pearson correlation coefficient between each parameter and the target read retention rate. In this embodiment, the absolute value of the Pearson correlation coefficient or Spearman correlation coefficient between each parameter and the target read retention rate can be calculated. Parameters with positive correlation and larger absolute values ​​are assigned higher weights, and normalization is performed so that the sum of the weights is 1. The first-level composite quality control correction factor can be expressed as: F1 = w1 × Q norm +w2×M norm +w3×S norm , Q norm M is the normalized value of the base mass mean. norm S is the complemented normalized value for the window mismatch rate. norm This represents the normalized ratio of read segments between the forward and reverse chains, where w1, w2, and w3 are weights. Normalization of the forward and reverse chain read ratio: For each original read cluster or target, count the forward chain reads F and the reverse chain reads R. Calculate the chain balance ratio: S raw =min(F,R) / max(F,R), If max(F,R)=0, then S raw =0.

[0050] S rawThe closer the ratio is to 1, the more balanced the amplification of the forward and reverse strands. It can be further normalized to the upper limit of the ratio, for example, S. norm =S raw .

[0051] Take the ratio of the smaller value of the forward and reverse link read segments to the larger value.

[0052] If max(F,R)=0, then S norm =0.

[0053] Reads below the rejection threshold are discarded, and the retained reads are classified as valid read clusters. The target read retention rate can be defined as the ratio of the number of valid read clusters to the number of original read clusters, reflecting the proportion of target reads lost during quality control. In subsequent step S3, the target is corrected based on this target read retention rate to compensate for the loss of molecules caused by quality control rejections.

[0054] The above three types of parameters are weighted and combined: For example, when Q norm When the correlation with retention rate is highest, we can set w1=0.5, w2=0.3, w3=0.2: w i =∣ρ i ∣ / ∑ j ∣ρ j |, If ρ i If <0, then let w i =0 and renormalize, and w1+w2+w3=1. The lower limit of judgment is F1≥0.4. For each target i, define the target read retention rate: ri=N eff,i / N raw,i , Where N eff,i N represents the number of effective read clusters for this target. raw,i This represents the original number of read segment clusters. If r i Too low, for example, r i If the value is less than 0.05, the target can be marked as a quality control risk target. It can still be calculated later, but a warning should be issued.

[0055] The weights w1, w2, and w3 can be determined based on the absolute value of the Pearson correlation coefficient or the Spearman correlation coefficient between each parameter and the target read retention rate. If Q norm The highest correlation with retention rate can be set as w1=0.5, w2=0.3, w3=0.2, and w1+w2+w3=1.

[0056] A lower threshold is set: read segments with F1 ≥ 0.4 are retained, while read segments below this threshold are discarded. The retained read segments are classified into a valid read segment cluster.

[0057] Reference Figure 5 S3. Construct a read similarity map based on unique molecular identifiers, cluster to obtain single-molecule families, and use the number of single-molecule families as the target single-molecule count, divided by the target read retention rate to obtain the target observed molecule count. The working mechanism is as follows: the single-molecule family count represents the actual number of original endogenous template molecules observed after merging by unique molecular identifiers; since S2 quality control removed some reads, the observed value is lower than the true value, therefore, the target read retention rate is used for back-correction to obtain the target observed molecule count. This target observed molecule count is used as the observation quantity in subsequent S4 and S6.

[0058] In this embodiment, the number of single-molecule families is used as the target single-molecule count. Since the S2 quality control removes some reads, the target read retention rate needs to be used for correction. O i =M i / r i , O i This represents the number of target molecules observed for target i, which will be used in subsequent S4 and S6.

[0059] Valid reads with the same target identifier and identical UMI are designated as nodes. If the base editing distance d between two reads does not exceed a preset threshold (e.g., d≤5), an edge is established between the two nodes. The edge weight can be set as follows: w=exp(-d / σ), Where σ can be 1-3.

[0060] Edit distance can be calculated using Levenshtein distance. To reduce computational complexity, groups can be created first by UMI exact match, and then edit distance can be calculated within each group.

[0061] In this implementation, a similarity graph is constructed using valid reads identified by the same target as nodes, with UMI correction distance and read edit distance as edge weights, and base edit distance between reads as edge weights. Edit distance, also known as Levenshtein distance, refers to the minimum number of insertion, deletion, or substitution operations required to convert one sequence into another; the smaller the edit distance, the more similar the two reads are. For reads within each single-molecule family, variational inference can be used to generate consensus sequences. Variational inference approximates the true posterior distribution with an optimizable approximate distribution, inferring the most likely bases site-by-site, thus obtaining a consensus sequence even in the presence of sequencing errors. Alternatively, a simple majority voting method can be used to generate consensus sequences.

[0062] A graph attention network is used to aggregate neighborhood features and classify single-molecule families based on a similarity threshold of the aggregated features. The working mechanism of the graph attention network is as follows: when aggregating neighbor information, each node automatically learns the importance of different neighbors through attention coefficients, thus grouping reads from the same original template molecule together even in the presence of sequencing errors or minor differences. Consistent sequences are generated for reads within a family using variational inference. The working mechanism of variational inference is to approximate the true posterior distribution with an optimizable approximate distribution, thereby inferring the consensus sequence that best represents the family when random errors exist in reads within the family.

[0063] The graph attention network is trained using a self-supervised learning approach. Specifically, connections between read pairs within the same UMI group are used as positive samples, while read pairs between different UMI groups are used as negative samples. The loss function is either contrastive loss or InfoNCE loss. The training epochs are 100-500, with a learning rate of 0.001-0.01. An early stopping strategy (patience=20) is used to prevent overfitting. The convergence criterion is that the validation set loss no longer decreases within 20 consecutive epochs. When the total number of valid reads is less than 1000, the graph attention network can be skipped, and edit distance connected component clustering can be used directly.

[0064] Node features may include one-hot encoding of bases, base quality, read position, chain direction, etc.

[0065] Edge features include UMI correction distance, inter-read base editing distance, and the number of mismatched bases in the primer region; node features include UMI sequence encoding, strand orientation, and quality value. The edge weights are: w=exp(−(αd edit +βd H +γm) / σ), Where d edit For base editing distance, d H denoted as UMI correction distance, m as window mismatch rate, α=0.5, β=0.3, γ=0.2, σ=1−3.

[0066] A graph attention network is used to aggregate neighborhood features. The graph attention network can have 1-3 layers, with 4-8 attention heads per layer and a hidden dimension of 32-128. The importance of different neighbors is learned through attention coefficients, and aggregated features are output. Single-molecule families are divided according to a similarity threshold of aggregated features, which can be 0.80-0.90, preferably 0.85.

[0067] The loss function is: L=L contrast +λ1L UMI +λ2L primer +λ3L size +λ4L strand , Where L contrast To compare the losses, L UMI For the UMI correction distance regularization term, L primer L is the primer mismatch regularization term. size For the family size sparse regularization term, L strand The chain balance regularization term λ1−λ4 is 0.1-1.0. The training epochs are 100-500, the learning rate is 0.001-0.01, and an early stopping strategy is employed. The similarity threshold is preferably 0.85. Number of target observed molecules: O i =M i / r i , In another implementation, UMI clustering, besides graph attention networks, can employ UMI-tools directed adjacency clustering, Starcode clustering, greedy edit distance clustering, or connected component clustering. Families with fewer than two reads within a single molecule family can be discarded or merged into the nearest family.

[0068] Reference Figure 6 S4. Divide the sample nucleic acid into several equal portions. Add the same known copy number of the internal control target. Add the synthetic template of each target at logarithmic intervals covering 0.25-4 times the estimated number of endogenous template molecules. Set up unlabeled equal portions. Amplify and sequence under the same conditions to obtain the observed number of target molecules in each portion. Specifically, divide the sample nucleic acid into several equal portions, 4-8 portions, preferably 6 portions. The volume and nucleic acid content of each portion should be as consistent as possible.

[0069] The internal reference target undergoes the same read cluster extraction, UMI merging, quality control, and competitive correction as the endogenous target to obtain its competitive correction copy number T. ref comp .

[0070] In this embodiment, the sample nucleic acid is divided into several equal portions, and each portion contains an internal reference target with the same known copy number. The internal reference target is an exogenous synthetic nucleic acid target, whose sequence is different from or distinguishable from the sample's endogenous target, and whose copy number is known, for subsequent S7 calculation of the global correction coefficient. Each portion contains an internal reference target with the same known copy number. The internal reference target is an exogenous synthetic nucleic acid, such as synthetic DNA or in vitro transcribed RNA, whose sequence is distinguishable from the endogenous target. The amount of internal reference target added can be 10... 3 -10 6 copies / μL, preferably 10 4 Add 1 μL per copy / μL.

[0071] Specifically, for each target to be quantified, unspecified aliquots and several spiked aliquots are prepared. The spiking amount covers 0.25-4 times the estimated number of endogenous template molecules for that target, using logarithmic interval gradients, such as 0.25×, 0.5×, 1×, 2×, 4×. If the estimated number of endogenous template molecules is T... est Then add 0.25T to each equal portion. est 0.5T est T est 2T est 4T est For the target synthesis template, when the number of equal parts is 8, intermediate gradient points can be added in the range of 0.25× to 4×, such as 0.35×, 1.5×, 3×, etc., to match the number of gradient points with the number of equal parts. Unlabeled equal parts are not included in the target synthesis template.

[0072] The preferred number of equal parts is 6; the amount of internal control target added is 10g per part. 4 copies; the logarithmic interval gradient is 0.25×, 0.5×, 1×, 2×, 4×; if the number of equal parts is 8, intermediate gradient points such as 0.35×, 1.5×, 3× can be added.

[0073] Each aliquot should be amplified, library constructed, and sequenced using the same multiplex amplification system, primer pool, polymerase, dNTPs, buffer, and cycling conditions, except for the different spiking amounts.

[0074] The internal control target was an exogenously synthesized nucleic acid target; each aliquot was amplified, library constructed, and sequenced using the same multiplex amplification system, primer pool, and cycling conditions. "Same conditions" means that, except for the amount of target synthesis template added, the polymerase, primer concentration, dNTPs, buffer, and cycle number were kept as consistent as possible to minimize the impact of batch effects on the quantitative results.

[0075] After sequencing, each aliquot was processed according to S1 to S3 to obtain the number of target molecules observed for each target in each aliquot. i (k) , where k represents the equivalence number. The number of observed molecules on the target includes contributions from both the number of endogenous template molecules and the contribution from the target synthetic template with known copy numbers, providing data for the joint solution of S6.

[0076] Reference Figure 7 S5. After normalizing and weighting the amplicon coverage uniformity, target site base mismatch ratio, and target read retention rate, a second-level composite quality control correction factor is generated. Targets below the judgment threshold are removed as low-quality targets. For each target, the sequencing depth at each position in the amplicon region is statistically analyzed, and the depth standard deviation σ is calculated. depth and mean μ depth Amplicon coverage uniformity can be determined using the coefficient of variation: CV = σ depth / μ depth express.

[0077] The smaller the CV, the more uniform the distribution. Normalization can be expressed as: U norm =1-min(CV,CV max ) / CV max , CV max It can be 1.0 or 1.5.

[0078] In this embodiment, amplicon coverage uniformity is calculated by dividing the standard deviation of sequencing depth by the mean; that is, the coefficient of variation (CV) = σ / μ. A smaller value indicates more uniform sequencing depth at each position of the amplicon; a larger value indicates uneven coverage, potentially indicating amplification bias. The proportion D of mismatched reads at the target site to the total reads is calculated. mis A higher proportion of base mismatches at the target site indicates a more severe sequencing error or nonspecific amplification. Normalization can be expressed as: D norm =1-min(D mis D max ) / D max , Dmax can be 0.2 or 0.3.

[0079] The target site mismatch ratio is calculated as the proportion of mismatched reads to the total reads; a higher target site mismatch ratio indicates a more severe sequencing error or nonspecific amplification at that target site. The target read retention rate follows the results from S2; a lower retention rate indicates greater loss of that target during quality control. The r value obtained from S2 is also used. i Normalization can be expressed as R. norm =min(r i ,1).

[0080] After normalizing the three parameters mentioned above, a weighted fusion is performed, with the weight coefficient determined based on the positive correlation portion of the Spearman correlation coefficient between each parameter and the repeatability of the target quantitative quantification. For example, multiple repeated spiking experiments can be conducted to calculate the Spearman correlation coefficient between each parameter and the repeatability of the target absolute copy number quantification value, assigning higher weights to parameters with a large positive correlation.

[0081] The weighting coefficients of the second-level composite quality control correction factor are determined based on the positive correlation component of the Spearman correlation coefficient between each parameter and the quantitative repeatability of the target; the second-level composite quality control correction factor can be expressed as: F2=v1×U norm +v2×D norm +v3×R norm , U norm D is the normalized value for amplicon coverage uniformity. norm R is the normalized value of the base mismatch ratio at the target site.norm is the normalized value of target read retention rate, with v1, v2, and v3 as weights. Targets with F2 below the decision threshold are judged as low-quality targets and do not proceed to S6 fitting and S7 recalibration.

[0082] First, calculate the Spearman correlation coefficient between each parameter and the target read retention rate; only retain the positive correlation coefficient; divide the positive correlation coefficient of a parameter by the sum of the positive correlation coefficients of all parameters to obtain the weight of that parameter; set the weight of parameters with a correlation coefficient of zero or negative to zero; if the correlation coefficients of all parameters are zero or negative, then the weights are evenly distributed or the preset default weights are used, and the sum of the weights is 1.

[0083] The weights w1, w2, and w3 can be determined based on the positive correlation portion of the Spearman correlation coefficient between each parameter and the target read retention rate. For example, the absolute value of the Pearson correlation coefficient between each parameter and absolute copy number repeatability is calculated through multiple repeated spiking experiments. Parameters with a positive correlation and a large absolute value are assigned higher weights, and v1+v2+v3=1.

[0084] Set a lower limit for judgment. For example, targets with F2≥0.3 are included in S6 fitting; targets with F2<0.3 are judged as low-quality targets and are not included in S6 fitting and S7 global correction coefficient calculation. However, their original read clusters, number of observed molecules, and F2 value can be retained and output with the final result as reference information.

[0085] In this embodiment, the original read clusters, the number of observed molecules and the second-level composite quality control correction factor values ​​of the low-quality targets removed in S5 can be retained and output along with the quantitative values ​​of the absolute copy number of each target. The values ​​of low-quality targets do not participate in the fitting in S6 and the calculation of the global correction coefficient in S7, but are only stored as reference information in the output results. This avoids low-quality targets interfering with the quantitative model and retains traceable information.

[0086] Reference Figure 8 Low-quality targets removed by S5 (S6 and S5) do not participate in the construction of the competitive network matrix; their corresponding rows and columns are deleted from the competitive network matrix. Connected component grouping is only performed on targets that pass the S5 quality control. A sliding window is set at 80%-200% of the primer length, and window similarity (number of identical bases / window length) and window mismatch rate (number of mismatched bases / window length) are calculated. Specifically, the sliding window works as follows: a fixed-length window is moved base-by-base along the primer-template binding region, and the ratio of identical bases to the window length within each window is calculated, i.e., window similarity. The window mismatch rate can be calculated as the ratio of mismatched bases to the window length. The sliding window length Lw is set at 80%-200% of the primer sequence length, preferably 100%-150%. The window is moved base-by-base along the target primer-template binding region, with a step size of 1-5 nt. For each pair of targets i and j, Calculate window similarity: S ij = Number of identical bases within the window / Lw; Window mismatch rate: e ij =Number of mismatched bases within the window / Lw.

[0087] The window similarity of the target primer and template binding region is compared. When the window similarity is higher than a preset threshold and the window mismatch rate is lower than a preset threshold, the initial value of the competition inhibition coefficient of the target pair is set to the product of the window similarity and a preset proportional coefficient; otherwise, it is set to zero. The working mechanism is as follows: the more similar the primer and template binding regions are, the more likely different target primers are to compete for the same template or share amplification resources, thus generating competition inhibition; if the similarity is low or the window mismatch rate is high, the competition can be ignored, and the initial value of the coefficient is set to zero. The preset similarity threshold S... th For example, 0.75-0.90, preferably 0.80; preset window mismatch rate threshold e th For example, 0.10-0.25, preferably 0.20.

[0088] When S ij ≥S th And e ij ≤e th Initial value of the competition inhibition coefficient: c ij 0 =S ij ×β; Otherwise c ij 0 =0.

[0089] Where β is a preset proportionality coefficient, which can be obtained by regression fitting of gradient-spiked amplification data through window similarity-measured amplification efficiency. For example, using the window similarity of different target pairs as the independent variable and the percentage decrease in measured amplification efficiency as the dependent variable, a linear regression is performed, and the regression slope is β.

[0090] Specifically, the fitting method for the preset proportionality coefficient β is as follows: using the number of target observation molecules in each gradient spiked equally divided in S4 as the measured value, and the number of target observation molecules in the unspiked equally divided as the benchmark, the measured amplification efficiency reduction ratio Δη between each target pair is calculated; using the window similarity Sij of each target pair calculated by the sliding window method in S6 as the independent variable, and the corresponding Δη as the dependent variable, linear regression is performed using the least squares method. Δη=β×S ij +ε, The regression slope is β. This fitting process is completed after obtaining the gradient-spiked data in S4 and before constructing the competition network matrix in S6. By fitting β using gradient-spiked data within the same experimental system, the influence of differences in amplification efficiency between different experimental systems on the initial estimation of the competition inhibition coefficient can be eliminated, thereby improving the accuracy and portability of the competition network matrix.

[0091] The competition suppression coefficients between targets are arranged in target pairs to form a competition network matrix C. Off-diagonal elements represent the absolute values ​​of the competition suppression coefficients for their respective target pairs, while diagonal elements are set to zero. Row-normalization is applied. The absolute value represents the competition intensity, regardless of direction; setting diagonal elements to zero indicates that a target does not compete with itself; row-normalization ensures the sum of all elements in the competition network matrix is ​​1, thus representing the competitive impact on each target as a relative weight. Off-diagonal elements represent the absolute values ​​of the competition suppression coefficients for their respective target pairs, while diagonal elements are set to zero. Row-normalization: c ij' =|c ij 0 | / Σ k |c ik 0 |, distinguishing the initial value c ij 0 With normalized matrix element c ik 0 .

[0092] Make the sum of the elements in each row equal to 1.

[0093] Targets are grouped according to the connected components of the competition network matrix. A connected component refers to a set of mutually reachable nodes in an undirected graph; targets within the same connected component have direct or indirect competitive relationships.

[0094] Targets are grouped according to the connected components of the competition network matrix. A connected component is a graph theory concept, referring to a set of mutually reachable nodes in an undirected graph. Targets within the same connected component have direct or indirect competitive relationships, while the competitive relationships between different connected components are weaker.

[0095] Reference Figure 9 Within each group, a competitive saturation response equation is established: the number of target observed molecules equals the product of the intrinsic amplification response coefficient and the number of endogenous template molecules divided by the competitive inhibition term. The competitive inhibition term is 1 plus the sum of the products of the intra-group competitive inhibition coefficient and the vector of the number of endogenous template molecules, which can be expressed as: O i = (A i ×T i ) / (1+∑ ij c ij ×T j )+ε i The correction equation for the intergroup coupling term is: O i = (A i×T i ) / (1+∑ j c ij T j +∑ m d im ×T m )+εi, where O i Let A be the number of target molecules observed for target i. i T represents the intrinsic amplification response coefficient of target i. i c is the number of endogenous template molecules for target i. ij T represents the competitive inhibition coefficient of target j against target i within the group. j c is an element of target j in the endogenous template molecule number vector. ii =0, ε i It represents the residual.

[0096] The competitive saturation response equation is established based on the following: In a multiplex amplification system, when multiple targets share the same primer pool, polymerase, and dNTP resources, the amplification efficiency between targets is competitively inhibited by template abundance. When T... j Much less than 1 / c ij When the equation degenerates into a linear response O, the equation degenerates into a linear response O. i ≈A i ×T i When T j Much greater than 1 / c ij When the amplification saturates, the equation tends to saturate. This saturation characteristic allows the equation to fit both low-abundance and high-abundance targets simultaneously over a wide dynamic range, avoiding quantitative bias caused by amplification saturation.

[0097] "1+∑" in the denominator j≠i c ij ×T j "That is, the competitive inhibition term. Its working mechanism is as follows: the larger the number of endogenous template molecules of the competing target, the stronger the competitive inhibition, and the relatively reduced observed response of target i, which is saturation or inhibition.

[0098] In another implementation, the competitive saturated response equation can be derived from variational Bayesian, maximum a posteriori estimation, or nonnegative least squares combined with sparse regularization, in addition to MCMC.

[0099] Under non-negative constraints, the number of target molecules observed in unspiked and spiked equal portions is used as the fitting target; the known amount of target synthetic template added in the spiked equal portions is included as a known term in the competitive saturation response equation, the second-level composite quality control correction factor is used as the residual weighting coefficient, and the competition inhibition coefficient is applied to the sparse Bayesian learning framework to make the irrelevant competition edge coefficients tend to zero, thereby avoiding overfitting.

[0100] Residual standard deviation σ i It can be set to 1 / F2i F2 i The higher the value, the greater the weight of the target in the likelihood.

[0101] A sparse Bayesian learning framework is used to impose a prior on the competition suppression coefficient, causing the coefficients of irrelevant competing edges to approach zero. The prior can be a Laplace prior, a horseshoe prior, or a spike-and-slab prior. For example: c ij ~Laplace(0,b), where b can be 0.1-1.0; T i ~Gamma(α,β), where α can be 2 and β can be 1, and is guaranteed to be non-negative.

[0102] The joint posterior distribution can be written as: p(T,c|O)∝Π iN (O i |A i T i / (1+Σ jc ij T j ),σ i 2 )×Π iGamma (T i |α,β)×Π {i,j} Laplace(c ij |0,b).

[0103] This paper proposes a Markov chain Monte Carlo posterior sampling method to iteratively solve for the joint posterior distribution of the endogenous template molecule number vector and the competition inhibition coefficient. The principle is to construct a Markov chain whose stationary distribution is the target joint posterior distribution, and then approximate this distribution through iterative sampling. The Markov chain Monte Carlo posterior sampling uses the Metropolis-Hastings algorithm or the Gibbs sampler to iteratively sample the joint posterior distribution of the endogenous template molecule number vector and the competition inhibition coefficient. For T... i Use Gibbs update or log-normal random walk (MH) update; for c ij The update uses a normal random walk (MH) method, with a proposed distribution standard deviation of 0.01-0.1. The number of iterations can be 10,000-50,000, the burn-in rate can be 20%-50%, and the thinning rate can be 5-10. The acceptance rate can be controlled at 20%-40%.

[0104] For scenarios with more than 50 targets, the following acceleration strategies can be adopted: Parallel MCMC is used to group different connected components and assign them to different computing nodes for independent sampling; The variational Bayesian approximation of the intragroup competition suppression coefficient is used to replace part of the MCMC sampling step; The total computation time for MCMC is estimated to be 2-8 hours on a standard server (16-core CPU, 64GB memory).

[0105] After each sampling, the weighted residual is calculated using the second-level composite quality control correction factor as the residual weighting coefficient. The convergence criterion can be: the Euclidean distance between the intrinsic template molecule number vectors of two adjacent samplings is less than 1% of their mean, or the weighted residual norm is less than a preset threshold, such as 0.05.

[0106] The reason for using the second-level composite quality control correction factor F2 as the residual weighting coefficient is as follows: F2 comprehensively reflects the amplicon coverage uniformity of the target, the base mismatch ratio of the target site, and the target read retention rate. A higher F2 value indicates a higher reliability of the observed molecule count for the target, and therefore it should be assigned a higher weight in the joint posterior distribution fitting. This makes F2... i Higher-quality targets contribute more to the likelihood function, thus prioritizing the fitting of data from high-quality targets when jointly solving for the endogenous template molecule number vector and the competition inhibition coefficient. This weighting strategy effectively reduces the interference of low-quality targets on the estimation of the competition inhibition coefficient and improves the robustness of solving for the endogenous template molecule number vector.

[0107] A cross-target group inter-group coupling term is introduced for global correction to obtain the competitive correction copy number for each target. The inter-group coupling coefficient is independently estimated based on inter-group window similarity, batch effect, or shared resource competition; the inter-group coupling term is d. im ×T m The working mechanism is as follows: even if the direct competition between different connected components is weak, global competition or batch effects may occur through sharing resources such as primer pools, polymerases, and dNTPs. Therefore, inter-group coupling terms are used to perform low-rank or global corrections on the intra-group equations to improve the consistency of multi-target quantification. The resulting target competition correction copy number can both remove intra-group competition inhibition and compensate for the influence of inter-group coupling. Row normalization ensures that the sum of the elements in each row of the competition network matrix is ​​1.

[0108] O i =A i ×T i / (1+Σ jc ij ×T j +Σ md im ×T m )+ε i , Where d im The inter-group coupling coefficient between targets grouped by different connected components can be estimated by inter-group window similarity, batch effect, or shared resource competition. This yields the competitive correction copy number T for each target. i comp .

[0109] In addition to this implementation, the method is applicable to tNGS multiplex amplification systems with 20-200 targets. For cases with more than 20 targets, the number of connected component groups in the competing network matrix may increase. In this case, the following strategy can be used to control computational complexity: The competitive network matrix is ​​truncated by a threshold, retaining only the edges whose competition suppression coefficient is greater than a preset threshold (e.g., 0.05), and setting the edges below the threshold to zero, thereby sparsifying the competitive network; For large components with more than 10 targets within a connected component, spectral clustering or community detection algorithms are used to further divide them into subgroups, and competitive saturation response equations are established for each subgroup. The number of MCMC sampling iterations increases proportionally to the number of targets. It is recommended to increase the number of iterations by 5000 for every 10 additional targets.

[0110] In another implementation, in addition to low-rank correction, batch effect correction models, ComBat, or graph Laplace regularization can also be used for inter-group coupling terms.

[0111] Reference Figure 10 S7. Using the ratio of the competitive correction copy number to the known copy number of the internal reference target in each segment as the response factor of each segment, abnormal segments that deviate from the mean ±30% are removed. After removal, the mean of the response factor is recalculated as the global correction coefficient. Using the mean of the response factor of each segment as the global correction coefficient, the quantitative value of the absolute copy number of each target is output.

[0112] In this embodiment, the ratio of the competitive correction copy number to the known copy number of the internal reference target in each fraction is used as the response factor for each fraction. The working mechanism is as follows: the internal reference target is an exogenous synthetic nucleic acid target, and its known copy number is determined before amplification. If there is no deviation in the amplification and sequencing process, the competitive correction copy number should be equal to the known copy number. The ratio of the two reflects the overall response efficiency of the fraction from amplification to sequencing, which can reduce the fluctuation of a single fraction.

[0113] Furthermore, the absolute copy number of each target is obtained by dividing the competitive correction copy number of each target by the global correction coefficient.

[0114] The spike recovery rate and linear regression coefficient of determination are calculated from the number of molecules observed in each spiked aliquot and the estimated value obtained according to the competitive saturation response equation. The spike recovery rate can be defined as the measured value divided by the theoretical value or the estimated value. The linear regression coefficient of determination can be the absolute value of the Pearson correlation coefficient or the linear regression coefficient of determination. When the preset acceptance criteria are not met, a quantification failure flag is output. In this embodiment, if the recovery rate exceeds the preset range or the linear regression coefficient of determination is lower than the preset threshold, it indicates that the quantification result is unreliable and a quantification failure flag is output.

[0115] The default acceptance criteria can be: recovery rate within the range of 70%-130%, R² ≥ 0.90. If these criteria are not met, a quantification failure flag will be output.

[0116] Quantitative values ​​of absolute copy number for each target: N i =T i comp / G.

[0117] N i This is the absolute copy number of target i, expressed in units of copies / μL or copies / reaction.

[0118] Taking five endogenous targets (A, B, C, D, E) and one internal control target (R) as an example, the extracted sample DNA was divided into six equal portions. 10 μL of [unspecified substance] was added to each portion. 4 The internal control target R was determined from copies. Target AE was spiked at different gradients: unspiked, 0.25×, 0.5×, 1×, 2×, and 4×. Each aliquot was amplified for 22 cycles using the same multiplex amplification system, followed by SE50 sequencing. The number of observed target molecules in each aliquot was obtained according to steps S1-S3 after sequencing.

[0119] According to S2 calculations, the read retention rates for target AE were 0.92, 0.88, 0.95, 0.60, and 0.90, respectively. The retention rate for target D was low. According to S5 calculations, target D had an F2 value of 0.22, which is below the lower limit of 0.30, thus it was judged as a low-quality target and not included in S6 fitting, but its observed molecule count and F2 were output with the results. Targets A, B, C, and E were included in S6.

[0120] In S6, the sliding window length was set to 1.2 times the primer length, with a step size of 1 nt. The calculated similarity for the AB window was 0.86, with a window mismatch rate of 0.12; the similarity for AC was 0.55, with a window mismatch rate of 0.30; the similarity for BC was 0.82, with a window mismatch rate of 0.15; the similarity of the remaining target pairs was below the threshold. β, after gradient-scaled regression fitting, was 0.35. Therefore, c... AB 0 =0.301, c BC 0 =0.287, the rest are 0. Construct a competitive network matrix and normalize it by row. Connected components are grouped into {A,B,C} and {E}. A competitive saturated response equation is established for {A,B,C}, and T is obtained by sampling 20,000 times using Laplace prior and MCMC, followed by 5,000 burn-in cycles. A T B T C And the competition inhibition coefficient.

[0121] After introducing the intergroup coupling term, the competing correction copy number T is obtained. Acomp T B comp T C comp T E comp .

[0122] In S7, the average ratio of the competing corrected copy number to the known copy number for the internal reference target R in each fraction was G = 0.95. The T of target A... A comp =4800 copies / reaction, then N A =4800 / 0.95≈5053 copies / reaction. The recovery rate was 96%, R0.95 2 =0.98, meeting the acceptance criterion, quantitative analysis successful. Target D outputs a reference value and marks low quality.

[0123] The recovery rate of target B was 92%, R 2 =0.96; the recovery rate of target C was 103%, R 2 =0.97; the recovery rate of target E was 89%, R 2 =0.93. Compared with digital PCR, the linear regression determination coefficient of the quantitative results of target AE in this embodiment is 0.95, and the deviation is within ±15%.

[0124] For each scalar division k: f k =T ref comp,k / C ref , Global correction factor: G=mean(f k ), If a certain equal part f k Targets deviating from the mean by ±30% are removed and recalculated. Absolute copy number of each target: N i =T i comp / G, The acceptance criteria for spiked recovery rate were 70%-130%, and the coefficient of determination for linear regression was ≥0.90.

[0125] In this embodiment, the method is implemented as a computer program product, including: a data input module, a comparison and read cluster extraction module, a UMI clustering module, a first-level quality control module, a spiking design module, a second-level quality control module, a competitive network construction module, a Bayesian solution module, a global correction module, and a result output module. The system may include a processor, a memory, and a display. The memory stores the above program instructions, and the processor executes them to implement steps S1-S7. The output results may include the absolute copy number of each target, confidence interval, recovery rate, and R0. 2 Quality control factors F1 / F2, low-quality target reference values, and indicators of successful or failed quantification.

[0126] The graph attention network uses two layers with four attention heads per layer and a hidden dimension of 64. The loss function weights are λ1=0.5, λ2=0.4, λ3=0.2, and λ4=0.1. It is trained for 200 rounds with a similarity threshold of 0.85.

[0127] MCMC iterations are performed 20,000 times, and burn-in iterations are performed 5,000 times. When there are 20-200 targets, the competitive network matrix threshold is truncated to 0.05, and large component spectral clustering is used. For every additional 10 targets, the MCMC iterations are increased by 5,000 times.

[0128] The working principle of this tNGS multi-target copy number absolute quantitative analysis method: The problem of PCR repeat counting is solved by using the unique molecular identifier of S1 and the original read cluster extraction; By using the first-level composite quality control correction factor of S2 and the target read retention rate, the differences in read quality, mismatch and chain balance are corrected. By using the read segment similarity map, graph attention network, and variational inference of S3, single-molecule family clustering and target observation molecule count calculation are achieved; Templates are synthesized using equal-splitting S4, internal reference targets, and log-interval gradient targets to provide data for competitive fitting and global calibration. The second-level composite quality control correction factor of S5 is used to eliminate low-quality targets and retain reference information. By using the competitive network matrix of S6, connected component grouping, competitive saturated response equation, sparse Bayesian prior and Markov chain Monte Carlo joint posterior sampling, the intrinsic template molecule number vector and competitive inhibition coefficient are solved, and the competitive correction copy number is obtained by introducing inter-group coupling terms. By recalibrating using the global correction coefficient of the S7 internal control target, the absolute copy number quantification values ​​of each target are output, and the quantification validity is judged by the spiked recovery rate and the coefficient of determination of linear regression. Thus, this method forms a closed loop in the stages of multiplex amplification, unique molecular identifier merging, quality control correction, competitive correction, and global calibration, improving the accuracy and feasibility of tNGS multi-target absolute copy number quantification.

[0129] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for absolute quantitative analysis of multi-target copy numbers using tNGS, employing target-specific primers with UMI to perform multiplex amplification and sequencing of sample nucleic acids, merging target single-molecule counts according to UMI, and adding an internal reference target with a known copy number for calibration, characterized by: S1. Target-specific primers with 5' end UMI were used for multiplex amplification and sequencing, and the original read clusters were extracted according to the primer binding sites; S2. Normalize and weight the mean base quality, window mismatch rate, and forward and reverse chain read ratio to generate the first-level composite quality control correction factor, and remove reads below the lower limit to obtain the effective read cluster and target read retention rate. S3. Construct a read segment similarity map according to UMI, cluster to obtain single molecule families, use the number of families as the target single molecule count, divide by the retention rate to obtain the target observed molecule number; S4. The nucleic acid samples were divided into several equal parts. Each part was added with a known copy number of the internal reference target to cover the estimated number of endogenous template molecules. The template was synthesized by adding each target at a logarithmic interval gradient of 0.25-4 times. An unlabeled part was set up. The samples were amplified and sequenced under the same conditions to obtain the number of target molecules observed in each part. S5. Normalize and weight the amplicon coverage uniformity, target site base mismatch ratio, and target read retention rate to generate a second-level composite quality control correction factor. S6. Set a sliding window with primer length of 80%-200%, calculate window similarity and window mismatch rate, and calculate competition inhibition coefficient. The preset proportional coefficient is fitted by regression of gradient-spiked data through window similarity and amplification efficiency; construct a row-normalized competition network matrix, with off-diagonal lines representing the absolute value of coefficients and diagonal lines representing zero; group by connected components, and calculate the number of observed molecules within each group; use the number of observed molecules in each group as the fitting target, the known amount of spiked template as the known term, and the second-level correction factor as the residual weight; under non-negative constraints, solve the joint posterior of the number of endogenous template molecules and competition inhibition coefficient using sparse Bayesian prior and MCMC; introduce inter-group coupling term correction to obtain the competition correction copy number; S7. The ratio of the competitive correction copy number of the internal reference target to the known copy number in each segment is used as the response factor for each segment. Abnormal segments that deviate from the mean by ±30% are removed. After removal, the mean of the response factor is recalculated as the global correction coefficient. The competitive correction copy number of each target is divided by the global correction coefficient to output the absolute copy number quantitative value.

2. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 1, characterized in that, In step S1, a unique molecular identifier is placed at the 5' end of the target-specific primer and incorporated with the primer during the first extension. The machine-readable segment is compared with the preset target primer sequence library, and the start and end sites of the primer binding region are recorded. Read segments with the same target identifier and whose fragment lengths conform to the amplicon length range are grouped into the original read cluster.

3. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 2, characterized in that, In S2, the average base quality is linearly scaled to 0-1 according to the quality value range; the window mismatch rate is normalized after taking the complement of the upper limit of the number of mismatched bases. The ratio of forward to reverse link reads is normalized to the upper limit of the ratio; the three types of parameters are summed according to preset weights, and the weights are determined by the positive correlation of the Spearman correlation coefficient between each parameter and the target read retention rate; the Spearman correlation coefficient between each parameter and the target read retention rate is calculated separately, and only the positive correlation coefficient is used as the basis for weighting; the normalized weight of a parameter is obtained by dividing the positive correlation coefficient of a parameter by the sum of the positive correlation coefficients of all parameters; the weight of parameters with a correlation coefficient of zero or negative is set to zero; if the correlation coefficients of all parameters are zero or negative, the weights of each parameter are evenly distributed or the preset default weights are used, and the sum of the weights is 1.

4. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 3, characterized in that, In S3, a similarity graph is constructed using valid reads identified by the same target as nodes, valid reads with UMI correction distance and read edit distance as edge weights, and base edit distance between reads as edge weights; neighborhood features are aggregated using a graph attention network, and single-molecule families are divided according to the similarity threshold of aggregated features; consistent sequences are generated for reads within a family using variational inference.

5. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 4, characterized in that, In step S4, target synthetic templates with known copy numbers are added at log intervals of 0.25-4 times the estimated number of endogenous template molecules for the target; the internal reference target is an exogenous synthetic nucleic acid target; each aliquot is amplified, library constructed, and sequenced using the same multiplex amplification system, primer pool, and cycling conditions.

6. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 5, characterized in that, In S5, the amplicon coverage uniformity is calculated by dividing the standard deviation of sequencing depth by the mean; the target site base mismatch ratio is calculated by the proportion of mismatched reads to the total reads; the weight coefficient of the second-level composite quality control correction factor is determined by the positive correlation part of the Spearman correlation coefficient between each parameter and the target quantitative repeatability; the Spearman correlation coefficient between each parameter and the target quantitative repeatability is calculated separately, and only the positive correlation coefficient is used as the weight basis; the normalized weight of a parameter is obtained by dividing the positive correlation coefficient of a certain parameter by the sum of the positive correlation coefficients of all parameters; the weight of parameters with a correlation coefficient of zero or negative is set to zero; low-quality targets that are removed are not included in S6 fitting and S7 recalibration.

7. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 6, characterized in that, In step S6, row-wise normalization is performed so that the sum of the elements in each row of the competition network matrix is ​​1.

8. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 7, characterized in that, In S6, the Markov chain Monte Carlo posterior sampling iteratively samples the joint posterior distribution of the endogenous template molecule number vector and the competitive inhibition coefficient using the Metropolis-Hastings algorithm or the Gibbs sampler. After each sampling, the weighted residual is calculated using the second-level composite quality control correction factor as the residual weighting coefficient; the convergence criterion is that the Euclidean distance between the intrinsic template molecule number vectors of two adjacent samplings is less than 1% of the mean of each element of the vector, or the weighted residual norm is less than a preset threshold; the inter-group coupling coefficient is independently estimated based on the inter-group window similarity, batch effect, or competition for shared resources.

9. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 8, characterized in that, In step S7, the ratio of the competing correction copy number to the known copy number of the internal reference target in each aliquot is used as the response factor for each aliquot, and the mean of the response factors for each aliquot is used as the global correction coefficient. The quantitative value of the absolute copy number of each target is obtained by dividing the competing correction copy number of each target by the global correction coefficient. The spike recovery rate and the linear regression determination coefficient are calculated from the observed number of target molecules in each spiked aliquot and the estimated value obtained according to the competitive saturation response equation. When the preset acceptance criterion is not met, a quantification failure flag is output.

10. The method for absolute quantitative analysis of tNGS multi-target copy number according to claim 9, characterized in that, The method also includes: for low-quality targets removed in S5, retaining their original read clusters, the number of observed molecules in the target, and the values ​​of the second-level composite quality control correction factor, and outputting them together with the quantitative values ​​of the absolute copy number of each target; the values ​​of low-quality targets are not involved in the fitting of S6 and the calculation of the global correction coefficient in S7, but are only stored as reference information in the output results.