Abnormality identification and tracing method based on coronavirus high-throughput detection
By using a dynamic learning-based identification model and noise-aware variant detection, combined with Bayesian networks and genetic distance calculation, the model bias and noise interference problems in high-throughput coronavirus detection were solved, enabling rapid identification and accurate tracing of emerging viruses.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YUNNAN KEYAO BIOTECHNOLOGY CO LTD
- Filing Date
- 2025-09-03
- Publication Date
- 2026-05-15
AI Technical Summary
Current high-throughput coronavirus detection suffers from several problems: training data not covering emerging or rare viral variants, leading to model bias; incomplete reference genomes, resulting in missed mutations; complex frequency distributions during co-infections, causing important low-frequency variants to be ignored; ignoring viral recombination in linkage relationships, leading to spurious associations; and sequencing noise, resulting in false positive linkages.
A dynamic learning-based identification model is adopted, which is trained through incremental learning, combined with noise-perceived mutation detection and Bayesian network construction, dynamically updates model parameters, integrates recombination rate correction and genetic distance calculation, and generates a source tracing report.
It improves the ability to identify emerging viruses, reduces sequencing error interference, preserves low-frequency important variants, accurately distinguishes genetic associations from noise, and generates accurate transmission path reports.
Smart Images
Figure CN121148469B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of virus tracing technology, and in particular to an anomaly identification and tracing method based on high-throughput coronavirus detection. Background Technology
[0002] With the development of neural network technology, high-throughput detection of coronaviruses is gradually beginning to use pre-defined recognition models (such as BERT) to determine the viral lineage. However, the following problems still exist:
[0003] 1. Training data may not cover emerging or rare virus variants, resulting in model bias or data imbalance. For example, if the training dataset is concentrated on known lineages, the accuracy of identifying unknown variant viruses (such as newly emerging viruses) may decrease significantly. In the context of an epidemic outbreak, the model may misjudge the lineage, leading to errors in subsequent analysis.
[0004] 2. Mutation site detection is based on comparison with a pre-set viral reference genome. However, the reference genome may be incomplete or outdated. In high-throughput sequencing, the coverage depth of certain genomic regions may be insufficient, which may lead to missed mutations. Moreover, linkage is based on co-occurrence frequency calculation (such as the distribution of the mutation site set in segmented data). However, high-throughput data often contains sequencing noise. If error correction algorithms are not integrated and only simple frequency thresholds are relied upon, false positive linkages may occur (such as random co-occurrences that are not genetically associated).
[0005] 3. When matching mutation sites, it is generally assumed that the expected frequency of each viral lineage can be preset or predicted. However, in reality, the frequency distribution is complex when there is mixed infection or coexistence of multiple lineages. If the difference between the measured frequency of a candidate mutation site and the expected frequency exceeds the threshold, the site is discarded. This may lead to the neglect of important low-frequency mutations (such as drug resistance-related mutations), especially in the scenario of dynamic evolution of viral quasi-species.
[0006] 4. Linkage is inferred solely from co-occurrence frequency (e.g., co-occurrence of variant sites in segmented data), but this ignores the common recombination in viral genetics, which may lead to spurious associations of non-linked sites. The scheme relies on high-frequency variant sites to screen haplotypes, but low-frequency linkages (e.g., <5% frequency) may reflect key evolutionary pathways, but are excluded due to abundance filtering. Summary of the Invention
[0007] To address the problems existing in the prior art, this invention provides an anomaly identification and tracing method based on high-throughput coronavirus detection.
[0008] The technical solution adopted in this invention is:
[0009] An anomaly identification and tracing method based on high-throughput coronavirus detection includes the following steps:
[0010] S1. Obtain high-throughput raw sequencing data of the target sample. The raw sequencing data consists of segmented data, and each segmented data represents a short sequence fragment in the nucleic acid sequence of the target sample.
[0011] S2. Input the raw sequencing data into the dynamic learning identification model. The model is trained through an incremental learning mechanism and outputs the lineage to which each virus in the target sample belongs and the confidence level of each lineage.
[0012] S3. Noise-sensing variant detection is performed on the raw sequencing data to determine each variant site and the variant frequency of each variant site contained in the raw sequencing data, as well as the linkage relationship between each variant site. The linkage relationship is used to reflect the genetic pattern of whether some variant sites in each variant site are passed on to the offspring as a whole.
[0013] S4. For each viral lineage, identify the mutation sites that match the preset specific mutation sites corresponding to that viral lineage from each mutation site, and use them as the mutation sites belonging to that viral lineage.
[0014] S5. Based on the viral lineage to which each mutation site belongs, the mutation frequency of each mutation site, and the linkage relationship between each mutation site, construct a Bayesian network to determine the genetic relationship between each mutation site, and generate viral haplotype sequences based on the genetic relationship.
[0015] S6. Mark mutation sites that meet the criteria of mutation frequency deviating from the historical distribution threshold of the global virus database or linkage confidence level lower than the preset risk threshold as anomalous mutation sites; calculate the genetic distance between anomalous mutation sites and the reference database, and infer the transmission path using the maximum likelihood method; output a source tracing report, which includes the geographical location of the anomalous mutation source and the transmission timeline.
[0016] Preferably, in step S1, the high-throughput raw sequencing data comes from an Illumina Novaseq 6000 or MGIDNBSEQ-T7 sequencing platform, and the sequencing mode is PE150.
[0017] Preferably, in step S2, the incremental learning mechanism triggers fine-tuning based on the data distribution offset of the original sequencing data. The offset is calculated using KL divergence based on the data distribution of the original sequencing data and the data distribution of the training sequencing data. When the offset exceeds the offset threshold, global abnormal virus variant data is injected to update the model parameters.
[0018] Preferably, in step S3, the noise-aware variation detection uses a convolutional autoencoder to denoise the segmented data, the loss function is constrained by reconstruction error and sparsity, and a hidden Markov model is applied to interpolate low-coverage areas with depth less than the trial threshold.
[0019] Preferably, in step S3, the loss function L is:
[0020] L = ||X noisy -X clean ‖2+λ·KL(q(z|X noisy )‖p(z))
[0021] Among them, X noisy X represents the segmented data sequence after alignment of the original sequencing data. clean For the denoised segmented data sequence, ||X noisy -X clean ||2 is X noisy and X clean The Euclidean distance between them, q(z|X) noisy ) represents the posterior distribution of the input segmented data sequence in the latent variable space, p(z) represents the preset prior distribution of the latent variables, and λ represents the weights of the translation reconstruction error and the distribution constraint.
[0022] Preferably, in step S3, the interpolation process of the Hidden Markov Model integrates recombination rate correction, and the state transition probability formula is:
[0023] P adj (s t |s t-1 )=P(s t |s t-1 )×(1+α·r recomb )
[0024] Among them, P adj (s t |s t-1 ) is the site s t-1 To s t The corrected transition probability, P(s) t |s t-1 ) is the site s t-1 To s t The standard state transition probability, r recomb α is the preset viral recombination rate, and α is the viral type coefficient.
[0025] Preferably, in step S5, during the construction of the Bayesian network, the edge weights W ij The calculation formula is:
[0026]
[0027] Where, d ij β represents the genomic distance between sites i and j, and β is set to 2.0 to enhance the detection sensitivity of abnormal linkage.
[0028] Preferably, in step S6, the genetic distance calculation adopts the Tamura-Nei model, the propagation path inference is combined with spatiotemporal metadata, and the anomaly threshold is set to genetic distance > 0.05 and confidence level < 75%.
[0029] The beneficial effects of this invention are:
[0030] 1. Employing a dynamic learning-based identification model enables the system to quickly adapt to emerging viruses. Through incremental learning, the entire network can be trained without retraining, significantly shortening the response cycle and ensuring continuous optimization of the identification capability for new variants.
[0031] 2. By using noise-sensing variant detection technology, the problem of sensitivity to sequencing noise is effectively solved, sequencing error interference is significantly reduced, and the accuracy of variant site detection is improved;
[0032] 3. By dynamically updating and retaining low-frequency but significant variant sites, key signals are ensured not to be filtered out;
[0033] 4. A recombination-aware Bayesian network was constructed. This network integrates viral recombination rate parameters and genomic distance decay coefficients to accurately quantify the genetic linkage strength between variant sites, effectively distinguishing between real genetic associations and random noise, and significantly reducing false positive results.
[0034] 5. By calculating the genetic distance between the virus sequence and the global database, and combining it with the sample's spatiotemporal metadata (such as geographical location and sampling time), a transmission path report can be generated. Attached Figure Description
[0035] Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention. Detailed Implementation
[0036] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0037] Example
[0038] S1. Obtain high-throughput raw sequencing data of the target sample. The raw sequencing data consists of segmented data, each segment representing a short sequence fragment in the nucleic acid sequence of the target sample.
[0039] High-throughput raw sequencing data were obtained from Illumina Novaseq 6000 or MGIDNBSEQ-T7 sequencing platforms, with the sequencing mode being PE150.
[0040] S2. Input the raw sequencing data into the dynamic learning identification model. The model is trained through an incremental learning mechanism and outputs the viral lineage to which each virus in the target sample belongs and the confidence level of each lineage.
[0041] The incremental learning mechanism triggers fine-tuning based on the data distribution offset of the original sequencing data. The offset is calculated using KL divergence based on the data distribution of the original sequencing data and the data distribution of the training sequencing data. KL (P new ||P old ), where P new P represents the data distribution of the raw sequencing data. old For training the sequencing data distribution, when the offset D KL When the value is greater than 0.3, inject global abnormal virus variant data (such as the GISAID database) to update the model parameters.
[0042] S3. Noise-sensing variant detection is performed on the raw sequencing data to determine each variant site and its frequency, as well as the linkage relationship between them. The linkage relationship reflects whether some variant sites in each variant site are passed on as a whole to the offspring in a genetic pattern.
[0043] Noise-aware variation detection uses a convolutional autoencoder to denoise segmented data. The loss function is constrained by reconstruction error and sparsity. Hidden Markov models are applied to interpolate low-coverage regions with depths less than the trial threshold.
[0044] The loss function L is:
[0045] L = ||X noisy -X clean ‖2+λ·KL(q(z|X noisy )‖p(z))
[0046] Among them, X noisy X represents the segmented data sequence after alignment of the original sequencing data. clean For the denoised segmented data sequence, ||X noisy -X clean ||2 is X noisy and X clean The Euclidean distance between them, q(z|X) noisy ) represents the posterior distribution of the input segmented data sequence in the latent variable space, p(z) represents the preset prior distribution of the latent variables, and λ represents the weights of the translation reconstruction error and the distribution constraint.
[0047] The interpolation process of a Hidden Markov Model integrates recombination rate correction, and the state transition probability formula is as follows:
[0048] P adj (s t |s t-1 )=P(s t |s t-1 )×(1+α·r recomb )
[0049] Among them, P adj (s t |s t-1 ) is the site s t-1 To s t The corrected transition probability, P(s) t |s t-1 ) is the site s t-1 To s t The standard state transition probability, r recomb The preset viral recombination rate is α, where α is the viral type coefficient (0.8 for RNA viruses).
[0050] S4. For each viral lineage, identify the mutation sites that match the preset specific mutation sites corresponding to that viral lineage from each mutation site, and use them as the mutation sites belonging to that viral lineage.
[0051] Specific variant sites are obtained from a dynamically updated global database of anomalous sites, which is synchronized every 48 hours to include emerging anomalous variant data.
[0052] S5. Based on the viral lineage to which each mutation site belongs, the mutation frequency of each mutation site, and the linkage relationship between each mutation site, construct a Bayesian network to determine the genetic relationship between each mutation site, and generate viral haplotype sequences based on the genetic relationship.
[0053] In the construction of a Bayesian network, the edge weight W ij The calculation formula is:
[0054]
[0055] Where, d ij β represents the genomic distance between sites i and j, and β is set to 2.0 to enhance the detection sensitivity of abnormal linkage.
[0056] S6. Mark mutation sites that meet the criteria of mutation frequency deviating from the historical distribution threshold of the global virus database or linkage confidence level lower than the preset risk threshold as anomalous mutation sites; calculate the genetic distance between anomalous mutation sites and the reference database, and infer the transmission path using the maximum likelihood method; output a source tracing report, which includes the geographical location of the anomalous mutation source and the transmission timeline.
[0057] Genetic distance was calculated using the Tamura-Nei model, and propagation path inference was combined with spatiotemporal metadata. The anomaly threshold was set as genetic distance > 0.05 and confidence level < 75%.
[0058] Using the above method for anomaly identification and tracing in high-throughput coronavirus detection has the following advantages:
[0059] 1. Improved model dynamism
[0060] A dynamic learning-based identification model is introduced, which automatically triggers an incremental learning mechanism by monitoring data distribution shifts in real time (e.g., when KL divergence > 0.3). This mechanism only fine-tunes the parameters of the fully connected layers, injecting new global variant data (e.g., the XBB.1.5 sequence from the GISAID database) while preserving historical knowledge. This design enables the system to adapt to virus evolution; for example, when unknown mutations appear in sewage samples, the model can be quickly updated, improving the accuracy of identifying emerging variants and avoiding the waste of computational resources caused by traditional retraining.
[0061] 2. Enhanced data reliability
[0062] Design a dual-track error correction mechanism:
[0063] Convolutional autoencoders: learn noise distribution characteristics through an encoder-decoder structure to separate the real signal from sequencing errors (such as base substitution errors in the Illumina platform) in the latent variable space.
[0064] Recombination-corrected HMM: Integrating the recombination rate parameter (r) into a Hidden Markov Model to target viral genome recombination characteristics. recomb ), optimize the state transition probability in low-coverage areas.
[0065] 3. Low-frequency signal integrity protection
[0066] A dynamic abnormal site database is constructed, synchronizing global emerging variant data (such as Spike:K444T resistance site) every 48 hours, and intelligent retention rules are set:
[0067] Variations located in functionally conserved regions (such as polymerase coding regions) are still marked as key sites even if their frequency is below a threshold;
[0068] Low-frequency variants that are strongly linked to known resistance / virulence sites (confidence >85%) were included in the analysis.
[0069] 4. In-depth optimization of genetic analysis
[0070] Integrating two types of key parameters in a Bayesian network:
[0071] Recombination rate r recomb : Quantify site association strength based on historical recombination hotspot data (such as the coronavirus S protein region);
[0072] Distance attenuation coefficient β: through genomic distance d ij Reduce random co-occurrence interference at distant sites.
[0073] This design reduces the false positive rate of genetic linkages and precisely locates cross-lineage co-evolutionary modules (such as Spike-RdRp linkage blocks) in Delta-Omicron recombinant analysis.
[0074] Finally, the genetic distance is calculated using the Tamura-Nei model, and the propagation network is constructed by combining the spatiotemporal metadata of the samples (sampling location / time). Based on this, a source tracing report containing a propagation path map and timeline can be further generated.
[0075] The above embodiments merely illustrate specific implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.
Claims
1. An anomaly identification and source tracing method based on high-throughput coronavirus detection, characterized in that, Includes the following steps: S1. Obtain high-throughput raw sequencing data of the target sample. The raw sequencing data consists of segmented data, and each segmented data represents a short sequence fragment in the nucleic acid sequence of the target sample. S2. Input the raw sequencing data into the dynamic learning identification model. The model is trained through an incremental learning mechanism and outputs the lineage to which each virus in the target sample belongs and the confidence level of each lineage. The incremental learning mechanism triggers fine-tuning based on the data distribution offset of the raw sequencing data. The offset is calculated using KL divergence based on the data distribution of the raw sequencing data and the data distribution of the training sequencing data. When the offset exceeds the offset threshold, global abnormal virus variant data is injected to update the model parameters. S3. Noise-aware variant detection is performed on the raw sequencing data to determine each variant site and its frequency, as well as the linkage relationship between them. The linkage relationship reflects whether some variant sites are passed on as a whole to the offspring. The noise-aware variant detection uses a convolutional autoencoder to denoise the segmented data. The loss function is constrained by reconstruction error and sparsity. Hidden Markov Model interpolation is applied to low-coverage regions with a depth less than the trial threshold. The interpolation process of the Hidden Markov Model integrates recombination rate correction. S4. For each viral lineage, identify the mutation sites that match the preset specific mutation sites corresponding to that viral lineage from each mutation site, and use them as the mutation sites belonging to that viral lineage. S5. Based on the viral lineage to which each mutation site belongs, the mutation frequency of each mutation site, and the linkage relationship between each mutation site, construct a Bayesian network to determine the genetic relationship between each mutation site, and generate viral haplotype sequences based on the genetic relationship. S6. Mark mutation sites that meet the criteria of mutation frequency deviating from the historical distribution threshold of the global virus database or linkage confidence level lower than the preset risk threshold as anomalous mutation sites; calculate the genetic distance between anomalous mutation sites and the reference database, and infer the transmission path using the maximum likelihood method; output a source tracing report, which includes the geographical location of the anomalous mutation source and the transmission timeline.
2. The anomaly identification and tracing method based on high-throughput coronavirus detection according to claim 1, characterized in that, In step S1, the high-throughput raw sequencing data comes from the Illumina Novaseq 6000 or MGI DNBSEQ-T7 sequencing platform, and the sequencing mode is PE150.
3. The anomaly identification and tracing method based on high-throughput coronavirus detection according to claim 2, characterized in that, In step S3, the loss function for: ； in, These are the segmented data sequences after alignment with the original sequencing data. This is the segmented data sequence after denoising. for and The Euclidean distance between them Let be the posterior distribution of the input segmented data sequence in the latent variable space. The prior distribution of the latent variables is pre-defined. The weights are assigned to the translation reconstruction error and the distribution constraints.
4. The anomaly identification and tracing method based on high-throughput coronavirus detection according to claim 3, characterized in that, In step S3, during the interpolation process of the Hidden Markov Model to integrate the recombination rate correction, the state transition probability formula is: ； in, For site to The corrected transition probability, For site to The standard state transition probability, The preset viral recombination rate, This represents the virus type coefficient.
5. The anomaly identification and tracing method based on high-throughput coronavirus detection according to claim 4, characterized in that, In step S5, during the construction of the Bayesian network, the edge weights... The calculation formula is: ； in, For site and Genomic distance, A value of 2.0 is used to enhance the detection sensitivity of abnormal linkages.
6. The anomaly identification and tracing method based on high-throughput coronavirus detection according to claim 5, characterized in that, In step S6, the genetic distance is calculated using the Tamura-Nei model, the propagation path inference is combined with spatiotemporal metadata, and the anomaly threshold is set to genetic distance > 0.05 and confidence level < 75%.