Intelligent identification method for transgenic crops based on high-throughput sequencing
The intelligent identification method based on high-throughput sequencing solves the problems of multi-platform compatibility, low efficiency, and rigid judgment logic in the identification of genetically modified crops, and achieves efficient, accurate, and reliable identification of genetically modified crops, which is applicable to scenarios such as cross-border customs clearance, food safety sampling inspection, and breeding line identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANJIN CUSTOMS IND PROD SAFETY TECH CENT
- Filing Date
- 2026-01-15
- Publication Date
- 2026-05-22
Smart Images
Figure CN121544285B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biological detection technology, and more specifically, to a method for intelligent identification of transgenic crops based on high-throughput sequencing. Background Technology
[0002] With the expansion of global agricultural trade, compliance identification of genetically modified crops has become crucial for food safety supervision and cross-border trade clearance. It is necessary to quickly identify prohibited strains to prevent biosafety risks, while also addressing the issues of data format differences and different processing logics brought about by high-throughput sequencing technologies from multiple platforms. Furthermore, current regulatory requirements for identification efficiency, accuracy, and traceability are increasing, but traditional identification methods are difficult to balance multi-platform compatibility, efficient processing, scenario-based judgment, and reliable traceability. They often result in problems such as delayed regulatory response, high misjudgment rates, and difficulty in resolving disputes over conclusions, failing to meet actual needs.
[0003] Existing technical solutions for identifying genetically modified crops have significant shortcomings. For example, in multi-platform data processing, traditional unified quality control lacks platform-specific rules, leading to large biases in the selection of qualified data and easy misjudgment of low-abundance samples. In terms of processing efficiency, there is a lack of intelligent parallel scheduling, resulting in uneven sample load and insufficient priority of emergency testing sample resources. In terms of judgment logic, the judgment rules are rigid, the misjudgment rate of low-abundance samples is high, boundary cases lack standardized expert review, and disputes over conclusions are difficult to resolve. Summary of the Invention
[0004] To overcome the aforementioned shortcomings of existing technologies and to achieve the above objectives, this invention provides the following technical solution: a smart identification method for transgenic crops based on high-throughput sequencing, comprising:
[0005] S1: Acquire raw high-throughput sequencing data and sample metadata from multiple platforms, conduct dynamic quality assessments for each platform, screen qualified data and divide them according to sample block rules, and generate a quality-controlled data block set.
[0006] S2: Based on the data block set after quality control, data blocks are allocated to multiple parallel computing nodes. Each computing node synchronously performs technical detection to obtain technical fact fragments, and then aggregates and supplements batch statistical information to generate a technical fact report.
[0007] S3: Based on the technical fact report, the judgment rule set is matched according to the pre-set decision tree, and the judgment is performed in parallel through the multi-instance rule engine. If the judgment result is clear, the regulatory conclusion is automatically generated. Expert review is triggered for boundary cases, and then the regulatory judgment conclusion set is integrated and generated.
[0008] S4: Based on the regulatory judgment conclusion set and the associated technical fact report, extract the original sequencing reads of all samples, generate a unique hash for each read, construct a processing trajectory chain, integrate metadata and visual evidence, generate a credible identification report and divide it into different scenario-specific versions for targeted distribution;
[0009] S5: Based on the credible identification reports accumulated within the cycle, extract identification cases for mining and clustering, and generate optimization suggestions to be fed back to the quality control rule base and the judgment rule base.
[0010] Furthermore, the sub-platform performs dynamic quality assessments and filters qualified data in the following ways:
[0011] We acquire raw high-throughput sequencing data and corresponding sample metadata from multiple platforms, call the platform-specific quality control rule library to conduct dynamic quality assessment for each platform, and then classify the data into qualified data, unqualified data, and low-abundance sample data that need to be sequenced to increase the depth through multi-dimensional qualification judgment.
[0012] For unqualified data, a retest is triggered. For low-abundance sample data that require increased sequencing depth, the sequencing depth is adjusted and a new dynamic quality assessment is performed.
[0013] Furthermore, the method of dividing the data according to the sample block rules to generate the quality control data block set includes:
[0014] Based on qualified data and associated sample metadata, a pre-set sample block rule library containing basic block rules and scenario adaptation rules is invoked.
[0015] Based on the total number of samples and the detection scenario, the corresponding sample segmentation rules in the sample segmentation rule library are matched. Then, based on the qualified data, the original sequencing data and sample metadata of the corresponding samples are extracted and packaged to generate data blocks, which are then integrated into a quality control data block set.
[0016] Furthermore, the method of allocating data blocks to multiple parallel computing nodes, with each computing node synchronously performing technical detection to obtain technical fact fragments, includes:
[0017] Based on the quality-controlled data block set, and combined with the currently available computing node resources, the data blocks are allocated to parallel computing nodes through load balancing scheduling.
[0018] After receiving the data block, each computing node synchronously performs sequence alignment, component identification, strain matching evaluation, and unknown sequence clustering according to the platform type.
[0019] Integrate the test results to generate standardized technical fact fragments.
[0020] Furthermore, the method for generating the technical fact report includes:
[0021] Using the unique identifier of a sample as the core key, technical fact fragments from each computing node are aggregated according to the sample dimension to form sample-level aggregated data; and batch statistical information is generated based on the sample-level aggregated data.
[0022] Integrate sample-level aggregated data with batch statistical information to generate standardized technical fact reports.
[0023] Furthermore, the method of performing parallel determination using a multi-instance rule engine based on a preset decision tree matching rule set includes:
[0024] Analyze technical fact reports and extract scene labels, sample size features, and special data markers as scene features;
[0025] The scene recognition engine loads a pre-set scene-rule set decision tree, judges the scene features in order of importance to match the unique corresponding judgment rule set, and generates a rule set-sample combination association table.
[0026] Based on the association table, a multi-instance rule engine is launched, an independent engine instance is assigned to each rule set, and samples associated with the same rule set are divided into different sample groups and evenly distributed to each engine instance.
[0027] Each engine instance synchronously performs judgment on the assigned sample group, obtaining automatically qualified samples, automatically unqualified samples, and judgment results pending review.
[0028] Furthermore, the method for generating the regulatory judgment conclusion set includes:
[0029] Automatically qualified and automatically unqualified samples are used as explicit automatic judgment results to directly generate regulatory conclusions;
[0030] The sample to be reviewed is used as a boundary case to trigger the expert review mechanism, generate an independent review task sheet for each sample to be reviewed, and then assign it to experts according to their areas of expertise for standardized review.
[0031] The expert review results are verified and standardized, and then integrated with the automatic judgment results to generate a standardized set of regulatory judgment conclusions.
[0032] Furthermore, the method of constructing the processing trajectory chain includes:
[0033] Based on the regulatory judgment conclusion set and technical fact report, the original sequencing reads of the sample and the judgment basis were extracted;
[0034] Generate a unique hash for each read segment and store it on the blockchain, and build a full-process processing trajectory chain in chronological order;
[0035] Integrate metadata and visual evidence from the entire sample process to generate a draft of a standardized and credible authentication report with digital signature.
[0036] Furthermore, the method of dividing different scenario-specific versions and distributing them in a targeted manner includes:
[0037] The initial draft of the credible assessment report undergoes multi-dimensional credibility verification to obtain a qualified credible assessment report. Then, according to the different needs of the users, the credible assessment report is adapted to generate a scenario-based version, either a simplified regulatory version or a complete technical version, as a scenario-based report.
[0038] Based on the principle of least privilege, scenario-based reports are distributed in a targeted, encrypted manner to different users, and distribution logs are recorded.
[0039] Furthermore, the methods for extracting and clustering the identified cases to generate optimization suggestions and feeding them back to the quality control rule base and the judgment rule base include:
[0040] Data related to identified cases is extracted in batches from accumulated credible identification reports according to a preset time period to construct a hierarchical dynamic knowledge base of basic cases, feature associations, and abnormal cases.
[0041] Multi-dimensional mining and clustering of the hierarchical dynamic knowledge base are performed to identify the direction of rule optimization. Based on the clustering analysis results, feasible optimization suggestions are generated and their effectiveness is verified. The quality control rule base and the judgment rule base are updated based on the effective optimization suggestions.
[0042] The technical effects and advantages of the intelligent identification method for transgenic crops based on high-throughput sequencing in this invention are as follows:
[0043] This invention designs exclusive quality control rules for different platforms and combines them with scenario-based block processing. This not only solves the screening bias problem of traditional unified quality control and improves the accuracy of qualified data, but also lays the foundation for subsequent parallel computing and ensures the orderly processing of data.
[0044] Secondly, data blocks are allocated according to node performance, and detection tools are called synchronously on different platforms to solve the problems of low processing efficiency and uneven load in traditional processing, which greatly improves the efficiency of batch and emergency sample detection. In addition, batch statistical information is supplemented when aggregating fragments, making the technical fact report more complete.
[0045] Then, based on scene features, exclusive judgment rules are matched, thresholds are dynamically adjusted, and expert review of boundary cases (including objection handling) is triggered to solve the problems of rigid traditional rules and drone collaboration, reduce the misjudgment rate, resolve disputes over conclusions, and improve the accuracy and credibility of judgments.
[0046] Next, a unique hash is generated for each read segment and stored on the blockchain to establish a full-process processing trajectory chain. Contextualized reports are generated and distributed according to user needs, solving the problems of insufficient credibility and traceability in traditional methods. This ensures that read segments are tamper-proof, the process is traceable, and the report is adaptable to different user needs, thus improving the report's practicality.
[0047] Finally, historical report data is extracted periodically to build a hierarchical knowledge base. Through multi-dimensional mining, optimization suggestions are generated to update the rule base. An effect tracking mechanism is established to solve the problem that traditional rules cannot be iterated, forming a "mining-optimization-tracking" closed loop, and continuously improving the system's adaptability to new product lines and new scenarios.
[0048] This solution is applicable to scenarios such as cross-border customs clearance, food safety sampling inspection, and breeding strain identification. It achieves multi-platform compatibility, efficient response, accurate judgment, and reliable traceability throughout the entire process, providing technical support for biosafety supervision. Attached Figure Description
[0049] Figure 1 This is a schematic diagram of the intelligent identification method for transgenic crops based on high-throughput sequencing according to the present invention.
[0050] Figure 2 This is a flowchart illustrating the expert review process for boundary cases in the intelligent identification method for transgenic crops based on high-throughput sequencing of the present invention.
[0051] Figure 3 This is a flowchart illustrating the intelligent identification system for transgenic crops based on high-throughput sequencing of the present invention. Detailed Implementation
[0052] 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, and 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.
[0053] Example 1
[0054] Please see Figure 1 and Figure 2 As shown in this embodiment, the intelligent identification method for transgenic crops based on high-throughput sequencing includes:
[0055] S1: Acquire raw high-throughput sequencing data and sample metadata from multiple platforms, conduct dynamic quality assessments for each platform, screen qualified data and divide them according to sample block rules, and generate a quality-controlled data block set.
[0056] S2: Based on the data block set after quality control, data blocks are allocated to multiple parallel computing nodes. Each computing node synchronously performs technical detection to obtain technical fact fragments, and then aggregates and supplements batch statistical information to generate a technical fact report.
[0057] S3: Based on the technical fact report, the judgment rule set is matched according to the pre-set decision tree, and the judgment is performed in parallel through the multi-instance rule engine. If the judgment result is clear, the regulatory conclusion is automatically generated. Expert review is triggered for boundary cases, and then the regulatory judgment conclusion set is integrated and generated.
[0058] S4: Based on the regulatory judgment conclusion set and the associated technical fact report, extract the original sequencing reads of all samples, generate a unique hash for each read, construct a processing trajectory chain, integrate metadata and visual evidence, generate a credible identification report and divide it into different scenario-specific versions for targeted distribution;
[0059] S5: Based on the credible identification reports accumulated within the cycle, extract identification cases for mining and clustering, and generate optimization suggestions to be fed back to the quality control rule base and the judgment rule base.
[0060] The methods for conducting dynamic quality assessments and screening qualified data across different platforms include:
[0061] It receives high-throughput sequencing raw data from multiple platforms such as Illumina (e.g., NovaSeq series) and Nanopore (e.g., MinION series), and simultaneously collects corresponding sample metadata (sample metadata is entered through a customized web form), including basic information (including sample ID (i.e., unique sample identifier), port of entry, crop, date, serial number, sampling batch number, country and region of origin, crop variety, etc.), testing requirements (including testing scenario labels (e.g., bulk cargo, emergency testing, etc.), priority (high, medium, low), whether low abundance testing is required, etc.) and sequencing association information (including sequencing platform model, preset sequencing depth).
[0062] Verify the uniqueness of the sample ID and the compatibility between the testing scenario and the sequencing platform (e.g., emergency testing requires association with the Nanopore platform); if the sample ID is duplicated, trigger the process of reallocating a unique ID; if the testing scenario and platform are incompatible, generate an error message and return the sample to the testing end for supplementary explanation, and re-verify after correction; after the verification is successful, establish an association between the metadata and the corresponding raw sequencing data through the sample ID, store it in a relational database, and generate a data-metadata association table;
[0063] Based on the data-metadata association table, the built-in quality control rule library specific to the corresponding sequencing platform is called according to the sequencing platform type;
[0064] The quality control rule base structure is a mapping relationship between platform, indicator, threshold, and tool, as exemplified by:
[0065] Sequencing platform (Illumina NovaSeq), [quality control indicators - threshold standards - association analysis tools]: [Q30 base percentage - ≥80% - FastQC (driving software) v0.11.9; read repeat rate - ≤15% - Picard MarkDuplicates (association analysis tool); adapter residue rate - ≤2% - Trimmomatic (association tool)];
[0066] Sequencing platform (Nanopore MinION), [quality control indicators - threshold standards - correlation analysis tools]: [Median QScore (median Q score, real-time value) - ≥10 - MinKNOW (driving operation software) v23.11.7; Percentage of single read lengths ≥500bp - ≥60% - NanoPlot (plotting analysis tool); Data output per 10 minutes - ≥1Gb - MinKNOW status log analysis];
[0067] The rule invocation logic is as follows: based on the sequencing platform model field in the sample metadata, the complete set of quality control rules and corresponding analysis tools for the corresponding platform are retrieved from the quality control rule base;
[0068] The dynamic quality assessment of the sub-platform includes offline batch assessment and online real-time assessment;
[0069] Offline batch evaluation of the Illumina platform: Batch analysis of all Illumina platform files was performed using analysis tools (such as FastQC) to obtain reports on indicators such as Q-value distribution (referring to Q distribution), GC content, and adapter residue for each sample; the individual sample indicator reports were integrated to form a batch-level quality control report, which statistically analyzed key indicators such as the average Q30 percentage and the maximum repetition rate of all samples in the batch; for samples marked as requiring low abundance detection, rapid k-mer counting was performed to preliminarily assess the coverage of exogenous gene characteristic k-mers; if the coverage was less than the coverage threshold (e.g., 5×), it was marked as requiring increased sequencing depth.
[0070] Nanopore platform online real-time evaluation: Automatically extracts data that has completed base identification at regular intervals (e.g., every 5 minutes), and calculates the Median Q Score and read length distribution within that time period using analysis tools (e.g., NanoPlot); For real-time data, it quickly compares it to a reference library of commonly used transgenic elements in the industry. If the number of matched reads is less than expected (e.g., less than 3 reads / 1Gb), it is marked as needing to increase sequencing depth (this indicator is a specific implementation of the Nanopore platform's low abundance fit determination, corresponding to the low abundance fit requirement in the multi-dimensional qualification determination);
[0071] The data is categorized into qualified data, unqualified data, and low-abundance sample data requiring increased sequencing depth through multi-dimensional qualification assessment.
[0072] The specific criteria for multi-dimensional qualification determination include basic indicator determination (the quality control indicators of data samples from each platform must meet the corresponding threshold standards in the quality control rule base; if any indicator fails to meet the standard, it is determined to be unqualified data) and low abundance adaptation determination.
[0073] The low abundance matching determination is as follows: if the initial detection signal of a sample marked as requiring low abundance detection is insufficient (each platform adopts corresponding evaluation indicators according to technical characteristics), it is determined to be a low abundance sample data that needs to increase sequencing depth (not unqualified data, only requiring supplemental sequencing); specifically, the Illumina platform uses "the coverage of exogenous gene characteristic k-mer < the coverage threshold (e.g., 5×)" as the standard, while the Nanopore platform uses "the number of matching reads in the common transgenic element reference library < the expected number (e.g., < 3 reads / 1Gb)" as the standard.
[0074] For non-compliant data, a retest is triggered, and the non-compliant quality control indicators are noted. If the data still fails to meet the requirements after multiple retests, sample resampling is triggered. For low-abundance sample data that requires increased sequencing depth, the ratio of the coverage threshold to the estimated low-abundance proportion is taken as the target coverage. The required additional sequencing depth is calculated based on the target coverage (e.g., if the estimated low-abundance proportion is 0.2% and the target coverage is 25×, if 10× has already been sequenced, 15× needs to be added). The depth adjustment command is sent to the sequencing platform through the API interface. After the additional sequencing is completed, the two sets of data are merged and dynamic quality assessment is performed again.
[0075] Mark the qualified data as eligible to proceed to the next step and associate it with the sample ID in the sample metadata.
[0076] The methods for generating quality control data block sets based on sample partitioning rules include:
[0077] Based on qualified data and associated sample metadata, a pre-set sample block rule library containing basic block rules and scenario adaptation rules is invoked.
[0078] The logical definition of the basic block division rule is as follows: Based on the total number of qualified data samples, set the ratio of the total number of samples to the number of data blocks (such as 10:1 or 5:1, for example, 50 qualified samples are divided into 5 to 10 data blocks), and divide the data blocks according to the set ratio, such as 5 samples / block or 10 samples / block, to ensure that the number of samples corresponding to each data block is balanced, which is convenient for subsequent parallel computing node load distribution.
[0079] The logical definition of the scenario adaptation rule is as follows: If the detection scenario is an emergency detection (e.g., a scenario with ≤10 samples), it is forced to be divided into 1 sample / block, and high-performance computing nodes are given priority to execute it; if the detection scenario is a low abundance mixed sample identification, it is divided into blocks according to the mixed sample group (assuming it is set to 5 samples / block, then the 5 sub-samples of the same mixed sample are divided into the same block) to avoid signal dispersion caused by cross-block splitting.
[0080] After determining the number of blocks, the block ID (including a list of sample IDs, block rule identifier, data size, etc., and the target computing node number will be added later after load balancing scheduling) is automatically recorded and stored in the block task table.
[0081] Based on the total number of samples and the detection scenario, the corresponding sample segmentation rules in the sample segmentation rule library are matched. Then, based on the sample ID list, the original sequencing data and sample metadata of the corresponding sample are extracted from the qualified data to generate a data block (containing core data (the original sequencing data of the corresponding sample) and metadata attachments (block-level metadata (including block ID, sample ID list, data volume, platform type, and segmentation timestamp))).
[0082] Finally, all data blocks are integrated to form a post-quality control data block set.
[0083] The methods for distributing data blocks to multiple parallel computing nodes, with each node synchronously performing technical detection to obtain fragments of technical facts, include:
[0084] Based on the quality-controlled data block set, and combined with the currently available computing node resources, the data blocks are allocated to parallel computing nodes through load balancing scheduling.
[0085] Specifically, the load balancing scheduling rule is defined as follows: data is allocated according to the principle of matching data volume with node performance (e.g., 10GB data blocks are allocated to 32-core CPU nodes, and 5GB data blocks are allocated to 16-core nodes) to ensure that the difference in processing time between nodes is as small as possible (e.g., the difference is ≤10%).
[0086] Data blocks marked as emergency detection are preferentially allocated to idle high-performance computing nodes (e.g., computing nodes with ≥128GB of memory are marked as high-performance computing nodes), and computing resources are locked until processing is complete;
[0087] After receiving the data block, each computing node calls the corresponding technical testing toolchain according to the platform type to synchronously perform operations such as sequence alignment, component identification, strain matching evaluation, and unknown sequence clustering.
[0088] The sequence alignment process is as follows: For Illumina platform data, tools such as BWA-MEM are used to align reads to the industry-standard transgenic core reference library (containing known strain whole genome sequences and exogenous element sequences), and the matching reference sequence ID and position information are output; for Nanopore platform data, tools such as Minimap2 (optimized for long reads) are used to align to the transgenic core reference library, retaining fusion reads that cross exogenous-endogenous boundaries (this is key identification evidence).
[0089] Element identification (transgenic element identification) is as follows: Based on the results of sequence alignment, sequence fragments that match characteristic elements such as promoters (e.g., CaMV 35S), terminators (e.g., NOS), and marker genes (e.g., nptII) in the reads are extracted, and then verified a second time using tools such as BLAST (usually the sequence similarity needs to be ≥95%). The element type, name, matching length, and position in the transgenic core reference library are recorded to generate an element identification list.
[0090] Strain matching assessment: Based on the element identification list, the identified element combinations are compared with known transgenic strain feature libraries (e.g., the element combination of MON87751 (insect-resistant soybean) is CaMV 35S+Cry1Ac (exogenous gene)+NOS). A weighted method is used to calculate the strain matching degree. The weight allocation principle is that the weight of core functional elements (e.g., exogenous genes) is higher than that of auxiliary regulatory elements (e.g., promoters, terminators). For example, in a normal scenario, the weight of core elements is 0.6 and the weight of auxiliary elements is 0.4. In low abundance or emergency scenarios, the weights can be adjusted according to element specificity. Finally, the best matching strain ID and strain matching degree score (within the range of 0~1) are obtained.
[0091] Unknown sequence clustering is performed as follows: For reads that do not match the transgenic core reference library (unknown sequences), clustering is performed according to platform characteristics: For the Illumina platform (short reads), tools such as CD-HIT are used to cluster based on sequence similarity (e.g., 95%), retaining clusters with the number of reads greater than a set large threshold (e.g., >50), and marking them with a unique unknown sequence cluster ID; For the Nanopore platform (long reads), after error correction using tools such as Canu, clustering is performed based on sequence similarity (the sequence similarity for long reads can be appropriately reduced, e.g., 90%), retaining clusters with the number of reads greater than a set small threshold (e.g., >30), and recording the common sequence length of the clusters;
[0092] After each node completes the testing, all technical test results are integrated into standardized technical fact fragments, including block ID, sequencing platform type and sample technical fact list (each sample includes sample ID, detected transgenic elements (including element type (e.g. promoter, exogenous gene, terminator), element name (e.g. CaMV 35S, Cry1Ac), sequence similarity (e.g. 0.98), strain matching results, and unknown sequence cluster information (including unknown sequence cluster ID and number of reads within the cluster)).
[0093] The methods for generating technical fact reports include:
[0094] The master compute node receives technical fact fragments uploaded by parallel compute nodes and then performs integrity checks (checking whether required fields such as block ID and sample ID in the technical fact fragment are complete; if missing, it is marked as an abnormal fragment, triggering the node to re-upload) and consistency checks (verifying whether the sequencing platform type in the technical fact fragment is consistent with the platform information in the sample metadata of the data block (e.g., fragments in Illumina data blocks must not be marked as Nanopore to avoid cross-platform confusion).
[0095] Using the unique identifier of each sample (i.e., sample ID) as the core key, all verified technical fact fragments are aggregated according to the sample dimension to form sample-level aggregated data. Specifically:
[0096] The detection results of the same sample in different data blocks (e.g., a sample spanning 2 data blocks) are merged, and duplicate records are removed (if the same element is detected by multiple computing nodes, the highest sequence similarity value is retained); for mixed detection samples (e.g., mixed group 1 contains 5 sub-samples), an additional association table of mixed group ID and sub-sample ID is established to mark the element co-occurrence relationship of each sub-sample (e.g., both sub-sample 1 and sub-sample 3 detected CaMV 35S).
[0097] Based on the sample-level aggregated data, batch statistical indicators are generated as batch statistical information, including basic statistical indicators, component and product distribution indicators, and abnormal indicators.
[0098] Basic statistical indicators include total number of samples, sample proportion of each platform (e.g., Illumina samples account for 60%, Nanopore samples account for 40%), and detection completion rate (number of successfully detected samples / total number of samples).
[0099] The component and strain distribution indicators include high-frequency detection components (a threshold for the number of occurrences can be set (e.g., 30 times). If the number of times a component is detected is greater than the threshold, it is marked as a high-frequency detection component, such as CaMV 35S appearing 80 times), high-frequency matching strains (similarly, a threshold for the number of samples matched can be set (e.g., 10). If the number of samples is greater than the threshold, it is marked as a high-frequency matching strain, such as MON87751 matching 25 samples), and average strain matching degree;
[0100] The abnormal indicators include the number of samples with unknown sequence clusters (e.g., 12) and the number of samples with low matching degree (i.e., the minimum matching degree threshold is associated with the preset basic threshold of the judgment rule set (the minimum matching degree threshold is 0.5 in normal scenarios, and can be adjusted synchronously in emergency or low abundance scenarios); samples with a matching degree greater than the strain matching degree threshold of the judgment rule set (e.g., 0.8 in normal scenarios) are considered high matching degree samples, samples with a matching degree less than or equal to the strain matching degree threshold and greater than the minimum matching degree threshold are considered low matching degree samples, and samples with a matching degree less than or equal to the minimum matching degree threshold indicate a mismatch. For example, samples with a strain matching degree of 0.5 to 0.8 can be marked as low matching degree samples).
[0101] Integrate sample-level aggregated data with batch statistical information to generate standardized technical fact reports.
[0102] The parallel decision-making method based on a pre-defined decision tree matching rule set and using a multi-instance rule engine includes:
[0103] The technical fact report is analyzed, and the following features are extracted: scenario tags (such as emergency detection and low abundance multitude) extracted from the sample metadata associated with the sample technical fact list, sample size features (different sample sizes are classified according to the total number of samples using a threshold segmentation method, for example, ≤10 samples are small batches, 11~50 samples are medium batches, and >50 samples are large batches) and special data markers (such as the presence of unknown sequence clusters and low matching degree samples) as scenario features;
[0104] The scene recognition engine loads a pre-set scene-rule set decision tree, judges the scene features according to the importance of the scene label, and determines the unique corresponding judgment rule set to generate a rule set-sample combination association table.
[0105] The decision tree's logic for determining the importance of scenario labels is as follows: Emergency detection labels are prioritized; if present, they are directly matched against the emergency scenario determination rule set (a pre-defined rule set whose logic typically involves relaxing thresholds, such as lowering the strain matching threshold by a certain percentage (e.g., to 0.75)). When no emergency label exists, low-abundance mixed sample labels are determined; if present, low-abundance scenario determination rule sets are matched against (whose logic typically requires three rounds of verification: component, strain, and boundary).
[0106] It should be noted that element verification checks the integrity (whether it contains a complete sequence fragment) and specificity (excluding interference from homologous sequences of non-target crops) of the detected transgenic elements (such as promoters and exogenous genes), and verifies the read coverage of low-abundance elements (e.g., it needs to meet ≥3×) to avoid misjudgment of a single read. Plexicon verification compares the element combination with the characteristic element combination of known transgenic varieties (e.g., the CaMV 35S promoter + Cry2Ab gene combination of MON87701), calculates the stability of the variety matching degree (e.g., the deviation ≤5% after multiple comparisons), and excludes variety mismatch caused by low-abundance signals. Boundary verification verifies the integrity of the boundary sequence of the exogenous gene inserted into the crop genome (it needs to contain the exogenous-endogenous fusion fragment), confirms the consistency of the boundary sequence with known varieties, and avoids the confusion of boundary sequences of different varieties in low-abundance mixtures.
[0107] When there are no emergency testing labels or low-abundance mixed sample labels, the corresponding rule set is matched according to the sample size. For example, large batches are matched with the rule set for bulk cargo scenarios (the core rule logic is: prioritize the judgment of high-priority product series samples), while small and medium batches are matched with the rule set for regular scenarios (the core rule logic is: full sample standard verification, no priority difference).
[0108] After matching is completed, the corresponding decision rule set (including rule ID, corresponding decision threshold, and core rule logic) is retrieved from the decision rule base to generate a rule set-sample combination association table;
[0109] The multi-instance rule engine is launched based on the association table. An independent engine instance (referring to a computing instance used to perform a specific task, such as 1 instance for emergency scenarios and 2 instances for large-scale scenarios) is assigned to each matched rule set. When the engine instance starts, all parameters of the corresponding judgment rule set are loaded. Then, the samples belonging to the same rule set are grouped (for example, 10 to 20 samples per group) and evenly distributed to the multiple engine instances corresponding to that rule set.
[0110] Each engine instance synchronously executes the judgment logic on the assigned sample group: if the sample strain matching degree is ≥ the strain matching degree threshold set by the rule set (e.g., 0.8 in a normal scenario) and matches a prohibited strain, it is judged as automatically unqualified, and the judgment basis is recorded (e.g., violating a certain regulatory rule, matching a certain prohibited strain, matching degree 0.93); if the sample has no external element detection results and the strain matching degree is ≤ the preset minimum matching degree threshold, it is judged as automatically qualified, and the basis is recorded (e.g., no transgenic element, strain matching degree 0.3); the judgment results of automatic qualified and automatic unqualified are clearly defined as the judgment results; in addition to the above-mentioned clear cases, other situations such as the sample strain matching degree being in the fuzzy range (e.g., 0.75~0.8) or containing unknown sequence clusters and no matching strains are marked as pending review, and key doubts are recorded (e.g., matching degree 0.7 close to the threshold, containing a certain unknown cluster); in low abundance scenarios, if a round of verification fails, it is directly marked as pending review, and verification failure details are attached (e.g., the second round of strain verification failed, missing left boundary sequence).
[0111] After all engine instances have completed the judgment, all results (including three types of status: automatic pass, automatic fail, and pending review) are aggregated by sample ID to generate a sample-judgment status-judgment basis comparison table.
[0112] Samples marked as pending review will be given a review priority (e.g., high priority for samples pending review in emergency scenarios and medium priority for samples pending review in normal scenarios).
[0113] The methods for generating regulatory judgment conclusion sets include:
[0114] Automatically qualified and automatically unqualified samples are used as explicit automatic judgment results to directly generate regulatory conclusions;
[0115] The sample to be reviewed is used as a boundary case to trigger the expert review mechanism. An independent review task sheet is generated for each sample to be reviewed, which includes the sample ID, the current judgment result, the summary of the associated technical fact report, and key visual evidence (such as the reading segment coverage distribution map).
[0116] The review tasks are automatically assigned based on the experts’ areas of expertise and review priorities (priority order: emergency scenario samples to be reviewed > low abundance mixed samples + special markers (such as unknown sequence clusters accounting for more than 30%) to be reviewed > low abundance mixed samples without special markers to be reviewed > regular scenario samples to be reviewed). Experts who are good at low abundance identification will process low abundance scenario samples first, and emergency scenario samples to be reviewed will be given priority to enter the review queue.
[0117] Verify the review results submitted by experts, including format verification (checking whether the judgment basis, conclusions and evidence are complete) and logical verification (verifying whether the conclusions and reasons are consistent).
[0118] Standardization involves converting the verified review results into a unified format, including fields such as sample ID, final conclusion, reviewer, review time, and core reasons.
[0119] Finally, the expert review results and the automatic judgment results are integrated by sample ID to ensure that each sample has a unique final conclusion. The conclusions are standardized and coded (e.g., qualified = 01, unqualified = 02) to generate a standardized set of regulatory judgment conclusions. At the same time, key metadata is added to the set of regulatory judgment conclusions, including the version of the judgment rule set, the proportion of automatic judgment (i.e., the proportion of automatic judgment results in the total results), the review pass rate, etc.
[0120] Methods for constructing processing trajectory chains include:
[0121] Based on the sample ID in the regulatory judgment conclusion set, the technical fact report and the data block set after quality control are retrieved together. The corresponding data block is located by sample ID, and the complete original sequencing read after quality control is extracted.
[0122] Meanwhile, key judgment criteria were extracted from technical fact reports and regulatory judgment conclusions, including lists of transgenic elements, strain matching results, information on unknown sequence clusters, final conclusions (qualified or unqualified), judgment rule versions, and summaries of review opinions; then, each sample's reading segment was matched with the judgment criteria one by one.
[0123] For all reads extracted from each sample, a unique hash is generated for each read (e.g., a SHA-256 hash is calculated by combining the read sequence, sequencing timestamp, and sequencing platform model number, with the sequencing platform model number matching the platform information in the sample metadata), and the hash value is stored in a blockchain node for evidence preservation.
[0124] Centered on the key links of the entire sample process, a full-process processing trajectory chain is constructed in chronological order. The processing trajectory chain records information of key links, including but not limited to: data acquisition (platform, raw data reception time), quality control block (quality control rule version, data block ID), technical testing (computation node, completion time), judgment review (judgment rule version, expert ID, conclusion time), read segment hash (generation time, blockchain evidence storage node ID), and subsequent report generation information.
[0125] Integrate basic sample information (such as source and variety), sequencing parameters (such as platform and depth), judgment parameters (such as rule version and review opinions), hash evidence information, and other metadata to generate read coverage maps (used to show the read coverage depth of exogenous gene regions, such as the CaMV 35S promoter region coverage ≥20×), transgenic element location maps (marking the position of elements in the reference genome, such as the Cry1Ac gene located at 12000-15000bp on chromosome 1), and strain matching heatmaps (showing the distribution of element matching similarity between samples and known strains, with red indicating high-matching regions) and other visual evidence.
[0126] This process generates a standardized, credible initial draft of the authentication report (in both PDF and XML formats).
[0127] At the same time, a digital signature (generated based on the expert's private key) is added to the initial draft of the trusted authentication report in PDF format to ensure the integrity of the initial draft of the trusted authentication report.
[0128] Different methods of dividing versions for different scenarios and distributing them in a targeted manner include:
[0129] For the initial draft of the credible authentication report, a multi-dimensional credibility verification process is performed, including content consistency verification, format compliance verification, and hash traceability verification. Specifically:
[0130] Content consistency verification involves calling the data comparison interface to cross-verify the core information in the initial draft of the credible assessment report with the upstream data source: comparing the final conclusion in the initial draft of the credible assessment report with the sample conclusion of the regulatory judgment conclusion set (if unqualified, they must be completely consistent); comparing the metadata in the initial draft of the credible assessment report with the metadata of the data block and the rule set version; comparing the visualized evidence in the initial draft of the credible assessment report with the detection data in the technical fact report. If the deviation exceeds the expectation (e.g., >5%), the content is marked as abnormal, triggering the regeneration of the initial draft of the credible assessment report.
[0131] For format compliance verification, the initial draft of the PDF trusted authentication report is checked for the validity of the digital signature (whether the signature of the private key certified by the national CA can be verified) and the standardization of the page layout (e.g., the order of the four chapters: sample information - trajectory chain - conclusion - evidence). For the XML format report, the report is validated using a preset XML Schema tool (checking whether any fields are missing). If the format does not meet the requirements, the format repair tool is automatically triggered (e.g., PDF reflow, XML field completion).
[0132] Hash traceability verification involves extracting the hash value of the read segment from the initial draft of the trustworthy authentication report, and verifying whether the hash value is completely consistent with the hash of the evidence through a blockchain node query interface (such as peer nodes in the Fabric consortium chain); at the same time, a certain proportion (such as 10%) of the read segments are randomly selected, their hashes are recalculated and compared with the records in the initial draft of the trustworthy authentication report to ensure that the read segments have not been tampered with. If the verification is successful, a hash traceability success mark is marked on the report.
[0133] After verification, the credible identification report is adapted to meet the different needs of different users (regulatory authorities, testing laboratories, and enterprises), generating a scenario-based version, either a simplified regulatory version or a complete technical version, as a scenario-based report.
[0134] Specifically, the adaptation process involves: generating a simplified regulatory scenario report for regulatory authorities by extracting core information (sample ID, final conclusion, key evidence thumbnails, hash traceability identifiers), removing technical details (such as read length distribution, clustering parameters, etc.), and adopting a streamlined layout (e.g., one page, including a QR code that can be scanned to view the complete hash record on the blockchain) to suit the rapid verification scenarios of regulatory authorities; generating a complete technical report for laboratories by retaining all the content of the initial draft of the credible identification report (including detailed trajectory chain nodes, read hash list, and high-definition images of the original visual evidence), adding a technical notes section (marking the version of the testing tool, such as BWA-MEM v0.7.17, and details of the judgment rules) to suit the technical review scenarios of laboratories; and converting the initial draft of the credible identification report in XML format into a standardized API data format (JSON), splitting the fields according to sample basic information, traceability information, and conclusion information, and providing it to external systems (such as regulatory traceability platforms and laboratory LIMS systems) for use.
[0135] Based on the principle of least privilege and role-based access control (RBAC), scenario-based reports are distributed in a targeted, encrypted manner to different users, and distribution logs are recorded. Specifically:
[0136] Three types of authorized roles and permissions are preset, including supervisors (who can obtain the simplified regulatory version of all samples + the technical full version of a specified sample), laboratory operators (who can only obtain the technical full version of samples tested by their own laboratory), and enterprise users (who can only obtain the simplified regulatory version of samples submitted by themselves). Role permissions are managed through the RBAC (role-based access control) mechanism, and adding a new role requires administrator approval.
[0137] For authorized users, reports are distributed in different ways: For internal system users (supervisors, operators), a report access link is generated within the system. The link contains a valid token with a time limit. Clicking the link requires secondary authentication (such as SMS verification code); For external users (enterprises), reports are sent via encrypted email (such as AES-256 encrypted attachments). The email body contains a report verification code. The attachment can only be decrypted after successful verification.
[0138] After the recipient opens the report, the system automatically obtains information such as the receiving time, receiving device IP, and report version, and generates a distribution log (example: 2024-10-01 15:30, regulator Zhang, IP: 10.0.0.5, received sample TJ-soybean-20241001-001 regulatory simplified version). The log is stored in association with the report ID, supporting subsequent audit traceability.
[0139] After distribution, the report is archived and managed long-term, while the index is optimized to improve subsequent query efficiency.
[0140] The methods for extracting and clustering identification cases to generate optimization suggestions and feeding them back to the quality control rule base and the judgment rule base include:
[0141] According to a preset time period (such as monthly or quarterly, which can be adjusted according to business needs), batch extract relevant data of identification cases from the accumulated archived credible identification report library. This includes basic case data (sample ID, crop type, origin region, sequencing platform, final judgment conclusion (qualified, unqualified, or requiring further testing)), technical characteristic data (detected transgenic elements, strain matching results (matching strain ID, strain matching degree), number of unknown sequence clusters, reasons for needing to be reviewed (such as matching degree 0.8~0.9, incomplete element boundary)), and process performance data (quality control pass rate (percentage of qualified data blocks), automatic judgment rate (percentage of samples that do not require review), review pass rate (percentage of samples that maintain the original conclusion after review)).
[0142] Then, the relevant data is preprocessed, including deduplication (deleting duplicate reports and keeping the latest version), completion (adding missing key fields (such as reasons for pending review not being recorded) and log records related to the review process), and standardization (converting unstructured data into structured format (such as classifying and encoding reasons for pending review according to low matching degree, missing components, and ambiguous boundaries)).
[0143] Based on the preprocessed relevant data, a hierarchical dynamic knowledge base is constructed, with the core levels and content including:
[0144] The basic case layer stores complete identification data for a single sample, indexed by the sample ID, linking technical facts, judgment results, and traceability information to form a case-by-case structure that supports case retrospective analysis. The feature association layer uses association analysis tools (such as Tableau Prep) to establish association tables between elements, strains, conclusions, or regions, crops, and abnormal patterns (e.g., "Brazilian soybean + CaMV 35S element → 80% match MON87751 strain → unqualified conclusion"). The abnormal case layer separately categorizes sample data requiring further review and testing, records abnormal trigger scenarios (e.g., blurred element boundaries in low-abundance mixed samples, unknown sequence clusters exceeding 30%), clarifies further testing directions (e.g., blurred element boundaries → supplemental boundary sequence sequencing, high proportion of unknown sequence clusters → full-field sequencing of unknown sequences), and marks the final solution (e.g., supplemental sequencing, adjustment of judgment thresholds).
[0145] Data mining algorithms or machine learning models can be used to perform multi-dimensional mining and clustering of hierarchical dynamic knowledge bases to identify the direction of rule optimization; among them, the core analysis dimensions include conventional feature clustering, abnormal pattern clustering, and performance bottleneck clustering;
[0146] Conventional feature clustering (used to optimize the transgenic core reference library and basic rules) includes high-frequency element and variety clustering, and region-crop feature clustering, specifically:
[0147] High-frequency component and strain clustering: Clustering algorithms such as K-means are used to cluster component-strain combinations to identify combinations that occur frequently within a time period (e.g., CaMV 35S+Cry1Ac→MON87701). If a new combination occurs more than a predetermined number of times (e.g., 5 times) and is not included in the existing transgenic core reference library, it is marked as a reference sequence that needs to be supplemented.
[0148] Region-crop characteristic clustering: Group by origin region-crop type, and statistically analyze the distribution of judgment conclusions for each group (e.g., Argentine maize → failure rate 25%, higher than the average of 10%), and mark it as a crop in a region that needs special attention, so that targeted testing items can be added for subsequent quality control;
[0149] Anomaly pattern clustering (used to optimize the details of judgment rules), including clustering of correlations between reasons to be reviewed and clustering of process failure points, specifically:
[0150] Clustering of reasons to be reviewed: Algorithms such as Apriori for learning association rules can be used to analyze the correlation between reasons to be reviewed and scene labels, and optimize the specific parameters of the judgment rules; for example, if it is found that "low abundance mixed samples + strain matching degree 0.8~0.85 → 90% of the time after review, it is corrected to qualified", then it is marked as "the low abundance scene strain matching degree threshold needs to be lowered to 0.8".
[0151] Process failure point clustering: Hierarchical clustering is used to analyze the quality control links and the failure links. For example, if the statistics show that "Nanopore platform + single read length < 500bp percentage > 40% → quality control failure rate exceeds 30%", it is marked as "Nanopore quality control read length percentage threshold needs to be adjusted (original 60% → 55%)".
[0152] Performance bottleneck clustering (used to optimize process efficiency) is as follows:
[0153] For clustering samples with low automatic judgment rate, such as finding "a large number of regular samples (>50) + unknown sequence clusters that are not empty → automatic judgment rate of only 40%", mark them as "the fast screening rules for unknown sequence clusters need to be optimized (for example, optimize the clusters with less than 20 read segments to not trigger the review)", thereby improving the batch processing efficiency;
[0154] Based on the cluster analysis results, implementable optimization suggestions are generated. Each optimization suggestion should include the optimization object (quality control rule base or judgment rule base), the original rule content, the optimized content, and the basis (cluster result data).
[0155] For example, the optimization target is the judgment rule base (low abundance scenario); the original rule is: the strain matching degree threshold is 0.85; after optimization, it is 0.8; the basis is: in the low abundance mixture, the samples with a matching degree of 0.8~0.85 are corrected to qualified after 90% review, and the original threshold is too strict.
[0156] The effectiveness of the optimization suggestions is verified through small-batch testing and expert review. For example, 10% of similar samples (such as low-abundance mixed samples) are selected, and the rules are adjusted according to the optimization suggestions before re-judgment. The increase in the automatic judgment rate and the decrease in the verification rate are statistically analyzed (e.g., the automatic judgment rate increases from 50% to 75%, and the verification is passed). A review group composed of testing technology experts and regulatory business experts is invited to score the compliance (whether it meets regulatory standards) and technical feasibility (whether it is compatible with the existing platform) of the suggestions. Only if the average score exceeds the expected score (e.g., ≥80 points) can it proceed to the subsequent update stage.
[0157] Validated optimization suggestions that have passed verification will be updated to the quality control rule base and judgment rule base of the preceding stages, and a scene effect tracking mechanism will be established.
[0158] Specifically, the quality control rule base is updated by: updating threshold parameters (e.g., lowering the threshold for the proportion of Nanopore reads ≥500bp from 60% to 55%), and supplementing reference sequences (e.g., adding reference genome fragments for high-frequency element-line combinations); the judgment rule base is updated by: adjusting judgment thresholds (e.g., the line matching threshold for low abundance scenarios from 0.85 to 0.8), and supplementing special scenario rules (e.g., "Argentine maize requires additional testing of Cry2Ab elements"). After the update, a rule base version number (e.g., V202411) is generated, and the updated content and associated clustering basis are recorded.
[0159] The scenario effect tracking mechanism is as follows: In the subsequent identification process over multiple (e.g., two) time periods, the changes in statistical optimization indicators are monitored, including: in the quality control stage, whether the retest rate of unqualified data decreases (e.g., from 15% to 10%); in the judgment stage, whether the automatic judgment rate (i.e., the proportion of automatic judgment results to total judgment results) increases (e.g., from 65% to 80%) and the review dispute rate (i.e., the proportion of samples awaiting review decreases (e.g., from 20% to 8%); and in the anomaly handling stage, whether the proportion of samples requiring further testing decreases (e.g., from 12% to 5%), thereby forming a closed-loop iteration of mining-optimization-tracking.
[0160] Example 2
[0161] Please see Figure 3 As shown, for parts not described in detail in this embodiment, please refer to the description in Embodiment 1. A high-throughput sequencing-based intelligent identification system for transgenic crops is provided, including:
[0162] Data quality control and block segmentation module: acquires raw high-throughput sequencing data and sample metadata from multiple platforms, performs dynamic quality assessments on each platform, filters qualified data and divides it according to sample segmentation rules, and generates a data block set after quality control.
[0163] Data block technology detection module: Based on the quality control data block set, data blocks are allocated to multiple parallel computing nodes. Each computing node synchronously performs technology detection to obtain technical fact fragments, and then aggregates and supplements batch statistical information to generate a technical fact report.
[0164] Technical Fact Determination and Review Module: Based on the technical fact report, it matches the judgment rule set according to the pre-set decision tree, performs parallel judgment through a multi-instance rule engine, and automatically generates regulatory conclusions if the judgment result is clear. Expert review is triggered for boundary cases, and then the regulatory judgment conclusion set is integrated and generated.
[0165] Credible identification report generation and distribution module: Based on the regulatory judgment conclusion set and related technical fact reports, extract the original sequencing reads of all samples, generate a unique hash for each read, construct a processing trajectory chain, integrate metadata and visualized evidence, generate credible identification reports and divide them into different scenario-specific versions for targeted distribution;
[0166] Iterative optimization module: Based on the credible identification reports accumulated within the cycle, it extracts identification cases for mining and clustering, and generates optimization suggestions to be fed back to the quality control rule base and the judgment rule base.
[0167] Example 3
[0168] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the intelligent identification system for transgenic crops based on high-throughput sequencing described above.
[0169] Since the electronic device described in this embodiment is the electronic device used to implement the intelligent identification method for transgenic crops based on high-throughput sequencing in the embodiments of this application, those skilled in the art can understand the specific implementation and various variations of the electronic device in this embodiment based on the intelligent identification method for transgenic crops based on high-throughput sequencing described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any electronic device used by those skilled in the art to implement the intelligent identification method for transgenic crops based on high-throughput sequencing in the embodiments of this application falls within the scope of protection of this application.
[0170] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.
[0171] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A smart identification method for transgenic crops based on high-throughput sequencing, characterized in that, include: S1: Acquire raw high-throughput sequencing data and sample metadata from multiple platforms, conduct dynamic quality assessments for each platform, screen qualified data and divide them according to sample block rules, and generate a quality-controlled data block set. The sub-platform performs dynamic quality assessments and filters qualified data in the following ways: We acquire raw high-throughput sequencing data and corresponding sample metadata from multiple platforms, call the platform-specific quality control rule library to conduct dynamic quality assessment for each platform, and then classify the data into qualified data, unqualified data, and low-abundance sample data that need to be sequenced to increase the depth through multi-dimensional qualification judgment. For unqualified data, a retest is triggered. For low-abundance sample data that require increased sequencing depth, the sequencing depth is adjusted and a dynamic quality assessment is performed again. S2: Based on the data block set after quality control, data blocks are allocated to multiple parallel computing nodes. Each computing node synchronously performs technical detection to obtain technical fact fragments, and then aggregates and supplements batch statistical information to generate a technical fact report. The method of allocating data blocks to multiple parallel computing nodes, with each computing node synchronously performing technical detection to obtain technical fact fragments, includes: Based on the quality-controlled data block set, and combined with the currently available computing node resources, the data blocks are allocated to parallel computing nodes through load balancing scheduling. After receiving the data block, each computing node synchronously performs sequence alignment, component identification, strain matching evaluation, and unknown sequence clustering according to the platform type. Integrate technical test results to generate standardized fragments of technical facts; The methods for generating the technical fact report include: Using the unique identifier of a sample as the core key, technical fact fragments from each computing node are aggregated according to the sample dimension to form sample-level aggregated data; and batch statistical information is generated based on the sample-level aggregated data. Integrate sample-level aggregated data with batch statistical information to generate standardized technical fact reports; S3: Based on the technical fact report, match the judgment rule set according to the preset decision tree, and retrieve the judgment rule set from the judgment rule library. Perform parallel judgment through the multi-instance rule engine. If the judgment result is clear, automatically generate the regulatory conclusion. Trigger expert review for boundary cases, and then integrate to generate a set of regulatory judgment conclusions. S4: Based on the regulatory judgment conclusion set and the associated technical fact report, extract the original sequencing reads of all samples, generate a unique hash for each read, construct a processing trajectory chain, integrate metadata and visual evidence, generate a credible identification report and divide it into different scenario-specific versions for targeted distribution; S5: Based on the credible identification reports accumulated within the cycle, extract identification cases for mining and clustering, and generate optimization suggestions to be fed back to the quality control rule base and the judgment rule base.
2. The intelligent identification method for transgenic crops based on high-throughput sequencing according to claim 1, characterized in that, The method of dividing the data according to the sample block rules to generate the quality control data block set includes: Based on qualified data and associated sample metadata, a pre-set sample block rule library containing basic block rules and scenario adaptation rules is invoked. Based on the total number of samples and the detection scenario, the corresponding sample segmentation rules in the sample segmentation rule library are matched. Then, based on the qualified data, the original sequencing data and sample metadata of the corresponding samples are extracted and packaged to generate data blocks, which are then integrated into a quality control data block set.
3. The intelligent identification method for transgenic crops based on high-throughput sequencing according to claim 2, characterized in that, The method of performing parallel determination using a multi-instance rule engine based on the pre-set decision tree matching rule set includes: Analyze technical fact reports and extract scene labels, sample size features, and special data markers as scene features; The scene recognition engine loads a pre-set scene-rule set decision tree, judges the scene features in order of importance to match the unique corresponding judgment rule set, and generates a rule set-sample combination association table. Based on the association table, a multi-instance rule engine is launched, an independent engine instance is assigned to each rule set, and samples associated with the same rule set are divided into different sample groups and evenly distributed to each engine instance. Each engine instance synchronously performs judgment on the assigned sample group, obtaining automatically qualified samples, automatically unqualified samples, and judgment results pending review.
4. The intelligent identification method for transgenic crops based on high-throughput sequencing according to claim 3, characterized in that, The methods for generating the regulatory judgment conclusion set include: Automatically qualified and automatically unqualified samples are used as explicit automatic judgment results to directly generate regulatory conclusions; The sample to be reviewed is used as a boundary case to trigger the expert review mechanism, generate an independent review task sheet for each sample to be reviewed, and then assign it to experts according to their areas of expertise for standardized review. The expert review results are verified and standardized, and then integrated with the automatic judgment results to generate a standardized set of regulatory judgment conclusions.
5. The intelligent identification method for transgenic crops based on high-throughput sequencing according to claim 4, characterized in that, The methods for constructing the processing trajectory chain include: Based on the regulatory judgment conclusion set and technical fact report, the original sequencing reads of the sample and the judgment basis were extracted; Generate a unique hash for each read segment and store it on the blockchain, and build a full-process processing trajectory chain in chronological order; Integrate metadata and visual evidence from the entire sample process to generate a draft of a standardized and credible authentication report with digital signature.
6. The intelligent identification method for transgenic crops based on high-throughput sequencing according to claim 5, characterized in that, The method of dividing different scenario-specific versions and distributing them in a targeted manner includes: The initial draft of the credible assessment report undergoes multi-dimensional credibility verification to obtain a qualified credible assessment report. Then, according to the different needs of the users, the credible assessment report is adapted to generate a scenario-based version, either a simplified regulatory version or a complete technical version, as a scenario-based report. Based on the principle of least privilege, scenario-based reports are distributed in a targeted, encrypted manner to different users, and distribution logs are recorded.
7. The intelligent identification method for transgenic crops based on high-throughput sequencing according to claim 6, characterized in that, The methods for extracting and clustering identification cases to generate optimization suggestions and feeding them back to the quality control rule base and the judgment rule base include: According to a preset time period, batch extraction of identification case data is carried out from the accumulated credible identification reports to construct a hierarchical dynamic knowledge base of basic cases, feature associations, and abnormal cases. Multi-dimensional mining and clustering of the hierarchical dynamic knowledge base are performed to identify the direction of rule optimization. Based on the clustering analysis results, feasible optimization suggestions are generated and their effectiveness is verified. The quality control rule base and the judgment rule base are updated based on the effective optimization suggestions.