Method for rapid quarantine of imported grain weeds based on DNA barcoding of compositae

Abstract

The application relates to the technical field of plant quarantine, in particular to an imported grain weed rapid quarantine method based on DNA barcodes of compositae plants. The method aims to solve the problems of high missing detection rate and insufficient identification capacity when identifying highly degraded DNA samples. The scheme comprises the following steps: performing low-depth shotgun sequencing on total DNA of grain samples and filtering background sequences; extracting sequence and evolution features from short fragments by using a pre-trained double-channel deep feature extraction network; decoupling the fragment features into conservation, position coding and degradation features by using a decoupling-collaborative reconstruction network, and reconstructing a synthetic feature sequence representing a complete barcode; and finally comparing the sequence with a standard feature database, calculating an identification confidence, and outputting an identification conclusion. The application is especially suitable for rapid and accurate identification of compositae weed seeds with severely degraded DNA in imported grains.

Description

Technical Field

[0001] This application relates to the field of plant quarantine technology, and in particular to a rapid quarantine method for imported grains and weeds based on DNA barcoding of Asteraceae plants. Background Technology

[0002] Imported grains (such as soybeans, wheat, and corn) are a significant pathway for the introduction and spread of invasive weeds. During port quarantine, rapid and accurate identification of small, broken, or poorly shaped weed seeds mixed in with grains is often necessary. DNA barcoding technology is a powerful tool for species identification, utilizing sequence differences in standard gene fragments (such as ITS2, psbA-trnH) for identification.

[0003] However, existing DNA barcoding-based quarantine methods have significant limitations when applied to imported grain samples. Grain samples often undergo high temperatures, high humidity, and mechanical damage during processing, storage, and transportation, leading to severe degradation of the DNA from mixed weed seeds, breaking it down into short fragments of less than a few hundred base pairs. Traditional DNA barcoding PCR methods, requiring relatively long amplified fragments (typically >300 bp), are prone to amplification failure when faced with highly degraded DNA templates, resulting in false negatives and missed detections of harmful weeds.

[0004] To address the issue of DNA degradation, some studies have proposed using "mini-barcodes" (<150bp) for amplification. However, this method still relies on the efficient amplification and sequencing of a single short fragment, which contains limited genetic information, has insufficient resolution for identifying closely related species, and the entire identification process is interrupted if a single experiment fails, resulting in poor robustness. Other studies have employed macro-barcoding technology, performing high-throughput sequencing of the total DNA in the sample and directly comparing it with a database. However, when applied to grain samples, the sequencing data largely originates from the grain crop itself, severely diluting the DNA signal from the target weeds, resulting in a low signal-to-noise ratio. Furthermore, the analysis of massive amounts of data is time-consuming, making it difficult to meet the needs of rapid customs clearance at ports.

[0005] Therefore, existing technologies lack an effective method specifically designed to address the characteristics of highly degraded DNA and complex backgrounds in imported grain samples, enabling rapid and accurate quarantine identification of weeds. Summary of the Invention

[0006] The purpose of this invention is to overcome the technical shortcomings of existing DNA barcoding quarantine technologies, such as high false negative rates, insufficient identification capabilities, and poor process robustness, when processing highly degraded weed DNA in imported grains. This invention provides a rapid quarantine method for imported grain weeds based on Asteraceae plant DNA barcoding. This method can intelligently reconstruct complete barcode features from complex backgrounds and fragmented information, achieving rapid, accurate, and reliable identification of quarantine weeds in the Asteraceae family. To achieve the above objective, the embodiments of this application disclose the following technical solutions:

[0007] In a first aspect, embodiments of this application provide a rapid quarantine method for imported grains and weeds based on DNA barcoding of Asteraceae plants, the method comprising:

[0008] Total degraded plant DNA was obtained from imported grain samples and shotgun sequencing was performed to obtain the original short read sequence set;

[0009] The original set of short read sequences was compared with the reference sequences of the genomes of major cereal crops, and the matching sequences were filtered out to obtain a set of target fragments enriched with DNA from non-cereal plants.

[0010] Each short DNA sequence in the target fragment set is input into a pre-trained dual-channel deep feature extraction network. This network extracts local sequence patterns and phylogenetic context features through parallel sequence feature channels and evolutionary feature channels, respectively, and fuses them to generate a unified high-dimensional feature vector for each fragment.

[0011] The unified high-dimensional feature vector set of all fragments from the same sample is input into a decoupled-cooperative reconstruction network. The network first decouples the features of each fragment into conservative features, positional encoding features, and degenerate features through a decoupled attention module. Then, the cooperative reconstruction unit based on the Transformer architecture uses the conservative features and positional encoding features to reconstruct a synthetic feature sequence representing the complete barcode.

[0012] The synthesized feature sequence is compared with entries in the standard feature database of quarantine weeds of the Asteraceae family using deep feature space similarity and matching score calculation.

[0013] Based on the matching score results and the identification confidence calculated from the network's internal consistency, the final identification conclusion for Asteraceae weed species is output.

[0014] Optionally, in the step of "aligning the original set of short read sequences with the reference sequences of major cereal crops", a fast alignment algorithm based on k-mer hashing is used. The reference sequences of major cereal crops include the genome assembly sequences of maize, soybean, wheat and rice. A matching threshold is set. When the best matching coverage of a short read sequence on the reference sequence exceeds the threshold, it is determined to be a cereal background sequence and filtered.

[0015] Optionally, in the pre-trained dual-channel deep feature extraction network, the sequence feature channel is composed of a convolutional neural network containing multiple layers of one-dimensional convolution, pooling, and nonlinear activation layers; the evolutionary feature channel is composed of a small neural network based on an attention mechanism. The pre-training task of this network is to predict the approximate location distribution of an input DNA fragment sequence in a broad plant phylogenetic tree.

[0016] Optionally, the decoupled attention module is implemented using a multi-head attention mechanism, which includes at least three functionally decoupled attention heads: the first attention head extracts the conserved features by calculating the feature stability weights across segments; the second attention head generates the position-encoded features by analyzing the cross-correlation between segment features; and the third attention head separates the degradation features by identifying abnormal feature patterns that correspond to common DNA degradation damage patterns.

[0017] Optionally, the common DNA degradation damage patterns identified by the third attention head specifically include cytosine deamination due to high temperature or oxidation, and guanine oxidation due to mold growth. These patterns are learned by training the decoupled attention module on a large amount of simulated degradation data.

[0018] Optionally, the collaborative reconstructor is a Transformer module with a decoder structure, which uses the learnable full-length barcode position encoding as the query vector, and the combination of the conservative features and position encoding features as the key and value, and iteratively generates each site feature of the synthetic feature sequence through a multi-layer attention and feedforward network.

[0019] Optionally, the deep feature space similarity comparison and matching score calculation are performed through a pre-trained Siamese network. One branch of the Siamese network is input to the synthetic feature sequence, and the other branch is input to candidate features from the standard feature database. The network outputs a matching score. The identification confidence is calculated by an independent confidence evaluation module. The inputs of this module include the attention weight entropy value of the decoupled-cooperative reconstruction network during the reconstruction process, the average proportion of the degenerate features in the total features, and the difference between the first-ranked matching score and the second-ranked matching score output by the Siamese network.

[0020] Optionally, the method for constructing the standard characteristic database of quarantine weeds in the Asteraceae family is as follows:

[0021] Collect the full-length standard DNA barcode sequences of the target Asteraceae weed species;

[0022] For each full-length sequence, a set of degradation fragments with different degrees and modes is generated through simulation.

[0023] A pre-trained dual-channel deep feature extraction network is used to extract features from all segments corresponding to each simulated segment set. The network has parallel sequence feature extraction channels and evolution feature extraction channels.

[0024] The extracted features are input into a decoder network based on the Transformer architecture for reconstruction. The decoder network uses learnable positional encoding as the query and fragment features as the key and value to generate a "standard synthetic feature sequence" corresponding to the standard sequence.

[0025] The "standard synthetic feature sequence" corresponding to each standard sequence is stored in the database to form the standard feature database of quarantine weeds of the Asteraceae family.

[0026] Optionally, the method further includes a verification step for low-confidence results: if the output identification confidence is lower than a preset first threshold but higher than a lower second threshold, the system automatically triggers a droplet digital PCR verification experiment using specific primers for one or two candidate species that rank high in the identification results, and confirms or corrects the preliminary identification conclusion based on the quantitative detection results of the experiment.

[0027] Optionally, the shotgun sequencing depth is controlled between 0.5X and 2X, and before feature extraction, the target fragment set is filtered based on sequence length and average sequencing quality value, retaining only fragments with a length between 60bp and 250bp and an average quality value higher than Q20 for subsequent analysis.

[0028] The beneficial effects of this application are as follows:

[0029] (1) This application does not rely on the successful PCR amplification of any DNA fragment of a specific length. It uses low-depth shotgun sequencing to capture all degraded fragments unbiasedly, which fundamentally avoids the "false negative" problem caused by template degradation and significantly enhances the detection capability of highly degraded samples.

[0030] (2) The scheme design deeply integrates the characteristics of the quarantine scenario for imported grains. By filtering the background sequence of grains, the signal-to-noise ratio of the target signal is greatly improved; by identifying and separating specific chemical damage patterns (such as C-to-T), the adaptability to real degradation samples is enhanced; by setting sequencing depth and quality filtering parameters, an optimal balance is achieved between cost and efficiency.

[0031] (3) A confidence assessment system based on multiple dimensions such as attention entropy, proportion of degradation features, and matching score difference was introduced, which can quantitatively evaluate the reliability of each identification result. Combined with the microdroplet digital PCR verification step, a closed loop of "intelligent initial screening - confidence assessment - accurate verification" was formed, which provides more scientific and reliable decision support for port quarantine personnel and reduces the risk of misjudgment. Attached Figure Description

[0032] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0033] Figure 1 The overall flowchart of the rapid quarantine method for imported grains and weeds based on DNA barcoding of Asteraceae plants provided in this embodiment of the invention;

[0034] Figure 2 This is a schematic diagram of the structure of the dual-channel depth feature extraction network in an embodiment of the present invention. Detailed Implementation

[0035] Specific embodiments of the invention will now be described in detail. Although the invention is described in conjunction with these specific embodiments, it should be understood that it is not intended to limit the invention to these specific embodiments. Rather, these embodiments are intended to cover alternative, modified, or equivalent embodiments that may be included within the spirit and scope of the invention as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. The invention may be practiced without some or all of these specific details.

[0036] When used in conjunction with the terms "comprising," "method comprising," or similar language in this specification and appended claims, the singular forms "a," "some," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0037] The following is in conjunction with the appendix Figure 1-2 The preferred embodiments of the present invention will be described in detail so that those skilled in the art can implement the present invention accordingly.

[0038] A rapid quarantine method for imported grains and weeds based on DNA barcoding of Asteraceae plants, the method comprising:

[0039] S100. Obtain total degraded plant DNA from imported grain samples and perform shotgun sequencing to obtain the original short read sequence set.

[0040] In this process, the depth of shotgun sequencing is controlled between 0.5X and 2X. Before feature extraction, the target fragment set is filtered based on sequence length and average sequencing quality value, and only fragments with a length between 60bp and 250bp and an average quality value higher than Q20 are retained for subsequent analysis.

[0041] Specifically, obtaining total degraded plant DNA from imported grain samples is the starting point of this method. Specifically, a representative sample of approximately 50-100g is taken from the imported grains to be tested (e.g., soybeans, wheat, corn, etc.). The sample typically contains grain particles, dust, and mixed small plant debris (such as possible weed seed fragments). Total DNA extraction is performed using a modified plant genomic DNA extraction kit (e.g., based on the CTAB method or silica membrane adsorption method). Considering the potentially highly degraded and low-content target weed DNA in grain samples, vector RNA can be added during extraction to reduce adsorption loss of trace DNA on the tube wall, and the vigorous vortexing step, which may exacerbate DNA fragmentation, can be omitted or shortened to obtain a DNA mixture containing degraded fragments. The extracted DNA needs to be preliminarily determined for concentration and purity (A260 / A280 ratio) using a fluorometer or micro-ultraviolet spectrophotometer.

[0042] The obtained degraded total DNA was used for shotgun sequencing library preparation. Since the DNA was partially degraded, the mechanical fragmentation step could be omitted or shortened, and end repair, 'A' tail addition, and sequencing adapter ligation could be performed directly. If the DNA fragments were too long, non-focused ultrasound could be used for mild fragmentation, aiming to obtain fragments with the main peak in the range of 200-500 bp.

[0043] Sequencing was performed using high-throughput sequencing platforms (such as Illumina NovaSeq or MGI DNBSEQ-T7) for paired-end sequencing. The sequencing strategy employed "low-depth" sequencing, strictly controlling the sequencing depth between 0.5X and 2X. Here, "X" refers to coverage relative to a presumed 1Gb plant genome equivalent. For example, choosing a 1X depth means generating approximately 1Gb of raw sequencing data for a single sample. Since the biomass of target quarantine weeds in imported grains is typically extremely low (often below 0.1%), their DNA is present in trace amounts in the total DNA library. Traditional metagenomics often uses depths above 10X to achieve species coverage, resulting in massive amounts of redundant background (grain) data, high costs, and a heavy analytical burden. This study experimentally validated that depths of 0.5X-2X are sufficient to capture degraded DNA fragments of low-abundance target weeds with a high probability, while using the majority of sequencing data to cover the basic structure of the background community, achieving an optimal balance between cost and detection sensitivity.

[0044] Unlike deep metagenomic sequencing, which aims to comprehensively analyze community composition, and unlike high-depth amplicon sequencing, which targets specific barcode fragments, this strategy is a cost-effective and targeted data acquisition solution for scenarios where "specific target signals are detected from high background noise".

[0045] After sequencing, the raw short reads undergo standard preprocessing, including using tools like Fastp or Trimmomatic to remove sequencing adapter sequences and eliminate low-quality terminal bases (typically below Q20). Following this, a crucial two-step filtering process is performed to obtain a high-quality set of target fragments. First, only sequences between 60 bp and 250 bp in length are retained. The lower limit of 60 bp is set to ensure that the retained fragments contain the minimum amount of effective information. Fragments that are too short (e.g., <50 bp) carry limited sequence information, making reliable alignment or feature extraction difficult, and are more susceptible to sequencing errors and random matching interference. Filtering these fragments improves the overall signal-to-noise ratio of subsequent data analysis. The upper limit of 250 bp is set based on the assessment of degradation status. After severe degradation during processing and transportation, DNA is mostly broken into shorter fragments. Setting the upper limit to 250 bp focuses on DNA fragments most likely from the target degrading weeds, while also helping to exclude interference from long sequences originating from the intact genome.

[0046] Next, only sequences with an average sequencing quality score higher than Q20 were retained. Q20 is a commonly used quality assessment metric in high-throughput sequencing, indicating a base recognition error probability of no more than 1% (accuracy ≥ 99%). The calculation formula is: ,in This represents the probability of a base identification error. Therefore, Q20 corresponds to the error rate. (i.e., 1%). Using the average quality value for filtering can systematically eliminate unreliable reads due to the high error rate of sequencing itself, ensuring the baseline quality of the data input into subsequent complex deep learning models and avoiding noise interference with the model's learning of real biological signals.

[0047] S200. The original set of short read sequences is compared with the reference sequences of the genomes of major cereal crops. Matching sequences are filtered out to obtain a set of target fragments enriched with DNA from non-cereal plants.

[0048] In the step of comparing the original set of short read sequences with the reference sequences of major cereal crops' genomes, a fast alignment algorithm based on k-mer hashing is used. The reference sequences of major cereal crops' genomes include the genome assembly sequences of maize, soybean, wheat, and rice. A matching threshold is set, and when the best matching coverage of a short read sequence on the reference sequence exceeds the threshold, it is identified as a cereal background sequence and filtered out.

[0049] In imported grain samples, plant-derived DNA primarily comes from the grain crops themselves (accounting for over 95% of the total plant DNA), while the DNA of the target quarantine weeds is typically present in very low concentrations and fragmented. Without preprocessing, directly analyzing all sequences would completely mask the weak weed signals with the massive background noise of the grain sequences, leading to a sharp decrease in subsequent detection sensitivity and a waste of computational resources. Therefore, this step is a preprocessing stage that improves the signal-to-noise ratio of the entire method and focuses the analytical target.

[0050] Specifically, to achieve accurate filtering, a localized "major cereal crop genome reference sequence database" needs to be pre-constructed. This database should include at least the following: maize, using authoritative reference genome assembly sequences such as B73 and Zhengdan 958; soybean, using reference genomes such as Williams 82 and Zhonghuang 13; wheat, given its large genome and allohexaploid nature, representative spliced ​​sequences of subgenomes A, B, and D or the UniGene set can be used; and rice, using reference genomes such as Nipponbare and 9311.

[0051] These sequences should be downloaded from authoritative databases such as NCBI and Ensembl Plants and integrated into a unified index file. These four crops were chosen because they constitute the most important bulk grain commodities traded globally and in China, covering the vast majority of quarantine scenarios for imported grains. The database can be dynamically expanded based on statistical data of actual imported goods at ports (e.g., adding barley, sorghum, etc.).

[0052] To address the efficiency requirements of aligning millions to tens of millions of short reads with large genomes, this step employs a fast alignment algorithm based on k-mer hashing, rather than traditional base-by-base alignment tools such as BLAST. The algorithm (such as Minimap2, BWA-MEM, or a specially optimized Kraken2) first segments the reference genome into short words (k-mers) of fixed length k (e.g., k=15 or 21) and establishes a hash index from the k-mers to their positions in the genome. When processing sequencing reads, these are similarly converted into a set of k-mers. Possible matching positions are quickly found by querying the hash table, and the final positions are confirmed through seed expansion and exact alignment.

[0053] This method transforms the sequence alignment problem into an efficient hash lookup problem, achieving alignment speeds tens to hundreds of times faster than traditional methods, thus meeting the timeliness requirements of port quarantine. For individual base variations caused by degradation or sequencing errors, the k-mer seed chain-based alignment algorithm can still achieve robust matching.

[0054] The "matching threshold" specifically refers to the optimal matching coverage. The "matching threshold" is the percentage of a short read sequence that can be continuously and with high quality aligned to the reference genome, out of the total sequence length. Through experimental verification and statistical analysis, the threshold was set to 90%. Setting the "matching threshold" to this value is sufficient to accommodate incomplete alignments at both ends of the short read sequence due to decreased sequencing quality or minor intraspecific genetic variations (<5%) with the reference genome, thus avoiding the accidental deletion of fragments belonging to the target species but occasionally showing partial similarity to the cereal genome.

[0055] Each short read sequence is aligned. If it finds an optimal matching position on any reference sequence from corn, soybean, wheat, or rice, and the matching coverage is ≥90%, the sequence is considered to be highly likely to originate from that cereal crop, is classified as a "cereal background sequence," and is filtered out from subsequent analyses. Conversely, if all alignment coverage is <90%, the sequence is retained and added to the "target fragment set." The final result is a target fragment set enriched with DNA from non-cereal plants.

[0056] Understandably, after preprocessing, the original short read sequence set can remove approximately 70%-95% of the original sequence data (the removal percentage depends on the sample impurity content), significantly increasing the relative abundance of the previously diluted target weed DNA fragments. This greatly reduces the amount of data that needs to be input into subsequent computationally intensive deep learning models, improving the speed and efficiency of the overall analysis process. It also avoids potential cross-similarity interference from cereal crop genome sequences during backend feature extraction and comparison, improving the specificity of weed identification.

[0057] S300. Input each short DNA sequence in the target fragment set into a pre-trained dual-channel deep feature extraction network. This network extracts local sequence patterns and phylogenetic context features through parallel sequence feature channels and evolutionary feature channels, respectively, and fuses them to generate a unified high-dimensional feature vector for each fragment.

[0058] In the pre-trained dual-channel deep feature extraction network, the sequence feature channel consists of a convolutional neural network containing multiple layers of one-dimensional convolution, pooling, and non-linear activation layers; the evolutionary feature channel consists of a small neural network based on an attention mechanism. The pre-training task of this network is to predict the approximate location distribution of an input DNA fragment sequence in a broad plant phylogenetic tree.

[0059] It should be noted that a single feature perspective (such as looking only at sequence patterns) is insufficient to address the challenges of identifying closely related species and degrading noise. Therefore, this approach uses a pre-trained dual-channel deep feature extraction network. This network learns features in parallel from two complementary biological perspectives. The sequence feature channel focuses on microscopic sequence patterns such as local base arrangements and motif combinations within a fragment, answering the question, "What unique arrangement patterns does this sequence appear to have?" The evolutionary feature channel focuses on the fragment's context within the macroscopic tree of life, answering the question, "With whom might this sequence be more evolutionarily related?" Through parallel processing and fusion of these two channels, the generated feature vector simultaneously encodes both "morphological" and "lineage" information, significantly enhancing its discriminative power.

[0060] Specifically, the sequence feature channel is responsible for automatically learning discriminative local patterns from the original base sequence. The input to this channel is the encoding. First, each short DNA sequence is converted into a two-dimensional matrix. A common method is one-hot encoding, where each base (A, T, C, G) is encoded as [1,0,0,0], [0,1,0,0], [0,0,1,0], [0,0,0,1]. A sequence of length L thus forms an Lx4 matrix.

[0061] This channel consists of a convolutional neural network containing multiple layers of one-dimensional convolutional, pooling, and non-linear activation layers. The one-dimensional convolutional layers use multiple kernels of varying widths (e.g., kernels with widths of 4, 6, and 8) that slide along the sequence direction. Each kernel acts as a "pattern detector," automatically learning to recognize specific short base motifs during training. Shallow kernels may learn to recognize simple dinucleotide or trinucleotide preferences, while deeper kernels can combine these simple patterns to identify more complex conserved regions or feature structures. Pooling layers typically follow the convolutional layers, performing max pooling or average pooling operations. Their role is to reduce data dimensionality, expand the receptive field of subsequent convolutional layers, and make the network somewhat invariant to small positional shifts in feature patterns within the sequence. This is particularly important for handling degradation fragments whose starting positions may differ slightly due to different breakpoints. Non-linear activation layers, such as the ReLU function, are applied to the output of each convolutional layer, introducing non-linear transformation capabilities into the network, enabling it to fit complex feature mapping relationships.

[0062] After several layers of such operations, the network ends with a flattening layer and a fully connected layer to compress the learned high-level sequence patterns into a fixed-dimensional vector (such as 128-dimensional), which is the sequence feature vector.

[0063] The network of evolutionary feature channels was first pre-trained on a large-scale dataset different from the final quarantine task. This dataset contained standard DNA barcode fragments from a large number of species across a broad plant phylogenetic tree (covering hundreds of families and thousands of genera). The pre-training task was to predict which major branch (e.g., family or genus-level node) in the phylogenetic tree a given DNA fragment sequence most likely originated from.

[0064] This network employs an attention-based architecture (such as a Transformer encoder or a BiLSTM with attention pooling). The attention mechanism allows the network to dynamically focus on the base regions in the input segment that are most critical for determining its phylogenetic position, rather than treating all positions equally. For example, when identifying plants in the Asteraceae family, attention might be more focused on certain hypervariable regions.

[0065] By completing the pre-training task described above, the network's internal parameters learn to map DNA sequences into a latent space that reflects their phylogenetic relationships. After pre-training, all parts of the network except the final classification layer (i.e., the "feature extractor") are retained. When a new, unknown short fragment is input, the hidden layer representation output by the network (e.g., a 64-dimensional vector) becomes the evolutionary feature vector. This vector is not directly a classification label, but a dense representation containing information such as "whether this fragment is more evolutionarily similar to plants in the Asteraceae family or to those in the Poaceae family."

[0066] Feature vectors from two channels (such as 128-dimensional sequence features and 64-dimensional evolutionary features) are passed to a fusion layer. Fusion is typically achieved through concatenation or weighted summation. For example, directly concatenating the two vectors yields a 192-dimensional fusion vector. This fusion vector may then undergo non-linear transformations and dimensionality adjustments through one or more fully connected layers, ultimately generating a uniform, high-dimensional (e.g., 256-dimensional) feature vector for each segment.

[0067] Understandably, the dual-channel design combines local fine-grained mode with global evolutionary context, making the generated features more robust to distinguishing closely related species and to single-channel noise (such as individual sequencing errors in sequence features). Pre-trained dual-channel deep networks solve the challenge of extracting highly discriminative features from short, degraded DNA fragments with limited information.

[0068] S400. Input a unified set of high-dimensional feature vectors from all segments of the same sample into a decoupled-cooperative reconstruction network. The network first decouples the features of each segment into conservative features, positional encoding features, and degenerate features through a decoupled attention module. Then, the cooperative reconstruction unit based on the Transformer architecture uses the conservative features and positional encoding features to reconstruct a synthetic feature sequence representing the complete barcode.

[0069] The decoupled attention module employs a multi-head attention mechanism, comprising at least three functionally decoupled attention heads: the first attention head extracts conserved features by calculating feature stability weights across segments; the second attention head generates position-encoded features by analyzing the cross-correlation between segment features; and the third attention head separates degenerate features by identifying anomalous feature patterns consistent with common DNA degradation damage patterns. The collaborative reconstructor is a Transformer module with a decoder structure. It uses a learnable full-length barcode position code as the query vector and a combination of conserved features and position-encoded features as keys and values. Through multi-layer attention and feedforward networks, it iteratively generates site features for each position in the synthesized feature sequence.

[0070] Furthermore, the common DNA degradation damage patterns identified by the third attention head include cytosine deamination due to high temperature or oxidation, and guanine oxidation due to mold growth. These patterns are learned by training the decoupled attention module on a large amount of simulated degradation data.

[0071] Specifically, the output of step S300 is a large number of discrete, unordered fragment feature vectors. The goal of step S400 is to virtually stitch together and complete the optimal digital representation of the original complete barcode in the feature space, based solely on the feature patterns and potential relative positional clues carried by these feature vectors, without relying on the physical overlap sequence between them. Its logic comprises two key stages: a decoupling stage and a reconstruction stage. The decoupling stage performs "component analysis" on each fragment (fragment feature), separating its mixed information into sub-components with different attributes. The reconstruction stage utilizes these effective sub-components to re-"draw" the complete barcode according to the inferred spatial relationships.

[0072] The decoupled attention module is implemented using a multi-head attention mechanism, which includes at least three functionally decoupled attention heads that process the input fragment feature vector set in parallel. .

[0073] The first attention focus is on conservative feature extraction. For the current segment features... The attention head calculates its features in relation to all other fragment features in set F. Similarity (as attention weight). Attention weight The calculation can be simplified to: in, and It is a learnable projection matrix. It refers to the feature dimension. If a certain feature pattern exists in multiple... Zhongdu and Highly similar, then these This will gain high weight. Then, the head will be applied to all... The features are weighted and summed, and the aggregation result is then compared with... The product itself is integrated, and the final output is... Corresponding conservative eigenvectors This process allows the network to "listen to the majority," reinforcing characteristic signals that are supported by multiple fragments.

[0074] The second attention head generates positional encoding features. This head learns the long-range dependencies and co-occurrence relationships between feature segments. The second attention head analyzes... and Features to determine and Whether they tend to appear at a specific relative distance within the complete sequence. For example, through learning, the second attention head discovers that feature patterns A and B are always about 100 "feature units" apart. When segments carrying both patterns A and B exist in the input, even if the feature segments do not overlap, the head can assign them position-encoded feature vectors that reflect this relative distance relationship. and This encoding is not a specific base index, but a direction vector in a high-dimensional space, used to guide spatial alignment in subsequent reconstruction.

[0075] The third attention head performs degradation feature separation. This head is specifically trained to act as a "damage pattern filter". During the training phase, the third attention head is exposed to a large amount of simulated degradation data, which generates cytosine deamination patterns (C-to-U / T) and ine oxidation patterns (G-to-A, etc. caused by 8-oxoG) by programmatically introducing specific chemical damage patterns into clean sequences.

[0076] The cytosine deamination pattern (C-to-U / T) simulates damage caused by high temperature, high pH, ​​or prolonged storage. In the training data, random Cs are replaced with Ts (because in sequencing, Us on DNA are read as Ts). The guanine oxidation pattern (G-to-A, etc., caused by 8-oxoG) simulates damage caused by mold growth and oxidative stress. In the training data, random Gs are replaced with As or other bases to simulate 8-oxoguanine mismatches.

[0077] Through learning, the parameters within this attention head form a feature "memory" of these anomalous patterns. During forward propagation, for the input features... The head calculates the "abnormal match" between itself and these internal memories, and separates the components with high match rates to form a degenerate feature vector. For example, if a fragment contains a T at a position that should be C in a conserved region, and this variation does not conform to any known single nucleotide polymorphism (SNP), the head can capture this anomalous feature. middle.

[0078] After decoupling, we obtain the purified set of conservative features. and corresponding position codes The task of the collaborative refactorer is to... Using raw materials, generate a length of Synthetic feature sequence of (target barcode feature sequence length) The collaborative refactorer is a Transformer module with a decoder structure, and its workflow is as follows:

[0079] The learnable query vector Q, i.e., the initialization... A learnable vector Each vector Corresponding to the complete sequence number The "problem" at the 1st site, namely the "1st" What characteristics should this location have? AND value It is the conservatism of each segment. Its positional encoding Add to form . indicates "a feature" Approximately located The indicated area.

[0080] For each target location Its query vector Will be key to all fragments Calculate attention weights. Location encoding. Plays a crucial role here: if Indicated virtual location and target location Similar, then and The higher the matching degree, the greater the attention weight will be.

[0081] The calculated attention weights are used for all values (Right now We perform a weighted summation to obtain a preliminary aggregate feature.

[0082] This aggregated feature undergoes deep processing in the decoder, including a feedforward neural network, layer normalization, and residual connections. It may also consider previously generated positional features (achieved through masked self-attention), ultimately outputting the synthetic feature for that position. .

[0083] For all Repeat this process at each position to generate a synthetic feature sequence representing the complete barcode. This synthetic characteristic sequence It is The matrix (D is the feature dimension) is the optimal continuous feature representation of the complete barcode derived from fragmented information.

[0084] Understandably, this step represents a paradigm shift from "overlap-based physical assembly" to "feature-based virtual reconstruction," fundamentally overcoming the technical obstacle of being unable to splice fragments due to their short length and lack of overlapping areas. Through decoupling operations, the limited mixed information (conservative information, location clues, and noise features) within short fragments is mined and utilized to the maximum extent, maximizing information utilization efficiency. A specially designed degradation feature separation module endows the method with inherent robustness to real damage (C deammoniation, G oxidation) during the storage and transportation of imported grains, significantly improving the identification accuracy on complex real-world samples.

[0085] S500. Perform deep feature space similarity comparison and matching score calculation between the synthesized feature sequence and the entries in the standard feature database of quarantine weeds of the Asteraceae family.

[0086] The deep feature space similarity comparison and matching score calculation are performed through a pre-trained Siamese network. One branch of the Siamese network is input with the synthetic feature sequence, and the other branch is input with candidate features from the standard feature database. The network outputs a matching score. The identification confidence is calculated by an independent confidence evaluation module. The input of this module includes the attention weight entropy value of the decoupled-cooperative reconstruction network during the reconstruction process, the average proportion of degenerate features in the total features, and the difference between the first-ranked matching score and the second-ranked matching score output by the Siamese network.

[0087] The construction method of the standard feature database for quarantine weeds in the Asteraceae family is as follows: Collect the full-length standard DNA barcode sequences of the target Asteraceae weed species; simulate and generate multiple sets of degradation fragments of different degrees and patterns for each full-length sequence; use a pre-trained dual-channel deep feature extraction network to extract features from all fragments corresponding to each simulated fragment set, with the network having parallel sequence feature extraction channels and evolutionary feature extraction channels; input the extracted features into a decoder network based on a Transformer architecture for reconstruction, where the decoder network uses learnable positional encoding as the query and fragment features as the key and value to generate the "standard synthetic feature sequence" corresponding to the standard sequence; store the "standard synthetic feature sequences" corresponding to each standard sequence in the database to form the standard feature database for quarantine weeds in the Asteraceae family.

[0088] Specifically, the Siamese network architecture employs a two-branch neural network with shared weights. The two branches have identical structures and parameters. One branch receives the synthetic feature sequence to be detected as input. (From S400), another branch inputs a standard synthetic feature sequence from the database. The two feature sequences are processed through the same multi-layer nonlinear transformation (such as fully connected layers and attention layers). The goal of the network is to learn a function that makes the features of the same species... The goal is to minimize the distance between pairs of different species in the high-dimensional space after mapping, while maximizing the distance between pairs of different species. Finally, the network outputs a matching score. It is usually normalized to the interval [0, 1], where 1 represents that they are exactly the same.

[0089] The network is pre-trained on a dataset consisting of positive sample pairs (characteristic sequences of the same species) and negative sample pairs (characteristic sequences of different species) of known species. The loss function often adopts contrastive loss or triplet loss to bring positive sample pairs closer and push negative sample pairs further apart.

[0090] To quantify the reliability of a single identification result, a separate confidence assessment module is introduced. Its input consists of three internal indicators that reflect the quality of this analysis from different perspectives: attention weight entropy value, average percentage of degenerate features, and the difference between the first-ranked and second-ranked matching scores.

[0091] The attention weight entropy value is derived from the attention distribution in the S400 collaborative reconstructor, and is calculated using the following formula:

[0092]

[0093] in, It represents the probability distribution of attention across all input segments when generating features at a certain location. Low entropy values ​​indicate highly focused attention (clear decision-making) during reconstruction, while high entropy values ​​indicate scattered attention (fuzzy decision-making).

[0094] The average percentage of degraded features is the ratio of the total number of degraded features decoupled from the S400 to the total number of features. A high percentage indicates severe DNA damage in the sample, poor input data quality, and may affect the reliability of the results. This is an indicator reflecting the quality of the input data.

[0095] The difference between the top-ranked and second-ranked matching scores represents the difference between the best and second-best matching scores given by the Siamese network. A large difference indicates that the best-matched species has a significant advantage, and the identification conclusion is clear; a small difference indicates the existence of multiple highly similar candidates, and the conclusion is ambiguous. This is an indicator reflecting the discriminative power of the results.

[0096] The confidence assessment module (which can be a regression model or a small neural network) integrates these indicators and outputs a comprehensive confidence score. For example, low entropy, low degradation ratio, and large score difference will correspond to a high confidence score (e.g., 0.9).

[0097] The database of standard characteristics of quarantine weeds in the Asteraceae family was generated offline. The construction process involved collecting full-length barcode sequences of morphologically confirmed target quarantine weeds in the Asteraceae family (such as ragweed, three-lobed ragweed, and Italian cocklebur) from authoritative databases (such as NCBI and BOLD). For each full-length sequence, bioinformatics tools were used to simulate the actual degradation it might undergo during the storage and transportation of imported grains: simulating mild, moderate, and severe degradation, corresponding to different random fragment length distributions (e.g., main peaks at 200bp, 150bp, and 100bp, respectively). Specific chemical damage was also introduced programmatically. For example, cytosine (C) was replaced with thymine (T) with a certain probability to simulate deamination; guanine (G) was replaced with adenine (A) or other bases to simulate oxidative damage. This generated a large number of short fragment mixtures that appeared "realistic." For each set of simulated degradation fragments, the process of S300 and S400 is completely replicated: using the same pre-trained dual-channel deep feature extraction network, features are extracted for each simulated fragment. The feature set is then input into a structurally identical, parameter-fixed "decoupled-cooperative reconstruction network" (i.e., a frozen model for generating standards) to generate a standard synthetic feature sequence. Next, the "standard synthetic feature sequence" corresponding to each species, along with its metadata (species name, barcode region, origin, etc.), is stored in a database.

[0098] S600: Based on the matching score results and the identification confidence calculated from the internal consistency of the network, output the final identification conclusion of the Asteraceae weed species.

[0099] Specifically, this step employs a hierarchical decision-making logic based on dual thresholds. The logic is as follows: when the best match score is greater than the species determination threshold (e.g., 0.8) and the identification confidence is greater than the high confidence threshold (e.g., 0.7), a definitive identification is made directly. When the best match score exceeds the species determination threshold, but the identification confidence is between the high and low thresholds (e.g., 0.3 ≤ confidence ≤ 0.7), the system provides a presumption and automatically triggers the verification process. When the best match score is lower than the species determination threshold, or the identification confidence is too low (e.g., <0.3), the system determines that a reliable identification cannot be given.

[0100] The final output is a structured report, including: core conclusions: clear species identification results or "not detected" statements; key evidence: best match score and identification confidence level; ancillary information: possibly listing 1-2 second-best match species and their scores for expert reference; and decision recommendations: standardized operational recommendations based on the conclusions (such as "release," "detention," or "manual review").

[0101] It should be noted that the various thresholds in the decision-making logic (species identification threshold, high / low confidence threshold) are not fixed values, but key parameters that need to be determined through extensive validation experiments and system optimization. The determination method involves testing the entire system using standard samples or simulated data containing known species and different degrees of degradation. The optimal matching score threshold, which maximizes the true positive rate and minimizes the false positive rate, is determined by plotting the receiver operating characteristic (ROC) curve. The confidence threshold is determined by analyzing the correlation between the confidence score and the actual error rate of the identification results; for example, a confidence score of 0.7 is set as the critical point for an error rate below 1%.

[0102] Furthermore, these thresholds allow management to adjust them based on different levels of quarantine risk tolerance. For example, during periods of high risk of an outbreak, the species determination threshold can be lowered to increase screening sensitivity (but may increase false positives), or the confidence threshold can be raised to require more conclusive evidence.

[0103] Furthermore, the above method also includes a verification step for low-confidence results: if the output identification confidence is lower than a preset first threshold but higher than a lower second threshold, the system automatically triggers a droplet digital PCR verification experiment for one or two candidate species ranked first in the identification results using their specific primers, and confirms or corrects the preliminary identification conclusion based on the quantitative detection results of the experiment.

[0104] Specifically, the first threshold represents the lower limit of "high confidence." A confidence level above this value indicates the algorithm's results are highly reliable and can be directly accepted without verification. The second threshold represents the upper limit of "extremely unreliable results." A confidence level below this value indicates the analysis is invalid due to extremely poor data quality or lack of clear matching; verification is meaningless, and the result should be directly judged as "unidentifiable" or resampling should be recommended. The verification procedure is initiated when the confidence level falls within the verification interval (i.e., between the first and second thresholds).

[0105] The entire validation process is highly automated, forming a command flow from digital signals to physical experiments. When the algorithm module outputs a result falling within the validation range, the system software automatically generates a validation task instruction, which includes the sample ID and the names of the top 1-2 candidate species. The system automatically retrieves species-specific primers and probes for these candidate species from its internal primer database (e.g., a primer-probe set designed for the ITS2 region of ragweed that produces an 80bp product). This ensures the validation experiment is highly targeted and avoids non-specific amplification. The automated liquid handling system receives the instruction, aliquots an appropriate amount from the original sample DNA retention solution, mixes it with the selected ddPCR premix, primers, and probes, generating tens of thousands of nano-liter water-in-oil droplets. PCR amplification is then performed. Each droplet acts as an independent reaction unit, and at the endpoint, the fluorescence signal of each droplet is read to determine whether it is "positive" (containing target DNA) or "negative".

[0106] ddPCR was chosen instead of conventional qPCR for validation due to its unique advantages in low-concentration, complex background samples. By counting the number of positive droplets, the absolute copy number of the target DNA template (in copies / µL) can be directly calculated, and the result is independent of the standard curve, making it objective and irrefutable. After splitting the reaction system, the PCR inhibitor is diluted in each droplet, significantly reducing its overall impact on amplification efficiency, making it particularly suitable for detecting trace targets in complex matrices (such as grain extracts). The quantitative result is an objective physical count of "present" or "absent". For example, a result showing "target DNA: 5.2 copies / µL, negative control: 0.1 copies / µL" clearly confirms the presence of the target species.

[0107] The quantitative results of ddPCR are the basis for the final decision. If the ddPCR test result for the top-ranked candidate species is positive (and the copy number is significantly higher than the preset positive threshold, such as >3 copies / µL), while the test result for the second-ranked species is negative, the system confirms the algorithm's preliminary identification conclusion. The final report will be updated to: "[Species X] confirmed to be detected by droplet digital PCR." However, two correction scenarios are possible. The first is: if the ddPCR result for the top-ranked species is negative, while the result for the second-ranked species is positive, the conclusion is revised to detect the second-ranked species. The second is: if the ddPCR results for all candidate species are negative, the algorithm's preliminary positive judgment is overturned, and the conclusion is revised to "not detected" or "the detected substance is inconsistent with expectations, further investigation is recommended."

[0108] The following specific embodiments will be used to illustrate a complete quarantine method according to an embodiment of the present invention, specifically including the following steps:

[0109] From grains (such as a batch of soybeans) intercepted at the port of entry, approximately 100g of impurity samples were randomly collected. Total DNA was extracted from the samples using a plant genomic DNA extraction kit (such as the CTAB method). The obtained DNA was randomly fragmented using an ultrasonic disruptor to construct a small DNA library. The library was sequenced using a shotgun sequencing platform (such as Illumina NovaSeq) with a paired-end 150bp (PE150) sequencing strategy, and the sequencing data volume of a single sample was controlled to approximately 1.5Gb, i.e., achieving 1X coverage of approximately 1.5Gb of the hypothetical genome. The raw sequencing data was processed using standard bioinformatics procedures such as adapter and low-quality end removal to obtain a high-quality raw short read sequence set.

[0110] The obtained raw short read sequences were aligned with locally constructed reference genomes of major cereal crops (maize (B73), soybean (Williams 82), wheat (Chinese Spring), and rice (Nipponbare)) using fast alignment tools based on k-mer hashing, such as Minimap2. A matching threshold of 90% was set. If more than 90% of a read could be continuously aligned to any cereal genome, it was considered a cereal background sequence and filtered out. This step can filter out approximately 85%-95% of the sequences.

[0111] The remaining "target fragment set" is then subjected to quality filtering: sequences shorter than 60 bp or longer than 250 bp are discarded; the average sequencing quality score (Phred score) for each sequence is calculated, and sequences with an average quality score not higher than Q20 (i.e., a base recognition error rate ≥1%) are discarded. This step yields a clean, high-quality set of short DNA fragments from non-cereal plants suitable for subsequent in-depth analysis.

[0112] Each filtered short DNA sequence (e.g., a 120bp sequence) is input into a pre-trained dual-channel deep feature extraction network.

[0113] The sequence feature channel is a one-dimensional convolutional neural network (1D-CNN). First, the DNA sequence is one-hot encoded (A: [1,0,0,0], T: [0,1,0,0], C: [0,0,1,0], G: [0,0,0,1]). The network contains three convolutional layers (kernel sizes of 6, 8, and 10), each followed by a ReLU activation function and a max-pooling layer. Finally, a fully connected layer outputs a 128-dimensional feature vector, representing the local sequence pattern of the fragment.

[0114] The evolutionary feature channel is a small network based on a Transformer encoder structure. Its pre-training task is to learn to predict the approximate node position (e.g., probability distribution of which family or genus it belongs to) of a short input sequence fragment on a dataset containing tens of thousands of plant ITS2 sequences and their corresponding phylogenetic tree positions. The channel outputs a 64-dimensional feature vector representing the phylogenetic context of that fragment.

[0115] The 128-dimensional and 64-dimensional vectors output from the two channels are concatenated to obtain a 192-dimensional vector, which is then mapped through a fully connected layer to the final unified high-dimensional feature vector (e.g., 256-dimensional). Thus, each short DNA fragment is transformed into a 256-dimensional mathematical vector with rich semantic information.

[0116] The 256-dimensional feature vectors of all segments from the same sample are aggregated and input into the decoupled-cooperative reconstruction network. The network first processes these features through a decoupled attention module. This module contains three functionally independent attention heads:

[0117] Conservative attention heads are used to calculate the similarity (stability) of features of each segment with the features of all other segments within the sample. High-similarity feature patterns that consistently appear across multiple segments are reinforced and extracted as conserved features (e.g., 128-dimensional). These features are considered to be the most likely to represent species-specific information.

[0118] A positional attention head analyzes the cross-correlation relationships between features of different segments. For example, feature pattern A and feature pattern B are typically a fixed distance apart in a known complete sequence. Based on this learned spatial constraint, the head generates a positional encoding feature (e.g., 64-dimensional) for each segment, indicating the segment's relative "virtual position" within the complete barcode.

[0119] A degenerate attention head is trained to identify anomalous feature patterns caused by specific chemical damage (such as C-to-T, G-to-A). It isolates these noisy features unrelated to biological variation, forming degenerate features (e.g., 64-dimensional). These features are discarded in subsequent reconstructions.

[0120] The decoupled conservative features and positional encoding features are combined as keys and values. A learnable query vector sequence is defined, with a length consistent with the feature sequence length of the target complete barcode (e.g., 500 query vectors for a 500bp barcode). These queries, keys, and values ​​are input into a Transformer decoder (co-reconstructor). The decoder uses multi-layer self-attention and cross-attention mechanisms to guide each query vector (corresponding to a position in the complete barcode) to "focus" on the fragment features most relevant to its position (guided by positional encoding), and aggregates the conservative information of these features. After transformation by a feedforward network, a complete synthetic feature sequence (e.g., a 500x256 dimensional matrix) is finally generated iteratively, representing the complete digital feature map of the barcode reconstructed from fragmented information.

[0121] The generated synthetic feature sequences were compared with entries in the standard feature database of quarantine weeds in the Asteraceae family. The database was constructed as follows: Full-length standard ITS2 barcode sequences of Asteraceae weeds (such as ragweed, three-lobed ragweed, and Italian cocklebur) listed in the "List of Quarantine Pests of Plants Entering the People's Republic of China" were collected. A computer program simulated grain storage and transportation environments (high temperature, high humidity, mold) to generate hundreds of thousands of simulated degradation fragments, ranging in length from 60-250 bp, with random C-to-T and G-to-A damage, for each full-length sequence. Using the same feature extraction and reconstruction network (with fixed parameters) as in steps three and four, these simulated fragments were processed to generate a corresponding "standard synthetic feature sequence" for each species, which was then stored in the database. During comparison, a pre-trained Siamese network was used. The two branches of this network were respectively input to the synthetic feature sequence to be tested and a standard feature sequence from the database, outputting a matching score between 0 and 1, with higher scores indicating higher similarity.

[0122] Simultaneously, a separate confidence assessment module begins operation. It receives three internal metrics: 1) the entropy of the attention weights during reconstruction (measuring the certainty of the reconstruction decision); 2) the average proportion of degraded features among all features (measuring the degree of sample degradation); and 3) the difference between the highest and second-highest matching scores output by the Siamese network (measuring the discriminative power of the identification result). This module (which can be a simple linear regression model or a small neural network) synthesizes these metrics and outputs an identification confidence score for this identification result (e.g., 0.85, ranging from 0 to 1).

[0123] The system combines the matching score and the identification confidence level to output the final identification conclusion. For example: "Identified as ragweed, matching score 0.92, confidence level 0.85".

[0124] As a closed loop for quality assurance, this method also includes a verification step: Confidence thresholds are pre-set, for example, a first threshold of 0.7 and a second threshold of 0.3. If the confidence level of an identification output is 0.55 (between 0.3 and 0.7), it is marked as a "low confidence result." The system automatically calls the laboratory automation system to perform droplet digital PCR experiments on the top two candidate species (such as ragweed and three-lobed ragweed) using their species-specific primers and probes. If the ddPCR results show that the number of positive droplets for the ragweed probe is significantly higher than the negative control background, while the three-lobed ragweed probe is negative, the system will correct the final conclusion to "Ragweed detected, confirmed by ddPCR verification." Conversely, if all ddPCR results are negative, it may be judged as a false positive or an unknown species, indicating the need for manual verification.

[0125] The method provided in this invention creatively solves the technical bottleneck caused by DNA degradation in the quarantine of imported grains by deeply integrating deep learning and bioinformatics technologies. The entire process is based on practical application scenarios and possesses high relevance, robustness, and automation potential, providing an effective technical solution for achieving rapid, accurate, and reliable intelligent plant quarantine.

[0126] The detailed description of the above specific embodiments fully illustrates the feasibility, preferred implementation, and technical effects achieved by the technical solution of the present invention. Those skilled in the art can make several modifications and substitutions based on the above description without departing from the principles and spirit of the present invention, and these modifications and substitutions should also be considered within the scope of protection of the present invention.

[0127] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention. The actual method is not limited to this. In conclusion, if those skilled in the art are inspired by this description and design similar methods and embodiments without departing from the spirit of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A rapid quarantine method for imported grains and weeds based on DNA barcoding of Asteraceae plants, characterized in that, The method includes: Total degraded plant DNA was obtained from imported grain samples and shotgun sequencing was performed to obtain the original short read sequence set; The original set of short read sequences was compared with the reference sequences of the genomes of major cereal crops, and the matching sequences were filtered out to obtain a set of target fragments enriched with DNA from non-cereal plants. Each short DNA sequence in the target fragment set is input into a pre-trained dual-channel deep feature extraction network. This network extracts local sequence patterns and phylogenetic context features through parallel sequence feature channels and evolutionary feature channels, respectively, and fuses them to generate a unified high-dimensional feature vector for each fragment. The unified high-dimensional feature vector set of all fragments from the same sample is input into a decoupled-cooperative reconstruction network. The network first decouples the features of each fragment into conservative features, positional encoding features, and degenerate features through a decoupled attention module. Then, the cooperative reconstruction unit based on the Transformer architecture uses the conservative features and positional encoding features to reconstruct a synthetic feature sequence representing the complete barcode. The synthesized feature sequence is compared with entries in the standard feature database of quarantine weeds of the Asteraceae family using deep feature space similarity and matching score calculation. Based on the matching score results and the identification confidence calculated from the network's internal consistency, the final identification conclusion for Asteraceae weed species is output.

2. The method according to claim 1, characterized in that, In the step of "aligning the original short read sequence set with the reference genome sequences of major cereal crops", a fast alignment algorithm based on k-mer hashing is used. The reference genome sequences of major cereal crops include the genome assembly sequences of maize, soybean, wheat and rice. A matching threshold is set. When the best matching coverage of a short read sequence on the reference sequence exceeds the threshold, it is determined to be a cereal background sequence and filtered.

3. The method according to claim 1, characterized in that, In the pre-trained dual-channel deep feature extraction network, the sequence feature channel consists of a convolutional neural network containing multiple layers of one-dimensional convolution, pooling, and nonlinear activation layers; the evolutionary feature channel consists of a small neural network based on an attention mechanism. The pre-training task of this network is to predict the approximate location distribution of an input DNA fragment sequence in a broad plant phylogenetic tree.

4. The method according to claim 1, characterized in that, The decoupled attention module is implemented using a multi-head attention mechanism, which includes at least three functionally decoupled attention heads: the first attention head extracts the conservative features by calculating the feature stability weights across segments; The second attention head generates the positional encoding features by analyzing the cross-correlation between fragment features; The third attention focus isolates the degradation features by identifying anomalous feature patterns that correspond to common DNA degradation damage patterns.

5. The method according to claim 4, characterized in that, The common DNA degradation damage patterns identified by the third attention head specifically include cytosine deamination due to high temperature or oxidation, and guanine oxidation due to mold growth. These patterns are learned by training the decoupled attention module on a large amount of simulated degradation data.

6. The method according to claim 1, characterized in that, The collaborative reconstructor is a Transformer module with a decoder structure. It uses the learnable full-length barcode position encoding as the query vector and the combination of the conservative features and position encoding features as the key and value. Through a multi-layer attention and feedforward network, it iteratively generates each site feature of the synthetic feature sequence.

7. The method according to claim 1, characterized in that, The deep feature space similarity comparison and matching score calculation are performed through a pre-trained Siamese network. One branch of the Siamese network is input to the synthetic feature sequence, and the other branch is input to the candidate features in the standard feature database. The network outputs a matching score. The identification confidence is calculated by an independent confidence evaluation module. The input of this module includes the attention weight entropy value of the decoupled-cooperative reconstruction network during the reconstruction process, the average proportion of the degenerate features in the total features, and the difference between the first-ranked matching score and the second-ranked matching score output by the Siamese network.

8. The method according to claim 1, characterized in that, The method for constructing the standard characteristic database of quarantine weeds in the Asteraceae family is as follows: Collect the full-length standard DNA barcode sequences of the target Asteraceae weed species; For each full-length sequence, a set of degradation fragments with different degrees and modes is generated through simulation. A pre-trained dual-channel deep feature extraction network is used to extract features from all segments corresponding to each simulated segment set. The network has parallel sequence feature extraction channels and evolution feature extraction channels. The extracted features are input into a decoder network based on the Transformer architecture for reconstruction. The decoder network uses learnable positional encoding as the query and fragment features as the key and value to generate a "standard synthetic feature sequence" corresponding to the standard sequence. The "standard synthetic feature sequence" corresponding to each standard sequence is stored in the database to form the standard feature database of quarantine weeds of the Asteraceae family.

9. The method according to claim 1, characterized in that, The method also includes a verification step for low-confidence results: if the output identification confidence is lower than a preset first threshold but higher than a lower second threshold, the system automatically triggers a droplet digital PCR verification experiment for one or two candidate species ranked first in the identification results using their specific primers, and confirms or corrects the preliminary identification conclusion based on the quantitative detection results of the experiment.

10. The method according to claim 1, characterized in that, The shotgun sequencing depth is controlled between 0.5X and 2X. Before feature extraction, the target fragment set is filtered based on sequence length and average sequencing quality value, and only fragments with a length between 60bp and 250bp and an average quality value higher than Q20 are retained for subsequent analysis.

Images