Ecological early warning-oriented red tide algae algal metagenome rapid screening method and system

Abstract

The application discloses an ecological early warning-oriented red tide alga algal metagenome rapid screening method and system, and relates to the technical field of gene detection. The method comprises the following steps: performing shotgun metagenome sequencing on seawater samples to obtain sequencing data, and performing quality control on the sequencing data to obtain clean sequencing sequences; comparing and sorting the clean sequencing sequences according to a present situation prior reference database to obtain a known sequence set and an unknown sequence set; inputting the known sequence set and the unknown sequence set into a known feature extraction channel and an unknown feature extraction channel respectively, obtaining double-channel output, performing feature splicing, and obtaining a comprehensive feature vector; and inputting the comprehensive feature vector into a prediction model to output a red tide occurrence probability value. The application effectively improves the screening efficiency of red tide alga algal metagenome, and realizes rapid, accurate and early warning of red tide ecology.

Description

Technical Field

[0001] This invention relates to the field of gene detection technology, specifically to a rapid screening method and system for the metagenomics of red tide algae for ecological early warning. Background Technology

[0002] Traditional red tide monitoring mainly relies on microscopic morphological identification and chemical toxin detection. Shotgun metagenomic sequencing can directly obtain the genetic information of all microorganisms in seawater samples, becoming a new technical approach for early red tide identification. Conventional metagenomic analysis mostly adopts standard procedures such as sequence alignment, species classification, and functional annotation for detection and judgment.

[0003] However, traditional monitoring methods suffer from poor timeliness and insufficient sensitivity, making it difficult to achieve early prediction of red tides and assessment of toxin potential. The massive data processing of metagenomic sequencing is cumbersome and time-consuming, resulting in low screening efficiency of red tide algae metagenomics and failing to meet the timeliness requirements for rapid red tide early warning. Summary of the Invention

[0004] This invention provides a rapid screening method and system for red tide algae metagenomics for ecological early warning, aiming to solve the technical problem that the screening efficiency of red tide algae metagenomics in the prior art is low and cannot meet the timeliness requirements of rapid red tide early warning.

[0005] In view of the above problems, the present invention provides a rapid screening method and system for the metagenomics of red tide algae for ecological early warning.

[0006] In a first aspect, this invention provides a rapid metagenomic screening method for red tide algae for ecological early warning, including:

[0007] Shotgun metagenomic sequencing was performed on seawater samples to obtain sequencing data, and the sequencing data was subjected to quality control to obtain clean sequencing sequences.

[0008] Based on the existing prior reference database, the clean sequencing sequences are compared and sorted to obtain a known sequence set and an unknown sequence set.

[0009] The known sequence set and the unknown sequence set are respectively input into the known feature extraction channel and the unknown feature extraction channel to obtain dual-channel output and feature concatenation to obtain a comprehensive feature vector.

[0010] The comprehensive feature vector is input into the prediction model, and the probability value of red tide occurrence is output.

[0011] Secondly, this invention provides a rapid metagenomic screening system for red tide algae for ecological early warning, comprising:

[0012] The sequencing quality control module is used to perform shotgun metagenomic sequencing on seawater samples to obtain sequencing data, and to perform quality control on the sequencing data to obtain clean sequencing sequences.

[0013] The sequence sorting module is used to compare and sort the clean sequencing sequences according to the existing prior reference database to obtain a known sequence set and an unknown sequence set;

[0014] The dual-channel feature fusion module is used to input the known sequence set and the unknown sequence set into the known feature extraction channel and the unknown feature extraction channel respectively, obtain the dual-channel output, and perform feature concatenation to obtain a comprehensive feature vector.

[0015] The red tide probability prediction module is used to input the comprehensive feature vector into the prediction model and output the red tide occurrence probability value.

[0016] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0017] This invention provides a rapid screening method and system for red tide algae metagenomics for ecological early warning. It purifies raw sequencing data through sequencing quality control to ensure sequence quality and analytical reliability; it achieves precise sorting of known / unknown sequences using a reference library, fully utilizing prior knowledge to improve the identification efficiency of known red tide algae while completely preserving potential signals of unknown algae, thus eliminating detection blind spots; through dual-channel feature extraction and splicing, it synergistically integrates dominant features of known sequences and potential features of unknown sequences, achieving efficient compression of massive amounts of data and integration of key information, improving processing speed; finally, the prediction model directly outputs the probability of red tide occurrence, transforming complex metagenomic analysis into intuitive early warning results, effectively improving the screening efficiency of red tide algae metagenomics and achieving rapid and accurate early warning of red tide ecology. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the rapid screening method for red tide algae metagenomics for ecological early warning provided in this embodiment of the invention;

[0019] Figure 2 This is a schematic diagram of the structure of a rapid metagenomic screening system for red tide algae for ecological early warning provided in an embodiment of the present invention;

[0020] The components represented by each number in the attached diagram are explained below:

[0021] Sequencing quality control module 11, sequence sorting module 12, dual-channel feature fusion module 13, red tide probability prediction module 14. Detailed Implementation

[0022] This invention provides a rapid screening method and system for red tide algae metagenomics for ecological early warning, which addresses the technical problem that the screening efficiency of red tide algae metagenomics in existing technologies is low and cannot meet the timeliness requirements of rapid red tide early warning.

[0023] Example 1, as Figure 1 As shown, this invention provides a rapid metagenomic screening method for red tide algae for ecological early warning, the method comprising:

[0024] S100: Perform shotgun metagenomic sequencing on seawater samples to obtain sequencing data, and perform quality control on the sequencing data to obtain clean sequencing sequences.

[0025] In this embodiment of the invention, shotgun metagenomic sequencing was performed on seawater samples to obtain sequencing data. The sequencing data was then subjected to quality control to obtain clean sequencing sequences. Raw data from red tide algae metagenomic sequencing is prone to contamination with low-quality bases, sequencing adapters, and host genomes. Furthermore, quantitative biases can easily occur during DNA extraction, library construction, and sequencing. Directly using these data for subsequent analysis would reduce sequence sorting accuracy and feature extraction reliability. Simultaneously, the biomass of red tide algae in seawater samples is extremely low, requiring microbial enrichment through filtration. Exogenous standard DNA can correct for losses throughout the process. Therefore, it is essential to first complete sample collection, sequencing correction, library construction, and data quality control to obtain clean sequencing sequences, providing a qualified data base for subsequent sequence sorting and feature extraction.

[0026] Step S100 in the method provided in this embodiment of the invention includes:

[0027] Seawater samples were collected from the surface of the target sea area, and the seawater samples were filtered using a filter membrane to obtain filter membrane samples.

[0028] Based on the recovery rate of exogenous standard DNA, the filter membrane sample was sequenced and corrected, and the extracted DNA was used to construct a library to obtain the sequencing data.

[0029] The sequencing data is subjected to quality control and preprocessing, including quality filtering, host contamination removal, and quality assessment, to obtain clean sequencing sequences.

[0030] First, surface seawater samples are collected in the target sea area. These samples are then filtered using a filter membrane to obtain filter membrane samples. The target sea area refers to a specific marine region requiring red tide ecological early warning, such as nearshore aquaculture areas, harbors, or areas prone to red tides. Surface seawater typically refers to the water layer 0.5-1m below the sea surface. This area has ample sunlight and a high density of plankton, making it a representative sampling layer for monitoring red tide occurrences. The filter membrane is a microporous filter membrane used to trap microorganisms and suspended particles in seawater. Commonly used materials are polycarbonate or mixed cellulose esters, with pore sizes typically of 0.22μm or 0.45μm, effectively trapping algal cells, bacteria, viruses, etc.

[0031] Specifically, at fixed monitoring stations in the pre-designated red tide monitoring area, surface seawater at a depth of 0.5-1m below the sea surface is collected; the seawater is filtered using a vacuum filtration method with a filter membrane of a specified pore size to trap microorganisms in the water onto the surface of the filter membrane; after filtration, the filter membrane is stored at low temperature to obtain a filter membrane sample.

[0032] For example, at a fixed red tide monitoring station, 10L of surface seawater at a depth of 0.8m below the sea surface was collected; vacuum filtration was performed using a 0.22μm polycarbonate filter membrane to completely trap red tide algae such as Alexandrium and Pseudo-Nyctaginus, as well as marine bacteria, on the filter membrane; after filtration, the filter membrane was placed in a sterile cryovial and stored in an ultra-low temperature freezer at -80℃ to obtain a filter membrane sample.

[0033] Secondly, based on the recovery rate of exogenous standard DNA, sequencing and correction were performed on the filter membrane samples, and library construction was conducted on the extracted DNA to obtain the sequencing data. Exogenous standard DNA (Spike-in DNA) refers to exogenous reference DNA with known copy numbers and well-defined sequences, used to correct recovery rate deviations throughout the extraction, library construction, and sequencing processes. DNA library construction refers to processing sample DNA into a set of DNA fragments adapted to high-throughput sequencing platforms, and is an essential pre-sequencing step.

[0034] Specifically, quantitative exogenous standard DNA is added to the filter membrane sample and simultaneously participates in the entire DNA extraction process; total DNA on the filter membrane is extracted using standard methods, and the sample data is corrected based on the sequencing recovery rate of the exogenous standard DNA; the corrected DNA is then subjected to end repair, adapter addition, and fragment screening to complete the construction of the sequencing library.

[0035] For example, 10 ppm was added to the above filter membrane sample. 6The copied PhiX174 phage genomic DNA was used as an exogenous standard; total DNA was extracted from the filter membrane using a modified CTAB method, and the PhiX174 DNA recovery rate was calculated after sequencing to correct for sample DNA quantification bias; the corrected DNA was constructed into a paired-end sequencing library with an insert fragment of 350 bp using the Illumina PCR-Free library preparation kit to obtain a qualified sequencing library.

[0036] Finally, the sequencing data undergoes quality control and preprocessing, including quality filtering, host contamination removal, and quality assessment, to obtain clean sequencing sequences. Quality filtering removes invalid data such as low-quality bases, short fragments, and sequencing adapters. Host contamination removal removes host genome sequences from the samples, including those from fish, zooplankton, and humans, to avoid interference with red tide algae analysis. Quality assessment checks the quality indicators of the filtered data to confirm its suitability.

[0037] Specifically, the raw sequencing data undergoes quality filtering to remove low-quality bases, short fragments, and adapter sequences; the filtered sequences are then aligned to the host genome database to remove matching host contamination sequences; and the processed data undergoes quality assessment, with the qualified sequences being considered clean sequencing sequences.

[0038] For example, the raw sequencing data of the above samples were processed using the FastP tool to remove low-quality bases with a base quality value (Q) < 20, short sequences with a length < 100 bp, and Illumina sequencing adapters. The sequences were then aligned to near-shore fish, brine shrimp, and human genome databases using the Bowtie2 gene sequence alignment tool to remove host contamination sequences. The FastQC gene data quality assessment tool was used to ensure that the data Q30 ≥ 85% and that there were no obvious adapter residues, resulting in clean sequencing sequences.

[0039] In this embodiment of the invention, target microorganisms such as red tide algae are efficiently obtained through standardized surface seawater sampling and membrane enrichment. Exogenous standard DNA is used to correct the quantitative deviation throughout the process, eliminating systematic errors caused by sample processing. Through three preprocessing steps of quality filtration, host contamination removal, and quality assessment, invalid data and impurities are completely removed, resulting in clean sequencing sequences with high purity, high quality, and quantifiable values. This provides a stable and reliable data foundation for subsequent sequence sorting and dual-channel feature extraction.

[0040] S200: Based on the existing prior reference database, the clean sequencing sequences are compared and sorted to obtain a known sequence set and an unknown sequence set.

[0041] In this embodiment of the invention, the clean sequencing sequences are compared and sorted according to a priori reference database to obtain a known sequence set and an unknown sequence set. The clean sequencing sequences after quality control contain a mixture of known red tide algae, closely related non-red tide algae, common marine microorganisms, toxin genes, and unknown biological sequences not included in the database. Direct feature extraction would result in inaccurate identification of known signals, obscuring unknown signals, and extremely low computational efficiency. Furthermore, comparison with a single database is prone to missing closely related species and toxin genes, leading to missed or false red tide warnings. Therefore, rapid screening and precise extended screening are needed to accurately classify the sequences into known and unknown sequence sets. This fully utilizes the prior knowledge of existing gene databases while completely preserving the unknown sequences of potential new red tide algae, providing a clearly categorized data source for subsequent dual-channel feature extraction.

[0042] Step S200 in the method provided in this embodiment of the invention includes:

[0043] The clean sequencing sequences are rapidly screened using the rapid classification database in the existing prior reference database to determine the high-confidence known group and the undetermined group;

[0044] The pending group is further compared and filtered using the extended reference database in the existing prior reference database, wherein the extended comparison and filtering includes consistency filtering and coverage filtering.

[0045] Based on the extended comparison and screening results, short read sequences in the undetermined group that meet the thresholds for both consistency and coverage are added to the high-confidence known group, and the output is the known sequence set;

[0046] Short read sequences outside the high-confidence known group are stored as the unknown sequence set.

[0047] First, the clean sequencing sequences are rapidly screened using the rapid classification database within the existing prior reference database to determine the high-confidence known group and the undetermined group. The existing prior reference database is a pre-constructed database containing all known gene information related to red tides, serving as the basis for sequence alignment. The rapid classification database is a subset of the existing prior reference database, specifically referring to the red tide algae reference genome, which includes all publicly available complete gene sequences of red tide algae, used for rapid identification of identified red tide algae sequences. The high-confidence known group refers to the set of short read sequences that precisely match the rapid classification database and are 100% identified as known red tide algae. The undetermined group refers to the set of short read sequences that cannot match the rapid classification database and whose attribution cannot be determined at this time, requiring further precise alignment.

[0048] Specifically, the rapid classification database is retrieved; a professional sequence classification tool is used to perform a full and rapid comparison between the clean sequencing sequences and the rapid classification database; short read sequences that can completely match the database are classified into the high-confidence known group; short read sequences that cannot match are classified into the pending group, awaiting the next step of precise screening.

[0049] For example, the clean sequencing sequences of red tide monitoring samples were quickly compared with a rapid classification database using the Kraken2 classification tool; short read sequences that successfully matched Alexandrium genes were classified into the high-confidence known group; the remaining unmatched short read sequences were classified into the undetermined group.

[0050] Secondly, the undetermined groups are further compared and screened using an extended reference database within the existing prior reference database. This extended comparison and screening includes consistency screening and coverage screening. The extended reference database is an extended subset of the existing prior reference database and contains three types of data: genomes of closely related non-red tide algae, genomes of common marine background microorganisms, and a toxin synthesis gene database. The toxin synthesis gene database is a database containing genes related to red tide toxins, including five gene families: the dams gene family (dominoic acid synthesis gene cluster); the sxt gene family (salicornic acid synthesis gene cluster); the PKS gene family (genus related to the synthesis of okata spongiformis acid and bryophylloides toxins, PKS = polyketide synthase, a key enzyme gene for toxin synthesis); and the hemolysin gene family (genus related to red tide algal hemolysins).

[0051] Among these, consistency screening refers to determining the proportion of gene base matches between the short read sequence and the database sequence; a higher matching proportion indicates higher consistency. Coverage screening refers to determining the proportion of gene fragments in the database that the short read sequence covers; the longer the coverage, the higher the coverage.

[0052] Specifically, the extended reference database is retrieved; a professional sequence alignment tool is used to accurately align the short read sequences of the undetermined group with the extended reference database; and a dual screening process is performed: first, consistency screening is performed, and then coverage screening is performed, retaining only sequences that simultaneously meet both thresholds.

[0053] For example, the undetermined group sequences of the above samples were compared with the extended reference database using the Minimap2 alignment tool; the screening thresholds were set: consistency ≥ 80% and coverage ≥ 50%; undetermined sequences matching closely related non-red tide algae, marine bacteria, and dams / sxt / PKS toxin genes were screened out.

[0054] Next, based on the expanded alignment and screening results, short read sequences from the undetermined group that meet both consistency and coverage thresholds are added to the high-confidence known group, and the output is the known sequence set. The known sequence set refers to the final determined collection of short read sequences belonging to known red tide-related organisms / genes, including qualified sequences from the high-confidence known group and the undetermined group. All undetermined group sequences that meet both consistency and coverage standards are merged into the high-confidence known group; the merged sequence set constitutes the known sequence set, used for subsequent known feature extraction channel analysis.

[0055] For example, sequences in the undetermined sample group that match closely related algae and toxin genes and meet the criteria (i.e., consistency ≥ 80% and coverage ≥ 50%) are all merged into the high-confidence known group; after merging, a set of known sequences containing red tide algae, closely related algae, and toxin genes is obtained.

[0056] Finally, short read sequences outside the high-confidence known groups are stored as the unknown sequence set. The unknown sequence set refers to the collection of short read sequences that cannot be matched with the fast classification database or screened through the extended reference database, including potential new red tide algae and rare microbial sequences that have not been included. All short read sequences that do not belong to the high-confidence known groups—that is, sequences that did not meet the criteria in the undetermined groups and invalid sequences that were removed—are uniformly stored as the unknown sequence set for subsequent unknown feature extraction channel analysis. For example, the remaining short read sequences in the above samples that were not included in the known sequence set are uniformly organized and stored to obtain the unknown sequence set.

[0057] In this embodiment of the invention, a two-layer sorting logic of rapid initial screening and precise re-screening is used, relying on four major databases: red tide algae, closely related algae, background microorganisms, and toxin genes, to complete the precise classification of short read sequences. This not only efficiently identifies known red tide algae and toxin gene sequences, but also completely preserves unknown biological sequences not included in the databases, solving the problems of missed detections and sequence confounding analysis in single database comparisons. Finally, a clear and non-overlapping set of known and unknown sequences is obtained, providing standardized classification data for subsequent dual-channel feature extraction, while avoiding the loss of unknown potential red tide algae signals.

[0058] S300: Input the known sequence set and the unknown sequence set into the known feature extraction channel and the unknown feature extraction channel respectively, obtain the dual-channel output and perform feature concatenation to obtain the comprehensive feature vector.

[0059] In this embodiment of the invention, the known sequence set and the unknown sequence set are respectively input into the known feature extraction channel and the unknown feature extraction channel to obtain dual-channel outputs and perform feature concatenation to obtain a comprehensive feature vector. The known sequence set after S200 sorting has clear classification information, and deterministic features such as species abundance and toxin coverage can be directly extracted; while the unknown sequence set contains potential new red tide algae / rare microorganisms that have not been included in the data and cannot be identified through conventional comparison. It needs to be reassembled to mine potential sequence features. If the unknown signals are directly processed separately or ignored, it will lead to missed red tide warnings and incomplete feature dimensions. Therefore, a dual-channel feature extraction system needs to be constructed: the known channel uses prior knowledge to extract explicit features, and the unknown channel mines potential features through assembly and noise reduction autoencoders. Finally, they are concatenated into a comprehensive feature vector to achieve a complete representation of full-dimensional features.

[0060] Step S300 in the method provided in this embodiment of the invention includes:

[0061] The known sequence set is input into the known feature extraction channel to obtain the known sequence feature vector, which includes species abundance, toxin gene coverage, and abundance statistical indicators.

[0062] The short read sequences in the unknown sequence set are reassembled to obtain an assembled fragment, and features are extracted from the assembled fragment through an unknown feature extraction channel to obtain an unknown sequence feature vector, wherein the unknown sequence feature vector includes a basic attribute class and a cropped short read sequence frequency class;

[0063] The unknown sequence feature vector is obtained by adaptively denoising the unknown sequence feature vector through a pre-trained dual-channel denoising autoencoder.

[0064] The known sequence feature vector and the unknown denoised sequence feature vector are concatenated to obtain the comprehensive feature vector.

[0065] First, the known sequence set is input into the known feature extraction channel to obtain feature vectors of the known sequences, including species abundance, toxin gene coverage, and abundance statistical indicators. The known feature extraction channel refers to a feature extraction module designed for known red tide algae and toxin gene sequences, used to extract three types of directly interpretable deterministic features.

[0066] Among these features, species abundance characteristics represent the relative / absolute abundance of each known red tide algae species in the sample, reflecting the proliferation level of red tide algae. Toxin gene coverage characteristics represent the coverage ratio of toxin synthesis genes (dams, sxt, PKS, hemolysin genes) in the sample, reflecting the toxin synthesis potential. Abundance statistical indicators are statistical indicators calculated based on species abundance, such as abundance range, coefficient of variation, and peak percentage, characterizing the distribution fluctuation characteristics of red tide algae.

[0067] Specifically, species annotation is performed on the known sequence set, the number of short read sequences of each red tide algae species is counted, and species abundance features are calculated; the toxin synthesis gene database is compared, the coverage of each toxin gene family is counted, and toxin gene coverage features are calculated; statistical indicators are calculated based on the abundance data to obtain abundance statistical indicator features; the three types of features are merged to generate a known sequence feature vector.

[0068] For example, for a known set of samples, Kraken2 annotation yielded an abundance of 15% for Alexandrium and 8% for Pseudo-Nyctaginus; comparison with a toxin database revealed a 60% coverage of the sxt gene and a 35% coverage of the dams gene; the abundance variation coefficient was calculated to be 0.4 and the peak percentage to be 70%; finally, these were merged into a known sequence feature vector containing 12 dimensions.

[0069] Secondly, the short read sequences in the unknown sequence set are reassembled to obtain assembled fragments. Features are then extracted from these assembled fragments using an unknown feature extraction channel to obtain unknown sequence feature vectors. These feature vectors include basic attribute classes and cropped short read sequence frequency classes. The unknown sequence set refers to short read sequences that, after S200 sorting, do not match any reference database, and includes potential new red tide algae and rare microbial sequences. Reassembly refers to the de novo splicing of unknown short read sequences, merging short fragments into longer assembled fragments (contigs) to improve sequence annotability. Basic attribute classes refer to the basic sequence features of the assembled fragments, including length, GC content, average quality value, and complexity. Cropped short read sequence frequency classes refer to the statistical frequency with which the assembled fragments are covered by cropped short read sequences of different lengths, reflecting the stability of the sequence structure.

[0070] Specifically, the unknown sequence set is reassembled using a de novo assembly tool to obtain contigs; the basic attributes of the contigs are extracted and basic attribute features are calculated; the contigs are truncated at multiple scales, the coverage frequency of the truncated short read sequences is statistically analyzed, and the frequency features of the truncated short read sequences are calculated; the two types of features are merged to generate the feature vector of the unknown sequence.

[0071] For example, for a set of unknown sample sequences, the Megahit tool is used to reassemble 500 assembled fragments; basic attributes such as fragment length, GC content, and average Q value are extracted; the coverage frequency of fragments by short read sequences pruned at 50bp, 100bp, and 150bp is statistically analyzed; and finally, a 20-dimensional feature vector of the unknown sequence is generated.

[0072] Next, the unknown sequence feature vector is adaptively denoised using a pre-trained dual-channel denoising autoencoder to obtain the unknown denoised sequence feature vector.

[0073] Specifically, an unknown denoised sequence feature vector is obtained by adaptively denoising the unknown sequence feature vector using a pre-trained dual-channel denoising autoencoder, including:

[0074] The unknown sequence feature vector is input into a pre-trained signal denoising autoencoder and a pre-trained noise separation autoencoder respectively to obtain signal reconstruction output and noise reconstruction output;

[0075] The signal weights are calculated based on the first reconstruction error between the unknown sequence feature vector and the signal reconstruction output, and the second reconstruction error between the unknown sequence feature vector and the noise reconstruction output.

[0076] Based on the signal weights, the signal reconstruction output and the noise reconstruction output are weighted and fused to obtain the fused reconstruction output;

[0077] Based on the difference between the unknown sequence feature vector and the fused reconstruction output, the reconstruction quality is calculated, wherein the reconstruction quality is used to characterize the noise reduction effect of the dual-channel noise reduction autoencoder;

[0078] Based on the difference between the unknown sequence feature vector and the sum of the signal reconstruction output and the noise reconstruction output, the decomposition consistency is calculated. The decomposition consistency is used to characterize the degree of separation and complementarity between the signal denoising autoencoder and the noise separation autoencoder.

[0079] Based on the reconstruction quality and the decomposition consistency, a comprehensive noise reduction quality index is calculated;

[0080] Based on the signal weights and the reconstruction quality, each unknown sequence feature vector is determined to be retained. The retention determination results are then combined with statistical analysis methods to process the unknown sequence features and obtain the unknown denoised sequence feature vector.

[0081] First, the unknown sequence feature vector is input into a pre-trained signal denoising autoencoder and a pre-trained noise separation autoencoder to obtain the signal reconstruction output and the noise reconstruction output.

[0082] The pre-training of the signal denoising autoencoder and the noise separation autoencoder includes:

[0083] Based on the reference genome sequences in the existing prior reference database, sequencing data is simulated and assembled, and feature vectors of the assembled fragments are extracted to form a signal training set.

[0084] Based on a known set of noise sequences, the sequences are assembled and the feature vectors of the assembled segments are extracted to form a noise training set.

[0085] A combination of noise is applied to each signal training sample in the signal training set to obtain a noisy signal training sample, wherein the combination of noise is composed of Gaussian noise and a random mask superimposed.

[0086] Mix a feature vector randomly selected from the signal training set into each noise training sample in the noise training set to obtain a mixed noise training sample;

[0087] Using the noisy signal training samples as input, and combining signal reconstruction loss and L1 sparse regularization loss, the signal denoising autoencoder is trained so that its reconstructed output approximates the signal training samples.

[0088] Using the mixed noise training samples as input, and combining the noise separation loss and L1 sparse regularization loss, the noise separation autoencoder is trained so that its reconstructed output approximates the noise training samples.

[0089] By minimizing the total loss function weighted by the decomposition consistency loss, signal reconstruction loss, and noise separation loss, the signal denoising autoencoder and the noise separation autoencoder are jointly fine-tuned to obtain the signal denoising autoencoder and the noise separation autoencoder.

[0090] First, based on the reference genome sequences in the existing prior reference database, sequencing data is simulated and assembled. Feature vectors of the assembled fragments are extracted to form a signal training set. The reference genome sequence refers to a complete red tide algae / marine microorganism genome sequence already included in the database. Simulated sequencing data refers to short-read sequence data generated by computer software based on the reference genome, consistent with the actual sequencing format. An assembled fragment refers to a long gene sequence (contig) obtained by splicing the simulated short-read sequences. A feature vector is a set of digital features formed after extracting sequence attributes such as length and GC content from the assembled fragments. The signal training set is a dataset composed of feature vectors of assembled fragments from real biological signals, used to train the signal denoising autoencoder.

[0091] Specifically, the process involves retrieving reference genome sequences of red tide algae and marine background microorganisms from the existing prior reference database; using a simulated sequencing tool to generate simulated sequencing data based on the reference genome; assembling the simulated sequencing data to obtain assembled fragments; extracting feature vectors from all assembled fragments; and summarizing all feature vectors to form a signal training set.

[0092] For example, retrieve the reference genome sequences of Alexandrium and Pseudo-Nyctaginus from the existing prior reference database; use the InSilicoSeq tool to simulate and generate Illumina format sequencing data; use the MEGAHIT tool to assemble 1000 assembled fragments; extract the length, GC content, coverage depth and other features of each fragment to form a feature vector; summarize the 1000 feature vectors to form a signal training set.

[0093] Secondly, based on the known set of noise sequences, feature vectors of the assembled fragments are extracted to form a noise training set. The noise sequence set is a collection of invalid sequences that are not real biological signals, including five categories: adapter sequences, synthetic random sequences, low-complexity sequences, host genome fragments, and laboratory contamination sequences. The noise training set is a dataset composed of feature vectors of assembled fragments from the noise sequences, used to train the noise separation autoencoder.

[0094] Specifically, five known noise sequences are selected; each noise sequence is assembled to obtain assembled segments; feature vectors of all noise assembled segments are extracted; and all feature vectors are summarized to form a noise training set.

[0095] For example, five types of noise sequences were selected: Illumina sequencing adapter sequences, artificially synthesized random DNA sequences, poly-A low-complexity sequences, large yellow croaker host genome fragments, and pUC19 plasmid contamination sequences; 800 noise assembly fragments were obtained by assembling them respectively; features were extracted to form feature vectors; and the results were summarized to form a noise training set.

[0096] Next, combined noise is applied to each signal training sample in the signal training set to obtain noisy signal training samples. The combined noise consists of a superposition of Gaussian noise and a random mask. Combined noise refers to simulated noise formed by the superposition of Gaussian noise and a random mask, used to reproduce errors and deletions in real sequencing. Gaussian noise refers to random errors that follow a normal distribution, simulating random base errors in the sequencing process. A random mask involves randomly setting some feature values ​​to zero, simulating partial sequence deletions in sequencing. Noisy signal training samples refer to signal training samples with combined noise applied, used to simulate noisy real biological signals.

[0097] Specifically, each signal training sample in the signal training set is extracted; Gaussian noise is applied to the sample to simulate random sequencing error; a random mask is applied to the sample to simulate partial sequence loss; the Gaussian noise and the random mask are superimposed to obtain combined noise; the sample after applying combined noise is the noisy signal training sample.

[0098] For example, Gaussian noise with a standard deviation of 0.1 is applied to the feature vector of the Alexandrium assembly fragment in the signal training set, and then a random mask with a 30% probability is applied; the two are superimposed to form combined noise, resulting in noisy signal training samples.

[0099] Furthermore, feature vectors randomly selected from the signal training set are mixed into each noise training sample in the noise training set to obtain mixed noise training samples. Mixed noise training samples refer to samples in which a small number of real signal feature vectors are mixed into pure noise samples to simulate the situation where noise and signal are mixed in a real scene.

[0100] Specifically, each noise training sample in the noise training set is extracted; a feature vector is randomly selected from the signal training set; the randomly selected signal feature vector is mixed into the current noise training sample at a fixed ratio; after mixing, a mixed noise training sample is obtained.

[0101] For example, for sequencing adapter noise samples in the noise training set, one feature vector is randomly selected from the signal training set and mixed into the noise sample at a ratio of 10%; after thorough mixing, a mixed noise training sample is obtained.

[0102] Subsequently, using the noisy signal training samples as input, the signal denoising autoencoder is trained by combining signal reconstruction loss and L1 sparse regularization loss, so that its reconstructed output approximates the signal training samples. A signal denoising autoencoder is a deep learning model specifically designed to learn the characteristics of real biological signals and restore clean signals from noisy data. Signal reconstruction loss is an indicator that measures the difference between the model's reconstructed signal and the original real signal; the smaller the loss, the better the restoration effect. L1 sparse regularization loss is a constraint indicator to prevent overfitting, ensuring that the model only learns the most critical signal features.

[0103] Specifically, the noisy signal training samples are used as the input of the signal denoising autoencoder; the model outputs the signal reconstruction result; the signal reconstruction loss and L1 sparse regularization loss are calculated; the model parameters are iteratively updated so that the model reconstruction output continuously approaches the original signal training samples; after the loss converges, the training of the signal denoising autoencoder is completed.

[0104] For example, the noisy Alexandrium feature vector is input into the signal denoising autoencoder; the model outputs the reconstructed signal vector; the signal reconstruction loss is calculated to be 0.08 and the L1 sparse regularization loss is calculated to be 0.01; the model is trained iteratively for 100 rounds until the difference between the reconstructed output and the original signal sample is less than 0.001, and the training is completed.

[0105] Furthermore, using the mixed noise training samples as input, and combining noise separation loss and L1 sparse regularization loss, the noise separation autoencoder is trained so that its reconstructed output approximates the noise training samples. A noise separation autoencoder is a deep learning model specifically designed to learn noise features and separate pure noise from mixed data. Noise separation loss is a metric that measures the difference between the noise separated by the model and the original pure noise; the smaller the loss, the better the separation effect.

[0106] Specifically, the mixed noise training samples are used as input to the noise separation autoencoder; the model outputs the noise reconstruction result; the noise separation loss and L1 sparse regularization loss are calculated; the model parameters are iteratively updated so that the model reconstruction output continuously approaches the original noise training samples; after the loss converges, the training of the noise separation autoencoder is completed.

[0107] For example, the noise sample from the connector, which contains signal features, is input into the noise separation autoencoder; the model outputs a reconstructed noise vector; the noise separation loss is calculated to be 0.09 and the L1 sparse regularization loss is calculated to be 0.01; the model is trained iteratively for 100 rounds until the difference between the reconstructed output and the original noise sample is less than 0.001, thus completing the training.

[0108] Finally, by minimizing the weighted total loss function of decomposition consistency loss, signal reconstruction loss, and noise separation loss, the signal denoising autoencoder and the noise separation autoencoder are jointly fine-tuned to obtain the signal denoising autoencoder and the noise separation autoencoder. Decomposition consistency loss is an indicator that measures whether the sum of the signal reconstruction result and the noise reconstruction result is close to the original input, ensuring that the two are complementary and separated. The total loss function is a comprehensive loss indicator obtained by weighted summation of decomposition consistency loss, signal reconstruction loss, and noise separation loss. Joint fine-tuning refers to simultaneously training the signal denoising autoencoder and the noise separation autoencoder, enabling them to work collaboratively and complementarily.

[0109] Specifically, the weights of the decomposition consistency loss, signal reconstruction loss, and noise separation loss are set; the total loss function is calculated by substituting the three types of losses; the parameters of the two autoencoders are iteratively updated simultaneously to minimize the total loss function; after the total loss converges, the final usable signal denoising autoencoder and noise separation autoencoder are obtained.

[0110] For example, the weights of decomposition consistency loss, signal reconstruction loss, and noise separation loss are set to 1.0; the total loss is calculated to be 0.025; after 50 rounds of joint fine-tuning, the total loss is reduced to 0.003 and remains stable; finally, a trained dual-channel denoising autoencoder is obtained.

[0111] Based on this, the unknown sequence feature vectors are input into a pre-trained signal denoising autoencoder and a pre-trained noise separation autoencoder, respectively, to obtain signal reconstruction output and noise reconstruction output. The unknown sequence feature vector refers to the set of digital features extracted from the assembled fragments (contigs) of the unknown sequence set, such as basic attributes and the frequency of truncated short reads; it is the input data for this step. The signal reconstruction output refers to the result of the signal DAE restoring the input feature vectors, representing the real biological signal identified by the model. The noise reconstruction output refers to the separation result of the noise DAE on the input feature vectors, representing the sequencing noise identified by the model.

[0112] For example, extract the feature vectors of 500 assembled segments from the unknown sequence set; input each feature vector into a pre-trained signal DAE and noise DAE respectively; the signal DAE outputs the signal reconstruction vector of each segment. The noise separation vector corresponding to the noise DAE output Two sets of reconstruction outputs were obtained.

[0113] Secondly, based on the first reconstruction error between the unknown sequence feature vector and the signal reconstruction output, and the second reconstruction error between the unknown sequence feature vector and the noise reconstruction output, the signal weights are calculated.

[0114] Among them, the first reconstruction error This refers to the relative difference between the original unknown sequence feature vector and the signal reconstruction output, measuring the accuracy of the signal DAE in reconstructing the true signal. The formula is: ,in X is the L2 norm squared, which is the sum of the squares of the elements of the vector, used to quantify the overall difference between two vectors. X is the feature vector of the original unknown sequence. The smaller the value, the more accurate the signal reconstruction. Second reconstruction error. This refers to the relative difference between the original unknown sequence feature vector and the noise reconstruction output, measuring the accuracy of noise separation in noise-based DAE (Discrete Algorithm). The formula is: The larger the value, the less likely the vector is to be noise.

[0115] Among them, signal weights The proportion of real biological signals in this feature vector, with a value range of [0,1], is represented by the formula: , where the natural exponential function exp( The exponent is a mathematical function used to convert errors into non-negative weights; the smaller the error, the larger the exponent value. The temperature parameter τ is a hyperparameter that controls the sharpness of weight allocation: the smaller τ is, the more the weights are tilted towards the model with smaller errors; the larger τ is, the more evenly the weight allocation. The noise weight β represents the proportion of noise in the feature vector, β = 1 − α, and is complementary to the signal weights.

[0116] Specifically, for each unknown sequence feature vector X, the first reconstruction error is calculated. Second reconstruction error Set the temperature parameter τ, usually between 0.1 and 1.0, and adjust it according to the scenario; substitute it into the formula to calculate the signal weight α and noise weight β; complete the weight calculation for all vectors.

[0117] For example, for the feature vector of an unknown sequence in a sample: the following is calculated: =0.10, =0.60; Set temperature parameter τ=0.5; Substitute into the formula: ;calculate That is, the signal accounts for approximately 73.1% of the vector, and the noise accounts for approximately 26.9%.

[0118] Next, based on the signal weights, the signal reconstruction output and the noise reconstruction output are weighted and fused to obtain a fused reconstruction output. Weighted fusion refers to summing the reconstruction outputs of the two models according to the signal weights and noise weights to obtain a fused result that balances signal restoration and noise suppression. (Fused reconstruction output) This refers to the final denoised feature vector after weighted fusion, and the formula is: .

[0119] Specifically, for each unknown sequence feature vector, the corresponding signal weight α and noise weight β are extracted; the corresponding signal reconstruction output is then extracted. and noise reconstruction output According to the formula Perform a weighted summation to obtain the fused and reconstructed output. .

[0120] For example, α = 0.731, β = 0.269. The signal is reconstructed into a vector. For the noise separation vector; weighted summation: ; Obtain the fused denoised feature vector It preserves the true signal while suppressing noise interference.

[0121] Furthermore, based on the difference between the unknown sequence feature vector and the fused reconstruction output, the reconstruction quality is calculated, whereby the reconstruction quality characterizes the denoising effect of the dual-channel denoising autoencoder. The reconstruction quality R measures the degree of matching between the fused reconstruction output and the original unknown sequence feature vector, directly characterizing the overall denoising effect of the dual-channel denoising autoencoder, and is calculated using the following formula: The value of R ranges from [0,1]. The closer R is to 1, the higher the quality of the fusion reconstruction and the better the noise reduction effect.

[0122] The value of reconstruction quality R directly corresponds to the denoising effect and subsequent processing method. When R ≥ 0.8, it means that the fusion reconstruction quality is high and the denoising effect of the dual-channel denoising autoencoder is excellent. When 0.5 ≤ R < 0.8, it means that the fusion reconstruction quality is medium and the denoising effect is at a general level. When R < 0.5, it means that the fusion reconstruction quality is poor and neither of the two autoencoders can effectively process the sequence. The sequence may be a highly novel biological signal or a severely noisy sequence, and it needs to go through a subsequent additional verification process.

[0123] Specifically, for each unknown sequence feature vector X, the corresponding fusion reconstruction output is extracted. Substitute the values ​​into the formula to calculate the reconstruction quality R; based on the range of R values, label the noise reduction quality level of the vector; complete the quality assessment of all vectors. For example, substituting into the formula yields R=0.88; R≥0.8 indicates high-quality noise reduction, and can be directly used as the quality level. As a feature after denoising; if a vector calculates R=0.62, it is of medium quality, retaining the feature but reducing the confidence; if R=0.35, it enters an additional verification process.

[0124] Subsequently, based on the difference between the unknown sequence feature vector and the sum of the signal reconstruction output and the noise reconstruction output, decomposition consistency is calculated. This decomposition consistency characterizes the degree of separation complementarity between the signal denoising autoencoder and the noise separation autoencoder. Decomposition consistency measures the degree of separation complementarity between the signal DAE and the noise DAE, i.e., whether the sum of the reconstructed outputs of the two models is consistent with the original input. The formula is: The value of C ranges from [0,1]. The closer C is to 1, the better the separation and complementarity of the two models, and the more reasonable the division between signal and noise.

[0125] Specifically, for each unknown sequence feature vector X, the corresponding signal is extracted and the output is reconstructed. and noise reconstruction output Calculate the sum of the two reconstructed outputs; substitute into the formula to calculate the decomposition consistency C; complete the complementarity evaluation of all vectors. For example, substituting into the formula, we get C = 1 − 0.07 = 0.93; C is close to 1, indicating that the separation complementarity of the signal DAE and the noise DAE is excellent, and the signal and noise separation is reasonable.

[0126] Then, combining the reconstruction quality and the decomposition consistency, a comprehensive noise reduction quality index (Q) is calculated. The comprehensive noise reduction quality index Q is a weighted composite index that integrates reconstruction quality R and decomposition consistency C, comprehensively quantifying the overall noise reduction effect of the dual-channel DAE. The formula is: ,in To reconstruct quality weights, To decompose the consistency weights, satisfying The closer Q is to 1, the better the overall noise reduction effect.

[0127] Specifically, for each unknown sequence feature vector, the corresponding reconstruction quality R and decomposition consistency C are extracted; weights are set. and Substitute the values ​​into the formula to calculate the overall noise reduction quality index Q; complete the overall quality assessment of all vectors. For example, R=0.88, C=0.93, set... =0.6, =0.4; Q=0.6×0.88+0.4×0.93=0.528+0.372=0.90. Q=0.90 indicates that the overall noise reduction effect of this vector is excellent.

[0128] Finally, based on the signal weights and the reconstruction quality, each unknown sequence feature vector is determined to be retained, and the unknown sequence features are processed by combining the retention determination results with statistical analysis methods to obtain the unknown denoised sequence feature vector.

[0129] Specifically, based on the signal weights and the reconstruction quality, a retention determination is made for each unknown sequence feature vector, and the unknown sequence features are processed using the retention determination results and statistical analysis methods to obtain the unknown denoised sequence feature vector, including:

[0130] Based on the signal weights and the reconstruction quality, a retention determination is made for each unknown sequence feature vector, including:

[0131] When the signal weight is not less than the first threshold and the reconstruction quality is not less than the second threshold, it is determined to be a high-confidence signal sequence, and the corresponding unknown sequence feature vector is retained.

[0132] When the signal weight is less than the third threshold, it is determined to be a high-confidence noise sequence, and the corresponding unknown sequence feature vector is discarded;

[0133] When either the signal weight is not less than the third threshold and is less than the first threshold, or the reconstruction quality is less than the second threshold, an additional verification process is initiated.

[0134] The additional verification process sequentially performs sequence quality index checks, assembly coverage depth checks, functional signal checks, and cross-sample consistency checks.

[0135] When any item in the additional verification process is determined to be a signal sequence, the corresponding unknown sequence feature vector is retained; when any item in the additional verification process is determined to be a noise sequence, the corresponding unknown sequence feature vector is discarded.

[0136] For all retained unknown sequence feature vectors, the encoder output of the signal denoising autoencoder is taken as the latent space representation, and cluster analysis is performed on the latent space representation;

[0137] Based on the cluster analysis results, cluster statistical features and global statistical features are extracted and concatenated to form the feature vector of the unknown denoised sequence.

[0138] First, when the signal weight is not less than the first threshold and the reconstruction quality is not less than the second threshold, it is determined to be a high-confidence signal sequence, and the corresponding unknown sequence feature vector is retained.

[0139] Secondly, when the signal weight is less than the third threshold, it is determined to be a high-confidence noise sequence, and the corresponding unknown sequence feature vector is discarded.

[0140] Furthermore, when either the signal weight is not less than the third threshold and is less than the first threshold, or the reconstruction quality is less than the second threshold, an additional verification process is initiated.

[0141] The first threshold is the lowest signal weight value for identifying high-confidence signal sequences. The second threshold is the lowest reconstruction quality value for identifying high-confidence signal sequences. The third threshold is the highest signal weight value for identifying high-confidence noise sequences. High-confidence signal sequences are assembled fragment sequences identified as genuine biological signals. High-confidence noise sequences are invalid sequences identified as sequencing noise or contamination. Sequences that cannot be directly identified and require additional validation are considered critical sequences.

[0142] Specifically, a first threshold, a second threshold, and a third threshold are set. If the signal weight is greater than or equal to the first threshold and the reconstruction quality R is greater than or equal to the second threshold, it is determined to be a high-confidence signal sequence and the feature vector is retained. If the signal weight is less than the third threshold, it is determined to be a high-confidence noise sequence and the feature vector is discarded. If the signal weight is greater than or equal to the third threshold and less than the first threshold or the reconstruction quality R is less than the second threshold, it is determined to be a critical sequence and enters the additional verification process.

[0143] For example, a first threshold of 0.8, a second threshold of 0.8, and a third threshold of 0.3 are set; a certain sequence signal with a weight of 0.85 and R=0.88 is a high-confidence signal and is retained; a certain sequence signal with a weight of 0.25 is high-confidence noise and is discarded; a certain sequence signal with a weight of 0.6 and R=0.7 is determined to be a critical sequence and enters the additional verification process.

[0144] Furthermore, the additional verification process sequentially performs sequence quality index checks, assembly coverage depth checks, functional signal checks, and cross-sample consistency checks.

[0145] Specifically, the following four checks are performed sequentially: First check: Sequence quality index check, based on the average quality score of the original sequence from which the assembled fragment originates, to determine whether it meets the preset minimum quality threshold; based on the length of the assembled fragment, to determine whether it meets the preset minimum length threshold; Second check: Assembly coverage depth check, based on the coverage depth of the assembled fragment, to determine whether it meets the preset minimum coverage depth threshold; Third check: Functional signal check, performing open reading frame prediction on the assembled fragment; if a complete open reading frame is predicted, a matching functional domain is searched in the functional domain database; if a known functional domain is found, it is determined to be a signal sequence; performing non-coding RNA prediction on the assembled fragment; if ribosomal RNA or transfer RNA is predicted, it is determined to be a signal sequence; Fourth check: Cross-sample consistency check, based on the repetition of short fragment frequency characteristics of the assembled fragment in samples at multiple time points, to determine whether it meets the preset repetition frequency threshold.

[0146] For example, for a critical sequence, the following checks are performed sequentially: average quality score 32 ≥ minimum quality threshold 30, length 1200bp ≥ minimum length threshold 1000bp, passing the sequence quality index check; coverage depth 45 times ≥ minimum coverage depth threshold 30 times, passing the assembly coverage depth check; predicting the complete ORF and matching the PKS toxin functional domain, passing the functional signal check; repeated occurrence of samples at 3 time points, passing the consistency check.

[0147] Secondly, when any step in the additional verification process is determined to be a signal sequence, the corresponding unknown sequence feature vector is retained; when any step in the additional verification process is determined to be a noise sequence, the corresponding unknown sequence feature vector is discarded. If the critical sequence is determined to be a signal sequence in any step of the additional verification, the unknown sequence feature vector is retained; if the critical sequence is determined to be a noise sequence in any step of the additional verification, the unknown sequence feature vector is discarded.

[0148] Based on this, for all retained unknown sequence feature vectors, the encoder output of the signal denoising autoencoder is taken as the latent space representation, and cluster analysis is performed on the latent space representation. The latent space representation refers to the low-dimensional core features output by the signal denoising autoencoder, preserving the biological essence of the sequence. Cluster analysis refers to grouping retained sequences with similar features into a cluster to uncover the patterns within the sequence group.

[0149] Specifically, for all the ultimately retained unknown sequence feature vectors, the encoder output of the signal denoising autoencoder is extracted as the latent space representation; cluster analysis is then performed on all latent space representations to divide the sequences into clusters. For example, latent space representations are extracted from 500 retained vectors, and HDBSCAN clustering is used to obtain 10 sequence clusters.

[0150] Then, based on the cluster analysis results, cluster statistical features and global statistical features are extracted and concatenated to form the feature vector of the unknown denoised sequence. Cluster statistical features refer to the statistical indicators within a single cluster. Global statistical features refer to the overall statistical indicators of all retained sequences and clusters. Extracting cluster statistical features involves obtaining the latent spatial coordinates of the cluster center, the number of fragments within the cluster, the average length, the average coverage depth, the average signal weight, and the proportion of suspected toxin gene fragments for each cluster. Extracting global statistical features involves obtaining the total number of clusters, the number of outliers not clustered, the average reconstruction quality of all retained fragments, the average decomposition consistency, and the average overall denoising quality index.

[0151] For example, a certain cluster has the following characteristics: center coordinates (0.2, 0.5…), number of fragments 45, average length 1100bp, average coverage depth 42 times, average signal weight 0.83, and toxin gene ratio 15%; global features include: total number of clusters 10, outlier fragments 28, average R=0.86, average decomposition consistency 0.91, and average overall noise reduction quality 0.88.

[0152] Furthermore, the unknown sequence features are processed by combining the retained judgment results with statistical analysis methods to obtain the unknown denoised sequence feature vector. The unknown denoised sequence feature vector refers to the set of clean, digitized features of the unknown sequence obtained after denoising, verification, and clustering. The extracted clustering statistical features and global statistical features are concatenated in dimensional order to form the final unknown denoised sequence feature vector. For example, concatenating the clustering statistical features of 10 clusters with the global statistical features yields a 120-dimensional unknown denoised sequence feature vector.

[0153] Finally, the known sequence feature vector and the unknown denoised sequence feature vector are concatenated to obtain the comprehensive feature vector. The known sequence feature vector refers to the set of features output from the known feature extraction channels, such as species abundance and toxin gene coverage. Feature concatenation refers to merging the known and unknown denoised feature vectors by dimension to form a full-dimensional feature vector. The comprehensive feature vector is a complete feature vector that integrates known and unknown signals, used for red tide probability prediction. For example, if the known sequence feature vector is 80-dimensional and the unknown denoised sequence feature vector is 120-dimensional, concatenation yields a 200-dimensional comprehensive feature vector.

[0154] In this embodiment of the invention, dual-channel feature extraction achieves the dual goals of accurate mining of known signals and in-depth mining of unknown signals: the known channel fully utilizes prior knowledge to extract deterministic features such as species abundance and toxin coverage, ensuring that no key red tide signals are missed; the unknown channel overcomes database limitations through reassembly, and combines pre-trained dual-channel DAE to achieve adaptive noise reduction. Through weight allocation, reconstruction quality assessment, and quadruple verification, it accurately distinguishes signals from noise, while cluster analysis uncovers potential sequence patterns to avoid losing unknown potential red tide algal signals. The final spliced ​​comprehensive feature vector fully covers both known red tide features and unknown potential features, providing comprehensive, accurate, and high-dimensional feature support for subsequent red tide occurrence probability prediction, thus improving the accuracy of red tide early warning.

[0155] S400: Input the comprehensive feature vector into the prediction model and output the probability value of red tide occurrence.

[0156] In this embodiment of the invention, the comprehensive feature vector is input into the prediction model, and the output is a red tide occurrence probability value. The comprehensive feature vector is merely a set of digitized features and cannot be directly used for red tide early warning decisions; the probability value of a single detection is easily affected by random fluctuations, resulting in insufficient early warning stability; furthermore, it is necessary to transform the abstract probability into an intuitive level to meet the practical requirements of ecological early warning. Therefore, a dedicated prediction model needs to be trained based on historical samples, the probability is smoothed, and the output is graded according to fixed rules to ultimately obtain stable and reliable red tide early warning results.

[0157] Step S400 in the method provided in this embodiment of the invention includes:

[0158] The comprehensive feature vector is standardized to obtain a standardized feature vector, which is then input into the prediction model to obtain the probability value of red tide occurrence.

[0159] The probability values ​​of red tide occurrence detected multiple times in succession are subjected to exponential smoothing to obtain smoothed probability values;

[0160] Based on the smoothed probability value, early warning levels are output, and the short-term rate of change of the smoothed probability value is calculated. When the short-term rate of change exceeds a preset rate threshold, a trend warning prompt is output.

[0161] The construction of the prediction model includes:

[0162] By combining signal denoising autoencoder, noise separation autoencoder and samples collected during the early warning window of historical red tide outbreaks, known sequence feature vectors and unknown denoised sequence feature vectors are extracted to obtain a positive sample feature set;

[0163] By combining signal denoising autoencoder, noise separation autoencoder and samples collected during the early warning window of historical non-red tide outbreak events, known sequence feature vectors and unknown denoised sequence feature vectors are extracted to obtain negative sample feature set;

[0164] The positive and negative sample feature sets are standardized, and the standardized feature sets are used as input, with the occurrence of red tide as the label, to train a logistic regression model as the prediction model.

[0165] First, by combining the signal denoising autoencoder, the noise separation autoencoder, and samples collected during the warning window period of historical red tide outbreaks, known sequence feature vectors and unknown denoised sequence feature vectors are extracted to obtain the positive sample feature set. Historical red tide outbreaks refer to confirmed nearshore red tide occurrences. The warning window period refers to the critical monitoring period of 24-72 hours before a red tide outbreak, during which samples have the highest warning value. The positive sample feature set is a feature set composed of known sequence feature vectors and unknown denoised sequence feature vectors from samples within the red tide outbreak warning window period, representing the red tide risk status.

[0166] Specifically, retrieve seawater sample data collected during the warning window period from historical red tide outbreaks in the target sea area over the past 5-8 years; call the trained signal denoising autoencoder and noise separation autoencoder, and extract the known sequence feature vector and unknown denoising sequence feature vector of the sample according to the aforementioned process; concatenate the two sets of features of all samples into comprehensive features, and summarize them to form a positive sample feature set.

[0167] For example, in coastal areas with high incidence of red tides, 31 verified red tide events over the past 7 years were retrieved, and samples were collected during the warning window period of 24-72 hours before the outbreak, resulting in 320 positive samples. Features were extracted and spliced ​​according to standard procedures to form a positive sample feature set containing 320 comprehensive features.

[0168] Secondly, by combining the signal denoising autoencoder, the noise separation autoencoder, and samples collected during the warning window period of historical non-red tide events, known sequence feature vectors and unknown denoised sequence feature vectors are extracted to obtain a negative sample feature set. Historical non-red tide events refer to regular monitoring periods in the same sea area where no red tides occur. The negative sample feature set is a set of features composed of known sequence feature vectors and unknown denoised sequence feature vectors from samples within the warning window period of non-red tide events, representing the normal marine ecological state.

[0169] Specifically, routine monitoring sample data from the same sea area and non-red tide outbreaks are retrieved; a dual-channel noise reduction autoencoder is also invoked to extract the known sequence feature vectors and unknown noise reduction sequence feature vectors of the samples; after being concatenated into comprehensive features, they are summarized to form a negative sample feature set.

[0170] For example, 1,860 samples from routine monitoring of the above-mentioned sea area were retrieved, covering spring, summer, autumn, and winter seasons and different hydrological environments; features were extracted and spliced ​​according to the same process to form a negative sample feature set containing 1,860 comprehensive features.

[0171] Next, the positive and negative sample feature sets are standardized. Using the standardized feature sets as input and the presence or absence of a red tide as the label, a logistic regression model is trained as the prediction model. Standardization eliminates differences in feature dimension, scale, and numerical value, mapping features to a standardized range to improve model training stability and generalization ability. The label refers to the target value for model training: 1 for a red tide outbreak and 0 for no outbreak. The logistic regression model is a binary classification model suitable for small samples with high-dimensional features and strong interpretability, outputting the probability of red tide occurrence between 0 and 1.

[0172] First, Z-score standardization is used to uniformly process the positive and negative sample feature sets, resulting in a standardized feature set. The prediction model structure adopts a single-layer generalized linear logistic regression model without hidden layers. Specifically, the input layer has the number of neurons equal to the total dimension of the comprehensive feature vector, which is consistent with the dimension of the concatenated known sequence features and unknown denoised sequence features. The output layer has one neuron, using the Sigmoid activation function, to output the probability value of red tide occurrence.

[0173] Secondly, the standardized feature set is used as the model input, with red tide outbreaks marked as 1 and non-red tides marked as 0 as labels. The L-BFGS quasi-Newton optimization algorithm is used for parameter iterative updates, which has fast convergence speed and high stability. Five-fold cross-validation is used to divide the training set and validation set to avoid model bias caused by a single partition. A class weight balancing strategy is introduced to give higher weights to positive samples with fewer samples to alleviate the problem of imbalance between positive and negative samples.

[0174] Furthermore, the total loss function is composed of a weighted sum of the binary cross-entropy loss and the L2 regularization term. The model convergence is determined and training is stopped if any of the following conditions are met: the loss function value decreases by less than 1e-6 for 10 consecutive iterations, and the loss tends to stabilize; the AUC value of the validation set fluctuates by less than 0.001 for 20 consecutive iterations, and the prediction performance no longer improves; or the preset maximum number of iterations of 1000 is reached, and training is forcibly stopped.

[0175] For example, 320 positive samples and 1860 negative samples were standardized using Z-score and then input into the logistic regression model. The L-BFGS optimizer and 5-fold cross-validation were used for training, with a regularization coefficient λ=1.0. By the 126th iteration, the loss had decreased by less than 1e-6 for 10 consecutive iterations, satisfying the convergence condition. The prediction model training was completed, and the validation set AUC reached 0.92, indicating stable model generalization performance. The trained prediction model was thus obtained.

[0176] Based on this, the comprehensive feature vector is standardized to obtain a standardized feature vector, which is then input into the prediction model to obtain the red tide occurrence probability value. The standardized feature vector refers to the feature data of the comprehensive feature vector to be tested, after being standardized in accordance with the training set, to meet the model's input requirements. The red tide occurrence probability value is a value between 0 and 1 output by the prediction model; a larger value indicates a higher probability of red tide occurrence.

[0177] Specifically, the comprehensive feature vector obtained from S300 undergoes the same standardization process as the training set to obtain a standardized feature vector. This standardized feature vector is then input into the trained prediction model, which directly outputs the probability value of red tide occurrence. For example, the prediction model outputs a red tide occurrence probability value of 0.76.

[0178] Then, the probability values ​​of red tide occurrence from multiple consecutive detections are exponentially smoothed to obtain smoothed probability values. Exponential smoothing refers to weighted averaging of the probability values ​​from multiple consecutive detections at the same station, reducing fluctuations in individual detections and obtaining more stable smoothed probability values. The smoothed probability value is the stable probability after smoothing calibration and serves as the sole basis for early warning classification.

[0179] Specifically, the probability values ​​of red tide occurrence from five or more consecutive measurements at the same monitoring station are retrieved; weighted using an exponential smoothing formula to eliminate random fluctuations; and a stable smoothed probability value is output. For example, the probabilities of six consecutive measurements at this station are: 0.68, 0.71, 0.74, 0.76, 0.75, and 0.77; after exponential smoothing, a stable smoothed probability value of 0.74 is obtained.

[0180] Finally, based on the smoothed probability value, a graded early warning output is performed, and the short-term rate of change of the smoothed probability value is calculated. When the short-term rate of change exceeds a preset rate threshold, a trend early warning is output. The short-term rate of change refers to the magnitude of the rise and fall of the smoothed probability value within a short period of time, reflecting the speed of change in red tide risk. The preset rate threshold is a critical value for determining rapid changes in risk; exceeding it triggers a trend early warning.

[0181] Specifically, the warning levels are as follows: a smoothing probability value < 0.3 indicates a normal level; 0.3 ≤ smoothing probability value < 0.6 indicates a level of concern; 0.6 ≤ smoothing probability value < 0.8 indicates a warning level; and a smoothing probability value ≥ 0.8 indicates an alarm level. The short-term rate of change of the smoothing probability value is calculated; if the rate of change exceeds a preset threshold, an additional trend warning is output.

[0182] For example, the smoothing probability value is 0.74, which corresponds to the output warning level; the short-term change rate is calculated to be 0.11 / time, which exceeds the preset threshold of 0.1; the final output is: warning level + warning message of rapid rise in red tide risk.

[0183] In this embodiment of the invention, a stable logistic regression prediction model is trained based on historical positive and negative samples, which transforms high-dimensional comprehensive features into intuitive red tide occurrence probabilities; probability fluctuations are eliminated through exponential smoothing to improve early warning stability; four-level early warning output is completed according to fixed thresholds, and trend early warning is achieved by combining short-term change rates; finally, the results of complex metagenomic analysis are transformed into intuitive ecological early warning instructions to meet the needs of red tide emergency monitoring and management decision-making.

[0184] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects:

[0185] This invention provides a rapid metagenomic screening method and system for red tide algae for ecological early warning. First, raw data is purified through sample sequencing and multi-dimensional quality control to eliminate processing biases and interference. Then, a reference library-guided two-layer sorting process accurately separates known and unknown sequences, efficiently identifying known red tide algae and toxin gene signals based on prior knowledge while completely preserving unknown sequences of potential red tide algae not yet included in the database. Subsequently, dual-channel feature extraction and adaptive noise reduction collaboratively mine dominant features of known sequences and potential features of unknown sequences, accurately separating biological signals from sequencing noise, achieving cross-scale compression and key information integration of massive metagenomic data. Finally, a prediction model trained with sufficient compliant samples completes probability calculation, smoothing calibration, and graded early warning output, transforming complex metagenomic analysis into intuitive red tide early warning results. Ultimately, this achieves rapid, accurate, comprehensive, and highly reliable early warning for red tide ecosystems.

[0186] Example 2, as Figure 2 As shown, this invention provides a rapid metagenomic screening system for red tide algae for ecological early warning, the system comprising:

[0187] Sequencing quality control module 11 is used to perform shotgun metagenomic sequencing on seawater samples to obtain sequencing data, and to perform quality control on the sequencing data to obtain clean sequencing sequences.

[0188] The sequence sorting module 12 is used to compare and sort the clean sequencing sequences according to the existing prior reference database to obtain a known sequence set and an unknown sequence set;

[0189] The dual-channel feature fusion module 13 is used to input the known sequence set and the unknown sequence set into the known feature extraction channel and the unknown feature extraction channel respectively, obtain the dual-channel output and perform feature concatenation to obtain a comprehensive feature vector.

[0190] The red tide probability prediction module 14 is used to input the comprehensive feature vector into the prediction model and output the red tide occurrence probability value.

[0191] In one embodiment, the sequencing quality control module 11 is further configured to:

[0192] Seawater samples were collected from the surface of the target sea area, and the seawater samples were filtered using a filter membrane to obtain filter membrane samples.

[0193] Based on the recovery rate of exogenous standard DNA, the filter membrane sample was sequenced and corrected, and the extracted DNA was used to construct a library to obtain the sequencing data.

[0194] The sequencing data is subjected to quality control and preprocessing, including quality filtering, host contamination removal, and quality assessment, to obtain clean sequencing sequences.

[0195] In one embodiment, the sequence sorting module 12 is further configured to:

[0196] The clean sequencing sequences are rapidly screened using the rapid classification database in the existing prior reference database to determine the high-confidence known group and the undetermined group;

[0197] The pending group is further compared and filtered using the extended reference database in the existing prior reference database, wherein the extended comparison and filtering includes consistency filtering and coverage filtering.

[0198] Based on the extended comparison and screening results, short read sequences in the undetermined group that meet the thresholds for both consistency and coverage are added to the high-confidence known group, and the output is the known sequence set;

[0199] Short read sequences outside the high-confidence known group are stored as the unknown sequence set.

[0200] In one embodiment, the dual-channel feature fusion module 13 is further configured to:

[0201] The known sequence set is input into the known feature extraction channel to obtain the known sequence feature vector, which includes species abundance, toxin gene coverage, and abundance statistical indicators.

[0202] The short read sequences in the unknown sequence set are reassembled to obtain an assembled fragment, and features are extracted from the assembled fragment through an unknown feature extraction channel to obtain an unknown sequence feature vector, wherein the unknown sequence feature vector includes a basic attribute class and a cropped short read sequence frequency class;

[0203] The unknown sequence feature vector is obtained by adaptively denoising the unknown sequence feature vector through a pre-trained dual-channel denoising autoencoder.

[0204] The known sequence feature vector and the unknown denoised sequence feature vector are concatenated to obtain the comprehensive feature vector.

[0205] Specifically, an unknown denoised sequence feature vector is obtained by adaptively denoising the unknown sequence feature vector using a pre-trained dual-channel denoising autoencoder, including:

[0206] The unknown sequence feature vector is input into a pre-trained signal denoising autoencoder and a pre-trained noise separation autoencoder respectively to obtain signal reconstruction output and noise reconstruction output;

[0207] The signal weights are calculated based on the first reconstruction error between the unknown sequence feature vector and the signal reconstruction output, and the second reconstruction error between the unknown sequence feature vector and the noise reconstruction output.

[0208] Based on the signal weights, the signal reconstruction output and the noise reconstruction output are weighted and fused to obtain the fused reconstruction output;

[0209] Based on the difference between the unknown sequence feature vector and the fused reconstruction output, the reconstruction quality is calculated, wherein the reconstruction quality is used to characterize the noise reduction effect of the dual-channel noise reduction autoencoder;

[0210] Based on the difference between the unknown sequence feature vector and the sum of the signal reconstruction output and the noise reconstruction output, the decomposition consistency is calculated. The decomposition consistency is used to characterize the degree of separation and complementarity between the signal denoising autoencoder and the noise separation autoencoder.

[0211] Based on the reconstruction quality and the decomposition consistency, a comprehensive noise reduction quality index is calculated;

[0212] Based on the signal weights and the reconstruction quality, each unknown sequence feature vector is determined to be retained. The retention determination results are then combined with statistical analysis methods to process the unknown sequence features and obtain the unknown denoised sequence feature vector.

[0213] Specifically, based on the signal weights and the reconstruction quality, a retention determination is made for each unknown sequence feature vector, and the unknown sequence features are processed using the retention determination results and statistical analysis methods to obtain the unknown denoised sequence feature vector, including:

[0214] Based on the signal weights and the reconstruction quality, a retention determination is made for each unknown sequence feature vector, including:

[0215] When the signal weight is not less than the first threshold and the reconstruction quality is not less than the second threshold, it is determined to be a high-confidence signal sequence, and the corresponding unknown sequence feature vector is retained.

[0216] When the signal weight is less than the third threshold, it is determined to be a high-confidence noise sequence, and the corresponding unknown sequence feature vector is discarded;

[0217] When either the signal weight is not less than the third threshold and is less than the first threshold, or the reconstruction quality is less than the second threshold, an additional verification process is initiated.

[0218] The additional verification process sequentially performs sequence quality index checks, assembly coverage depth checks, functional signal checks, and cross-sample consistency checks.

[0219] When any item in the additional verification process is determined to be a signal sequence, the corresponding unknown sequence feature vector is retained; when any item in the additional verification process is determined to be a noise sequence, the corresponding unknown sequence feature vector is discarded.

[0220] For all retained unknown sequence feature vectors, the encoder output of the signal denoising autoencoder is taken as the latent space representation, and cluster analysis is performed on the latent space representation;

[0221] Based on the cluster analysis results, cluster statistical features and global statistical features are extracted and concatenated to form the feature vector of the unknown denoised sequence.

[0222] The pre-training of the signal denoising autoencoder and the noise separation autoencoder includes:

[0223] Based on the reference genome sequences in the existing prior reference database, sequencing data is simulated and assembled, and feature vectors of the assembled fragments are extracted to form a signal training set.

[0224] Based on a known set of noise sequences, the sequences are assembled and the feature vectors of the assembled segments are extracted to form a noise training set.

[0225] A combination of noise is applied to each signal training sample in the signal training set to obtain a noisy signal training sample, wherein the combination of noise is composed of Gaussian noise and a random mask superimposed.

[0226] Mix a feature vector randomly selected from the signal training set into each noise training sample in the noise training set to obtain a mixed noise training sample;

[0227] Using the noisy signal training samples as input, and combining signal reconstruction loss and L1 sparse regularization loss, the signal denoising autoencoder is trained so that its reconstructed output approximates the signal training samples.

[0228] Using the mixed noise training samples as input, and combining the noise separation loss and L1 sparse regularization loss, the noise separation autoencoder is trained so that its reconstructed output approximates the noise training samples.

[0229] By minimizing the total loss function weighted by the decomposition consistency loss, signal reconstruction loss, and noise separation loss, the signal denoising autoencoder and the noise separation autoencoder are jointly fine-tuned to obtain the signal denoising autoencoder and the noise separation autoencoder.

[0230] In one embodiment, the red tide probability prediction module 14 is further configured to:

[0231] The comprehensive feature vector is standardized to obtain a standardized feature vector, which is then input into the prediction model to obtain the probability value of red tide occurrence.

[0232] The probability values ​​of red tide occurrence detected multiple times in succession are subjected to exponential smoothing to obtain smoothed probability values;

[0233] Based on the smoothed probability value, early warning levels are output, and the short-term rate of change of the smoothed probability value is calculated. When the short-term rate of change exceeds a preset rate threshold, a trend warning prompt is output.

[0234] The construction of the prediction model includes:

[0235] By combining signal denoising autoencoder, noise separation autoencoder and samples collected during the early warning window of historical red tide outbreaks, known sequence feature vectors and unknown denoised sequence feature vectors are extracted to obtain a positive sample feature set;

[0236] By combining signal denoising autoencoder, noise separation autoencoder and samples collected during the early warning window of historical non-red tide outbreak events, known sequence feature vectors and unknown denoised sequence feature vectors are extracted to obtain negative sample feature set;

[0237] The positive and negative sample feature sets are standardized, and the standardized feature sets are used as input, with the occurrence of red tide as the label, to train a logistic regression model as the prediction model.

[0238] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A rapid metagenomic screening method for red tide algae for ecological early warning, characterized in that, include: Shotgun metagenomic sequencing was performed on seawater samples to obtain sequencing data, and the sequencing data was subjected to quality control to obtain clean sequencing sequences. Based on the existing prior reference database, the clean sequencing sequences are compared and sorted to obtain a known sequence set and an unknown sequence set. The known sequence set and the unknown sequence set are respectively input into the known feature extraction channel and the unknown feature extraction channel to obtain dual-channel output and feature concatenation to obtain a comprehensive feature vector. The comprehensive feature vector is input into the prediction model, and the probability value of red tide occurrence is output. Specifically, the known sequence set and the unknown sequence set are input to the known feature extraction channel and the unknown feature extraction channel, respectively, to obtain dual-channel outputs, which are then concatenated to obtain a comprehensive feature vector, including: The known sequence set is input into the known feature extraction channel to obtain known sequence feature vectors including species abundance, toxin gene coverage, and abundance statistical indicators. The short read sequences in the unknown sequence set are reassembled to obtain an assembled fragment, and features are extracted from the assembled fragment through an unknown feature extraction channel to obtain an unknown sequence feature vector, wherein the unknown sequence feature vector includes a basic attribute class and a cropped short read sequence frequency class; The unknown sequence feature vector is obtained by adaptively denoising the unknown sequence feature vector through a pre-trained dual-channel denoising autoencoder. The known sequence feature vector and the unknown denoised sequence feature vector are concatenated to obtain the comprehensive feature vector; Specifically, an unknown denoised sequence feature vector is obtained by adaptively denoising the unknown sequence feature vector using a pre-trained dual-channel denoising autoencoder, including: The unknown sequence feature vector is input into a pre-trained signal denoising autoencoder and a pre-trained noise separation autoencoder respectively to obtain signal reconstruction output and noise reconstruction output; The signal weights are calculated based on the first reconstruction error between the unknown sequence feature vector and the signal reconstruction output, and the second reconstruction error between the unknown sequence feature vector and the noise reconstruction output. Based on the signal weights, the signal reconstruction output and the noise reconstruction output are weighted and fused to obtain the fused reconstruction output; Based on the difference between the unknown sequence feature vector and the fused reconstruction output, the reconstruction quality is calculated, wherein the reconstruction quality is used to characterize the noise reduction effect of the dual-channel noise reduction autoencoder; Based on the difference between the unknown sequence feature vector and the sum of the signal reconstruction output and the noise reconstruction output, the decomposition consistency is calculated. The decomposition consistency is used to characterize the degree of separation and complementarity between the signal denoising autoencoder and the noise separation autoencoder. Based on the reconstruction quality and the decomposition consistency, a comprehensive noise reduction quality index is calculated; Based on the signal weights and the reconstruction quality, each unknown sequence feature vector is determined to be retained. The retention determination results are then combined with statistical analysis methods to process the unknown sequence features and obtain the unknown denoised sequence feature vector.

2. The rapid metagenomic screening method for red tide algae for ecological early warning as described in claim 1, characterized in that, Shotgun metagenomic sequencing was performed on seawater samples to obtain sequencing data. The sequencing data underwent quality control to obtain clean sequencing sequences, including: Seawater samples were collected from the surface of the target sea area, and the seawater samples were filtered using a filter membrane to obtain filter membrane samples. Based on the recovery rate of exogenous standard DNA, the filter membrane sample was sequenced and corrected, and the extracted DNA was used to construct a library to obtain the sequencing data. The sequencing data is subjected to quality control and preprocessing, including quality filtering, host contamination removal, and quality assessment, to obtain clean sequencing sequences.

3. The rapid metagenomic screening method for red tide algae for ecological early warning as described in claim 1, characterized in that, Based on the existing prior reference database, the clean sequencing sequences are compared and sorted to obtain a known sequence set and an unknown sequence set, including: The clean sequencing sequences are rapidly screened using the rapid classification database in the existing prior reference database to determine the high-confidence known group and the undetermined group; The pending group is further compared and filtered using the extended reference database in the existing prior reference database, wherein the extended comparison and filtering includes consistency filtering and coverage filtering. Based on the extended comparison and screening results, short read sequences in the undetermined group that meet the thresholds for both consistency and coverage are added to the high-confidence known group, and the output is the known sequence set; Short read sequences outside the high-confidence known group are stored as the unknown sequence set.

4. The rapid metagenomic screening method for red tide algae for ecological early warning as described in claim 1, characterized in that, Based on the signal weights and the reconstruction quality, a retention determination is made for each unknown sequence feature vector, and the unknown sequence features are processed by combining the retention determination results with statistical analysis methods to obtain the unknown denoised sequence feature vector, including: Based on the signal weights and the reconstruction quality, a retention determination is made for each unknown sequence feature vector, including: When the signal weight is not less than the first threshold and the reconstruction quality is not less than the second threshold, it is determined to be a high-confidence signal sequence, and the corresponding unknown sequence feature vector is retained. When the signal weight is less than the third threshold, it is determined to be a high-confidence noise sequence, and the corresponding unknown sequence feature vector is discarded; When either the signal weight is not less than the third threshold and is less than the first threshold, or the reconstruction quality is less than the second threshold, an additional verification process is initiated. The additional verification process sequentially performs sequence quality index checks, assembly coverage depth checks, functional signal checks, and cross-sample consistency checks. When any item in the additional verification process is determined to be a signal sequence, the corresponding unknown sequence feature vector is retained; when any item in the additional verification process is determined to be a noise sequence, the corresponding unknown sequence feature vector is discarded. For all retained unknown sequence feature vectors, the encoder output of the signal denoising autoencoder is taken as the latent space representation, and cluster analysis is performed on the latent space representation; Based on the cluster analysis results, cluster statistical features and global statistical features are extracted and concatenated to form the feature vector of the unknown denoised sequence.

5. The rapid metagenomic screening method for red tide algae for ecological early warning as described in claim 1, characterized in that, The pre-training of the signal denoising autoencoder and the noise separation autoencoder includes: Based on the reference genome sequences in the existing prior reference database, sequencing data is simulated and assembled, and feature vectors of the assembled fragments are extracted to form a signal training set. Based on a known set of noise sequences, the sequences are assembled and the feature vectors of the assembled segments are extracted to form a noise training set. A combination of noise is applied to each signal training sample in the signal training set to obtain a noisy signal training sample, wherein the combination of noise is composed of Gaussian noise and a random mask superimposed. Mix a feature vector randomly selected from the signal training set into each noise training sample in the noise training set to obtain a mixed noise training sample; Using the noisy signal training samples as input, and combining signal reconstruction loss and L1 sparse regularization loss, the signal denoising autoencoder is trained so that its reconstructed output approximates the signal training samples. Using the mixed noise training samples as input, and combining the noise separation loss and L1 sparse regularization loss, the noise separation autoencoder is trained so that its reconstructed output approximates the noise training samples. By minimizing the total loss function weighted by the decomposition consistency loss, signal reconstruction loss, and noise separation loss, the signal denoising autoencoder and the noise separation autoencoder are jointly fine-tuned to obtain the signal denoising autoencoder and the noise separation autoencoder.

6. The rapid metagenomic screening method for red tide algae for ecological early warning as described in claim 1, characterized in that, The comprehensive feature vector is input into the prediction model, and the output red tide occurrence probability value is included: The comprehensive feature vector is standardized to obtain a standardized feature vector, which is then input into the prediction model to obtain the probability value of red tide occurrence. The probability values ​​of red tide occurrence detected multiple times in succession are subjected to exponential smoothing to obtain smoothed probability values; Based on the smoothed probability value, early warning levels are output, and the short-term rate of change of the smoothed probability value is calculated. When the short-term rate of change exceeds a preset rate threshold, a trend warning is output.

7. The rapid metagenomic screening method for red tide algae for ecological early warning as described in claim 1, characterized in that, The construction of the prediction model includes: By combining signal denoising autoencoder, noise separation autoencoder and samples collected during the early warning window of historical red tide outbreak events, known sequence feature vectors and unknown denoised sequence feature vectors are extracted to obtain a positive sample feature set; By combining signal denoising autoencoder, noise separation autoencoder and samples collected during the early warning window of historical non-red tide outbreak events, known sequence feature vectors and unknown denoised sequence feature vectors are extracted to obtain negative sample feature set; The positive and negative sample feature sets are standardized, and the standardized feature sets are used as input, with the occurrence of red tide as the label, to train a logistic regression model as the prediction model.

8. A rapid metagenomic screening system for red tide algae for ecological early warning, characterized in that, The method for rapid metagenomic screening of red tide algae for ecological early warning, as described in any one of claims 1-7, comprises: The sequencing quality control module is used to perform shotgun metagenomic sequencing on seawater samples to obtain sequencing data, and to perform quality control on the sequencing data to obtain clean sequencing sequences. The sequence sorting module is used to compare and sort the clean sequencing sequences according to the existing prior reference database to obtain a known sequence set and an unknown sequence set; The dual-channel feature fusion module is used to input the known sequence set and the unknown sequence set into the known feature extraction channel and the unknown feature extraction channel respectively, obtain the dual-channel output, and perform feature concatenation to obtain a comprehensive feature vector. The red tide probability prediction module is used to input the comprehensive feature vector into the prediction model and output the red tide occurrence probability value.

Images