Intelligent identification and real-time early warning system for unknown pathogens in respiratory tract sample of fever outpatient service

Abstract

The invention relates to the field of public health, in particular to an intelligent identification and real-time early warning system for unknown pathogens in respiratory tract samples of fever outpatient service, which comprises a data quality control processing module for performing quality control and standardized preprocessing on metagenome original sequencing data to obtain pure microorganism reading length; the known pathogen identification and annotation module is used for identifying known pathogens, classifying non-host reads, constructing a classification database according to a classification result, and annotating reads with low classification confidence or unclassified reads in the classification result to obtain an annotation table; the recognition model training module is used for constructing a positive and negative sample data set and training a recognition model; and the unknown pathogen identification module is used for identifying potential unknown pathogens in the metagenome of the pure microorganism read length by adopting the trained identification model according to the classification database and the annotation list. According to the invention, identification, classification and dynamic monitoring of known and unknown respiratory pathogens can be realized.

Description

Technical Field

[0001] This invention relates to the field of public health, and in particular to an intelligent identification and real-time early warning system for unknown pathogens in respiratory samples from fever clinics. Background Technology

[0002] Early warning and pathogen surveillance of respiratory infectious diseases are crucial components of modern public health prevention and control systems. In recent years, with the development of metagenomic next-generation sequencing (mNGS) technology, the identification and quantitative analysis of pathogens in complex samples using high-throughput sequencing platforms has become a research and application hotspot. Common methods for microbial identification and trend monitoring include:

[0003] 1. Metagenomic Sequencing-Based Pathogen Classification Methods. Known pathogen sequences are identified using rapid classification tools based on k-mers, such as Kraken2. For sequences that fail to be identified by Kraken2, sequence alignment-based tools are typically used for supplementary annotation. BLASTn (Basic Local Alignment Search Tool for nucleotides) is a tool based on a local alignment algorithm that can perform one-to-one alignment of the query nucleic acid sequence with sequences in a known nucleic acid database, making it suitable for detecting the similarity of conserved nucleic acid regions. DIAMOND, on the other hand, is a high-performance alignment tool for protein sequences that can translate the input nucleic acid sequence into six open reading frames and perform rapid alignment with large-scale protein databases (such as NR), making it suitable for identifying low-similarity or distantly homologous protein sequences.

[0004] To discover potential pathogen sequences not annotated in databases, de novo assembly strategies are often used to assemble high-throughput sequencing reads. Commonly used tools include MEGAHIT, an assembly algorithm optimized based on the DBG (De-Bruijn Graph) data structure, which can efficiently process large-scale metagenomic data on a single-node server. This method constructs multi-scale assembly graphs by traversing different k-mer lengths, extracting longer contig sequences, which helps identify complete genomic regions of rare pathogens, co-infectious fragments, and candidate novel species.

[0005] 2. Gene Structure Identification and Functional Annotation. Prokka is a commonly used tool for annotating contig sequences obtained from splicing. It is an automated annotation pipeline designed for prokaryotic genomes, integrating modules such as gene prediction, rRNA / tRNA detection, and database alignment. Prokka can quickly output annotation files from GFF, GenBank, and protein sequences, identifying CDS, functional enzymes, and putative proteins, providing a foundation for subsequent functional classification, virulence factor prediction, or drug resistance gene prediction.

[0006] 3. Change point detection methods in time series. The Bayesian change point detection method was proposed by Barry and Hartigan (1993). It assumes that the time series consists of multiple segmented normal distributions and uses Bayesian inference to calculate the posterior probability that each time point is a change point. It is widely used in expression analysis, environmental monitoring and other non-stationary sequence modeling.

[0007] In terms of tool implementation, Ruptures is an analysis library in Python that supports various change point detection models. Its PELT (Pruned Exact Linear Time) algorithm has linear time complexity, making it suitable for processing large-scale time series. When using the RBF (Radial Basis Function) model, Ruptures can identify structural change points that simultaneously contain abrupt changes in both mean and variance, making it suitable for nonlinear mutations commonly found in pathogen abundance, especially in the early stages of trace pathogen outbreaks.

[0008] It is evident that in traditional methods, tools such as Kraken2, BLASTn, and DIAMOND operate in isolation, easily leading to redundant results or annotation conflicts, thus reducing the accuracy of pathogen identification. Furthermore, traditional pathogen identification methods rely on homologous sequence alignment and annotation, failing to effectively identify novel and unknown pathogens. Moreover, traditional prediction methods predict individual pathogens independently, ignoring inter-pathogen relationships, resulting in limited prediction accuracy. Summary of the Invention

[0009] This invention provides a system for analyzing and warning of unknown pathogens in respiratory specimens from patients with fever based on metagenomic data. This system can identify, classify, and dynamically monitor known and unknown respiratory pathogens, overcoming the limitations of traditional detection methods in the discovery and early warning of emerging pathogens.

[0010] In the technical solution adopted in the embodiments of the present invention, the intelligent identification and real-time early warning system for unknown pathogens in respiratory samples from fever clinics includes the following modules:

[0011] The data quality control and processing module is used to perform quality control and standardized preprocessing on raw metagenomic sequencing data to obtain pure microbial reads.

[0012] The known pathogen identification and annotation module is used to identify known pathogens, classify non-host reads, construct a classification database based on the classification results, and annotate reads with low classification confidence or unclassified reads in the classification results to obtain a standard annotation table.

[0013] The identification model training module is used to construct a positive and negative sample dataset and train the identification model for identifying unknown pathogens.

[0014] The unknown pathogen identification module uses a trained identification model to identify potential unknown pathogens in the metagenomics of pure microbial reads, based on the classification database and standard annotation table constructed by the known pathogen identification annotation module.

[0015] Preferably, the data quality control processing module uses the FastQC tool to assess the quality of the raw sequencing data for each sample. The assessment includes base quality distribution, GC content, adapter contamination, sequence repetition rate, and N base content. The Trimmomatic tool is used to trim the raw sequencing data, removing adapter sequences and low-quality base fragments. The Bowtie2 tool is used to align the trimmed paired reads with the human reference genome GRCh38. Using the --un-conc-gz parameter, paired reads that do not align with the host genome are output and retained as non-host reads, resulting in pure microbial reads.

[0016] Preferably, the known pathogen identification and annotation module uses the Kraken2 tool to quickly classify non-host reads; the Kraken2 tool uses the k-mer exact matching strategy to align sequencing sequences, obtain preliminary species attribution and LCA depth, and construct a classification database.

[0017] Furthermore, the BLASTn nucleic acid sequence alignment tool and the DIAMOND protein sequence alignment tool were used to annotate reads with low classification confidence or those that were not classified in the Kraken2 tool classification results. The annotation results were standardized through a unified format and then integrated to output a standard annotation table with unique attribution.

[0018] Preferably, in the unknown pathogen identification module, the process for identifying unknown pathogens is as follows: for unaligned reads, the MEGAHIT tool is used to perform de novo splicing and assembly to obtain spliced ​​sequence contigs, and the spliced ​​sequence contigs are used for species classification and functional annotation; then, the trained identification model is used to predict the spliced ​​sequence.

[0019] Furthermore, the unknown pathogen identification module employs a deep learning-based unknown pathogen potential scoring model to screen non-host contigs obtained using the MEGAHIT tool. The screening process includes:

[0020] Structural and statistical features are extracted from each contig of spliced ​​sequences;

[0021] The features are input into the trained recognition model, and the output is a probability score that the contiguous sequence is a potential viral sequence.

[0022] Contigs with a probability score higher than a set threshold are marked as candidate unknown pathogens.

[0023] Preferably, the early warning system further includes:

[0024] The dynamic monitoring and early warning module is used to summarize pathogens detected in samples across time points and construct an abundance time series; it uses the Transformer model to perform simultaneous prediction of multiple pathogens; and it introduces a change point detection model to automatically identify trend changes and outbreak signals.

[0025] The present invention has the following advantages and effects compared with the prior art:

[0026] 1. Improve the accuracy of annotations.

[0027] This invention unifies and standardizes the output results of multiple tools, selects high-reliability results, significantly improves the accuracy of pathogen annotation, and increases data processing efficiency and analysis speed, making it particularly suitable for large-scale sample batch analysis scenarios.

[0028] 2. Improve the early warning capabilities for emerging pathogens.

[0029] This invention introduces a deep convolutional neural network (CNN) model, which automatically extracts features based on sequence features. It can screen potential emerging pathogens without explicit homology references, and realizes automated and efficient identification of unknown potential viruses without clear homology sequences. It effectively improves the detection sensitivity of unknown pathogens and enables early detection and rapid response to emerging and sudden pathogens. It is especially suitable for pathogen response and monitoring in global public health events.

[0030] 3. Make full use of the synergistic characteristics among pathogens to improve prediction accuracy.

[0031] This invention effectively captures potential collaborative transmission and dynamic interaction patterns among pathogens by simultaneously inputting multi-pathogen sequence information, significantly improving the accuracy of simultaneous multi-pathogen prediction. Through batch simultaneous prediction of future trends of multiple pathogens using models such as the Transformer, prediction efficiency and real-time performance are greatly improved, meeting the public health system's need for rapid and accurate prediction. Furthermore, the prediction results, combined with multi-pathogen variable point detection strategies, can quickly identify potential abnormal outbreak trends and trigger timely public health early warnings. Attached Figure Description

[0032] Figure 1 This is a structural block diagram of the intelligent identification and real-time early warning system in an embodiment of the present invention. Detailed Implementation

[0033] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0034] Example

[0035] This embodiment proposes an intelligent identification and real-time early warning system for unknown pathogens in respiratory samples from fever clinics, such as... Figure 1 As shown, it includes a data quality control processing module, a known pathogen identification and annotation module, an identification model training module, and an unknown pathogen identification module.

[0036] In the data quality control module, quality control and standardization preprocessing are performed on the raw metagenomic sequencing data. FastQC is used to assess the quality of the raw sequencing data, identifying issues such as adapter contamination, base quality degradation, and GC shifts. Trimmomatic is used to perform an automated trimming process: removing low-quality tails and adapter sequences, filtering short sequences below 50 bp, and standardizing sequencing data quality. Bowtie2 is used to align the cleaned reads with the human reference genome, removing host background sequences to obtain clean microbial reads, providing uniform, high-quality data input for subsequent analysis. Specifically, this includes:

[0037] (1) Quality assessment of raw sequencing data.

[0038] The FastQC tool was used to assess the quality of the raw sequencing data (.fastq.gz format) for each sample. The main assessment parameters included: base quality distribution, GC content, adapter contamination, sequence repetition rate, and N base content. After execution, quality report files in HTML and ZIP formats were generated for analysts to confirm whether the sequencing quality was acceptable. A graphical quality assessment report in HTML format and a compressed statistical data report were also generated to determine whether the obtained raw sequencing data met the preset quality control standards. When the assessment results met the set thresholds (e.g., Q30 ratio ≥ 85%, GC content between 50% and 60%, low adapter contamination rate, etc.), the adapter removal and low-quality fragment trimming steps were initiated.

[0039] (2) Joint removal and low-quality fragment trimming

[0040] To remove adapter sequences and low-quality base fragments, the Trimmomatic tool was used to trim the raw sequencing data.

[0041] (3) Host sequence removal

[0042] To eliminate interference from the human host background, the Bowtie2 tool was used to align the pruned paired reads with the human reference genome GRCh38. The alignment parameter was set to `--very-sensitive` to enhance the ability to identify reads with potentially low similarity. Simultaneously, the `--un-conc-gz` parameter was used to output paired reads that did not align with the host genome; these were retained as non-host reads for subsequent pathogen identification analysis.

[0043] The above process can be automated under the Snakemake workflow management system and is suitable for large-scale sample analysis tasks in sentinel hospitals, research institutions or disease control centers.

[0044] The known pathogen identification and annotation module is used to identify known pathogens, classify non-host reads, construct a classification database based on the classification results, and accurately annotate reads with low classification confidence or those that are not classified, thereby obtaining a standard annotation table.

[0045] In this embodiment, the aforementioned identification and annotation module uses the Kraken2 tool to rapidly classify non-host reads. The Kraken2 tool employs a k-mer exact matching strategy to align sequencing sequences, obtaining preliminary species attribution and LCA depth, and constructing a classification database. This database includes NCBI RefSeq data on viruses, bacteria, fungi, and prokaryotes, and is regularly updated to improve identification coverage.

[0046] To further validate and supplement the classification results of the Kraken2 tool, the BLASTn nucleic acid sequence alignment tool and the DIAMOND protein sequence alignment tool were used to accurately annotate reads with low classification confidence or those that were unclassified in the Kraken2 tool classification results. When using the BLASTn tool, the -evalue was set to 1e-5 and the -perc_identity was set to 90. When using the DIAMOND tool, the -evalue was also set to 1e-5 to maintain consistency with the BLASTn tool, and the -k was set to 1 to reduce redundant comparisons and thus speed up processing.

[0047] All annotation results were standardized using a unified format (JSON / TSV) and then integrated based on the following methods: annotations at lower LCA levels (species / genus level) were prioritized; results with high consistency (identity > 90%, length coverage > 70%) were assigned higher scores; and for multiple matches of the same read length, only the highest-scoring entry was retained. Finally, a uniquely attributed standard annotation table was output, and a "sample × pathogen" abundance matrix was constructed to store the standard annotation table.

[0048] In summary, the known pathogen identification and annotation module utilizes the Kraken2 tool to construct a rapid and accurate k-mer species classification system, and combines a customized database (integrating RefSeq, nt, and virus databases) to annotate reads. BLASTn is used to confirm suspected reads through nucleic acid sequence alignment, and Diamond is used to perform protein-level alignment of unannotated reads, improving the sensitivity for identifying pathogens with mutations or structural variations. A module for generating pathogen composition maps and abundance tables is established to facilitate subsequent comparison with historical monitoring data and achieve dynamic change analysis.

[0049] This embodiment uses a convolutional neural network (CNN) to build a recognition model for identifying unknown pathogens, and trains it using a recognition model training module. To train the CNN model for identifying unknown pathogens, this embodiment uses a manually constructed dataset of positive and negative samples, including the following:

[0050] Positive set: For known viral sequences from the public databases NCBI RefSeq and Virus-Host DB, select contigs with clear viral tags (such as influenza virus, coronavirus, rotavirus, etc.). Their characteristics include: containing typical viral structural proteins (such as RdRp, capsid), no host-origin contamination, and a length range of 500bp-10kb (simulating real contigs).

[0051] Negative set: Non-viral sequences from humans or common microorganisms in the environment (such as bacteria, fungi, prokaryotes), as well as false-positive fragments generated by background splicing. Characteristics include: lack of viral functional domain annotation, commonly found in background host or sample contamination sequences.

[0052] By constructing a balanced dataset (50% positive samples + 50% negative samples), the identification model can learn key sequence features to distinguish between viral and non-viral contigs.

[0053] The structural design of convolutional neural networks is shown in Table 1:

[0054] Table 1

[0055]

[0056] The one-hot mapping relationship is as follows:

[0057]

[0058] In this embodiment, when training the CNN model, the loss function is binary cross-entropy; the learning rate of the optimizer Adam is initially set to 1e-4; 5-fold cross-validation is used to evaluate the model's generalization ability, and the validation metrics are AUC (Area Under ROC Curve) and F1-score. If the validation AUC does not improve for 5 consecutive rounds, the training is terminated.

[0059] The unknown pathogen identification module is used to identify potential unknown pathogens in the metagenomics of pure microbial reads based on the classification database and standard annotation table constructed by the known pathogen identification annotation module and the trained identification model.

[0060] The main process for identifying unknown pathogens is as follows: for unaligned reads, the MEGAHIT tool is used to perform de novo splicing to obtain spliced ​​sequence contigs, which are then used for species classification and functional annotation; and finally, a trained convolutional neural network is used to predict the spliced ​​sequences.

[0061] Specifically, to identify potential unknown pathogens in the metagenomics, after removing the host sequence, the MEGAHIT tool was used to perform de novo splicing of non-host reads with a k-mer of 21, obtaining contigs based on read overlap. The meta-large preset was then used to optimize the splicing quality. After splicing, a script was used to screen the contigs, with the filtering criteria being a contig length of at least 500 bp and an average sequencing depth of at least 10×.

[0062] In this embodiment, to overcome the limitation of traditional annotation tools in identifying non-homologous sequences, a deep learning-based unknown pathogen potential scoring model is introduced to screen non-host contigs obtained using the MEGAHIT tool. The process is as follows:

[0063] (1) Extract structural and statistical features from each contig of spliced ​​sequences, including: k-mer spectrum, GC content and length, number and length distribution of open reading frames (ORFs), and whether it contains virus-related protein domains (such as RdRp, capsid, etc.).

[0064] (2) Input the above features into the trained recognition model, i.e., the convolutional neural network (CNN) model, and output the probability score of the contig of the spliced ​​sequence being a potential virus sequence;

[0065] (3) Mark the contigs with a probability score higher than a set threshold (e.g., 0.9) as candidate unknown pathogens.

[0066] Furthermore, to identify the abundance trends of pathogens in fever samples at different time points, improve the expressive power of downstream influenza trend modeling, and achieve early detection of potential epidemics, this embodiment also includes a dynamic monitoring and early warning module. Through this module, pathogens detected in samples are aggregated across time points to construct an abundance time series; a Transformer model is used for simultaneous prediction of multiple pathogens; and a ChangePoint Detection model is introduced to automatically identify trend abrupt changes and outbreak signals.

[0067] A standardized three-dimensional tensor representation structure is introduced into the dynamic monitoring and early warning module to perform dynamic trend modeling and early warning analysis on the abundance time series of each pathogen.

[0068] The three-dimensional tensor has the following dimensions: a sample dimension, representing the number of samples (N); a pathogen dimension, representing the number of pathogen species (M); and a time dimension, representing the duration of time (T). The original tensor can then be represented as:

[0069] Tensor normalization: To avoid the impact of differences in the abundance of different pathogens on training, tensor normalization is required, specifically by z-score standardization of the original CPM values.

[0070]

[0071] Batch partitioning: Divide the dataset into training, validation, and test sets in a 7:1:2 ratio, and further partition the input data on each dataset with a step size of 1 and a window size of 52.

[0072]

[0073] Dimensional transformation: To match the dimensions expected by the Transformer interface, the tensor undergoes dimensional transformation:

[0074]

[0075] Model Adjustment: Since the Transformer itself supports multi-dimensional feature input, this embodiment only needs to adapt the linear layer of the output layer to achieve simultaneous prediction of multiple pathogens. The Transformer output remains (52, N, M), and the last time point (N, M) is passed into the linear layer, followed by another dimensionality transformation.

[0076]

[0077] For the predicted abundance data of each pathogen over the next 13 weeks, change point monitoring will be used for early warning:

[0078] (1) For each pathogen, extract its CPM sequence from multiple samples;

[0079] (2) The Pelt model in the ruptures package is used as a one-dimensional time series input;

[0080] (3) Set the detection model to "rbf" (radial basis kernel function model), which is suitable for capturing changes in both mean and variance simultaneously;

[0081] (4) Set the penalty parameter pen=10 as a constraint on the number of variable points to control the balance between detection sensitivity and overfitting risk;

[0082] (5) Execute the .predict(pen=10) method to output a list of variable point positions;

[0083] (6) If the change point detection results show that the abundance of a certain pathogen increases significantly after a certain time point and remains at a level higher than the background 3σ (3 times the standard deviation), the system will trigger an early warning.

[0084] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A smart identification and real-time early warning system for unknown pathogens in respiratory samples from fever clinics, characterized in that, Includes the following modules: The data quality control and processing module is used to perform quality control and standardized preprocessing on raw metagenomic sequencing data to obtain pure microbial reads. The known pathogen identification and annotation module is used to identify known pathogens, classify non-host reads, construct a classification database based on the classification results, and annotate reads with low classification confidence or unclassified reads in the classification results to obtain a standard annotation table. The identification model training module is used to construct a positive and negative sample dataset and train the identification model for identifying unknown pathogens. The unknown pathogen identification module uses a trained identification model to identify potential unknown pathogens in the metagenomics of pure microbial reads, based on the classification database and standard annotation table constructed by the known pathogen identification annotation module.

2. The intelligent identification and real-time early warning system according to claim 1, characterized in that, In the data quality control module, the FastQC tool is used to assess the quality of the raw sequencing data for each sample. The assessment includes base quality distribution, GC content, adapter contamination, sequence repetition rate, and N base content. The Trimmomatic tool is used to trim the raw sequencing data, removing adapter sequences and low-quality base fragments. The Bowtie2 tool is used to align the trimmed paired reads with the human reference genome GRCh38. Using the --un-conc-gz parameter, paired reads that do not align with the host genome are output as non-host reads and retained to obtain pure microbial reads.

3. The intelligent identification and real-time early warning system according to claim 1, characterized in that, The known pathogen identification and annotation module uses the Kraken2 tool to quickly classify non-host reads; the Kraken2 tool uses the k-mer exact matching strategy to align sequencing sequences, obtain preliminary species attribution and LCA depth, and construct a classification database.

4. The intelligent identification and real-time early warning system according to claim 3, characterized in that, The BLASTn nucleic acid sequence alignment tool and the DIAMOND protein sequence alignment tool were used to annotate reads with low classification confidence or those that were not classified in the Kraken2 tool classification results. The annotation results were standardized by a unified format and then integrated to output a standard annotation table with unique attribution.

5. The intelligent identification and real-time early warning system according to claim 4, characterized in that, The integration method is as follows: priority is given to retaining annotations with lower LCA levels; results with high consistency are assigned higher scores; and only the highest-scoring item is retained for multiple matches of the same read length.

6. The intelligent identification and real-time early warning system according to claim 1, characterized in that, When the recognition model training module constructs the positive and negative sample datasets, the positive samples are contigs with clear viral labels selected from known viral sequences from public databases; the negative samples are non-viral sequences from humans or common microorganisms in the environment, as well as false positive fragments generated by background splicing.

7. The intelligent identification and real-time early warning system according to claim 6, characterized in that, The recognition model is built using a convolutional neural network. When training the recognition model, the loss function is binary classification cross-entropy, and the learning rate of the optimizer is initially set to 1e-4. Five-fold cross-validation is used to evaluate the model's generalization ability.

8. The intelligent identification and real-time early warning system according to claim 1, characterized in that, In the unknown pathogen identification module, the process for identifying unknown pathogens is as follows: for unaligned reads, the MEGAHIT tool is used to perform de novo splicing and assembly to obtain spliced ​​sequence contigs, which are then used for species classification and functional annotation; the trained identification model is then used to predict the spliced ​​sequences.

9. The intelligent identification and real-time early warning system according to claim 8, characterized in that, The unknown pathogen identification module employs a deep learning-based unknown pathogen potential scoring model to screen non-host contigs obtained using the MEGAHIT tool. The screening process includes: Structural and statistical features are extracted from each contig of spliced ​​sequences; The features are input into the trained recognition model, and the output is a probability score that the contiguous sequence is a potential viral sequence. Contigs with a probability score higher than a set threshold are marked as candidate unknown pathogens.

10. The intelligent identification and real-time early warning system according to claim 1, characterized in that, The early warning system also includes: The dynamic monitoring and early warning module is used to summarize pathogens detected in samples across time points and construct an abundance time series; it uses the Transformer model to perform simultaneous prediction of multiple pathogens; and it introduces a change point detection model to automatically identify trend changes and outbreak signals.

Images