A pathogenic microorganism detection method and system based on low-coverage genome sequencing and multi-model fusion and application thereof

Abstract

The application belongs to the field of bioinformatics, and particularly relates to a pathogenic microorganism detection method and system based on low-coverage genome sequencing and multi-model fusion and application thereof. In view of the problems of low detection rate of low-abundance pathogens, high false negative and false positive rates and lack of group early warning capability in the traditional process, the process is preprocessed (quality filtering, adapter trimming and host removal), and then reference database alignment and CNN model are used in parallel to identify microorganisms. The CNN uses low-quality sequences to improve the detection rate. The random forest model (read length, GC content and the like) is used for accurate classification. Through z1 / z2 two-dimensional normalization combined with GMM modeling, pathogenic bacteria and background bacteria are distinguished, and group abnormalities are captured. The drug resistance information is also mined by integrating assembly technology. The process realizes end-to-end automation, improves the diagnosis accuracy, supports large-scale screening and hospital infection early warning, and is suitable for clinical diagnosis and environmental microbiology research.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to a method, system, and application for detecting pathogenic microorganisms based on low-coverage genome sequencing and multi-model fusion. Background Technology

[0002] Metagenomic next-generation sequencing (mNGS) technology, with its advantages of not requiring pre-defined pathogens and being able to simultaneously detect multiple microorganisms, has become a core tool for pathogen detection in urine samples and for microbiome research, playing a crucial role in clinical urinary tract infection (UTI) diagnosis, hospital infection control, and environmental microbiology research. Currently, traditional mNGS bioinformatics workflows primarily rely on sequence alignment technologies (such as BLAST and Bowtie2) combined with reference databases like RefSeq to identify microorganisms, while using simple statistical methods like Z-scores for data normalization to determine the presence of pathogens in the sample.

[0003] However, in the context of low-coverage sequencing of urine samples, traditional procedures face multiple technical bottlenecks: On the one hand, low-coverage data itself suffers from problems such as high noise, short effective sequences, and a low proportion of reads from low-abundance pathogens, which makes identification methods that rely on sequence alignment prone to missing low-abundance pathogens (resulting in false negatives), and at the same time, it is difficult to distinguish the sequence differences between background bacteria, colonizing bacteria, and pathogens (resulting in false positives); on the other hand, traditional normalization methods such as Z-score can only perform simple standardization on microbial abundance data in a single dimension, and cannot capture the complex distribution patterns of microorganisms within the sample (at the individual level) and between samples (at the population level), further reducing the accuracy of pathogen identification.

[0004] In recent years, machine learning (such as random forests) and deep learning (such as convolutional neural networks CNNs) have shown significant potential in the field of bioinformatics, demonstrating superior performance compared to traditional methods in tasks such as sequence feature extraction and complex data classification. However, current technologies have not yet integrated these advanced computational models into the entire workflow of mNGS analysis of low-coverage urine samples. Most approaches only attempt to apply machine learning methods in a single stage (such as microbial identification), lacking an end-to-end integrated design from raw data preprocessing to pathogen identification and result output, thus failing to fully leverage the advantages of multi-model collaboration.

[0005] Therefore, developing a dedicated bioinformatics workflow for low-coverage mNGS data from urine samples, by integrating deep learning, machine learning, and statistical modeling methods, can overcome the limitations of traditional workflows in detecting low-abundance pathogens, resolving interference, and capturing complex distribution patterns. This is of great significance for improving the accuracy of clinical pathogen diagnosis, meeting the needs of large-scale screening, and realizing early warning of hospital infections. It can also provide technical reference for the analysis of low-coverage sequencing data in the field of environmental microbiology. Summary of the Invention

[0006] To address the aforementioned shortcomings, this invention provides a bioinformatics workflow for microbiome analysis of low-coverage metagenomic next-generation sequencing (mNGS) data from urine samples. This workflow integrates a deep learning CNN model (to improve the detection rate of low-abundance pathogens), a machine learning random forest (to achieve accurate taxonomic classification), and a Gaussian mixture model (GMM) (combining z1 / z2 dual-dimensional normalization indices to distinguish pathogens from background bacteria and capture population anomalies). It also integrates drug resistance information mining capabilities, effectively solving the problems of high false negative and false positive rates, low automation, and lack of population early warning capabilities in traditional workflows, thus adapting to the needs of clinical diagnosis and large-scale screening.

[0007] The technical solution of this invention is as follows:

[0008] On the one hand, this invention provides a method for detecting pathogenic microorganisms based on low-coverage genome sequencing and multi-model fusion, comprising the following steps:

[0009] (1) Data preprocessing: The raw data of low-coverage metagenomic high-throughput sequencing were subjected to quality control, adapter trimming and host sequence removal to obtain non-human sequences;

[0010] (2) Microbial identification: Non-human sequences are input into the reference database alignment system and convolutional neural network model in parallel to complete the preliminary identification and obtain the identification results;

[0011] (3) Taxonomic classification: The identification results were classified at the family, genus, and species levels using a random forest model;

[0012] (4) Pathogen identification: Calculate the z1 / z2 two-dimensional normalized value, and analyze it by Gaussian mixture model. The pathogenic microorganism is identified based on the posterior probability. z1 is the value of each microorganism read segment count in a single sample after logarithmic transformation and Z-score standardization, and z2 is the Z-score standardization result of the z1 value of the microorganism in all samples.

[0013] (5) Results output: Integrate pathogen identification results and drug resistance information to output a test report.

[0014] Specifically, the data preprocessing in step (1) uses the FastP software to perform quality control and connector trimming. The parameters include enabling automatic connector detection and removal, sliding window size of 4, average quality threshold of 20, low complexity threshold of 30%, and retaining reads with a length ≥ 50bp. The Bowtie2 software is used in high sensitivity mode to remove human host sequences.

[0015] Specifically, the input of the convolutional neural network model in step (2) is a k-mer frequency matrix with k=6, which is converted into a numerical matrix by one-hot encoding. The probability distribution of microbial species is output through ReLU activation function, max pooling and Softmax classifier, and trained using Adam optimizer and cross-entropy loss function.

[0016] Specifically, the classification features of the random forest model in step (3) include read length, alignment score, GC content, unique read ratio and k-mer frequency features. Based on Gini impurity split nodes, the majority voting results and probability estimates are output.

[0017] Specifically, the calculation of z1 in step (4) satisfies: Where count is the microbial read count, μ and σ are the mean and standard deviation of all microbial read counts in the sample; z2 is calculated according to: , where μ and σ are the mean and standard deviation of the z1 value of the microorganism in all samples.

[0018] Specifically, the probability density function of the Gaussian mixture model in step (4) is: ,in For mixed weights, and The parameter is Gaussian; a posterior probability < 0.05 indicates a positive result for pathogenic microorganisms.

[0019] Specifically, the reference database in step (2) is the NCBI RefSeq database, and the convolutional neural network model is pre-trained using simulated data and the NCBI database. The simulated data covers common pathogens and human genome sequences.

[0020] Specifically, the raw data for low-coverage metagenomic high-throughput sequencing in step (1) comes from urine samples, blood samples, sputum samples or environmental samples.

[0021] In another aspect, the present invention provides a system for detecting pathogenic microorganisms based on low-coverage genome sequencing and multi-model fusion, including a preprocessing module, a microorganism identification module, a taxonomic classification module, a pathogen determination module, and a result output module, for performing the aforementioned method.

[0022] Specifically, it also includes a drug resistance information mining unit and a cluster infection early warning unit. The early warning unit captures abnormal distribution patterns of microorganisms through a Gaussian mixture model and triggers a cluster infection early warning.

[0023] In another aspect, the present invention provides applications of the aforementioned methods or systems, selected from the following:

[0024] (1) Support for in vitro detection of low-abundance pathogenic microorganisms in human samples;

[0025] (2) Detection of homologous strains in hospital environmental samples and early warning of infection outbreaks;

[0026] (3) Analysis of the environmental microbial community structure from low-coverage sequencing data;

[0027] (4) Pathogenic microorganism tracing and screening of environmental samples in public health surveillance;

[0028] (5) Qualitative analysis of low abundance strains in laboratory microbial sample banks.

[0029] The beneficial effects of this invention are as follows:

[0030] (1) High sensitivity and accuracy: By introducing a deep learning CNN model, the low-quality or marginal sequences discarded by traditional comparison are effectively utilized, which improves the detection rate of low-abundance pathogens and reduces false negatives.

[0031] (2) Strong anti-interference and discrimination ability: The Gaussian mixture model (GMM) is used to model the multi-dimensional distribution of the normalized z1 / z2 values, which can more reliably distinguish pathogens from background bacteria and colonizing bacteria, and significantly reduce false positives.

[0032] (3) Intelligent and automated: The entire process integrates machine learning and deep learning models, realizing end-to-end automated analysis from raw data to pathogen identification, reducing the dependence on operator experience, and is particularly suitable for large-scale screening.

[0033] (4) High clinical guidance value: It can not only answer "whether there is a pathogen", but also answer "who is the main pathogen" through quantitative indicators (z1, z2, posterior probability) and combine assembly technology to mine drug resistance information, providing direct and powerful evidence for precision diagnosis and treatment.

[0034] (5) Possesses population early warning function: By establishing a population baseline (z2) and anomaly detection model (GMM), the present invention has the ability to discover abnormal patterns of microbial distribution, which can be used for hospital infection control and epidemiological monitoring, which is a blind spot of traditional methods. Attached Figure Description

[0035] Figure 1 This is a flowchart of the detection method of the present invention.

[0036] Figure 2 This is a graph showing the diagnostic efficacy results. Detailed Implementation

[0037] The present invention will be further clearly and completely illustrated below through embodiments. These embodiments are only some examples of the present invention and are not intended to limit the present invention, but are only for illustrating the present invention. Unless otherwise specified, the experimental methods used in the following embodiments are all conventional experiments, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0038] Example 1

[0039] 1.1 Standard Preprocessing Steps

[0040] A: Quality Control & Trimming: The raw data 1.fastq.gz and 2_fastq.gz (paired-end sequencing data) were processed using fastp software (version 0.23.0 or above).

[0041] Specific parameters:

[0042] ① Enable automatic connector detection and removal function (--detect_adapter_for_pe);

[0043] ② Quality filtering: Set the sliding window size to 4 (--cut_window_size 4), and the average quality threshold within the window to 20 (--cut_mean_quality 20), to remove low-quality bases at the 3' and 5' ends;

[0044] ③ Low complexity filtering: Enable low complexity sequence filtering and set the complexity threshold to 30% (--complexity_threshold 30).

[0045] ④ Length filtering: Remove short reads with a length less than 50bp after trimming (--length_required 50);

[0046] ⑤ Output clean data after processing: -o -O (paired-end sequencing data).

[0047] B: Removal of human host sequence:

[0048] Align the clean data above to the human reference genome (e.g., GRCh37 / hg19 version). Use Bowtie2 software for alignment with the following parameters:

[0049] ① Use a high-sensitivity mode (--very-sensitive-local) to ensure the best possible match with the host sequence;

[0050] ② Extract unmapped reads and generate non-human sequence files, which will serve as input data for subsequent microbial identification.

[0051] 1.2 Deep Learning Convolutional Neural Networks (CNNs) for Microbial Identification

[0052] In the microbial identification process, in addition to traditional comparison, a CNN model is introduced to process the read sequences. The input data is the k-mer frequency matrix (k=6) of the read sequences, which is converted into a numerical matrix through one-hot encoding. The CNN architecture includes:

[0053] Input layer: A matrix of size L×4 (L is the read length, and 4 represents nucleotides A, C, G, and T).

[0054] Convolutional layer: Uses multiple filters to extract local features. The formula for convolution operation is:

[0055]

[0056] Where W is the filter weight, b is the bias, and σ is the ReLU activation function.

[0057] Pooling layer: Max pooling reduces dimensionality.

[0058] Fully connected layer: Features are flattened and then input into the fully connected network.

[0059] Output layer: The Softmax function outputs the probability distribution of microbial species, and the loss function uses cross-entropy.

[0060]

[0061] in, For real labels, To predict probabilities.

[0062] CNN models are trained using backpropagation and the Adam optimizer, and are pre-trained using simulated data and public databases such as NCBI.

[0063] 1.3 Machine Learning: Random Forest for Taxonomy

[0064] In the taxonomic classification stage, a random forest model was used to classify reads and contiguous groups at the family, genus, and species levels. Features included:

[0065] Segment length;

[0066] Alignment score;

[0067] GC content;

[0068] Unique ratio;

[0069] k-mer frequency characteristics;

[0070] Random forests consist of multiple decision trees, each trained using bootstrap sampling. Splitting nodes utilize Gini impurity.

[0071]

[0072] in, p k The proportion of category k. The model output is the majority vote result, along with a probability estimate.

[0073] 1.4 Gaussian Mixture Model (GMM) for Pathogenic Microorganism Identification

[0074] The distribution of the normalized z1 and z2 values ​​was modeled using a GMM. z1 was calculated as the logarithmic transformation and Z-score normalization of the read count for each microorganism in a single sample.

[0075]

[0076] Where μ and σ are the mean and standard deviation of the counts of all microbial reads in the sample. z2 is calculated as the Z-score normalized to the z1 value of a particular microorganism across all samples.

[0077] Calculation of z2: For a specific microorganism (e.g., microorganism A), collect its z1 values ​​from all samples to form a distribution. Then, standardize this distribution again using the Z-score to obtain the z2 value. The formula is:

[0078]

[0079] in, and Z1 represents the mean and standard deviation of the z1 values ​​for this microorganism across all samples. Z2 also reflects the relative anomalousness of this microorganism within the sample population, helping to identify microorganisms that are anomalously enriched in a specific sample.

[0080] GMM assumes that the z1 and z2 values ​​come from multiple Gaussian distributions, with the probability density function being:

[0081]

[0082] in, For mixed weights, and The parameters are Gaussian. Pathogenic microorganisms are identified as outliers with low posterior probability, with a threshold set such that the posterior probability is <0.05.

[0083] Example 1

[0084] To monitor prevalent strains of nosocomial urinary tract infections (UTIs) and achieve early warning, a large general hospital collected routine urine samples from 200 hospitalized patients for one week. The recruitment criteria were:

[0085] (1) Age 18 years or older;

[0086] (2) Patients who are clinically diagnosed with urinary tract infection (with symptoms such as urinary frequency, urgency, dysuria or fever) or who are scheduled to undergo urinary system surgery and need to undergo preoperative infection screening;

[0087] (3) Agree to provide a midstream urine sample and sign the informed consent form.

[0088] Exclusion criteria:

[0089] (1) The patient had received standard treatment with broad-spectrum antibiotics within 2 weeks prior to sampling (which may lead to false negatives);

[0090] (2) The sample volume is less than 2 mL or the sample is obviously contaminated with feces or menstrual blood;

[0091] (3) Patients whose clinical data are missing and cannot be followed up.

[0092] Sample collection and processing: Collect clean-catch midstream urine from patients in the morning, strictly following aseptic techniques. Divide the sample in two: one copy is immediately sent to the clinical microbiology laboratory for standard culture (as a gold standard control); the other copy is subjected to low-coverage mNGS sequencing.

[0093] Clinical follow-up: All patients were followed up clinically for 7 days. Based on culture results, blood routine indicators (WBC, CRP) and treatment outcomes, two senior clinicians jointly diagnosed the "final infection status" to calculate sensitivity and specificity.

[0094] The goal of this study is to rapidly and accurately screen positive samples for pathogenic microorganisms from a large number of samples and identify potential outbreak trends. The process is as follows: Figure 1 As shown.

[0095] 1. Data and Methods

[0096] Preprocessing and Identification: Strict quality and host sequence filtering were performed, referring to Example 1.1. Subsequently, microbial identification was performed in parallel using conventional alignment (Bowtie2+RefSeq) and the deep learning CNN model of this invention. The CNN model was pre-trained using 100,000 simulated sequences containing 500 common urinary tract pathogens. Details and specific methods of CNN model pre-training are as follows:

[0097] (1) Data generation: Using the InSilicoSeq simulator and based on the NCBI RefSeq database, 500 common urinary tract pathogens (covering Enterobacteriaceae, Staphylococcus, Enterococcus, etc.) were selected. According to the Illumina HiSeq sequencing error model, simulated reads of 50-150 bp in length were generated, totaling 100,000 positive samples; at the same time, 100,000 negative samples were generated from the human genome.

[0098] (2) Input features: The DNA sequence is converted into a k-mer frequency matrix (k=6) and one-hot encoded. The input dimension is...

[0099]

[0100] (3) Model architecture:

[0101] ① Convolutional layer (Conv1D): 64 convolutional kernels, kernel size = 5, stride = 1, activation function ReLU;

[0102] ② Pooling layer (MaxPooling): Window size = 2;

[0103] ③ Fully connected layer (Dense): 128 neurons, Dropout rate 0.5 (to prevent overfitting);

[0104] ④ Output layer: The Softmax classifier outputs the probability of belonging to each pathogen.

[0105] (4) Training parameters:

[0106] The Adam optimizer was used (learning rate 0.001), with cross-entropy loss function, batch size set to 64, and training was performed for 50 epochs.

[0107] (5) Validation results: The model achieved a classification accuracy of 96.2% on the independent test set, ensuring the ability to identify low-quality or short sequences.

[0108] (b) Taxonomic classification: For the identified microorganisms, a random forest model is used to perform fine classification at the "genus" level, with features including read length, GC content, and proportion of unique reads.

[0109] (c) Normalization and Modeling: Calculate the z1 value for each microorganism in each sample. Calculate the z2 value for Escherichia coli detected in all samples. Fit the z1 and z2 distributions of all samples using a Gaussian mixture model (GMM), and set a posterior probability threshold of P < 0.05 as pathogen positivity.

[0110] Traditional metagenomics alignment process (control group)

[0111] To compare the effectiveness of this invention, a traditional mNGS analysis procedure was also run. The specific experimental steps and detailed parameters are as follows:

[0112] Step A: FastP processes the raw data: parameters are set to automatically remove connectors, homogenize the 4bp sliding window to remove less than 20bp, and retain length >50bp;

[0113] Step B: Remove host sequences (Bowtie2) Align the clean data to the human reference genome (GRCh37) with high-sensitivity local alignment mode, retaining unmapped reads for subsequent analysis;

[0114] Step C: Microbial alignment (Bowtie2 + RefSeq) aligns non-human sequences to the NCBI RefSeq microbial database;

[0115] Step D: Result Judgment (Traditional Threshold Method) Count the number of reads aligned to each species. The traditional criteria for a positive read are set as follows: Reads Count > 50; Coverage > 1%; The relative abundance (RPM) of the top-ranked species is significantly higher than that of other species.

[0116] 2. Results

[0117] Efficiency and sensitivity: The CNN model improved the utilization rate of non-human sequences with low quality and short length by 15%, and additionally detected 3 rare pathogens (such as hemolytic Staphylococcus) that were missed by traditional comparison.

[0118] Precise screening: The GMM model successfully divided 200 samples into three clusters:

[0119] Cluster 1 (low risk, 165 samples): Both z1 and z2 values ​​are low, with a high posterior probability (P>0.9), and are identified as colonizing bacteria or background noise.

[0120] Cluster 2 (medium risk, 28 samples): z1 value is moderate, z2 value is dispersed, suggesting possible infection or carrier status, which needs to be reviewed in conjunction with clinical indicators.

[0121] Cluster 3 (high risk, 7 samples): z1 and z2 values ​​were significantly high, with extremely low posterior probability (P<0.01). These 7 samples were definitively identified as positive for pathogenic microorganisms.

[0122] Outbreak Warning: Of the 7 positive samples, 5 were identified as Escherichia coli ST131 (a highly pathogenic, multidrug-resistant clone), and their z1 and z2 values ​​formed a tight cluster of anomalous points in the GMM distribution. The system automatically issued a warning of "suspected outbreak of homologous strains." Subsequent epidemiological investigation by the hospital's infectious disease department confirmed that these 5 patients were all located in the same ward, indicating cross-infection.

[0123] Diagnostic efficacy table 1 and Figure 2 :

[0124] Table 1

[0125]

[0126] In summary, the process of this invention can not only achieve high-throughput and automated pathogen identification in large-scale screening scenarios in hospitals, but its core GMM modeling method can also keenly capture abnormal microbial distribution patterns at the population level, thereby achieving early outbreak warning, which is something that traditional single threshold methods (such as fixing z1>2) cannot achieve.

[0127] Example 2

[0128] A large tertiary hospital is conducting research on pathogen identification and viral-bacterial co-infection mechanisms in patients with culture-negative sterile pyuria. Pyuria is usually caused by bacteria and typically requires antibiotic prescriptions. However, traditional urine cultures are time-consuming and have a high risk of false negatives. Furthermore, the role of urinary viruses in pyuria remains unclear. Metagenomics, by directly sequencing the urinary microbiota, can improve the accuracy and efficiency of diagnosis. This embodiment aims to retrospectively analyze the urinary microbiota characteristics of such patients using the multi-model fusion system of this invention, revealing potential bacterial and viral pathogens.

[0129] 1. Data and Methods

[0130] Inclusion criteria:

[0131] (1) Clinical diagnosis of pyuria (white blood cells >5 / HPF in urine sediment microscopy);

[0132] (2) The standard midstream urine bacterial culture results during the same period were negative (no bacterial growth or colony count <10).3 (CFU / mL)

[0133] (3) Whole genome sequencing (WGS / mNGS) was performed.

[0134] Grouping:

[0135] Case group (pyuria group): 120 patients who met the above criteria and were diagnosed with urinary tract infection symptoms.

[0136] Control group (non-puuria group): individuals undergoing physical examinations without pyuria or symptoms, or patients with non-infectious urinary tract diseases (120 cases).

[0137] Process reference:

[0138] Step A: Deep learning-enhanced microbial identification (CNN model)

[0139] To address the potential presence of low-abundance fastidious bacteria (such as Atopobium) and non-bacterial pathogens (such as Polyomavirus) in cultured negative samples, this invention utilizes a specifically trained CNN model. Training details are provided in Example 1. The model is configured with a k-mer frequency matrix of Reads as input. The training set is further enhanced with simulated sequences of difficult-to-culture microorganisms such as Mycoplasma genitalium, Ureaplasma urealyticum, Polyomaviridae (BK / JC viruses), and Gardnerella vaginalis. The CNN successfully identifies viruses and short bacterial fragments that are easily missed by traditional alignments from a large amount of host background noise.

[0140] Step B: Random Forest Classification and Feature Selection

[0141] The detected microbial genera were classified using a random forest model, and the top 10 dominant genera were extracted as key features. The model automatically screened out the microbial biomarkers that showed the most significant differences between the pyuria and non-pyuriria groups by calculating the Mean Decrease Gini.

[0142] Step C: GMM Normalization Modeling and Risk Assessment

[0143] Calculate z1 and z2 for key microorganisms. Analyze the distribution patterns of these microorganisms in the two population groups using a GMM model:

[0144] Judgment logic: If a microorganism is significantly enriched in the high-risk cluster of GMM (Cluster 3, z1, z2 both high) in the pyuria group, but in the low-risk cluster in the control group, it is defined as "an independent risk microorganism associated with pyuria".

[0145] 2. Results Analysis

[0146] (1) Discovery of the microbial community lineage

[0147] This system identified 29 genera of microorganisms in all samples, including 27 bacterial genera and 2 viral genera. Compared with the traditional process, the present invention improves the detection rate of polyomavirus by 40% and the detection rate of atopobium by 35%, thanks to the ability of CNN to capture atypical sequences.

[0148] (2) Association between key pathogens and risk (multivariate analysis)

[0149] Based on the quantitative indicators (z1, z2) output by this invention, multivariate logistic regression analysis was performed using clinical data, revealing the following significant associations:

[0150] Independent risk factors (pathogens):

[0151] Atopobium: OR = 11.688 [95% CI: 2.190-62.362], p = 0.004. This suggests that this bacterium is one of the important causes of culture-negative pyuria.

[0152] Gardnerella: OR = 9.904 [95% CI: 2.180-45.005], p=0.003.

[0153] Polyomavirus: OR = 5.205 [95% CI: 1.295-20.919], p=0.020. This is a pathogen that is almost undetectable by traditional culture methods.

[0154] Protective factors (probiotics):

[0155] Lactobacillus: Significantly associated with a reduced risk of pyuria (p=0.031). In the GMM model, Lactobacillus was mainly distributed in the "high abundance-steady-state cluster" in the healthy control group, while its abundance was significantly reduced in the pyuria group.

[0156] In summary, the CNN model of this invention effectively solves the problem of difficult sequence identification of viruses (polyomaviruses) and mycobacteria, filling the blind spots of traditional detection. Through GMM modeling and statistical analysis, mycobacteria, Gardnerella vaginalis, and polyomaviruses have been successfully identified as important risk factors independent of traditional urinary tract pathogens (such as Escherichia coli). This process can not only answer the question "Is there a pathogen?", but also provide more precise basis for antibiotic or antiviral treatment in clinical practice by revealing the imbalance of the microecology (reduction of lactobacilli vs. increase of pathogens).

[0157] The above detailed description is a specific illustration of one feasible embodiment of the present invention, and this embodiment is not intended to limit the patent scope of the present invention. It should be noted that all equivalent implementations or modifications that do not depart from the present invention should be included within the scope of the technical solution of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A method for detecting pathogenic microorganisms based on low-coverage genome sequencing and multi-model fusion, characterized in that, Includes the following steps: (1) Data preprocessing: The raw data of low-coverage metagenomic high-throughput sequencing were subjected to quality control, adapter trimming and host sequence removal to obtain non-human sequences; (2) Microbial identification: Non-human sequences are input into the reference database alignment system and convolutional neural network model in parallel to complete the preliminary identification and obtain the identification results; (3) Taxonomic classification: The identification results were classified at the family, genus, and species levels using a random forest model; (4) Pathogen identification: Calculate the z1 / z2 two-dimensional normalized value, and analyze it by Gaussian mixture model. The pathogenic microorganism is identified based on the posterior probability. z1 is the value of each microorganism read segment count in a single sample after logarithmic transformation and Z-score standardization, and z2 is the Z-score standardization result of the z1 value of the microorganism in all samples. (5) Results output: Integrate pathogen identification results and drug resistance information to output a test report; The calculation of z1 in step (4) satisfies: Where count is the microbial read count, μ and σ are the mean and standard deviation of all microbial read counts in the sample; z2 is calculated according to: , where μ and σ are the mean and standard deviation of the z1 value of the microorganism in all samples; The probability density function of the Gaussian mixture model in step (4) is: ,in For mixed weights, and For Gaussian parameters; A posterior probability < 0.05 indicates a positive result for pathogenic microorganisms.

2. The method according to claim 1, characterized in that, The data preprocessing in step (1) uses the fastp software to perform quality control and connector trimming. The parameters include enabling automatic connector detection and removal, sliding window size of 4, average quality threshold of 20, low complexity threshold of 30%, and retaining reads with a length ≥50bp. The Bowtie2 software is used in high sensitivity mode to remove human host sequences.

3. The method according to claim 1, characterized in that, The input to the convolutional neural network model in step (2) is a k-mer frequency matrix with k=6, which is converted into a numerical matrix by one-hot encoding. The model outputs the probability distribution of microbial species through ReLU activation function, max pooling and Softmax classifier, and is trained using Adam optimizer and cross-entropy loss function.

4. The method according to claim 1, characterized in that, The classification features of the random forest model in step (3) include read length, alignment score, GC content, proportion of non-redundant effective alignment reads and k-mer frequency features of k=6. Based on the Gini impurity split nodes, the majority voting results and probability estimates are output.

5. The method according to claim 1, characterized in that, The reference database in step (2) is the NCBI RefSeq database. The convolutional neural network model is pre-trained using simulated data and the NCBI database. The simulated data covers common pathogens and human genome sequences.

6. The method according to claim 1, characterized in that, The raw data for the low-coverage metagenomic high-throughput sequencing in step (1) comes from urine samples, blood samples, sputum samples or environmental samples.

7. A system for detecting pathogenic microorganisms based on low-coverage genome sequencing and multi-model fusion, characterized in that, It includes a preprocessing module, a microbial identification module, a taxonomic classification module, a pathogen determination module, and a result output module, used to execute the method described in any one of claims 1-6.

8. The system according to claim 7, characterized in that, It also includes a drug resistance information mining unit and a cluster infection early warning unit. The early warning unit uses a Gaussian mixture model to capture abnormal distribution patterns of microorganisms and triggers a cluster infection early warning.

9. The application of the method according to any one of claims 1-6 or the system according to any one of claims 7-8, characterized in that, Selected from the following aspects: (1) Support for in vitro detection of low-abundance pathogenic microorganisms in human samples; (2) Detection of homologous strains in hospital environmental samples and early warning of infection outbreaks; (3) Analysis of the environmental microbial community structure from low-coverage sequencing data; (4) Pathogenic microorganism tracing and screening of environmental samples in public health surveillance; (5) Qualitative analysis of low abundance strains in laboratory microbial sample banks.

Images