A method for binning metagenomic contigs using a reference database

Abstract

The invention discloses a method for binning a metagenomic contig using a reference database, comprising the following steps: S1. obtaining reads fragments from a plurality of metagenomic samples, and collecting the obtained reads fragments into a gene library; S2. assembling the reads fragments collected into the gene library to obtain contig sequence fragments; S3. performing an unsupervised clustering process on the obtained contig sequence fragments to obtain binning clustering results containing contig sequences of different sizes; S4. performing quality detection on the binning clustering results containing contig sequences of different sizes, and judging whether the binning meets the quality threshold, if not, storing the binning in a set S; if so, placing the binning in a final result Results set; S5. performing marker gene analysis on the contig sequences in the set S in combination with the reference database and constructing an integrated SVM classification model to complete the binning operation. The invention can obtain a higher accuracy rate and improve the binning effect.

Description

Technical Field

[0001] The present invention relates to the field of metagenome technology, and in particular to a metagenome contig binning method using a reference database. Background Art

[0002] Metagenomics can directly extract all the genetic material of microorganisms from the environment and perform functional and sequence analysis. Compared with traditional methods, its advantage is that it does not need to provide an environment for the survival of microorganisms for separate cultivation, but directly extracts their genetic material from the environment for analysis. Since most microorganisms cannot be cultured in the laboratory, the emergence of this method provides scientists with an important method to understand the structure and function of microbial communities. It also has great significance for the diagnosis and treatment of diseases, environmental governance, and the understanding of life. Through the second-generation sequencing technology, we can obtain a large number of metagenomic samples from the environment. These samples contain DNA genetic sequences of different lengths from different strains. How to bin the DNA sequences in metagenomic samples at a lower classification level is a hot spot and difficulty in metagenomic research.

[0003] The current metagenomic data classification methods are divided into two categories based on whether a reference database is used in the classification process: 1. Binning with a reference database; 2. Binning without a reference database.

[0004] Classification with a reference database can be divided into classification methods based on sequence homology and sequence composition according to the binning characteristics, and can be divided into sequence alignment and classification model construction according to different binning methods. The method based on sequence homology requires the complete genes of related organisms for alignment, which requires a complete and comprehensive reference database. Due to the limited reference database, this method is very limited. At the same time, the current sequencing technology can produce GB-sized metagenomic samples. If the sequence homology-based alignment is used, the computational cost is very high. On the other hand, the method based on sequence composition uses the K-mers feature of the sequence for sequence alignment or modeling. The running speed of this method will be faster than the method based on sequence homology. Among them, the method based on k-mers alignment requires the establishment of a specific reference database, and the modeling requires the knowledge of training data and classification categories in advance. At the same time, in the face of complex metagenomic samples, not only is the modeling speed slow, but it is also difficult to classify more accurately.

[0005] Another classification without a reference database (i.e., a metagenomic clustering method) does not rely on a constructed reference database. They cluster based on the composition characteristics or coverage characteristics of the sequence, or a combination of the above two characteristics. The clustering results are often accurate to the species level or even the strain level. Clustering methods rarely use reference databases and are limited by the characteristics of the sequences used and the clustering algorithms. Therefore, the present invention provides a metagenomic contig binning method using a reference database. Summary of the invention

[0006] The purpose of the present invention is to provide a method for binning metagenomic contigs using a reference database in order to address the deficiencies of the prior art.

[0007] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0008] A method for binning metagenome contigs using a reference database comprises the following steps:

[0009] S1. Obtain read fragments from multiple metagenomic samples and pool the obtained read fragments into a gene library;

[0010] S2. Assemble the reads fragments collected in the gene library to obtain contiguous sequence fragments;

[0011] S3. performing an unsupervised clustering process on the obtained contig sequence fragments to obtain binned clustering results containing contig sequences of different sizes;

[0012] S4. Perform quality detection on the bin clustering results containing contig sequences of different sizes, and determine whether the bins meet the quality threshold. If not, store the bins in the set S; if yes, put the bins into the final result Results set;

[0013] S5. Perform marker gene analysis on the overlapping group sequences in set S in combination with the reference database and construct an integrated SVM classification model to complete the binning operation.

[0014] Furthermore, the unsupervised clustering process in step S3 adopts the MetaBAT method for unsupervised clustering of metagenomic contigs, wherein the MetaBAT method utilizes sequence composition features and joint abundance in combination with a pre-trained probability model for clustering.

[0015] Furthermore, the step S3 is specifically as follows:

[0016] S31. Obtain an initial k-medoid from the sample data containing the contiguous sequence fragments;

[0017] S32. Setting the cutoff distance of other contig sequences to the obtained initial k-medoid, and screening the contig sequences within the k-medoid cutoff distance, and generating clusters of the screened contig sequences;

[0018] S33. Re-screen a new k-medoid from the generated clusters;

[0019] S34. Repeat steps S32-S33 until k-medoid stops updating, obtains the contig sequences in the cluster, and generates a bin for the obtained contig sequences;

[0020] S35. Repeat steps S31-S34 until all overlapping group sequences are executed, and all bins of different sizes are obtained.

[0021] Furthermore, after step S35, the following steps are further included:

[0022] S36. Filter all bins whose quality is less than the threshold, and decompose the filtered bins into free contig sequences.

[0023] Furthermore, the quality detection of the obtained clustering results in step S4 is specifically as follows:

[0024] If it is a real data set, use the binning quality assessment tool to evaluate the quality of the clustering results and obtain bins with quality information; if it is a simulated data set, calculate the quality information of the bins based on the label information.

[0025] Furthermore, the quality information includes the precision and recall of each bin.

[0026] Furthermore, the step S5 is specifically as follows:

[0027] S51. Using the reference database to identify marker genes for the contig sequences in set S;

[0028] S52. Classify marker genes using reference databases;

[0029] S53. Classify the contig sequences corresponding to the marker genes;

[0030] S54. Selection of classification units when building an integrated SVM model;

[0031] S55. Balancing the training data for ensemble SVM model training;

[0032] S56. Build a linear integrated support vector machine model, vote on the final binning results, and put the results into the Result set.

[0033] Furthermore, the metagenomic simulation dataset calculates the quality information of the bins according to the following calculation method of accuracy and recall, expressed as:

[0034]

[0035]

[0036] Where Precision represents the accuracy; Recall represents the recall rate; n represents the number of species in the metagenomic dataset; C represents the number of clusters obtained after the algorithm runs; B ij represents the number of overlapping groups of species j assigned to cluster i; max j represents the maximum number of contigs from the same species in cluster i, max i Represents the cluster with the largest number of contigs from a given species.

[0037] Compared with the prior art, the present invention proposes a binning method for metagenomic contigs using unsupervised clustering and reference databases, which fully utilizes the advantages of unsupervised clustering and reference databases constructed by scientists to improve the overall classification effect. This method is compared with other state-of-the-art clustering algorithms including CONCOCT, Metabin2.0, and MetaBAT on simulated and real data sets, and the results show that this method can obtain higher accuracy and improve the binning effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of a method for binning metagenomic contigs using a reference database provided in Example 1;

[0039] Figure 2 This is a schematic diagram of the integrated SVM construction process provided in Example 1;

[0040] Figure 3 is a comparison chart of the accuracy and recall of the four methods on the 'Strain' dataset provided in Example 2;

[0041] Figure 4 This is a comparison chart of the accuracy and recall of the four methods on the 'species' dataset provided in Example 2;

[0042] Figure 5 This is a comparison chart of the accuracy and recall rates of the four methods on the 'sharon' dataset provided in Example 2. DETAILED DESCRIPTION

[0043] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0044] Existing metagenomics can directly extract the genetic material of all microorganisms from the environment and obtain metagenomic samples with a large number of unknown DNA sequences. Clustering of metagenomic samples is a hot issue in metagenomics research. The current clustering problems are mainly: 1. Unsupervised clustering methods rarely use reference databases, resulting in a certain degree of resource waste; 2. Unsupervised clustering methods are limited by the sequence features and clustering algorithms used, and the clustering effect is limited.

[0045] Therefore, the present invention provides a method for binning metagenome contigs using a reference database, and the whole process is divided into two stages. In the first stage, we first collect the reads fragments in multiple metagenome samples into a gene library and then assemble them with a reads assembly tool to assemble them into contig fragments. Then perform an unsupervised clustering process. In the second stage, the clustering results generated in the first stage are quality tested. It is determined whether the obtained bins meet the quality threshold, and the bins that meet the threshold are placed in the Results set, and the bins that do not meet the threshold are placed in the set S as input data for the next analysis. The present invention uses a method for directly obtaining sample-derived data and classification information for training using a reference database and a metagenome sample. For the construction of an integrated SVM classification model, five linear support vector machine models are used to respectively use the 4-mers, 5-mers, 6-mers, 7-mers, and 8-mers features of the sequence for modeling, and then the final results are voted on and the results are placed in the Result set.

[0046] Embodiment 1

[0047] This embodiment provides a method for binning metagenome contigs using a reference database, such as Figure 1 As shown, the following steps are included:

[0048] S1. Obtain read fragments from multiple metagenomic samples and pool the obtained read fragments into a gene library;

[0049] S2. Assemble the reads fragments collected in the gene library to obtain contiguous sequence fragments;

[0050] S3. performing an unsupervised clustering process on the obtained contig sequence fragments to obtain binned clustering results containing contig sequences of different sizes;

[0051] S4. Perform quality detection on the bin clustering results containing contig sequences of different sizes, and determine whether the bins meet the quality threshold. If not, store the bins in the set S; if yes, put the bins into the final result Results set;

[0052] S5. Perform marker gene analysis on the overlapping group sequences in set S in combination with the reference database and construct an integrated SVM classification model to complete the binning operation.

[0053] In step S2, the reads fragment rows collected in the gene library are assembled to obtain contiguous sequence fragments.

[0054] The reads fragment rows in the gene library are assembled using reads assembly tools.

[0055] In step S3, an unsupervised clustering process is performed on the obtained contig sequence fragments to obtain binned clustering results containing contig sequences of different sizes.

[0056] The unsupervised clustering process uses MetaBAT, one of the most representative methods for unsupervised clustering of metagenomics, which performs particularly well on complex data sets. MetaBAT uses sequence composition features (quadruple oligonucleotide frequency) and joint abundance combined with a pre-trained probability model for clustering, and proposes a new K-medoids algorithm to adaptively determine the number of strains contained in the sample.

[0057] In this embodiment, the first stage of the binning process requires binning the metagenomic samples using an unsupervised metagenomic clustering method. First, the data preparation stage requires multiple metagenomic samples, where the samples contain reads data. The reads data of different samples are pooled into a gene library, and then assembled using a reads assembly tool to assemble the reads fragments into overlapping group fragments.

[0058] MetaBAT pre-trains a probabilistic model that can calculate the probability distance of the quadruple oligonucleotides between sequences and the probability distance of the joint abundance, combining the two distances into a combined distance, and using the combined distance to determine the similarity of a pair of contig sequences. The combined distances between all contig data form a feature matrix, which is applied to the improved k-medoid clustering algorithm. MetaBAT proposes an improved k-medoid clustering algorithm.

[0059] The specific process of performing unsupervised clustering is as follows:

[0060] S31. Obtain an initial k-medoid from the sample data containing the contiguous sequence fragments;

[0061] Find an initial k-medoid in metagenomic sample data.

[0062] S32. Setting the cutoff distance of other contig sequences to the obtained initial k-medoid, and screening the contig sequences within the k-medoid cutoff distance, and generating clusters of the screened contig sequences;

[0063] The cutoff distance from other contig sequences to the k-medoid is set, and the contig sequences within the k-medoid cutoff distance are recruited as a cluster.

[0064] S33. Re-screen a new k-medoid from the generated clusters;

[0065] A new k-medoid is reselected from the generated clusters based on the k-medoid update criterion in the cluster of the traditional k-medoid algorithm.

[0066] S34. Repeat steps S32-S33 until k-medoid stops updating, obtains the contig sequences in the cluster, and generates a bin for the obtained contig sequences;

[0067] Repeat S32-S33 until k-medoid is no longer updated, and these contig sequences form a bin.

[0068] S35. Repeat steps S31-S34 until all overlapping group sequences are executed, and all bins of different sizes are obtained.

[0069] For the remaining contig sequences, S31-S34 are repeated to form more bins until no contig sequences are left.

[0070] In this embodiment, the process further includes: S36. screening bins smaller than the mass size in all bins, and decomposing the screened bins into free overlapping group sequences.

[0071] Retain bins of a certain size (e.g., greater than 200 kb) and decompose small bins into free contig sequences. The advantage of this method is that it does not require the number of species in the metagenomic sample to be selected, and is suitable for clustering metagenomic samples with high complexity.

[0072] In step S4, the quality of the binned clustering results containing overlapping group sequences of different sizes is checked, and it is determined whether the bins meet the quality threshold. If not, the bins are stored in the set S; if so, the bins are placed in the final result Results set.

[0073] The quality test of the clustering results is as follows: In the first stage of the binning process, the MetaBAT metagenomic clustering method is used to cluster the metagenomic samples, and eventually bins of different sizes will be obtained. Each bin represents that the overlapping group sequences in it are derived from the same gene, while the overlapping group sequences in different bins represent different genes. The quality of the bins is evaluated. Use binning quality assessment tools such as CheckM to perform quality assessment on the binning results, and eventually bins with quality information will be obtained. CheckM is generally suitable for binning quality detection on real data sets, because the overlapping group data in real data sets have no labels, and tools such as CheckM are needed to analyze single copy genes to obtain the accuracy and recall of each bin. The overlapping group data in the simulated data set has classification labels, and the accuracy and recall of each bin can be calculated based on the labels.

[0074] The quality information obtained in this embodiment includes the accuracy and recall rate of each bin, and a quality threshold is set. The bins that meet the threshold in the binning results obtained after clustering are placed in the Results set (the set stores the final binning results), and the bins that do not meet the quality threshold are placed in the set S. In the second stage, the overlapping group data in the set S is combined with the reference database to obtain sample-derived training data by the training data selection method, and then an integrated SVM classification model is constructed for further binning. Here, we only focus on whether the overlapping group data from the same species are divided into the same bin, and do not care whether the classification annotation is appropriate.

[0075] In step S5, the contig sequences in set S are combined with the reference database to perform marker gene analysis and build an integrated SVM classification model to complete the binning operation.

[0076] The acquisition of training data for the integrated SVM classifier is extremely critical. Good training data is of great significance to the construction and training of the classification model. However, most of the metagenomic data extracted from the environment come from unknown microorganisms. The selection of training data mainly depends on the metagenomic data itself. The specific method used in this embodiment is as follows:

[0077] S51. Using the reference database to identify marker genes for the contig sequences in set S;

[0078] Marker gene identification for metagenomic data: DNA sequences in bins obtained after clustering are translated into protein sequences using all six reading frames. Regions similar to 34 marker genes with classification information are identified in translated and untranslated sequences. DNA sequences corresponding to the identified marker genes are used for further analysis.

[0079] S52. Classify marker genes using reference databases;

[0080] Assignment of marker genes: Use a composition-based naive Bayes classifier to assign labels with classification information to marker genes, and assign them at the lowest classification level possible.

[0081] S53. Classify the contig sequences corresponding to the marker genes;

[0082] Assignment of DNA sequences corresponding to marker genes: If a sequence contains multiple marker genes, the classification labels can be assigned by setting confidence zone thresholds.

[0083] S54. Selection of classification units when building an integrated SVM model;

[0084] Selection of classification units during model construction: Each classification unit needs to have more than 100 kb of sample-derived data for training. Classification units with less than 100 kb of training data do not participate in the construction of the classification model (parameter setting, minimum training data size = 100,000).

[0085] S55. Balancing the training data for ensemble SVM model training;

[0086] Balanced training data: For each classification unit, the training data should be no larger than 400 kb. Those larger than 400 kb should be screened according to the confidence value and length of the contigs (parameter setting, maximum training data size: 400,000).

[0087] S56. Build a linear integrated support vector machine model, vote on the final binning results, and put the results into the Result set.

[0088] like Figure 2 As shown in the figure, the construction method of the integrated SVM classification model is as follows:

[0089] Metagenomic classification methods based on sequence composition generally use nonlinear classifiers, and some use linear classifiers. This embodiment uses a linear classifier. Assume that e j ∈R D (D=4 k ) is the input data of the classifier, j∈{1,…,N}, where N is the number of sequences in the training set. The objective equation of the linear SVM is as follows:

[0090]

[0091]

[0092] Among them, c j ∈{-1,+1} is the species label of the jth sequence. j = +1, indicating that the jth sequence belongs to the cth category, otherwise c j = -1 means that the jth sequence does not belong to the cth category. The ω parameter controls the penalty intensity, while ξ j is the relaxation factor, which is used to express the classification loss of the jth sequence. W and b are the weight vector and bias, respectively. This embodiment uses the one-vs-rest strategy to transform the complex classification problem into a binary classification problem. Here, the jth class is used as an example to divide the samples into those belonging to the jth class and those not belonging to the jth class. Here, the linear SVM with L2-lossd is used, and the objective equation is as follows:

[0093]

[0094] Among them, c j Indicates label; e j Represents input data.

[0095] This embodiment uses the 4-mers, 5-mers, 6-mers, 7-mers, and 8-mers features of the sequence to train five linear SVM classifiers respectively, and finally uses voting to determine the attribution of the sequence. Some methods of integrated support vector machines based on different k-mers have been proposed. It can be seen that when k is less than 8, the binning effect increases with the growth of k. When k is greater than or equal to 8, the effect begins to decline. When k is too large, the dimension of the k-mers vector is very high, which will cause dimensional disaster to a certain extent and reduce the effect of the classifier. Therefore, this embodiment uses the 4-mers, 5-mers, 6-mers, 7-mers and 8-mers features of the sequence to build a model.

[0096] Performance evaluation of classification results:

[0097] The method is evaluated using precision, recall and F1 metrics. Because each overlapping group in the simulated data has its corresponding label, the precision and recall of the clustering results can be easily calculated. The precision describes the ability of overlapping groups in each bin to come from the same species, while the recall describes the completeness of overlapping groups from the same species in each bin. The calculation method of precision and recall can be conveniently expressed as follows:

[0098]

[0099]

[0100] Where Precision represents the accuracy; Recall represents the recall rate; n represents the number of species in the metagenomic dataset; C represents the number of clusters obtained after the algorithm runs; B ij represents the number of overlapping groups of species j assigned to cluster i; max j represents the maximum number of contigs from the same species in cluster i, max i Represents the cluster with the largest number of contigs from a given species.

[0101] Accuracy and recall are contradictory metrics. Generally speaking, when the accuracy is high, the recall tends to be low. When the accuracy is high, the recall tends to be low. In the classification of metagenomic data, this embodiment prefers to obtain bins with high accuracy. When measuring the classification effect of the classification method, the proportion of accuracy should be greater than the recall, so we use F β To express different preferences for precision and recall, it is expressed as:

[0102]

[0103] Among them, F β Indicates the different preferences for accuracy and recall; P represents accuracy; R represents recall; β represents the weights of the different preferences for accuracy and recall. When β>0, it measures the relative importance of accuracy and recall. When β=1, it degenerates into the standard F1 metric. When β>1, the recall rate accounts for a larger proportion, and vice versa, the accuracy rate accounts for a larger proportion. Here, β=0.8 is set.

[0104] This embodiment proposes a binning method for metagenomic contigs using unsupervised clustering and a reference database, which fully utilizes the advantages of unsupervised clustering and a reference database constructed by scientists to improve the overall classification effect.

[0105] Embodiment 2

[0106] The method for binning metagenome contigs using a reference database provided in this embodiment is different from that in the first embodiment in that:

[0107] In order to prove the effectiveness of this method, this method was compared with MetaBAT, Maxbin2.0, and CONCOCT on simulated data sets and real data sets. MetaBAT, Maxbin2.0, and CONCOCT are the most representative and most commonly used methods for metagenomic data clustering. The inputs of these three methods are the frequency and joint abundance characteristics of the sequence's quadruple oligonucleotides. Among them, MetaBAT proposed a new K-medoids algorithm for this method, which can adaptively determine the number of clusters during the algorithm operation. Maxbin2.0 uses the EM algorithm for clustering, and uses the marker gene identification screening method to initialize the algorithm. CONCOCT uses the principal component analysis method for dimensionality reduction and then uses the Gaussian mixture model. The strategy proposed by this method is to combine unsupervised clustering with a reference database, focusing on improving the accuracy and recall of binning rather than the annotation of overlapping group sequences. Use an existing reference database to generate a reference sequence set. The reference taxonomy was NCBI taxonomy, and the following reference databases from NBCI were merged to generate the reference sequence (RS) set: NCBI genomes, NCBI draft bacterial genomes, NCBI human microbiome project, NCBI RefSeq microbial version. Duplicate sequences were removed to ensure that the RS set was not redundant. The RS set contained sequences from 841 different genera, 2543 different species, and 4516 different strains. The total size of the RS set was 16Gb. For the marker gene (MG) analysis, the following MG sets and HMM annotations were used: Bacterial and archaeal reference sequences from the SILVA database were used for MG analysis of 16S and 23S. The corresponding taxonomic identifiers were mapped to the NCBI taxonomy. The final set contained 126,742 sequences corresponding to 39,291 different species (a total of 199Mb). In addition, MG data from NCBI by Malott et al. were used for MG analysis of 5S, and the sequence set included 12,424 sequences corresponding to 1278 species (a total of 5.8Mb). In addition, reference sequences of 31 bacterial marker gene families were obtained from NCBI. The MG collection contains 63,530 sequences corresponding to 1380 different species (a total of 52Mb). HMM maps of 16S, 23S and 5S marker genes were obtained from Huang et al. HMM annotations trained on protein families of 31 bacterial MGs were obtained from Wu & Scott.

[0108] This method uses existing simulated datasets to simulate two groups of microbial communities, called the 'Strain' and 'Species' datasets. Both synthetic datasets are based on 16S rRNA samples participating in the Human Microbiome Project. These samples have been subjected to noise reduction and the OUT criterion is the presence of 3% sequence difference with the similar species. Contigs are assembled by pooling read fragments from each sample, and the number of reads mapped to each sample for each contig is the coverage information. For both simulated datasets, Ray was used to assemble reads from all samples. Ray was used to generate co-assembled contigs because it can handle large metagenomic datasets. The 'Strain' dataset contains 9417 co-assembled contigs from 64 samples, which contains a total of 20 species of microorganisms, and each contig has its corresponding label, which can be used as a basis for clustering evaluation.

[0109] The 'Species' dataset has 101 different species, including 37628 contigs assembled from 96 samples. The purpose of this dataset is mainly to test the classification ability of the proposed method at the species level.

[0110] Real data This method uses the 'Sharon' dataset, which is a microbial sample from premature infants, including 18 data samples. Each contig in the real data does not have its corresponding label. Here, TAXAassign is used to annotate each contig sequence in the dataset. After filtering out the ambiguous labels at the species level, we obtained 7 species, containing 2599 contig sequences.

[0111] In the process of running this method, the threshold is set to be greater than or equal to 50% for both precision and recall. It is then tested and compared with other state-of-the-art clustering methods CONCOCT, Maxbin2.0, and MetaBAT on simulated data sets and real data sets. The results are as follows Figure 3 , 4, 5. It can be seen that this method has a high accuracy and recall rate. On the 'Strain' dataset, the accuracy and recall rate of this method are higher than those of MaxBin2.0 and MetaBAT, with an accuracy rate of 85.3% and a recall rate of 92.3%. CONCOCT performs best on the 'Strain' dataset. On the 'Species' dataset, this method performs best in terms of accuracy, with an accuracy rate of 99.1% and a recall rate of 93.6%. Maxbin2.0 and MetaBAT perform better when facing more complex datasets, while the clustering effect of CONCOCT will be significantly reduced. On the real dataset 'Sharon', similar to the performance on the simulated dataset, this method has the highest accuracy.

[0112] F of different classification methods on simulated data sets and real data sets 0.8 The comparison of the F 0.8 Only CONCOCT performs better than this method on the 'Strain' dataset.

[0113] F of four different classification methods on simulated data sets and real data sets 0.8 The comparison is shown in Table 1:

[0114] Methods 'Strain' dataset 'Species' dataset 'Sharon' dataset CONCOCT 0.939 0.881 0.829 Maxbin2.0 0.792 0.939 0.776 MetaBAT 0.741 0.928 0.847 Our method 0.879 0.968 0.924

[0115] Table 1 The running time of CONCOCT, Maxbin2.0, MetaBAT and this method is shown in Table 2:

[0116]

[0117] Table 2

[0118] This method proposes a method that combines unsupervised clustering and reference database, and compares it with three other state-of-the-art clustering methods CONCOCT, Maxbin2.0, and MetaBAT on simulated data and real data sets, and achieves good results. This method mainly uses the advantages of unsupervised clustering methods, and uses integrated support vector machines to process the parts of unsupervised clustering results that are not good. When facing metagenomic data samples extracted from the environment, due to the large number of unknown species in the samples, if supervised classification is performed directly, there are the following difficulties: 1. Lack of label data, supervised training cannot be performed. 2. If there are many species in the sample, there will be many categories that need to be modeled, modeling will take a long time, and it is difficult to classify at a low level. 3. If it is a method based on reference database comparison, due to the large number of unknown species in the metagenomic samples, this method has great limitations. Unsupervised methods can overcome these difficulties. This type of method mainly clusters based on the characteristics of the sequence itself, does not rely on reference databases and label data, and can be classified to the species level. The disadvantage is that it is not easy to obtain more high-quality bins from metagenomic samples due to the limitations of sequence features and methods, and the existing reference database cannot be used. Therefore, based on the above points, this method uses the MetaBAT clustering method to perform quality detection on the clustering results, and then selects the bins whose quality conditions do not meet the threshold for integrated support vector machine modeling. Due to the lack of direct label information, this method uses a method of directly obtaining training data and labels from samples and combining them with reference databases, so that there is no need to know the training data and labels in advance. Then this method constructs an integrated support vector machine classification model, models different sequence k-mers features, and obtains a total of five linear support vector machine classifiers. The reason why different k-mes features are used for training is that for metagenomic data, there is no definite k-mers feature that can achieve the best classification effect. Therefore, 4-mers, 5-mers, 6-mers, 7-mers, and 8-mers are used here to train five linear SVM classifiers. Finally, a voting method is used. The classification with a voting result greater than 50% is the classification of the contig, otherwise it is classified as unclassifiable. The purpose of this method is to integrate support vector machine classification for the parts that cannot be well classified by unsupervised clustering, so that the reference database constructed in supervised classification can be used to further classify the results of unsupervised clustering, thereby improving the classification effect. This method does not focus on the annotation of species in supervised classification methods, but focuses on classifying overlapping groups of the same species into the same bin.

[0119] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for binning metagenome contigs using a reference database, characterized in that: The following steps are involved: S1. Obtain read fragments from multiple metagenomic samples and pool the obtained read fragments into a gene library; S2. Assemble the reads fragments collected in the gene library to obtain contiguous sequence fragments; S3. performing an unsupervised clustering process on the obtained contig sequence fragments to obtain binned clustering results containing contig sequences of different sizes; S4. Perform quality detection on the bin clustering results containing contig sequences of different sizes, and determine whether the bins meet the quality threshold. If not, store the bins in the set S; if yes, put the bins into the final result Results set; S5. Perform marker gene analysis on the contig sequences in set S in combination with the reference database and construct an integrated SVM classification model to complete the binning operation; The step S5 is specifically as follows: S51. Using the reference database to identify marker genes for the contig sequences in set S; S52. Classify marker genes using reference databases; S53. Classify the contig sequences corresponding to the marker genes; S54. Selection of classification units when building an integrated SVM model; S55. Balancing the training data for ensemble SVM model training; S56. Build a linear integrated support vector machine model, vote on the final binning results, and put the results into the Results collection.

2. A method for binning metagenome contigs using a reference database according to claim 1, characterized in that: The unsupervised clustering process in step S3 adopts the unsupervised clustering MetaBAT method for metagenomic contigs, wherein the MetaBAT method uses sequence composition features and joint abundance in combination with a pre-trained probability model for clustering.

3. A method for binning metagenome contigs using a reference database according to claim 2, characterized in that: The step S3 is specifically as follows: S31. Obtain an initial k-medoid from the sample data containing the contiguous sequence fragments; S32. Setting the cutoff distance of other contig sequences to the obtained initial k-medoid, and screening the contig sequences within the k-medoid cutoff distance, and generating clusters of the screened contig sequences; S33. Re-screen a new k-medoid from the generated clusters; S34. Repeat steps S32-S33 until k-medoid stops updating, obtains the contig sequences in the cluster, and generates a bin for the obtained contig sequences; S35. Repeat steps S31-S34 until all overlapping group sequences are executed, and all bins of different sizes are obtained.

4. A method for binning metagenome contigs using a reference database according to claim 3, characterized in that: After step S35, the following steps are also included: S36. Filter all bins whose quality is less than the threshold, and decompose the filtered bins into free contig sequences.

5. A method for binning metagenome contigs using a reference database according to claim 3, characterized in that: The quality detection of the obtained clustering results in step S4 is specifically as follows: If it is a real data set, use the binning quality assessment tool to evaluate the quality of the clustering results and obtain bins with quality information; if it is a simulated data set, calculate the quality information of the bins based on the label information.

6. A method for binning metagenome contigs using a reference database according to claim 5, characterized in that: The quality information includes the precision and recall of each bin.

7. A method for binning metagenome contigs using a reference database according to claim 6, characterized in that: The metagenomic simulation dataset calculates the quality information of the bins according to the following calculation method of precision and recall, expressed as: Where Precision represents the accuracy; Recall represents the recall rate; n represents the number of species in the metagenomic dataset; C represents the number of clusters obtained after the algorithm runs; B ij represents the number of overlapping groups of species j assigned to cluster i; max j represents the maximum number of contigs from the same species in cluster i, max i Represents the cluster with the largest number of contigs from a given species.

Images