A method for predicting bacterial phenotypic characteristics from bacterial genomic data based on machine learning

Abstract

The present invention discloses a method for predicting bacterial phenotypic characteristics based on bacterial genome data of machine learning, comprising the following steps: (1) performing protein domain analysis on bacterial genome data; (2) reconstructing and combining the bacterial protein domains obtained by analysis, counting the frequencies and converting them into matrices, and obtaining a protein domain feature data set reconstructed by bacteria; (3) dividing the integrated bacterial data in step (2) into a training set and a prediction set according to whether they are known feature types; (4) performing model training on the training set data using a machine learning algorithm to select the best prediction model, and using the best prediction model to perform feature prediction on the prediction set data. The method can perform batch prediction of bacterial features on bacteria that are difficult to separate and culture and bacterial genomes assembled using metagenomes, obtain a large amount of bacterial feature information, and save a large amount of experimental research costs and time.

Description

Technical Field

[0001] The present invention relates to the technical field of biological high-throughput data analysis, and in particular to a method for predicting bacterial phenotypic characteristics based on bacterial genome data based on machine learning. Background Art

[0002] Microorganisms are almost everywhere, including bacteria, archaea, algae, fungi and protozoa, as well as viruses, and play an important role in maintaining good human, animal and plant health, but some microorganisms are also the direct cause of many diseases. The cultivation and study of microorganisms using biological experiments is indispensable for understanding the role of microorganisms in the world around us, and it remains essential even in the age of bioinformatics.

[0003] At present, when researchers use culturable methods to study microorganisms, only a small number of microorganisms can be cultured in pure culture, because too many microbial culture conditions are unknown or there are related dependencies between microorganisms, which seriously hinders the study of the laws of microbial life activities and the development and utilization of microbial resources. In addition, metagenomic sequencing data, especially second-generation and third-generation metagenomic sequencing data, can obtain some complete metagenome assembled genomes (MAG-s) using metagenomic binning technology, and the recently announced bacterial single-cell sequencing technology can obtain a large number of bacterial reference genomes. With the increasing number of bacterial genomes obtained through high-throughput sequencing technology, it is difficult to obtain phenotypic characteristics such as bacterial aerobicity of many new species represented by new genomes that are difficult to identify by ribosomal sequences. Although there have been a few reports on the use of naive Bayesian algorithms to predict bacterial characteristics in the past, such as using bacterial genomes to predict bacterial aerobicity, the prediction accuracy is 88% for anaerobic bacteria, 87% for aerobic bacteria, and 35% for facultative bacteria, respectively. The accuracy needs to be improved and it is difficult to apply in practice. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a method for predicting bacterial phenotypic characteristics based on bacterial genome data based on machine learning through technical improvements.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0006] The present invention provides a method for predicting bacterial phenotypic characteristics based on bacterial genome data based on machine learning, comprising the following steps:

[0007] (1) Protein domain analysis of bacterial genome data;

[0008] (2) Reconstruct and combine the bacterial protein domains obtained by analysis, count the frequencies and convert them into matrices to obtain a bacterial reconstructed protein domain dataset;

[0009] (3) The obtained bacterial reconstructed protein domain feature dataset is divided into a training set and a prediction set according to whether it is a known feature type;

[0010] (4) Use machine learning algorithms to train the training set data to select the best prediction model, and use the best prediction model to perform feature prediction on the prediction set data.

[0011] As a further optimization scheme of the present invention, in step (1), the bacterial genome data is derived from a bacterial genome database.

[0012] As a further optimization scheme of the present invention, in step (1), the open source software pfam_scan is used to perform protein domain prediction analysis on the bacterial genome data, using the default parameters: pfam_scan.pl-fasta protein.fa-dir / data3 / pfam / Pfam-A.hmm>protein.pfam.outfile.

[0013] As a further optimization scheme of the present invention, in step (2), based on the obtained bacterial protein domain characteristics, the protein domain composition is reconstructed according to each gene and the frequency of the reconstructed domain in each bacterium is analyzed to construct a frequency matrix of the protein domain composition reconstructed for each bacterium.

[0014] As a further optimization scheme of the present invention, in step (3), the characteristic types include but are not limited to aerobic type and optimal growth temperature.

[0015] As a further optimization scheme of the present invention, in step (4), the machine learning algorithm includes but is not limited to random forest, support vector machine, decision tree, gradient boosting tree, naive Bayes and conditional inference tree.

[0016] As a further optimization scheme of the present invention, in step (4), the machine learning algorithm selects a random forest algorithm.

[0017] The beneficial effect of the present invention is that the method can accurately predict unknown characteristics of bacteria in batches, which can not only save a lot of experimental research costs and time, but also obtain a large number of phenotypic characteristics of bacteria with difficult-to-cultivate bacterial genomes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A technical solution roadmap provided for the present invention;

[0019] Figure 2The best prediction model provided in Example 1 of the present invention predicts the results of 1571 known aerobic bacteria;

[0020] Figure 3 The best prediction model provided in Example 1 of the present invention predicts the results of 1542 unknown aerobic bacteria;

[0021] Figure 4 The result of predicting the optimal growth temperature of bacteria using the random forest algorithm provided in Example 2 of the present invention;

[0022] Figure 5 This is the result of evaluating the accuracy of the prediction method of the present invention using the absolute value of the error provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0023] The present application is further described in detail below in conjunction with the accompanying drawings. It is necessary to point out here that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technical personnel in this field can make some non-essential improvements and adjustments to the present application based on the above application content.

[0024] Example 1

[0025] like Figure 1 As shown, this embodiment provides a method for predicting bacterial aerobic types using bacterial genomes based on machine learning, comprising the following steps:

[0026] (1) Download a large amount of bacterial genome data from the bacterial genome database and perform protein domain analysis on it: use the open source software pfam_scan with default parameters: pfam_scan.pl-fasta protein.fa-d ir / data3 / pfam / Pfam-A.hmm>protein.pfam.outfile.

[0027] (2) Integration of bacterial data: The protein domains of bacteria obtained by analysis were reconstructed and combined according to each gene coding sequence and the frequencies were counted to construct the frequency matrix of protein domain composition of each bacteria. The reconstructed protein domain dataset pfam.txt containing multiple bacteria was obtained.

[0028] (3) Search for bacterial phenotypic data: Use the standardized bacterial information database Bac Dive to query the aerobic type and optimal growth temperature of bacteria one by one, or directly collect and organize the data published in existing research, and divide the bacteria with known aerobic type and optimal growth temperature into the training set, and divide the bacteria with unknown aerobic type and optimal growth temperature into the prediction set.

[0029] (4) Create an experimental design table design.txt, first divide the data of known phenotypic types into a training set (group 1) and a test set (group 2), and adjust the model ratio (1:1 to 3:1) to achieve the best prediction model.

[0030] Taking the random forest model loaded in R language as an example, the prediction results of the test set are stored in RF_prediction_binary.txt:

[0031] design=read.table("design.txt",header=T,row.names=1)

[0032] otu_table=read.table("pfam.txt", header=T, row.names=1)

[0033] design_sub=subset(design,Group%in%c("group1"))

[0034] design_sub$Type=as.factor(design_sub$Type)

[0035] idx=rownames(design_sub)%in%colnames(otu_table)

[0036] design_sub=design_sub[idx,]

[0037] otu_sub=otu_table[,rownames(design_sub)]

[0038] library(randomForest)

[0039] set.seed(315)

[0040] rf=randomForest(t(otu_sub),design_sub$Type,importance=TRUE,proximity=T,ntree=1000)

[0041] print(rf)

[0042] set.seed(315)

[0043] result=rfcv(t(otu_sub),design_sub$Type,cv.fold=5)

[0044] result$error.cv

[0045] imp=as.data.frame(rf$importance)

[0046] imp=imp[order(imp[,1],decreasing=T),]

[0047] write.table(imp,file="importance_class.txt",quote=F,sep='\t',row.names=T,col.names=T)

[0048] design_test=subset(design,Group%in%c("group2"))

[0049] design_test$Type=as.factor(design_test$Type)

[0050] idx=rownames(design_test)%in%colnames(otu_table)

[0051] design_test=design_test[idx,]

[0052] otu_sub=otu_table[,rownames(design_test)]

[0053] otutab_t=as.data.frame(t(otu_sub))

[0054] otutab_t$Type=design[rownames(otutab_t),]$Type

[0055] set.seed(315)

[0056] otutab.pred=predict(rf,t(otu_sub))

[0057] pre_tab=table(observed=otutab_t[,"Type"],predicted=otutab.pred)

[0058] predict=data.frame(Type=otutab_t[,"Type"],predicted=otutab.pred)

[0059] write.table("SampleID\t",file=paste("RF_prediction_binary.txt",sep=""),ap-pend=F,quote=F,eol="",row.names=F,col.names=F)

[0060] write.table(predict,file="RF_prediction_binary.txt",append=T,quote=F,row.names=T,col.names=T,sep="\t")

[0061] Taking the 3113 bacterial dataset as an example, the dataset of known aerobic types is 1571, and the test set prediction results of the best model are as follows: Figure 2 As shown in the figure, the recall rates of aerobic bacteria, anaerobic bacteria, and facultative bacteria predicted by the optimal random forest algorithm are 88.87%, 96.43%, and 88.03%, respectively. The Kappa coefficient of the random forest algorithm is 0.87, and the prediction accuracy is very high. The best model is used to predict 1542 unknown aerobic bacteria. The results are as follows Figure 3 shown.

[0062] Example 2

[0063] This embodiment provides a method for predicting the optimal growth temperature of bacteria using bacterial genome based on machine learning. The prediction implementation process of the specific prediction method is consistent with the steps given in Example 1.

[0064] In this embodiment, the constructed machine learning model includes 378 training sets and 377 test sets of bacterial samples. By optimizing multiple algorithms, it is found through testing that the method of using the random forest algorithm to predict the optimal growth temperature of bacteria is the most accurate. The results are as follows Figure 4 As shown;

[0065] In the evaluation based on the absolute values ​​of the errors of 15°C, 10°C, and 5°C, the accuracy of the test set prediction reached 100%, 97%, and 78%, respectively. It can be seen that the prediction method provided by the present application can more accurately predict the optimal growth temperature of bacteria. The results are as follows Figure 5 shown.

[0066] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for predicting bacterial phenotypic characteristics based on bacterial genome data based on machine learning, characterized in that: The following steps are involved: (1) Use pfam_scan to analyze protein domains of bacterial genome data; (2) The protein domains of bacteria obtained by analysis were reconstructed and combined according to each coding sequence and the frequencies were counted to construct the composition frequency matrix of the protein domains reconstructed by each bacteria, thus obtaining the characteristic dataset of the protein domains reconstructed by bacteria; (3) For the protein domain feature dataset reconstructed by bacteria obtained in step (2), based on whether it is a known feature type, bacteria with known aerobic type and optimal growth temperature are divided into a training set, and bacteria with unknown aerobic type and optimal growth temperature are divided into a prediction set; (4) Use the random forest algorithm to train the training set data to select the best prediction model, and use the best prediction model to perform feature prediction on the prediction set data.

2. The method for predicting bacterial phenotypic characteristics based on bacterial genome data based on machine learning according to claim 1, characterized in that: In step (1), the bacterial genome data comes from the bacterial genome database.

Images