CN106778070A - A kind of human protein's subcellular location Forecasting Methodology - Google Patents

Abstract

The invention discloses a method for predicting the subcellular location of a human protein, which uses the sequence of a human protein to predict the subcellular location of the protein, and optimizes the human protein subcellular classification algorithm based on Gene Ontology (GO) features and conserved domain correlations . Firstly, the sequence residue statistical features of the protein (amino acid composition features, normalized specificity scoring matrix features), conserved domain features and GO features are obtained through the protein sequence; secondly, CFS feature selection is used for the sequence residue statistical features The method extracts feature subsets, calculates the similarity measures of the conservative domain features and GO features, and uses the KNN method with weights to calculate the probability information, and then integrates the obtained features and uses the SVM classifier for classification.

Description

A method for predicting the subcellular location of human proteins

technical field

The invention belongs to the technical field of biological information, in particular to a method for predicting the subcellular location of human proteins.

Background technique

Understanding the subcellular location of proteins is of great significance for understanding protein functions, protein-protein interactions, and drug-targeted therapy. However, it takes a lot of time and cost to obtain the subcellular location of proteins by using experimental methods. Therefore, it is of great significance to use protein subcellular location prediction tools to predict a large number of proteins. According to our statistics, there are a total of 550,552 proteins in the SWISS-PROT protein database released in February 2016, of which only 10.4% of the proteins have experimentally verified subcellular locations, and the remaining proteins with unknown subcellular locations urgently need to pass a A reliable forecasting method to forecast.

So far, there are many tools that can predict the subcellular location of proteins. Common web servers include BaCeLlo, YLoc, MultiLoc, GOASVM, WoLF PSORT, CellPLoc, HSLPred, etc. These predictive tools have brought great convenience to biologists in related fields.

The subcellular location information of proteins is often used in gene therapy and drug-targeted therapy of diseases. For example, the role of the Hippo/YAP pathway in the evolution of pediatric hepatocellular carcinoma was investigated by examining the expression and subcellular localization of the protein YAP in tumors. Therefore, an easy-to-use high-precision prediction tool will be very helpful for these laboratories to conduct clinical research. Our previously released web server Hum-mPLoc2.0 is specifically designed for predicting human protein localization. The number of uses per year has increased from 20,000 in 2010 to more than 80,000 in 2015. This shows that in order to provide better prediction services, it is of great significance to further enhance the prediction ability based on new technologies and more comprehensive and accurate annotation databases.

Generally, computational methods for predicting protein subcellular localization can be divided into two categories, namely, homology search-based methods and machine learning-based methods. Homology search-based methods can be thought of as making predictions using nearest neighbor methods, where the distance between two proteins is usually measured by their sequence homology. By calculating the homology between the query protein and a large number of sequences that have subcellular location annotation information, the method finds the top K most similar proteins, and transfers their annotation information to the protein to be predicted as the classification result. The method based on homology search is a relatively straightforward prediction method, but its performance depends significantly on whether it can find homologous sequences with high similarity and annotated subcellular location information. In addition, sometimes the difference between two protein sequences The similarity between them is high but they may have very different structures or functions, which will lead to the failure of this method.

Machine learning-based predictors are a class of flexible models for protein subcellular location prediction. They require so-called training data sets and then learn classification rules through algorithms based on statistical learning. Therefore, the quality of the training data is closely related to the quality of the statistical rules learned. Benefiting from the increasing and reliable annotation of subcellular location information in protein databases, we can more fully train classification models by collecting large-scale training data. Another important issue in machine learning models is how to encode protein sequences, because most algorithms need to extract feature vectors as input, how to extract features from original protein sequences and associated existing knowledge is crucial to the final performance of classifiers. important. Existing machine learning tools for predicting subcellular locations use various features as follows:

(1) Residue-based statistical features, pseudo-amino acid composition and position-specific scoring matrices.

(2) Characterization of functional domains based on signal peptides.

(3) Features based on database annotations, such as Gene Ontology (GO) features.

Since GO features are high-level abstractions of domain knowledge, they usually have higher accuracy than sequence-based features when they have sufficient annotations. However, large amounts of annotated data bring new algorithmic challenges. For example, by using a Bernoulli event model for each GO feature, i.e. binary encoding the presence or absence of that GO feature, often results in extremely high-dimensional feature spaces. As the GO database is regularly expanded and updated, the dimensions will continue to increase as our knowledge about proteins expands. High-dimensional feature vectors increase the complexity of the machine learning process, and we also take into account the impact of potential noise in the annotated database. Although the entire GO database is huge, each protein actually only contains a few GO features. According to our statistics, those proteins with at least one GO feature in the SWISS-PROT database have an average of 6 GO annotations. That is to say, the GO feature of a protein is a sparse feature vector, which has thousands of dimensions, but only about 6 GO annotations. At present, different methods have been proposed to deal with this problem in the field. For example, YLoc selects only GO annotations and PROSITE patterns that are clearly relevant for specific subcellular locations. Therefore, it reduces unnecessary features and makes the results easier to understand, but this also leads to loss of information. WegoLoc assigns weights to each GO feature to highlight useful GO features.

Contents of the invention

The invention provides a method for predicting the subcellular location of human proteins, with the purpose of improving the prediction accuracy of a human protein subcellular classifier by using potential correlation information between annotation features.

A method for predicting the subcellular location of a human protein, which predicts the subcellular location of a protein based on a human protein sequence, comprising the following steps:

Step 1: Use human protein sequence information to extract the statistical characteristics of residues of the full-length sequence, N-terminal and C-terminal protein sequence fragments, including amino acid composition characteristics and specificity scores obtained by using protein homology information Matrix features and normalize the features, and use Correlation-based Feature Selection, a supervised feature selection algorithm, to perform dimensionality reduction after combining these two features;

Step 2: By extracting the GO features of all human proteins in the protein database, use GOSSTO to obtain three similarity matrices in the GO (BP, MF, CC) feature space;

Step 3: search for homologous proteins in the Swiss-Prot database by the blast method, extract the GO features of the homologous proteins, and use the same method to obtain the GO features of the proteins in the training set;

Step 4: Divide the three parts of the protein GO feature (BP, MF, CC) into 7 parts (BP, MF, CC), (BP&MF, BP&CC, MF&CC) through one-tuple, two-tuple and three-tuple ,(BP&MF&CC);

Step 5: Through the correlation of protein GO features, divide it into seven parts to calculate the correlation of two proteins, and through parameter optimization, extract ten highly correlated proteins in the training set to do the weighted KNN method to obtain the The probability value of the protein at each subcellular location;

Step 6: Obtain the conserved domain features of all human proteins in the Swiss-Prot database through rps-blast, and calculate the correlation between features through the information difference to obtain the similarity matrix of conserved domain features, and then obtain it through rps-blast The conserved domain features of the target protein are used to calculate the correlation between the two proteins, and through parameter optimization, ten highly correlated proteins are extracted from the training set to do the weighted KNN method to obtain the probability of the protein at each subcellular location value;

Step 7: Merge the obtained sequence features, the probability features of the seven parts of GO, and the probability features of the conserved domain. Using the Binary Relevance strategy to build can predict the centrosome, cytoplasm, cytoskeleton, endoplasmic reticulum, endosome, secretory pathway, Golgi SVM classifiers for 12 subcellular locations of body, lysosome, mitochondria, nucleus, peroxisome and membrane.

A method for predicting the subcellular location of a human protein, which predicts the subcellular location of a protein based on a human protein sequence, comprising the following steps:

S101, using human protein sequence information to extract the amino acid composition characteristics of the full-length sequence, the first 10 to 60 at the N-terminus, and the first 10 to 100 at the C-terminus, and the normalized PSSM matrix features, and use CFS to reduce the dimension , where the formula for normalizing the PSSM matrix and converting it into a 20-dimensional feature in each part is:

Where S _i,j represents the probability score of the amino acid appearing at the i-th (1≤i≤L) position of the sequence to evolve into the j-th (1≤j≤20) amino acid during the evolution process, and L represents the protein sequence length.

Represents the score of this specificity scoring matrix after normalization, and N of this represents the number of amino acids, so N is equal to 20 in Formula 2.

in Indicates the value after adding and averaging the scores of each column;

It is the normalized PSSM matrix feature we got.

S102, by extracting the GO features of all human proteins in the Swiss-Prot database, using GOSSTO to obtain three similarity matrices in the GO (BP, MF, CC) feature space;

S103, search for homologous proteins in the Swiss-Prot database by the blast method, extract their GO features, and use the same method to obtain the GO features of the proteins in the training set;

S104, divide the three parts (BP, MF, CC) of the protein GO feature into 7 parts (BP, MF, CC), (BP&MF, BP&CC, MF&CC), ( BP&MF&CC);

S105, through the correlation of protein GO features, it is divided into seven parts to calculate the correlation of two proteins:

Among them, Cor( _xi ,K) represents the correlation between the GO annotation features represented by _xi and the Kth protein under this section.

where Sim _k represents the correlation between the Kth protein in the training set and our predicted protein.

After obtaining the correlation between all the proteins in the training set and the predicted protein, we extract ten highly correlated proteins in the training set as the weighted KNN method to obtain the probability value of the protein at each subcellular location :

Among them, num _a and num represent the number of proteins in the a-th subcellular position in the training set and the total number of proteins in the training set, respectively. And pro _a represents the probability that the predicted protein is in the ath subcellular position.

S106, use rps-blast to obtain the conserved domain features of all human proteins in the Swiss-Prot database, and calculate the correlation between features through the information difference:

in Denotes the entropy of the i-th CDD feature, Indicates the probability that the i-th CDD feature exists in the protein training set. Represents the differential entropy of the i-th feature and the j-th feature, represents the correlation between the i-th CDD feature and the j-th CDD feature.

Obtain the similarity matrix of the conserved domain features, and then use rps-blast to obtain the conserved domain features of the target protein to calculate the correlation between the two proteins, and extract ten highly correlated proteins from the training set as the weighted KNN method to obtain The probability value of the protein at each subcellular location;

S107, integrate the obtained sequence features, the probability features of the seven parts of GO, and the probability features of the conserved domains, use the Binary Relevance strategy to build 12 SVM classifiers to predict the subcellular location of the protein, and the probability of each subcellular location.

In the present invention, feature vectors are encoded by GO related information instead of using the frequency of GO in annotations. As we all know, GO features can be roughly divided into three parts, namely biological process (BP), molecular function (MF) and cellular composition (CC). These three parts feature a hierarchical structure. According to this hierarchy, many methods have been proposed in the field to define the semantic similarity between GO features, such as entropy-based methods and graph-theory-based methods. However, to our knowledge, few current prediction algorithms for protein subcellular locations take into account the correlations between these GO features. This motivates us to obtain a better similarity measure between two high-dimensional but sparse GO feature vectors through the hidden correlation between GO features. We propose a novel approach to exploit hidden correlations between annotated features of proteins. In order to deal with the lack of GO annotations for some proteins that need to be predicted due to the incompleteness of the GO database, we also combined statistical protein sequence residue features and peptide-based functional domain features extracted from the Conserved Domain Database (CDD) to construct A new predictor, called Hum-mPLoc3.0, is named after our previously developed predictor for human protein localization prediction, but endowed with a completely new feature representation.

Compared with methods in the prior art, the present invention has significant advantages:

(1) The potential correlation between annotation features is used in the model, which effectively improves the prediction accuracy of human protein subcellular location;

(2) The statistical features of sequence residues, conserved domain features and GO features are integrated, which effectively improves the prediction accuracy of human protein subcellular location.

Description of drawings

Fig. 1 is a system structure diagram of the human protein sequence prediction method of the present invention:

detailed description

The present invention will be further described below in conjunction with the accompanying drawings.

Fig. 1 has provided the human protein sequence prediction method system structural diagram of the present invention:

Firstly, the sequence residue statistical features, conserved domain features and GO features of the protein are obtained through the protein sequence; secondly, the CFS feature selection method is used to extract the feature subset for the sequence residue statistical features, and the conserved domain features and GO features are calculated separately The similarity measure of these features is obtained, and the probability information is calculated using the KNN method with weights, and then the obtained features are integrated and classified using the SVM classifier. The following is a detailed description:

S101, using human protein sequence information to extract the amino acid composition characteristics of the full-length sequence, the first 10 to 60 at the N-terminus, and the first 10 to 100 at the C-terminus, and the normalized PSSM matrix features, and use CFS to reduce the dimension , where the formula for normalizing the PSSM matrix and converting it into a 20-dimensional feature in each part is:

Where S _i,j represents the probability score of the amino acid appearing at the i-th (1≤i≤L) position of the sequence to evolve into the j-th (1≤j≤20) amino acid during the evolution process, and L represents the protein sequence length.

Represents the score of this specificity scoring matrix after normalization, and N of this represents the number of amino acids, so N is equal to 20 in Formula 2.

in Indicates the value after adding and averaging the scores of each column;

It is the normalized PSSM matrix feature we got.

S102, by extracting the GO features of all human proteins in the Swiss-Prot database, using GOSSTO to obtain three similarity matrices in the GO (BP, MF, CC) feature space;

S103, search for homologous proteins in the Swiss-Prot database by the blast method, extract their GO features, and use the same method to obtain the GO features of the proteins in the training set;

S104, divide the three parts (BP, MF, CC) of the protein GO feature into 7 parts (BP, MF, CC), (BP&MF, BP&CC, MF&CC), ( BP&MF&CC);

S105, through the correlation of protein GO features, it is divided into seven parts to calculate the correlation of two proteins:

Among them, Cor( _xi ,K) represents the correlation between the GO annotation features represented by _xi and the Kth protein under this section.

where Sim _k represents the correlation between the Kth protein in the training set and our predicted protein.

After obtaining the correlation between all the proteins in the training set and the predicted protein, we extract ten highly correlated proteins in the training set as the weighted KNN method to obtain the probability value of the protein at each subcellular location :

Among them, num _a and num represent the number of proteins in the a-th subcellular position in the training set and the total number of proteins in the training set, respectively. And pro _a represents the probability that the predicted protein is in the ath subcellular position.

S106, use rps-blast to obtain the conserved domain features of all human proteins in the Swiss-Prot database, and calculate the correlation between features through the information difference:

in Denotes the entropy of the i-th CDD feature, Indicates the probability that the i-th CDD feature exists in the protein training set. Represents the differential entropy of the i-th feature and the j-th feature, represents the correlation between the i-th CDD feature and the j-th CDD feature.

Obtain the similarity matrix of the conserved domain features, and then use rps-blast to obtain the conserved domain features of the target protein to calculate the correlation between the two proteins, and extract ten highly correlated proteins from the training set as the weighted KNN method to obtain The probability value of the protein at each subcellular location;

S107, integrate the obtained sequence features, the probability features of the seven parts of GO, and the probability features of the conserved domains, use the Binary Relevance strategy to build 12 SVM classifiers to predict the subcellular location of the protein, and the probability of each subcellular location.

Example:

There is an input sequence with the following data:

>query protein 1；example of multiple subcellularlocationsMSAVGAATPYLHHPGDSHSGRVSFLGAQLPPEVAAMARLLGDLDRSTFRKLLKFVVSSLQGEDCREAVQRLGVSANLPEEQLGALLAGMHTLLQQALRLPPTSLKPDTFRDQLQELCIPQDLVGDLASVVFGSQRPLLDSVAQQQGAWLPHVADFRWRVDVAISTSALARSLQPSVLMQLKLSDGSAYRFEVPTAKFQELRYSVALVLKEMADLEKRCERRLQD

This is a sequence to be tested, and the software output results using the method of the present invention are as follows:

From the results, it can be seen that this method is effective and accurately predicts the subcellular location of this protein in humans.

The above embodiments do not limit the present invention in any way, and all technical solutions obtained by means of equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims (2)

1. A human protein subcellular position prediction method, based on human protein sequence prediction protein subcellular position, is characterized in that, comprising the following steps: Step 1: Use human protein sequence information to extract the statistical characteristics of residues of the full-length sequence, N-terminal and C-terminal protein sequence fragments, including amino acid composition characteristics and specificity scores obtained by using protein homology information Matrix features and normalize the features, and use Correlation-based Feature Selection, a supervised feature selection algorithm, to perform dimensionality reduction after combining these two features; Step 2: By extracting the GO features of all human proteins in the protein database, use GOSSTO to obtain three similarity matrices in the GO (BP, MF, CC) feature space; Step 3: search for homologous proteins in the Swiss-Prot database by the blast method, extract the GO features of the homologous proteins, and use the same method to obtain the GO features of the proteins in the training set; Step 4: Divide the three parts of the protein GO feature (BP, MF, CC) into 7 parts (BP, MF, CC), (BP&MF, BP&CC, MF&CC) through one-tuple, two-tuple and three-tuple ,(BP&MF&CC); Step 5: Through the correlation of protein GO features, divide it into seven parts to calculate the correlation of two proteins, and through parameter optimization, extract ten highly correlated proteins in the training set to do the weighted KNN method to obtain the The probability value of the protein at each subcellular location; Step 6: Obtain the conserved domain features of all human proteins in the Swiss-Prot database through rps-blast, and calculate the correlation between features through the information difference to obtain the similarity matrix of conserved domain features, and then obtain it through rps-blast The conserved domain features of the target protein are used to calculate the correlation between the two proteins, and through parameter optimization, ten highly correlated proteins are extracted from the training set to do the weighted KNN method to obtain the probability of the protein at each subcellular location value; Step 7: Merge the obtained sequence features, the probability features of the seven parts of GO, and the probability features of the conserved domain. Using the Binary Relevance strategy to build can predict the centrosome, cytoplasm, cytoskeleton, endoplasmic reticulum, endosome, secretory pathway, Golgi SVM classifiers for 12 subcellular locations of body, lysosome, mitochondria, nucleus, peroxisome and membrane. 2. A human protein subcellular position prediction method, based on human protein sequence prediction protein subcellular position, characterized in that, comprising the following steps: S101, using human protein sequence information to extract the amino acid composition characteristics of the full-length sequence, the first 10 to 60 at the N-terminus, and the first 10 to 100 at the C-terminus, and the normalized PSSM matrix features, and use CFS to reduce the dimension , where the formula for normalizing the PSSM matrix and converting it into a 20-dimensional feature in each part is: Where S _i,j represents the probability score of the amino acid appearing at the i-th (1≤i≤L) position of the sequence to evolve into the j-th (1≤j≤20) amino acid during the evolution process, and L represents the protein sequence length, ${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}}{\sqrt{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\underset{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$ S ⁰ _{i, j} represents the score of this specificity scoring matrix after normalization, N represents the number of amino acids, in formula (2), N=20, $\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$ in Indicates the value after adding and averaging the scores of each column; $\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; m</span>}}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \&lsqb;</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; ...</span><span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; 20</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}}<span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$ It is the PSSM matrix feature after normalization processing; S102, by extracting the GO features of all human proteins in the Swiss-Prot database, using GOSSTO to obtain three similarity matrices in the GO (BP, MF, CC) feature space; S103, search for homologous proteins in the Swiss-Prot database by the blast method, extract their GO features, and use the same method to obtain the GO features of the proteins in the training set; S104, divide the three parts (BP, MF, CC) of the protein GO feature into 7 parts (BP, MF, CC), (BP&MF, BP&CC, MF&CC), ( BP&MF&CC); S105, through the correlation of protein GO features, it is divided into seven parts to calculate the correlation of two proteins: $\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; x</span>}}\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$ Where Cor( _xi ,K) represents the correlation between the GO annotation features represented by _xi and the Kth protein under this section, ${\mathrm{<span\; class="notranslate">\; Sim</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; r</span>}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$ where Sim _k represents the correlation between the Kth protein in the training set and our predicted protein, After obtaining the correlation between all proteins in the training set and the predicted protein, extract ten highly correlated proteins in the training set as the weighted KNN method to obtain the probability value of the protein at each subcellular location: ${\mathrm{<span\; class="notranslate">\; pro</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}}}{\mathrm{<span\; class="notranslate">\; sim</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; num</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}}{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; m</span>}}}{{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}}{\mathrm{<span\; class="notranslate">\; sim</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$ Among them, num _a and num represent the number of proteins in the a-th subcellular position in the training set and the total number of proteins in the training set, respectively. And pro _a represents the probability that the predicted protein is in the ath subcellular position. S106, use rps-blast to obtain the conserved domain features of all human proteins in the Swiss-Prot database, and calculate the correlation between features through the information difference: $\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \}</span>}{<span\; class="notranslate">\; \&Sigma;</span>}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; log</span>}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$ ${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 11</span>}<span\; class="notranslate">\; )</span>$ Where H(f _i ^cdd ) represents the entropy of the i-th CDD feature, and p(f _i ^cdd =1) represents the probability that the i-th CDD feature exists in the protein training set. H(f _j ^cdd ,f _i ^cdd ) represents the differential entropy of the i-th feature and the j-th feature, S _i,j ^cdd represents the correlation between the i-th CDD feature and the j-th CDD feature, Obtain the similarity matrix of the conserved domain features, and then use rps-blast to obtain the conserved domain features of the target protein to calculate the correlation between the two proteins, and extract ten highly correlated proteins from the training set as the weighted KNN method to obtain The probability value of the protein at each subcellular location; S107, integrate the obtained sequence features, the probability features of the seven parts of GO, and the probability features of the conserved domains, use the BinaryRelevance strategy to build 12 SVM classifiers to predict the subcellular location of the protein, and the probability of each subcellular location.