A method and device for mining and screening functional proteins

Abstract

The present invention discloses a method and device for mining and screening functional proteins, which combines multi-dimensional bioinformatics analysis such as protein structure prediction, protein function annotation, and protein cluster mining with artificial intelligence methods based on protein big language, and constructs a process for mining and screening functional proteins. Protein structural analysis is introduced, and the association between protein structure and function is utilized to construct a structural similarity map to discover proteins with similar functions. The protein big language model can include characteristic parameters related to protein function such as protein structure, post-translational modification, and biophysical properties, and reveal the intrinsic functional relevance of proteins in the form of vector matrices. By introducing these two analysis methods and combining them with traditional sequence-based protein function annotation methods, proteins with lower sequence similarity and affinity but corresponding functions are mined and screened, so that biological macromolecules can be used more efficiently and accurately.

Description

Technical Field

[0001] The invention belongs to the field of bioinformatics and proteomics, and in particular relates to a method and device for mining and screening functional proteins. Background Art

[0002] With the upgrading of artificial intelligence technology and the development of high-throughput bioinformatics, the integration of artificial intelligence and life sciences has brought great innovation to the analysis of biological macromolecules. As an important macromolecule in biological systems, protein is the most basic component in living organisms. It participates in a large number of life activities and plays a huge role in many biochemical reactions. It can be said that proteomics analysis is the first step to understand biological functions and utilize biological products. As a biological macromolecule composed of 26 amino acids, the sequence, structure and function of protein are mapped to each other. The sequence analysis of proteins has become increasingly mature with the development of various technologies such as genomics, protein sequencing and computer technology. At the same time, in terms of protein structure, the introduction of deep learning protein three-dimensional structure prediction methods such as AlphaFold, RoseTTAFold and ESMFold can accurately and precisely analyze the structure of proteins. As an emerging development direction of artificial intelligence analysis, large language models can integrate multi-dimensional data information. Large language models based on proteomics (such as ESM, Protbert, ProtT5, etc.) parameterize the information closely related to proteomics and protein function (such as secondary and tertiary structures, physical and chemical properties, regulatory pathways, etc.), providing multi-faceted reference information for protein function annotation.

[0003] Traditional protein function mining methods based on sequence analysis rely on the conservation of sequences. For example, sequence alignment and comparison of sequence landmark regions such as functional domains and structural domains require a certain degree of sequence similarity. Therefore, there are defects in protein function mining in protein groups with lower similarity (such as species with distant kinship). Secondly, proteins with similar sequences do not have the same functions. Changes in a few sites in the sequence may bring about huge structural changes or even make them inactive. Therefore, proteins mined only based on sequence information also lack cross-validation results. Protein mining methods based on protein structure and protein large language models, on the basis of sequence analysis, introduce multi-dimensional protein properties such as structural similarity as the basis for screening and verification, and then verify the integrity of biochemical reaction pathways based on genomic analysis, so as to achieve mutual verification of analysis results from multiple angles, and also allow proteomic analysis to not be limited to sequences with high similarity. As a computational process that integrates multiple proteomic analyses and artificial intelligence analyses, the invention can quantitatively characterize protein functions from multiple angles, thereby effectively improving the accuracy of functional protein mining. Summary of the invention

[0004] The purpose of the present invention is to provide a method and device for mining and screening functional proteins in view of the deficiencies of the prior art, thereby solving the problem that mining of functional proteins is limited by sequence conservation and species affinity.

[0005] The object of the present invention is achieved through the following technical scheme: a method for mining and screening functional proteins, comprising:

[0006] Integrate several protein databases and obtain the protein species sequence number information mapping file to build a protein comparison database; compare the target functional protein reference sequence with the NR library for protein-protein sequence comparison, and output the first protein sequence aligned to the NR library and its species name and species sequence number;

[0007] Based on the functional domains and domain fragments of the reference sequence annotated in the protein functional domain and structural domain database, the functional domains and structural domains that are related to the target function shared by the reference sequence and whose fragment coverage of the sequence is greater than a first preset threshold are selected, and a second protein sequence covering the selected functional domains and structural domain fragments is screened from the first protein sequence;

[0008] Predicting the tertiary structures of the second protein and the reference protein and clustering them, and selecting proteins from the second protein that can be clustered into the same category as the reference protein structure as the first candidate protein;

[0009] Obtain a classification information file with the species sequence number as an index, obtain the classification sequence numbers of the species corresponding to the reference protein and the second protein based on the file, vectorize the second protein sequence and the reference sequence based on the protein large language model and merge them into a vector matrix, map the vector matrix to the classification sequence number, remove features in the vector matrix with eigenvalues ​​greater than or equal to a second preset threshold, and obtain a new vector matrix;

[0010] The new vector matrix is ​​clustered using a clustering model, and proteins with the same label as the reference protein are selected as the second candidate proteins;

[0011] Protein clusters of target functional pathways are mined from reference species, and candidate species are screened based on the protein clusters of target functional pathways. The candidate proteins corresponding to the screened candidate species are the target proteins.

[0012] Furthermore, a randomized decision tree algorithm is used to map the vector matrix to the classification sequence number.

[0013] Furthermore, the clustering model is selected and parameter optimized, including:

[0014] Three de novo clustering algorithms were selected for the clustering model: 1) density-based noise application spatial clustering, namely DBSCAN, 2) clustering by sorting points to determine cluster structure, namely OPTICS, and 3) hierarchical clustering based on Ward connection, namely WardHierarchical Clustering. The clustering parameters were optimized and evaluated respectively, including 1) eps value and minimum sample number in DBSCAN model function, 2) minimum sample number and xi value in OPTICS model function, 3) number of categories and distance threshold in connection function in Ward Hierarchical Clustering clustering model using Agglomerative Clustering algorithm; the reference protein and the second protein were taken as category 1, and the others as category 0 as the reference value matrix to evaluate the difference between the reference value and the clustering result; the clustering effect and model were evaluated by the homogeneity score, completeness score and V evaluation score between the reference value and the clustering result, as well as the silhouette coefficient score of the clustering result, and the parameter group and model with the best score were selected.

[0015] Furthermore, protein clusters of target functional pathways were mined from reference species, including:

[0016] First, the sequence of the target functional pathway protein is obtained from the proteome of the reference species and a comparison database is built for it, and then nucleic acid-protein sequence alignment is used; each genome sequence in the reference species is aligned to the database of the target functional pathway protein in turn, and the alignment output file is output in outfmt6 format, and contains the start and end sites of the query object, i.e., the genome scaffold, the coverage of the target, i.e., the functional pathway protein, and the translated genome alignment region sequence; the alignment results of each reference species are sorted according to the priority genome scaffold at the start site, and the genome scaffold with the most adjacent protein sequences mined on the same genome scaffold is selected, and then these protein sequences are built into a library, and compared with the CDS on the same genome scaffold for nucleic acid-protein sequence alignment, and the alignment results that can completely cover the target functional pathway protein and have the highest percentage of pident are selected, and the corresponding CDS translated sequence is output; the genome fragments containing these CDS are the protein clusters of the target functional pathway.

[0017] Furthermore, the genome sequence and proteome sequence are obtained, including: the species sequence number corresponds to the genome assembly sequence number of the genome assembly database on NCBI, and then the genome and its sequence, coding region sequence and proteome and its sequence are downloaded through the genome assembly sequence number.

[0018] Furthermore, candidate species are screened based on the protein clusters of the functional pathway, including:

[0019] Candidate species were screened according to any one or more of the following criteria:

[0020] Contains all necessary target functional pathway proteins, and they are distributed in one protein cluster;

[0021] Contains all necessary target functional pathway proteins, but distributed in non-adjacent locations in the same genome scaffold;

[0022] Contains all necessary target functional pathway proteins, but distributed in different locations in the genome.

[0023] Furthermore, if no candidate species are screened out when screening candidate species based on the protein cluster of the functional pathway, the sequence alignment of the protein is queried with the NR library position database based on the target functional pathway protein position, and the species list is restricted to the candidate species to obtain the species containing all the necessary target functional pathway proteins.

[0024] The present invention also provides a functional protein mining and screening device, comprising:

[0025] The NR library comparison module is used to integrate several protein databases and obtain the protein species sequence number information mapping file to build a protein comparison database; compare the target functional protein reference sequence with the NR library for protein-protein sequence comparison, and output the first protein sequence that is compared to the NR library and its species name and species sequence number;

[0026] The first screening module is used to annotate the functional domains and domain fragments of the reference sequence based on the protein functional domain and structural domain database, select the functional domains and structural domains that are related to the target function shared by the reference sequence and whose fragment coverage of the sequence is greater than a first preset threshold, and screen out a second protein sequence covering the selected functional domain and structural domain fragments from the first protein sequence;

[0027] The second screening module is used to predict the tertiary structures of the second protein and the reference protein and perform clustering, and select proteins that can be clustered into the same category as the reference protein structure from the second protein as the first candidate protein;

[0028] A vector matrix acquisition module is used to obtain a classification information file with the species sequence number as an index, obtain the classification sequence numbers of the species corresponding to the reference protein and the second protein based on the file, vectorize the second protein sequence and the reference sequence based on the protein large language model and merge them into a vector matrix, map the vector matrix to the classification sequence number, remove features in the vector matrix with eigenvalues ​​greater than or equal to a second preset threshold, and obtain a new vector matrix;

[0029] The third screening module is used to cluster the new vector matrix using a clustering model and select proteins with the same label as the reference protein as the second candidate protein;

[0030] The fourth screening module is used to mine protein clusters of target functional pathways from reference species, screen candidate species based on the protein clusters of target functional pathways, and the candidate proteins corresponding to the screened candidate species are the target proteins.

[0031] The present invention also provides a functional protein mining and screening device, comprising one or more processors, for implementing the above-mentioned functional protein mining and screening method.

[0032] The present invention also provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, it is used to implement the above-mentioned functional protein mining and screening method.

[0033] Compared with the prior art, the beneficial effects of the embodiments of the present invention are: the tertiary structure of proteins and the large language model of proteins do not rely on the conservation of protein sequences, can greatly expand the scope of natural functional protein mining, and can introduce information closely related to the function of proteins (such as physical and chemical properties, etc.) into the analysis of protein functional characterization, so that protein mining and screening are more accurate and efficient. However, the tertiary structure of proteins is expressed in atomic spatial position information, and the large language model of proteins is presented in the form of mathematical vectors. The interpretability of the two is much worse than that of sequences, and therefore their application scenarios are subject to certain restrictions. The invention provides an innovative calculation process, which uses machine learning techniques such as clustering algorithms to perform multi-angle combined quantitative analysis based on the protein functional characterization results of three-dimensional structures and large language models, and then uses genomics-based protein clusters and evolutionary tree analysis as a reference to achieve efficient screening of specific functional proteins, thereby improving the efficiency of using natural biological products in industrial production, pharmaceutical research and development, and environmental protection, and further promoting the development of the bio-industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0035] Figure 1 The flowchart of the screening based on protein structure clustering.

[0036] Figure 2 Characterizing a flow chart for large protein-based language models.

[0037] Figure 3 Schematic diagram of the distribution of protein feature vectors.

[0038] Figure 4 Flowchart of the model and parameter optimization approach for protein vector clustering.

[0039] Figure 5 Flowchart of the screening process based on protein vector clustering.

[0040] Figure 6 This is an example diagram of protein clusters for functional pathways.

[0041] Figure 7 Flowchart for protein cluster mining.

[0042] Figure 8 The figure shows the process and example diagram of the construction method of protein structure similarity map.

[0043] Fig. 9 The flowchart and example diagram of the method for constructing evolutionary relationships of candidate proteins.

[0044] Fig.10 A flow chart of a functional protein mining and screening method.

[0045] Fig.11 This is the spatial distribution of the characterization vectors of luciferase candidate proteins and background proteins.

[0046] Fig.12 Characterize the distribution of vectors for proteins that pass the clustering filter.

[0047] Fig.13 A schematic diagram of a hardware structure provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0049] In a method for mining and screening functional proteins of the present invention, a reference protein and its sequence (i.e., a target functional protein reference sequence, referred to as a reference sequence) are known, such as Fig.10 As shown, the following steps are included:

[0050] (1) Integrate several protein databases and obtain the protein species sequence number information mapping file to build a protein comparison database; compare the target functional protein reference sequence with the NR library for protein-protein sequence comparison, and output the first protein sequence aligned to the NR library and its species name and species sequence number;

[0051] In one embodiment, protein mining based on a non-redundant protein database includes: Non-redundant Protein Database (link https: / / ftp.ncbi.nlm.nih.gov / blast / db / FASTA / nr.gz) is a standard protein database maintained by the National Center for Biotechnology Information (NCBI), which contains basically all protein sequence information recorded in the data center, and integrates other major protein databases such as SwissProt, PIR and PDB, etc. It is a mainstream database containing protein information of all species. Therefore, the results obtained by mining based on this database can cover a relatively wide range of sequence space and species. In order to be able to associate species information when establishing a protein comparison database, it is also necessary to Get the protein species sequence number information mapping file. After getting the corresponding file, construct a protein comparison database (NR library for short) for its sequence. Compare the target functional protein reference sequence (can be one or more) with the NR library for protein-protein sequence comparison (i.e. blastp comparison method). When comparing, limit the target species category (such as animals, plants, etc., including the species source of the reference sequence) to improve the comparison efficiency. The comparison threshold is evalue less than or equal to 0.0001, the comparison mode is very sensitive, and the number of targets is not limited to capture proteins with distant sequence similarity. The output selection of the comparison results is in the form of a table, and the protein sequence and its species name and species sequence number that are compared to the NR library are output at the same time.

[0052] In one embodiment, in order to better screen the target protein, it is necessary to obtain the classification information and evolutionary relationship of its species. The corresponding files can be obtained through the database NCBI Taxonomy Database (https: / / ftp.ncbi.nlm.nih.gov / pub / taxonomy / new_taxdump / new_taxdump.tar.gz): 1) classification information fullnamelineage.dmp indexed by species name, 2) classification information taxidlineage.dmp indexed by species sequence number, which is used for subsequent analysis processes. The species sequence number of the NR library comparison result can be mapped to its species classification information. Considering that one species can correspond to multiple comparison results, the comparison results are sorted with the species sequence number as the index.

[0053] (2) Based on the functional domains and domain fragments of the reference sequence annotated in the protein functional domain and structural domain database, the functional domains and structural domains that are related to the target function shared by the reference sequence and whose fragment coverage of the sequence is greater than a first preset threshold are selected, and a second protein sequence covering the selected functional domains and structural domain fragments is screened from the first protein sequence;

[0054] In one embodiment, by characterizing the functional domains and structural domains on the protein sequence, it can be preliminarily determined whether the protein has a characteristic fragment expressing the target function. First, based on the functional and structural domain databases of protein functional domains such as Pfam, Prosite, CDD, and CATH-3d, the functional and structural domain fragments of the reference sequence are annotated, and the functional domains and structural domains that are related to the target function shared by the reference sequence and whose fragment coverage of the sequence is greater than the first preset threshold (such as 0.3) are selected. Then, the functional domain and structural domain annotation tool is used to scan the sequence mined by the NR library comparison to identify whether it has the above-mentioned characteristic functional domains or structural domain fragments, and the sequences that can cover the characteristic fragments are screened out for subsequent analysis.

[0055] (3) Predicting the tertiary structures of the second protein and the reference protein and clustering them, and selecting the protein from the second protein that can be clustered into the same category as the reference protein structure as the first candidate protein;

[0056] In one embodiment, in order to compare the structural similarity between the second protein and the reference protein, it is necessary to obtain the tertiary structure of the protein. Currently, computational tools based on artificial intelligence methods can already make relatively accurate predictions. In this embodiment, a tertiary structure prediction tool based on a large protein language model is used to predict the tertiary structure of the second protein and the reference protein, and the three-dimensional atomic structure of the protein is identified by the structural domain and functional domain. Each sequence is output as a corresponding PDB format file. In order to screen proteins with high structural consistency with the reference sequence, the tertiary structures of the second protein and the reference protein are clustered, and protein sequences that can be clustered into one category with the reference structure are selected for subsequent evaluation, such as Figure 1 shown.

[0057] Specifically, all PDB structure files of all sequences were placed in the same folder to cluster all structures. Set-Cover was selected as the cluster mode. The purpose of clustering was to distinguish reference sequences from other sequences as much as possible, such as clustering all (or most) structures into one or a few categories. Therefore, multiple sets of parameters needed to be tried for clustering. The parameter selection range was coverage between 0.75 and 0.9, tmscore threshold above 0.4, and minimum sequence identity between 0.6 and 0.9. Proteins that could be clustered with the reference structure were selected as the first candidate proteins.

[0058] (4) Obtaining a classification information file with the species sequence number as an index, obtaining the classification sequence numbers of the species corresponding to the reference protein and the second protein based on the file, vectorizing the sequences of the second protein and the reference protein based on the protein large language model and merging them into a vector matrix, mapping the vector matrix to the classification sequence number, removing features whose eigenvalues ​​are greater than or equal to a second preset threshold in the vector matrix, and obtaining a new vector matrix;

[0059] In one embodiment, in order to better evaluate the screening effect based on the protein large language model, random protein sequences are required as background and control. Here, you can choose to randomly select proteins from a non-redundant protein database, and the number of selected proteins can be between 0.5-2 times the second protein. The source of the background protein sequence needs to cover the species classification under the Superkingdom of the target species (for example, if the target species is mammals, the background protein needs to cover three Kingdom levels in eukaryotes, including green plants Viridiplantae, animals Metazoa, and fungi Fungi), and the functional annotation of the background protein is also not restricted.

[0060] In order to facilitate the machine learning analysis of the second protein, the second protein sequence is converted into a vector that can represent its structure, physicochemical characteristics and other properties. Here, the protein language model is used to characterize the second protein and output it as a 1024-dimensional protein vector, such as Figure 2 As shown. The sequences of the reference protein, the second protein, and the background protein are divided into different files, and their sequences are input into the large language model in turn to calculate the protein vectorization results. In this process, the backend for computing embeddings is selected as protbert_bdf, the pool mode is mean, the batch size is 1, and each sequence is output as a 1024-dimensional representation vector. The calculation results of each batch of sequences are output in the form of pickle file compression.

[0061] Principal component analysis (PCA) can be used to preliminarily observe the distribution of protein vectors, so as to determine whether proteins that meet the conditions can be screened by clustering the feature vectors, such as Figure 3 As shown. Here, the result of vectorization through the protein large language model, that is, the pickle file, is merged into the first matrix, the format of the vector in the matrix is ​​converted into double format, and the serial numbers corresponding to the reference protein, the second protein and the background protein are recorded. The PCA algorithm is used to convert the first matrix into a new 2-dimensional vector matrix composed of two principal components. Then the KMeans de novo clustering algorithm is called to cluster the first matrix and the new matrix respectively. The init method selects k-means++, the number of init is 4, and the number of categories can be selected from multiple groups of values ​​between 6-12 to determine whether the reference protein and the second protein can be distinguished from the background protein. Here, a scatter plot can be drawn with the two principal component values ​​of the new matrix as the horizontal and vertical coordinate values ​​to determine whether the second protein is spatially distinguished from other background proteins, which is used to verify whether the method of characterizing vectors based on the protein large language model is feasible.

[0062] In one embodiment, the species evolution process has an impact on the protein sequence, and therefore can cause certain interference to the clustering results of its vector. In order to better screen out proteins with similar functions, it is necessary to evaluate and filter the impact of species classification on the characterization vector matrix. Take out the characterization vector matrix of the reference protein and the second protein, and obtain the classification sequence number of the subclass level of the species corresponding to the reference protein and the second protein through the classification information file taxidlineage.dmp indexed by the species sequence number as the grouping information. Here, the randomized decision tree algorithm is used to map the vector matrix to the classification sequence number, and the feature value with a large correlation with the species classification in the vector matrix can be found. Select the features with feature weights (feature importances) greater than or equal to the second preset threshold (such as 0.001) as the features with a large correlation with the species classification, and calculate and record their total value as the size of the species classification impact. After removing these features, output the new vector matrix and perform PCA analysis on it again to observe the impact of removing this part of the features on the distribution of the vector matrix.

[0063] (5) Use the clustering model to cluster the new vector matrix and select proteins with the same label as the reference protein as the second candidate protein;

[0064] In one embodiment, in order to determine the clustering effect of the machine learning method on the vector matrix, and to select appropriate parameters and models for clustering and screening, it is necessary to use different clustering algorithms and parameters on the matrix for testing, and then quantitatively evaluate the clustering effect, such as Figure 4 As shown. Here, three de novo clustering algorithms are selected for the clustering model, including 1) density-based spatial clustering with noise (DBSCAN), 2) point sorting to determine cluster structure clustering (OPTICS), and 3) Ward Hierarchical Clustering, and their clustering parameters are optimized and evaluated respectively, including 1) eps value and minimum number of samples (minimum samples) in DBSCAN model function, 2) minimum number of samples (minimum samples) and xi value in OPTICS model function, 3) number of categories and distance threshold in Ward Hierarchical Clustering clustering model using Agglomerative Clustering algorithm, and kneighbors graph. In order to have enough protein samples for evaluating model and parameter performance, the reference protein and the second protein are used as category 1, and the others are used as category 0 as the reference value matrix to evaluate the difference between the reference value and the clustering result.

[0065] The clustering effect and model were evaluated by the homogeneity score (Homogeneity Score), completeness score (Completeness Score) and V measure score (V Measure Score) between the reference value and the clustering result, as well as the silhouette score (Silhouette Score) of the clustering result. The parameters of the three clustering models were tested for their clustering effect on the protein vector matrix by continuous equidistant sampling within a reasonable parameter range, and the above clustering coefficient scores were calculated and recorded. The parameter group and model that can get the best score were selected so that: 1) the homogeneity score is close to 1; 2) the completeness score is close to 1; 3) the V measure score is close to 1; 4) the silhouette coefficient score is close to 0, which is used as the final parameter and model for clustering the protein vector matrix.

[0066] Based on the optimized model and parameter set, the vector matrix represented by the large language model of proteins is clustered de novo to screen proteins with high similarity to the representation vector of the reference protein, such as Figure 5 First, after clustering is completed, the label value and number of labels of the reference protein after clustering are recorded, and the second protein with the same label as the reference protein is screened out as the second candidate protein.

[0067] The label values ​​of the second protein screened by the protein tertiary structure and their overlap range with the label values ​​of the reference protein were recorded, and the consistency of the screening results of the two methods was evaluated using the homogeneity scoring function. The vector matrices of the screened proteins and the reference proteins were separated, and the PCA analysis of the three principal components was performed (n_components=3), and a three-dimensional scatter distribution diagram with its three principal components as the coordinate axes was drawn. The spatial distribution of the cluster labels on the protein vector was observed to evaluate the effect of the second candidate protein screening.

[0068] (6) Mining protein clusters of the target functional pathway from the reference species, screening candidate species based on the protein clusters of the target functional pathway, and the candidate proteins corresponding to the screened candidate species are the target proteins. The reference species is the species to which the reference protein belongs, and the candidate species is the species to which the candidate protein belongs; in this application, the candidate proteins include the first candidate protein and the second candidate protein;

[0069] In one embodiment, mining protein clusters of target functional pathways from reference species includes:

[0070] Acquisition of genome and proteome sequences: The synthesis or regulatory pathways that perform biological functions usually require the participation of multiple proteins or biological molecules. The higher the integrity of the pathway, the greater the possibility that the protein will perform the corresponding function, such as Figure 6 As shown. In order to mine protein clusters of protein expression pathways, it is necessary to obtain the genome of the corresponding species, that is, to obtain the genome assembly sequence number of the genome assembly database (Assembly) on NCBI corresponding to the species sequence number by referring to the sequence number of the species to which the candidate protein belongs. Then, the genome and its sequence (genome), coding region sequence (CDS) and proteome and its sequence (protein) are downloaded through the genome assembly sequence number. The sequence information of each species is packaged in a compressed file.

[0071] The sequence of the target functional pathway protein is obtained from the proteome of the reference species and a comparison database is built for it. Then, each genome sequence in the reference species is aligned to the database of the target functional pathway protein in turn using nucleic acid-protein sequence alignment (alignment condition evalue is less than or equal to 0.001, and the alignment mode is very-sensitive). The alignment output file is output in outfmt6 format and contains the query object, i.e. the start site (query start) and the end site (query end) of the genome scaffold, the coverage of the target, i.e. the target functional pathway protein, and the translated genome alignment region sequence (translated query sequence). The alignment results of each reference species are sorted according to the priority genome scaffold and then the starting site. The genome scaffold with the most adjacent protein sequences mined on the same genome scaffold is selected, and then these protein sequences are built into a library and compared with the CDS on the same genome scaffold for nucleic acid-protein sequence alignment (alignment condition evalue is less than or equal to 0.00001). The alignment results that can completely cover the target functional pathway protein and completely match, that is, the percentage pident is the highest, are selected, and the corresponding CDS translated sequence is output. The genomic fragments containing these CDS are the protein clusters of the target functional pathway, and their positions and orders in the corresponding protein clusters are annotated based on the position information of the CDS. The process is as follows: Figure 7 shown.

[0072] In one embodiment, candidate species are screened based on protein clusters of target functional pathways, including:

[0073] Generally, if the functional pathway has a relatively high expression efficiency, its proteins are more likely to form protein clusters in order to better perform the transcription and translation processes. Therefore, the function and efficiency of the corresponding pathway of the candidate species can be inferred by evaluating the existence of protein clusters. After obtaining the protein cluster of the target functional pathway, the candidate species can be screened in the following order: 1) containing all the necessary target pathway proteins, and they are distributed in a protein cluster, 2) containing all the necessary target pathway proteins, but distributed in non-adjacent positions in the same genome scaffold, 3) containing all the necessary target pathway proteins, but distributed in different positions in the genome. If no candidate species passes the above three criteria, some species may not be fully mined due to incomplete genome sequencing and assembly. Therefore, when querying the sequence alignment of the protein with the target pathway protein position in the NR library position database, the species list (taxon list) is restricted to the sequence number of the candidate species, the alignment condition evalue is less than or equal to 0.001, and the very sensitive alignment mode is selected, and the species name and sequence number containing all the necessary target pathway proteins after alignment are output.

[0074] It should be noted that the protein sequence in the present application is the amino acid sequence of the protein; by obtaining a protein sequence, the corresponding protein can be obtained.

[0075] In one embodiment, protein information integration and evaluation are also included

[0076] 7.1 Protein structure comparison

[0077] After obtaining the tertiary structure of the candidate protein, the similarity of the candidate protein structure is compared by structural alignment, and the similarity parameters between all the aligned structures are obtained. The PDB structure files of all candidate proteins are placed in the same folder and the protein structure library and index are built in the folder. After the structure library is built, the structure in the folder is compared with the built structure library (the alignment method is 3Di+AA, and the tm score threshold tmscorethreshold is 0.4). The output file format is BLAST-TAB table format and includes alignment tm score (alntmscore), structure consistency score (fident) and structure identity probability score (prob).

[0078] 7.2 Construction of protein structure similarity map

[0079] Based on the structural similarity maps of the candidate proteins that have passed the above screening, proteins with representative tertiary structures are screened out, such as Figure 8As shown. First, based on the results of the pairwise structural alignment of proteins in step 7.1 (and limiting the minimum length of the sequence of the aligned structure to alnlen to 50, and the structural consistency score fident to be greater than or equal to 0.8), the candidate proteins after the alignment results are taken as nodes, and the fident of the structural alignment results of a pair of candidate proteins are taken as edges. The distance of the edges is calculated according to 1 / fident * 10, and a structural similarity map is constructed for the candidate proteins. The node colors are grouped into label values ​​based on the vector represented by the large language model (from step 5). By observing the distribution of node connections and their label distribution in the map, the difference between the two methods of protein screening based on structure and screening based on the representation of the large language model can be judged. For example, the nodes that are clustered are similar, and the labels also represent the mutual verification between the two methods. Secondly, the nodes that can be clustered in the protein map also represent the stability and conservation of the structure during the evolution process, and representative proteins can be screened from them.

[0080] 7.3 Cluster analysis based on protein sequence similarity

[0081] By comparing the similarity of candidate protein sequences and cross-validating with structural similarity, protein sequences are aligned and clustered. The sequences of the selected candidate proteins are placed in a file, and sequence clustering analysis is performed based on the results of pairwise alignment of the sequences. The proteins that can be clustered into one category meet the following criteria: 1) the global identity threshold of the aligned sequences is 0.8, and 1) the alignment coverage threshold for the shorter sequence in the pairwise alignment is 0.75. After clustering, the protein sequences are created into Fasta files according to the categories to which they belong, and the categories are used as labels to map to the nodes of the protein structure similarity map constructed in step 7.2. By observing and comparing the clusters formed by the protein nodes in the map and their label classification, it can be used to compare the relationship between the candidate protein structure and sequence similarity, and to assist in screening out functional proteins with low sequence similarity but characteristic structures.

[0082] 7.4 Protein evolution analysis and construction of evolutionary tree

[0083] By constructing an evolutionary tree of candidate proteins and other proteins in their pathways and comparing species classification relationships, the evolutionary process of functional pathways can be analyzed, providing a biological theoretical basis for the screening of functional proteins. Fig. 9As shown in the figure, according to the results of protein function annotation, the protein sequences on the protein cluster mined in step 6 are divided into different Fasta files, the candidate proteins are also put into the same Fasta file, and the proteins on these pathways are aligned in turn, and the aligned sequence files are output. Then these sequence files are sequentially constructed based on the maximum likelihood method (MaximumLikelihood) evolutionary tree, and the corresponding maximum likelihood method-based evolutionary distance file (MaximumLikelihood Distance) and evolutionary tree file (treefile) are obtained. After that, the evolutionary tree file is visualized and a tree map is constructed. A heatmap is drawn with the evolutionary distance as the value, so as to visualize the distance between the pathway proteins of each pair of species. By comparing the evolutionary tree map and the evolutionary distance heatmap, the evolutionary process of the functional pathway during the species differentiation process and the conservation of its function can be inferred, which is used to provide evolutionary biological information for protein screening.

[0084] Embodiment 1:

[0085] In order to verify the actual effect of this scheme, the luciferase protein (Luciferase) in the fungal luminescence pathway (Fungal Bioluminescence Pathway, abbreviated as FBP) was used as an example test. The 10 reference luciferase proteins (from 10 different fungal species) reported by the reference literature (Building customizable auto-luminescent luciferase-based reporters in plants) on the study of fungal luminescence mechanism were obtained, and their protein sequences and the genome sequences of the species to which they belonged were obtained from the NCBI database. Based on this information, the process was used to mine more luciferase proteins of fungal species.

[0086] The mining and screening process is briefly described as follows. The reference genome was aligned to the non-redundant protein database and the protein source was restricted to fungi. 2861 first proteins with certain sequence similarity were obtained. By scanning the domain and function domain of the first protein sequence, 2823 second proteins were obtained. The 2823 second proteins came from 1633 species containing the key functional domain PTHR38695. First, the 2823 second proteins were structurally screened. ESMFold was used to predict the tertiary structure of the 2823 second proteins and reference proteins. Among them, 2757 proteins of the 2823 second proteins had successfully resolved structures, and the reference proteins were successfully resolved. After structural clustering, 33 first candidate proteins (from 32 newly mined species) were obtained. Secondly, protein screening based on the protein large language model was performed. The Protbert_bdf large language model was used to characterize the 2823 second proteins and obtain 2773 representation vectors. Then 1728 background proteins were randomly obtained and characterized using the same large language model to obtain their characterization results. The reference proteins were characterized using the same large language model to obtain their characterization results. PCA analysis of the candidate proteins and the reference proteins showed that the vectors of the candidate proteins could be clustered in a relatively concentrated area, such as Fig.11 As shown. The species relationship analysis of the candidate protein vectors found 382 features with high correlation with species classification (contributing a total of 53.85% of the correlation). After removing these features, the filtered vector matrix was used for clustering analysis. First, multiple clustering models and continuous clustering tests of multiple groups of parameters were performed. It was found that Ward Hierarchical Clustering had the best performance, and the most suitable parameters were n_clusters=36 and n_neighbors=6. Based on this model and parameters, 309 second candidate proteins (from 291 newly mined species) were clustered for the filtered vector matrix, as shown Fig.12 As shown. The luminescence pathway proteins of the newly discovered species were mined, and it was found that 117 species had relatively complete luminescence pathway proteins on their genomes, of which 33 species (corresponding to 38 candidate proteins) could mine protein clusters of the luminescence pathway on their genomes. Through this protein mining and screening process, 31 new luciferase proteins were discovered in fungi, of which only 33.71% had the lowest consistency with the reference protein sequence, verifying that this process can mine functional proteins with low species affinity and low protein sequence conservation.

[0087] The present invention also provides a functional protein mining and screening device, comprising:

[0088] The NR library comparison module is used to integrate several protein databases and obtain the protein species sequence number information mapping file to build a protein comparison database; compare the target functional protein reference sequence with the NR library for protein-protein sequence comparison, and output the first protein sequence that is compared to the NR library and its species name and species sequence number;

[0089] The first screening module is used to annotate the functional domains and domain fragments of the reference sequence based on the protein functional domain and structural domain database, select the functional domains and structural domains that are related to the target function shared by the reference sequence and whose fragment coverage of the sequence is greater than a first preset threshold, and screen out a second protein sequence covering the selected functional domain and structural domain fragments from the first protein sequence;

[0090] The second screening module is used to predict the tertiary structures of the second protein and the reference protein and perform clustering, and select proteins that can be clustered into the same category as the reference protein structure from the second protein as the first candidate protein;

[0091] A vector matrix acquisition module is used to obtain a classification information file with the species sequence number as an index, obtain the classification sequence numbers of the species corresponding to the reference protein and the second protein based on the file, vectorize the second protein sequence and the reference sequence based on the protein large language model and merge them into a vector matrix, map the vector matrix to the classification sequence number, remove features in the vector matrix with eigenvalues ​​greater than or equal to a second preset threshold, and obtain a new vector matrix;

[0092] The third screening module is used to cluster the new vector matrix using a clustering model and select proteins with the same label as the reference protein as the second candidate protein;

[0093] The fourth screening module is used to mine protein clusters of target functional pathways from reference species, screen candidate species based on the protein clusters of target functional pathways, and the candidate proteins corresponding to the screened candidate species are the target proteins.

[0094] It should be noted that the device embodiment shown in this embodiment matches the content of the above method embodiment, and the content of the above method embodiment can be referred to, which will not be repeated here.

[0095] Corresponding to the aforementioned embodiment of a method for mining and screening functional proteins, the present invention also provides an embodiment of a device for mining and screening functional proteins.

[0096] See also Fig.13 A functional protein mining and screening device provided in an embodiment of the present invention includes one or more processors for implementing a functional protein mining and screening method in the above embodiment.

[0097] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0098] An embodiment of a functional protein mining and screening device of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, the processor of any device with data processing capabilities reads the corresponding computer program instructions in the non-volatile memory into the memory and runs them. From the hardware level, if Fig.13 As shown, it is a hardware structure diagram of any device with data processing capability in which a functional protein mining and screening device of the present invention is located. Fig.13 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiments is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0099] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0100] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0101] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon, and when the program is executed by a processor, a functional protein mining and screening method in the above embodiment is implemented.

[0102] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capability and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0103] It should be noted that the device embodiment shown in this embodiment matches the content of the above method embodiment, and the content of the above method embodiment can be referred to, which will not be repeated here.

[0104] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A method for mining and screening functional proteins, characterized in that: include: Integrate several protein databases and obtain protein species sequence number information mapping files to build a protein comparison database; Perform protein-protein sequence comparison between the target functional protein reference sequence and the protein comparison database, and output the first protein sequence and its species name and species sequence number that are compared to the protein comparison database; the reference protein and its sequence are known, and the reference protein sequence is the target functional protein reference sequence; Based on the functional domains and domain fragments of the reference sequence annotated in the protein functional domain and structural domain database, the functional domains and structural domains that are related to the target function shared by the reference sequence and whose fragment coverage of the sequence is greater than a first preset threshold are selected, and a second protein sequence covering the selected functional domain and structural domain fragments is screened from the first protein sequence, thereby obtaining a second protein; Predicting the tertiary structures of the second protein and the reference protein and clustering them, and selecting proteins from the second protein that can be clustered into the same category as the reference protein structure as the first candidate protein; Obtain a classification information file with the species sequence number as an index, obtain the classification sequence numbers of the species corresponding to the reference protein and the second protein based on the file, vectorize the second protein sequence and the reference sequence based on the protein large language model and merge them into a vector matrix, map the vector matrix to the classification sequence number, remove features in the vector matrix with eigenvalues ​​greater than or equal to a second preset threshold, and obtain a new vector matrix; The new vector matrix is ​​clustered using a clustering model, and proteins with the same label as the reference protein are selected as the second candidate proteins; Protein clusters of target functional pathways are mined from reference species, and candidate species are screened based on the protein clusters of target functional pathways. The candidate proteins corresponding to the screened candidate species are the target proteins.

2. The method according to claim 1, characterized in that Use the randomized decision tree algorithm to map the vector matrix to the classification sequence number.

3. The method according to claim 1, characterized in that The clustering model is selected and parameter optimized, including: Three de novo clustering algorithms were selected for the clustering model: 1) density-based noise application spatial clustering, namely DBSCAN, 2) clustering by sorting points to determine cluster structure, namely OPTICS, and 3) hierarchical clustering based on Ward connection, namely WardHierarchical Clustering. The clustering parameters were optimized and evaluated respectively, including 1) eps value and minimum sample number in DBSCAN model function, 2) minimum sample number and xi value in OPTICS model function, 3) number of categories and distance threshold in connection function in Ward Hierarchical Clustering clustering model using Agglomerative Clustering algorithm; the reference protein and the second protein were taken as category 1, and the others as category 0 as the reference value matrix to evaluate the difference between the reference value and the clustering result; the clustering effect and model were evaluated by the homogeneity score, completeness score and V evaluation score between the reference value and the clustering result, as well as the silhouette coefficient score of the clustering result, and the parameter group and model with the best score were selected.

4. The method according to claim 1, characterized in that: Mining protein clusters of target functional pathways from reference species, including: First, the sequence of the target functional pathway protein is obtained from the proteome of the reference species and a comparison database is built for it, and then nucleic acid-protein sequence alignment is used; each genome sequence in the reference species is aligned to the database of the target functional pathway protein in turn, and the alignment output file is output in outfmt6 format, and contains the start and end sites of the query object, i.e., the genome scaffold, the coverage of the target, i.e., the functional pathway protein, and the translated genome alignment region sequence; the alignment results of each reference species are sorted according to the priority genome scaffold at the start site, and the genome scaffold with the most adjacent protein sequences mined on the same genome scaffold is selected, and then these protein sequences are built into a library, and compared with the CDS on the same genome scaffold for nucleic acid-protein sequence alignment, and the alignment results that can completely cover the target functional pathway protein and have the highest percentage of pident are selected, and the corresponding CDS translated sequence is output; the genome fragments containing these CDS are the protein clusters of the target functional pathway.

5. The method according to claim 4, characterized in that The acquisition of genome sequence and proteome sequence includes: the species sequence number corresponds to the genome assembly sequence number of the genome assembly database on NCBI, and then the genome and its sequence, coding region sequence and proteome and its sequence are downloaded through the genome assembly sequence number.

6. The method according to claim 1, characterized in that Candidate species are screened based on protein clusters in this functional pathway, including: Candidate species were screened according to any one or more of the following criteria: Contains all necessary target functional pathway proteins, and they are distributed in one protein cluster; Contains all necessary target functional pathway proteins, but distributed in non-adjacent locations in the same genome scaffold; Contains all necessary target functional pathway proteins, but distributed in different locations in the genome.

7. The method according to claim 1, characterized in that If no candidate species are screened out when screening candidate species based on the protein cluster of the functional pathway, the protein alignment database is used to query the sequence alignment of the target functional pathway protein, and the species list is restricted to the candidate species to obtain species containing all the necessary target functional pathway proteins.

8. A functional protein mining and screening device, characterized in that: include: The protein comparison database comparison module is used to integrate several protein databases and obtain the protein species sequence number information mapping file to build a protein comparison database; compare the target functional protein reference sequence with the protein comparison database on a protein-protein basis, and output the first protein sequence and its species name and species sequence number that are compared to the protein comparison database; the reference protein and its sequence are known, and the reference protein sequence is the target functional protein reference sequence; A first screening module is used to annotate the functional domains and domain fragments of the reference sequence based on the protein functional domain and structural domain database, select the functional domains and structural domains that are related to the target function shared by the reference sequence and whose fragment coverage of the sequence is greater than a first preset threshold, and screen out a second protein sequence covering the selected functional domain and structural domain fragments from the first protein sequence, thereby obtaining a second protein; The second screening module is used to predict the tertiary structures of the second protein and the reference protein and perform clustering, and select proteins that can be clustered into the same category as the reference protein structure from the second protein as the first candidate protein; A vector matrix acquisition module is used to obtain a classification information file with the species sequence number as an index, obtain the classification sequence numbers of the species corresponding to the reference protein and the second protein based on the file, vectorize the second protein sequence and the reference sequence based on the protein large language model and merge them into a vector matrix, map the vector matrix to the classification sequence number, remove features in the vector matrix with eigenvalues ​​greater than or equal to a second preset threshold, and obtain a new vector matrix; The third screening module is used to cluster the new vector matrix using a clustering model and select proteins with the same label as the reference protein as the second candidate protein; The fourth screening module is used to mine protein clusters of target functional pathways from reference species, screen candidate species based on the protein clusters of target functional pathways, and the candidate proteins corresponding to the screened candidate species are the target proteins.

9. A functional protein mining and screening device, characterized in that: The method comprises one or more processors for implementing a functional protein mining and screening method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, it is used to implement a functional protein mining and screening method according to any one of claims 1 to 7.

Images