Protein function prediction method and system based on Contig perception

Abstract

The invention relates to a Contig perception-based protein function prediction method and system. The method comprises the following steps: obtaining a protein amino acid sequence, a nucleotide sequence corresponding to Contig and arrangement information of CDS on the Contig; splicing the protein-level features corresponding to the CDS sequence on the same Contig with the k-mer frequency vector of the Contig nucleotide sequence to generate enhanced protein-level features; setting the length of a sliding window, acquiring a plurality of fragments with fixed window lengths, which have protein function tags at the central positions of enhanced protein-level features according to a CDS sequence on the same Contig, taking one fragment as a sample, and training a function prediction network consisting of a bidirectional long-short-term memory network BiLSTM and a multilayer perceptron, and obtaining a target protein prediction probability value of each fragment. And the accuracy of protein function annotation is obviously improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of biological information engineering, and specifically relates to a protein function prediction method and system based on Contig perception, which realizes protein function prediction through multi-level feature fusion. BACKGROUND

[0002] Proteins perform key biological processes including immune response, signal transduction, and cell communication. Understanding their functions helps elucidate disease mechanisms, develop therapeutic methods, and drive the development of precision medicine. The Cluster of Orthologous Groups (COG) database classifies protein functions based on evolutionary conservation, dividing them into 26 functional categories such as transcription, cell motility, and cytoskeleton. Although the COG database provides a standardized classification system, its coverage is still limited. For example, UniProtKB contains over 200 million protein sequences, of which less than 0.1% have COG annotations. This limitation is mainly due to its reliance on laboratory-verified prokaryotic data. High-throughput sequencing technology has further exacerbated this imbalance, as the rate of newly discovered protein sequences far exceeds the progress of functional annotation. This disparity highlights the need for efficient and accurate computational methods to predict protein COG functions using existing biological data.

[0003] The current method with the highest accuracy for predicting protein function is the structure-based protein function prediction method, such as DeepFRI, which combines laboratory-resolved protein structures with sequence embedding features to construct a protein graph and process it using a graph convolution network (GCN). GAT-GO uses predicted protein structures and combines sequence features from the ESM-1b protein pre-training model to construct a protein graph and process it using a graph attention network (GAT). The ESM-2 protein pre-training model is the latest improved version of ESM-1b, with better protein sequence feature expression and a more lightweight feature. Graph isomorphism network (GIN) has been proven to be particularly effective in graph classification tasks and is suitable for protein function prediction tasks. In the above protein function prediction methods, only single protein features are analyzed, without considering protein-protein relationships, and the accuracy of protein function annotation prediction needs to be improved. In addition, existing technologies lack a protein-protein relationship that can be applied to all annotation systems. Therefore, it is necessary to explore a protein-protein relationship that can be applied to all annotation systems to promote more accurate protein function annotation. SUMMARY

[0004] To address the deficiencies of the prior art, the present application aims to solve the technical problem of providing a protein function prediction method and system based on Contig perception.

[0005] The technical problem solved by the present application is solved by the following technical solution:

[0006] In a first aspect, the present application provides a Contig-aware protein function prediction method, which comprises the following steps:

[0007] obtaining the protein amino acid sequence, the corresponding Contig nucleotide sequence and the arrangement information of CDS on the Contig; obtaining the sequence feature of the protein using a protein pre-training model according to the protein amino acid sequence, and predicting the structure of the protein using a protein structure prediction model; representing the Contig nucleotide sequence as a k-mer frequency vector;

[0008] constructing a protein graph using the sequence feature of the protein obtained by the protein pre-training model and the structure of the protein predicted by the protein structure prediction model, and obtaining the protein-level feature using a graph isomorphism network (GIN);

[0009] splicing the protein-level features corresponding to the CDS in sequence on the same Contig and the k-mer frequency vector of the Contig nucleotide sequence to generate enhanced protein-level features;

[0010] setting a sliding window length, obtaining a plurality of fragments with a fixed window length of protein function labels at the center position of the enhanced protein-level features in sequence on the same Contig according to the arrangement information of CDS on the Contig, taking one fragment as a sample, training a function prediction network composed of a bidirectional long short-term memory network (BiLSTM) and a multilayer perceptron, and obtaining a target protein prediction probability value of each fragment by extracting the relationship between the CDS encoded proteins on the same Contig.

[0011] Further, the protein pre-training model is ESM-2, and the protein structure prediction model is ESMFold.

[0012] Further, the Contig nucleotide sequence is represented as a 3-mer frequency vector f 3-mer , the sliding window length is 3-5, preferably the sliding window length is an odd number, and the protein function label is located at the center element.

[0013] Further, the protein-level feature is obtained by the following process:

[0014] based on the predicted protein structure x esmfold constructing a protein graph, the sequence feature x esm2 of the protein obtained by the protein pre-training model is initialized as a residue-level node feature: assuming that G(V, E) represents the protein graph, wherein v i ∈V represents a residue r i in the predicted protein structure x esmfold , ei,j ∈E represents residue pair (r i ,r j There are edges connecting them, and one edge e i,j There exists if and only if two residues r i ,r j The central carbon atom C α The Euclidean distance is less than

[0015] V = {v} i |r i ∈x esmfold} (5)

[0016]

[0017] Then for each residue r i The corresponding protein sequence characteristics As a protein graph node v i Initial node characteristics;

[0018] For each protein map, node features are propagated through multi-layer GIN to obtain enhanced node features with structure awareness; the node features output by each layer of GIN are aggregated by average pooling to generate hierarchical protein features; all hierarchical protein features are concatenated and processed by MLP to finally obtain integrated protein-level features.

[0019] Furthermore, the loss function L during the training process of the function prediction network is:

[0020]

[0021] Where, N contig m represents the total number of contigs. i This represents the number of fragments on the i-th contig. y represents the predicted probability value of the target protein in the j-th fragment of the i-th contig. ij This represents the true label of the target protein in the j-th fragment of the i-th contig.

[0022] Secondly, the present invention provides a protein function prediction system based on contig perception, the prediction system comprising:

[0023] Protein pre-training models are used to obtain the sequence features of proteins;

[0024] Protein structure prediction models are used to predict the structure of proteins.

[0025] a k-mer frequency vector calculation module configured to represent the contig nucleotide sequence into a k-mer frequency vector;

[0026] a protein graph construction module configured to take the sequence features of the protein obtained by the protein pre-training model and the structure of the protein predicted by the protein structure prediction model as inputs, construct a protein graph from the predicted structure of the protein according to the Euclidean distance between the central carbon atoms C α of the residues, and take the sequence features of the protein obtained by the protein pre-training model as the node features of the protein graph;

[0027] a protein-level feature extraction module configured to propagate the node features of the protein graph through multiple layers of GIN, and finally combine the graph-level features of each GIN layer to obtain the protein-level features

[0028] a protein feature enhancement module configured to take the protein-level feature array corresponding to the CDS sequence on the same contig and the k-mer frequency vector of the contig nucleotide sequence as inputs, splice the k-mer frequency vector and the protein-level features obtained by the protein-level feature extraction module to generate enhanced protein-level features;

[0029] a contig-level topology-aware feature extraction module configured to take the enhanced protein-level feature array corresponding to the CDS sequence on the same contig as input, divide the enhanced protein-level feature array into several equal-length segments using a sliding window, and then aggregate the protein features using a bidirectional long short-term memory network (BiLSTM) to obtain topology-aware contig-level protein features by extracting the relationship between the CDS encoded proteins on the same contig;

[0030] a classification module configured to input the contig-level protein features into a multi-layer perceptron (MLP) to predict the COG functional category of the protein.

[0031] Compared with the prior art, the present application has the following beneficial effects:

[0032] 1. The present application uses the relationship between proteins, i.e., the topology information of CDS on the contig and the contig sequence information, to assist protein function annotation, achieving more accurate protein function prediction. Specifically, first, the protein-level features are spliced with the 3-mer frequency vector of the contig sequence, and then the enhanced protein-level features corresponding to multiple CDS on the contig are aggregated using BiLSTM as the contig-level protein features of the intermediate protein. The contig-level protein features are used for protein function classification, which has achieved better results in all evaluation indicators compared with the currently highest-accuracy structure-based protein prediction method, significantly improving the accuracy of protein function annotation.

[0033] 2. The present application constructs a protein graph by using the ESM-2 sequence feature and the protein structure predicted by ESMFold, and adopts the multi-layer GIN aggregation node feature, so as to obtain the protein-level feature of the protein. ESM-2 has better protein sequence feature expression and lighter characteristics, and GIN has been proved to be particularly effective in graph classification tasks and is suitable for the task of classifying protein graphs in protein function prediction tasks. Therefore, the protein sequence and structure information are fully integrated into the protein-level feature, so that the protein-level feature has better distinguishability.

[0034] 3. The present application has data set independence. Unlike the traditional gene ontology GO special function prediction method, the method of the present application is not limited by the prior knowledge of a specific data set, and can be flexibly applied to various annotation systems such as COG and GO, and effectively establishes the correlation bridge between the genomic environment and the protein function. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a flowchart of an embodiment of the protein function prediction method based on Contig awareness of the present application.

[0036] Figure 2 is an architecture diagram of a protein function prediction model of an embodiment of the present application.

[0037] Figure 3 is a detailed description diagram of the benchmark data set in the embodiment of the present application.

[0038] Figure 4 is a comparison diagram of the classification performance of the model of the present application and the existing protein function prediction model.

[0039] Figure 5 is a confusion matrix diagram of 23 COG categories of the model of the present application. DETAILED DESCRIPTION

[0040] The specific embodiments are described below with reference to the accompanying drawings, which are only used to clearly and completely introduce the technical solutions of the present application, and do not limit the protection scope of the present application.

[0041] Contig on the coding sequence (CDS, can be expressed as protein) topology information, CDS represents a certain nucleotide sequence that can express protein, which has many characteristics that help protein function annotation, which can be used in various annotation systems. A contig represents a continuous DNA sequence obtained by genome assembly, which is an intermediate product of genome assembly. In the same contig, adjacent coding sequences CDS may exhibit functional relevance. This correlation is derived from various biological mechanisms, including operon structure, metabolic gene cluster, and horizontal transfer unit. For example, in prokaryotes, functionally related genes are often co-localized in the form of operons. Enzymes involved in a particular metabolic pathway are often distributed in clusters, and resistance genes or mobile elements are usually present in the form of complete functional modules in a single contig. Therefore, the present application enhances protein function annotation by integrating the topological relationship of CDS within the contig.

[0042] The present application is based on a Contig-aware protein function prediction method (referred to as method, see Figure 1 ), which integrates the sequence features, structure features of the protein and the topological information of CDS on the contig to comprehensively predict the protein function through multi-level feature fusion, including the following steps:

[0043] Step one: obtain the protein amino acid sequence, and the nucleotide sequence corresponding to the contig and the arrangement information of CDS on the contig.

[0044] Step two: according to the protein amino acid sequence, use EMS-2 to obtain the sequence features of the protein, and use ESMFold to predict the structure of the protein; represent the contig nucleotide sequence as a 3-mer frequency vector.

[0045] Step three: construct a protein function prediction model and train the model:

[0046] The protein function prediction model includes five main modules, as shown in Figure 2 :

[0047] (1) Protein graph construction module: taking the sequence features of the protein obtained by ESM-2 and the protein structure predicted by ESMFold as input, constructing a protein graph from the predicted protein structure according to the Euclidean distance between the central carbon atoms C α of the residues, and taking the sequence features of the protein obtained by ESM-2 as the node features of the protein graph;

[0048] (2) Protein-level feature extraction module: propagate the node features of the protein graph through multiple layers of GIN, and finally combine the graph-level features of each GIN layer as the protein-level features

[0049] (3) Protein feature enhancement module: taking the protein-level feature array corresponding to the CDS sequence on the same Contig and the 3-mer frequency vector of the Contig nucleotide sequence as input, splicing the protein-level feature array with the 3-mer frequency vector to generate enhanced protein-level features;

[0050] (4) Contig-level topology-aware feature extraction module: taking the enhanced protein-level feature array corresponding to the CDS sequence on the same Contig as input, dividing the enhanced protein-level feature array into several fragments with a length of 3 using a sliding window, and then using a bidirectional long short-term memory network (BiLSTM) to aggregate protein features, obtaining topology-aware Contig-level protein features by extracting the relationship between CDS encoded proteins on the same Contig;

[0051] (5) Classification module: inputting the Contig-level protein features into a multi-layer perceptron (MLP) for prediction to obtain the COG functional category of the protein.

[0052] According to the arrangement information of the CDS on the Contig, a plurality of fragments with a fixed window length of the center position of the enhanced protein-level feature in the CDS sequence on the same Contig are obtained, one fragment is taken as a sample, and a large number of samples constitute an enhanced fragment dataset, and the functional prediction network composed of BiLSTM and MLP is trained by using the enhanced fragment dataset,

[0053] The trained functional prediction network, combined with the protein graph construction module, the protein-level feature extraction module, and the protein feature enhancement module, constitutes a protein function prediction model, which is used for protein function prediction after overall training.

[0054] Step four: in actual prediction, the amino acid sequence of the protein to be predicted, the nucleotide sequence of the Contig where the protein is located, and the CDS arrangement information on the Contig where the protein is located are obtained; the sequence features of the protein are obtained using ESM-2, the structure of the protein is predicted using ESMFold, and the nucleotide sequence of the Contig where the protein is located is represented as a 3-mer frequency vector; the sequence features, predicted structure, and 3-mer frequency vector of the nucleotide sequence of the Contig where the protein is located of the protein to be predicted are input into the protein function prediction model for prediction to obtain the COG functional category of the protein to be predicted.

[0055] Example 1

[0056] The embodiment is based on a Contig-aware protein function prediction method, which includes the following steps:

[0057] ​Step 1: Obtain the protein amino acid sequence, the corresponding contig nucleotide sequence, and the arrangement information of the CDS on the contig. One protein amino acid sequence corresponds to one contig nucleotide sequence and one contig CDS arrangement information.

[0058] Step 2: For each protein, use the esm2_t30_150M_UR50D model in ESM-2 to extract the protein's sequence features x. esm2 Predict protein structure using the esmfold_3B_v1 model in ESMFold. esmfold Represent the contig nucleotide sequence as a 3-mer frequency vector f 3-mer ;

[0059] k-mers are continuous substrings of length k in a nucleotide or amino acid sequence, and are the basic unit in bioinformatics analysis. By segmenting a sequence into overlapping k-mers, local sequence features can be quantified. Encoding the nucleotide sequence of a contig into a k-mer frequency vector characterizes its sequence features. The specific process is as follows:

[0060] Given a nucleotide sequence of length L, S = s1s2...s L , where each s i ∈{A,T,C,G}, where A,T,C,G are four nucleotides, and all possible k-mers are extracted using a sliding window method; specifically, the k-mer starting at position i is represented by w. i It can be defined as:

[0061] w i =s i s i+1 ...s i+k-1 ,i∈{1,2,...,L-k+1} (1)

[0062] This process generates N = L - k + 1 overlapping k-mers;

[0063] Arrange all possible k-mers into a set in lexicographical order. The total number of possible k-mers is 4. k ;

[0064] k-mer frequency vector The construction method is: its j-th component f j Statistics on k-mer w (j) The number of times it appears in sequence S;

[0065]

[0066] where w (j) is the jth element of the set W;

[0067] Finally, the k-mer frequency vector can be represented as To eliminate the bias caused by sequence length, the frequency vector is divided by the total number of k-mers N = L - k + 1, and then normalized to the range [0, 1] by a normalization layer; finally, the normalized frequency vector is input into the MLP to obtain the final k-mer frequency vector f k-mer .

[0068]

[0069] where Norm denotes the normalization operation.

[0070] In this embodiment, k is set to 3, that is, the 3-mer frequency vector f 3-mer of the Contig nucleotide sequence is obtained.

[0071] Step three: construct a protein graph and obtain protein-level features

[0072] Based on the predicted protein structure x esmfold , the sequence feature x esm2 is constructed into a protein graph, and the sequence feature x i is initialized as the node feature at the residue level: assume that G (V, E) represents a protein graph, where v esmfold ∈ V represents a residue r i in the structure x i,j , e i ∈ E represents that there is an edge connected between the residue pair (r j , r i,j ), and an edge e i exists if and only if the Euclidean distance between the center carbon atoms C j of the two residues r α , r

[0073] V = {v i |r i ∈ x esmfold} (5)

[0074]

[0075] Then for each residue r i , its corresponding sequence feature x is taken as the initial node feature of the protein graph node v i .

[0076] For each protein graph, enhanced node features with structure-aware ability are obtained by node feature propagation through multi-layer GINs; the node features output by each layer of GIN are aggregated by average pooling to generate hierarchical protein features; to comprehensively obtain the multi-level structure information of the protein, all hierarchical protein features are spliced and processed through an MLP to finally obtain the integrated protein-level feature.

[0077] The processing process of the multi-layer GIN is as follows:

[0078] The node feature updating process of the (l+1)th layer GIN follows the following calculation rules:

[0079]

[0080] wherein represents the output feature of node v at the (l+1)th layer, is the feature of node v at the lth layer, is the feature of node u at the lth layer, MLP (l+1) is the MLP of the (l+1)th layer, ∈ (l+1) is the learnable parameter of the (l+1)th layer, is the 1st-order neighbor node set of node v;

[0081] The node features of each GIN layer are aggregated by average pooling to generate the hierarchical protein features h p of the protein p; in order to capture diversified structure information, all hierarchical protein features are spliced and processed through an MLP to finally obtain the protein-level feature

[0082]

[0083] wherein is the feature of node v at the lth layer, V p is the node set in the protein graph G p , |V p | is the number of nodes, represents the protein feature generated by the lth layer through average pooling, N h is the number of GIN layers, and CONCAT is the splicing operation.

[0084] Step four: obtaining enhanced protein-level features

[0085] In order to further enhance the protein-level feature and make full use of the functional correlation between proteins located in the same Contig, the protein-level feature is combined with the 3-mer frequency vector f 3-merperforming splicing;

[0086]

[0087] wherein represents a splicing operator, is an enhanced protein-level feature of protein p.

[0088] According to the genomic coordinates of the CDSs performing sorting, constructing the enhanced protein-level features corresponding to the CDS order, and capturing the functional and topological relationships between the proteins encoded by the same Contig:

[0089] Given n CDSs on a Contig and their corresponding enhanced protein-level feature vectors d is the dimension of the enhanced protein-level feature vector, and the CDSs are regarded as similar text arrangements to capture the context relationship of the enhanced protein-level features of adjacent CDSs, then the construction of the input sequence matrix F is as follows:

[0090]

[0091] Subsequently, F is divided into m segments using a sliding window of length 3, and the protein located at the center in the length-3 segment is the target protein, where the i-th segment can be represented as:

[0092]

[0093] wherein, is the target protein.

[0094] These segmented segments are represented in matrix form as:

[0095]

[0096] wherein, F segment represents the segment matrix.

[0097] Step five: function prediction:

[0098] Subsequently, F segment is input into BiLSTM, and BiLSTM is used to capture the features of adjacent CDSs to assist protein function annotation, so as to aggregate the protein features in each segment;

[0099] O = BiLSTM (F segment ) (13)

[0100] The output of BiLSTM is the predicted target protein feature of each segment, and the dimension of o m is d1;

[0101] Subsequently, the O is input to the multi-layer perception machine (MLP) for prediction:

[0102]

[0103] The loss function is calculated as follows:

[0104]

[0105] where N contig represents the total number of contigs, m i represents the total number of labeled proteins (i.e., the number of fragments) on the ith Contig, represents the predicted probability value of the target protein of the jth fragment on the ith Contig, y ij represents the true label of the jth fragment on the ith Contig.

[0106] Step six: Obtain the amino acid sequence of the target protein, the nucleotide sequence of the Contig where it is located, and the CDS arrangement information of the Contig where it is located; use ESM-2 to obtain the sequence feature x esm2 of the target protein, use ESMFold to predict the structure x esmfold of the target protein, and represent the nucleotide sequence of the Contig where the target protein is located as a 3-mer frequency vector f 3-mer ; input the sequence feature x esm2 of the target protein, the structure x esmfold , and the 3-mer frequency vector f 3-mer of the nucleotide sequence of the Contig where the target protein is located into the trained protein function prediction model for prediction to obtain the COG function category of the target protein. Simulation experiment:

[0107] The construction of the benchmark dataset is as follows:

[0108] Human assembled data were downloaded from the MGnify database, which is a microbial genome database under EMBL-EBI, providing standardized storage, functional annotation and interactive analysis of metagenomic data; the downloaded assembled data contains 15,000 Contigs, covering the nucleotide sequence of Contigs, the amino acid sequence of CDS and the arrangement information of CDS on Contigs; first, the Contigs with an average nucleotide similarity (ANI) > 85% are removed using FastANI, and then the CDS with a sequence similarity > 85% are removed by CD-HIT clustering; at the same time, the maximum length of CDS sequence is limited to 1,022 amino acids to ensure that all protein sequence features can be generated by ESM-2; to add COG classification annotations to CDS, the COGorg24.faa.gz file is downloaded from the NCBI COG database, and DIAMOND is used for sequence alignment, with a sequence similarity > 50% and a coverage > 80% as the threshold to assign COG classification; in the final processed data set, 9,000 Contigs are randomly selected as the training set, 1,000 Contigs as the validation set, and 2,000 Contigs as the test set; the training set contains 58,803 CDS (of which 28,136 are labeled), the validation set contains 6,779 CDS (of which 3,139 are labeled), and the test set contains 13,019 CDS (of which 6,229 are labeled); it needs to be specially pointed out that only the labeled CDS is used to calculate the model loss function, and the unannotated CDS is used as the CDS sequence of Contig to input BiLSTM for feature learning; the detailed description of the benchmark data set is shown in Table 1.

[0109] Table 1 Details of the benchmark data set

[0110]

[0111] Figure 3 The distribution of proteins in the COG functional categories in the benchmark data set is shown. The benchmark data set contains 23 COG categories (categories A and Y are excluded due to lack of training samples; and category S is removed because its function is unknown). The benchmark data set shows a clear uneven distribution: categories G and E each contain more than 3,000 proteins, while categories X and B each contain less than 200 proteins (the number of statistics is the sum of the training set, the validation set and the test set); from the proportion of proteins in each category, the training and validation sets and the test set both show uneven distribution characteristics, and the proportion of proteins in each category in the training and validation sets and the test set is basically consistent with the proportion of proteins in each category in the whole data set; for ease of experimental analysis, the categories are further divided into three groups according to the number of proteins: high abundance (> 2900 proteins), medium abundance (600-2900 proteins) and low abundance (< 600 proteins).

[0112] To verify the effectiveness of the method of the present application, the model of the present application and the existing protein function prediction model are compared in four evaluation indexes of accuracy, precision, recall and F1 score on the benchmark dataset; to verify the prediction performance of the method of the present application in different protein function categories and evaluate its robustness on the protein distribution unbalanced dataset, the model of the present application is compared with the existing protein function prediction model in terms of precision, recall and F1 score three indexes, and the confusion matrix of the model of the present application in 23 COG function categories is analyzed to verify the performance of the model of the present application in different protein function categories in detail.

[0113] The model of the present application is compared with two sequence-based protein function prediction models and two structure-based protein function prediction models on the benchmark dataset, and the prediction accuracy, precision, recall and F1 score are shown in Table 2.

[0114] Table 2 Performance comparison of each comparative method and the model of the present application

[0115]

[0116] As shown in Table 2, the model of the present application achieves the optimal performance in the four evaluation indexes, and exceeds the suboptimal model by 11.24%, 12.43%, 13.59% and 13.30% respectively in the four evaluation indexes. The results show that the method has higher accuracy and stability for protein function prediction, and can achieve better performance in the protein function prediction task.

[0117] The model of the present application is compared with two sequence-based protein function prediction models and two structure-based protein function prediction models in the benchmark dataset, and the prediction accuracy, precision, recall and F1 score are shown in Table 2. Figure 4

[0118] As shown in Table 2, the model of the present application achieves the optimal performance in the four evaluation indexes, and exceeds the suboptimal model by 11.24%, 12.43%, 13.59% and 13.30% respectively in the four evaluation indexes. The results show that the method has higher accuracy and stability for protein function prediction, and can achieve better performance in the protein function prediction task. Figure 4 ​It can be seen that the model of the present application exhibits superior performance in almost all categories, and only in the U category, its precision is slightly lower than Struct2GO (1.73% difference). Performance analysis by abundance level shows that: (1) In high abundance categories (such as G and E), the model of the present application obtains F1 scores of 0.9625 and 0.9676, respectively, which are improved by 8.81% to 43.26% and 11.15% to 51.88% compared with the comparison methods; (2) In medium abundance categories (such as O and W), the model of the present application achieves F1 scores of 0.9601 and 0.9466, and the performance advantage reaches 11.26% to 81.47% and 11.01% to 48.68%; (3) Especially in low abundance categories (such as X and B), the model of the present application still maintains robust performance, with F1 scores of 0.8980 and 0.9333, which exceed the comparison models by 15.29% to 77.50% and 16.41% to 83.33%. This comprehensive evaluation proves the outstanding ability of the model of the present application in different abundance protein function prediction tasks, especially its outstanding performance in the low abundance category of few sample learning scenarios.

[0119] Figure 5 The confusion matrix of the model of the present application in 23 COG categories is shown. Although most categories achieve high prediction accuracy, it can be observed that R category (only with general function prediction) proteins exhibit slightly high misclassification rates, with an F1 score of 0.8601, while the overall F1 score is 0.9250. This phenomenon is particularly evident in function-related categories such as G category (carbohydrate transport and metabolism). This phenomenon may be due to the functional ambiguity of R category proteins, although their exact function has not been determined, they are often involved in basic cellular processes that intersect or are regulated by carbohydrate metabolism pathways. These inherent functional similarities pose significant challenges to classification models, leading to inevitable performance degradation (for example, the F1 score of the R category of the Struct2GO model is 0.7047, while the overall F1 score is 0.7920).

[0120] The above results show that the model of the present application achieves the most advanced protein function prediction performance in the overall and all abundance categories. Its innovative combination of protein sequence and structure features, as well as protein CDS functional association, is proven to be particularly effective in achieving robust performance on unbalanced data sets.

[0121] The unmentioned part of the present application is applicable to the prior art.

Claims

1. A contig-aware based protein function prediction method, characterized in that, The prediction method comprises the following contents: Obtaining a protein amino acid sequence, a nucleotide sequence corresponding to a Contig, and arrangement information of CDS on the Contig; obtaining sequence features of the protein by using a protein pre-training model according to the protein amino acid sequence, and predicting a structure of the protein by using a protein structure prediction model; representing the nucleotide sequence of the Contig into a k-mer frequency vector; Constructing a protein graph by using the sequence features of the protein obtained by the protein pre-training model and the structure of the protein predicted by the protein structure prediction model, and obtaining protein-level features by using a graph isomorphism network (GIN); Splicing the protein-level features corresponding to the CDS in sequence on the same Contig and the k-mer frequency vector of the nucleotide sequence of the Contig to generate enhanced protein-level features; Setting a sliding window length, obtaining a plurality of fragments with a fixed window length of protein function labels at the center position of the enhanced protein-level features in sequence of the CDS on the same Contig according to the arrangement information of the CDS on the Contig, taking one fragment as a sample, training a function prediction network composed of a bidirectional long short-term memory network (BiLSTM) and a multilayer perceptron (MLP), and obtaining a target protein prediction probability value of each fragment by extracting the relationship between the CDS coding proteins on the same Contig.

2. The prediction method of claim 1, wherein, The protein pre-training model is ESM-2, and the protein structure prediction model is ESMFold.

3. The prediction method of claim 1, wherein, The Contig nucleotide sequence is represented as a 3-mer frequency vector f 3-mer with a sliding window length of 3 to 5, preferably with an odd sliding window length, and the protein functional tag is located on the central element.

4. The prediction method of claim 1, wherein, The protein-level feature obtaining process is as follows: Based on predicted protein structure x esmfold Construct a protein map by combining the sequence features of proteins obtained from the protein pre-training model. esm2 Initialization is performed using node features at the residue level: assuming G(V,E) represents the protein graph, where v i ∈V represents the predicted protein structure x esmfold There are residues r i e i,j ∈E represents residue pair (r i ,r j There are edges connecting them, and one edge e i,j There exists if and only if two residues r i ,r j The central carbon atom C α The Euclidean distance is less than V = {v i | r i ∈x esmfold} (5) After that for each residue r i Sequence features of the corresponding protein As protein graph node v i Initial node features; For each protein graph, node feature propagation is performed through a multilayer GIN to obtain enhanced node features with structure perception ability; The node features output by each layer of GIN are aggregated by average pooling to generate hierarchical protein features; All hierarchical protein features are spliced and processed by an MLP to finally obtain integrated protein-level features.

5. The prediction method of claim 1, wherein, The loss function L in the function prediction network training process is as follows: where N contig denotes the total number of contigs, m i denotes the number of fragments on the i-th contig, denotes the predicted probability value of the target protein of the j-th fragment on the i-th contig, y ij denotes the true label of the target protein of the j-th fragment on the i-th contig.

6. A Contig-aware based protein function prediction system, characterized in that, The prediction system comprises: A protein pre-training model for obtaining sequence features of a protein; A protein structure prediction model for predicting a structure of the protein; A k-mer frequency vector calculation module for representing a nucleotide sequence of a Contig into a k-mer frequency vector; The protein graph construction module takes the sequence features of the protein obtained by the protein pre-training model and the structure of the protein predicted by the protein structure prediction model as inputs, constructs a protein graph from the predicted structure of the protein according to the Euclidean distance of the central carbon atoms C α of residues, and takes the sequence features of the protein obtained by the protein pre-training model as node features of the protein graph. The protein level feature extraction module: the node features of the protein graph are propagated through multiple layers of GIN, and the graph level features of each GIN layer are finally combined as the protein level features A protein feature enhancement module: inputting an array of protein-level features corresponding to the CDS in sequence on the same Contig and the k-mer frequency vector of the nucleotide sequence of the Contig, splicing the k-mer frequency vector and the protein-level features obtained by a protein feature extraction module to generate enhanced protein-level features; A Contig-level topology perception feature extraction module: taking an array of enhanced protein-level features corresponding to the CDS in sequence on the same Contig as input, dividing the array of enhanced protein-level features into a plurality of equal-length fragments by using a sliding window, then performing protein feature aggregation by using a bidirectional long short-term memory network (BiLSTM), and obtaining Contig-level protein features with topology perception by extracting the relationship between the CDS coding proteins on the same Contig. A classification module inputs the Contig-level protein features into a multi-layer perceptron (MLP) to predict the COG functional category of the protein.

Images