A method and system for predicting protein essentiality across different cell lines

Abstract

The application provides a method and system for predicting protein essentiality in different cell lines, the method comprising: collecting protein sequence data; collecting exon information to assist in predicting protein essentiality in different cell lines; setting a DeepExon network structure, the DeepExon network structure comprising: a coding module, a CNN, a multi-head self-attention module, a Bi-LSTM and a prediction module; in the coding module, the amino acid position weight and the dipeptide combination probability weight are integrated into the coding operation; the CNN network is used to extract local features from the original sequence and the exon information; the obtained sequence information features and exon features are spliced and sent to the multi-head attention module; the LSTM captures and stores long-term information from the spliced features by introducing memory cells and a gating mechanism; and the protein essentiality is predicted. The application solves the technical problems of not fully considering cell line specificity, insufficient information mining, and low prediction accuracy and efficiency for essential proteins.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of gene feature recognition, and in particular to a method and system for predicting protein essentiality under different cell lines. BACKGROUND

[0002] Genes are the key to any organism, and essential genes are essential components for cell survival. The expression of the same gene is different under different cell environments. For different cell types, since they perform different biological functions, the essentiality of the same gene will vary greatly. Essential proteins are the products of essential genes, so the essentiality of proteins will also change with changes in the cell environment. Therefore, if the essentiality of proteins can be more accurately predicted under different cell environments and conditions, it is particularly important for precision medicine and personalized treatment. Since the essentiality of proteins is affected by environmental factors, the essentiality of the same protein will be different under different environments. The features obtained by predicting only the sequence are very limited.

[0003] Existing methods for determining essential genes include single gene knockout or multiple gene knockout, mutagenesis screening, and RNA interference screening. There are also many methods for identifying essential proteins, including transposon mutagenesis, single gene knockout, RNA interference, and CRISPR gene editing technology. However, whether it is the determination of essential genes or the identification of essential proteins, they all require more manpower, material resources, and financial resources. Therefore, it is necessary to develop a more effective, faster, and more accurate computational method to predict essential proteins.

[0004] Current computational methods can be broadly divided into two categories. One is based on network centrality measures, with the earliest being the centrality-lethality rule proposed by Jeong et al. Then came the betweenness centrality (BC), closeness centrality (CC), eigenvector centrality (EC), local average centrality (LAC), and maximum neighborhood component (MNC) to identify essential proteins. The other category is based on deep learning methods. With the development of high-throughput sequencing technology, many related studies have emerged. Kuang et al. (2021) developed a machine learning model that combines gradient boosting trees, support vector machines, and multilayer perceptrons (MLP) to predict essential genes. Zeng et al. (2021) developed an integrated deep learning model by integrating multiple gradient boosting decision tree (GBDT) base classifiers to achieve accurate prediction. Li et al. (2021) developed an integrated deep learning model, EP-EDL, which applies convolutional neural networks (CNN) to predict human essential proteins from evolutionary information. Li et al. (2023) proposed a deep learning model, DeepCellEss, which can predict the essentiality of proteins under different cell line environments.

[0005] Although there are many methods to realize the prediction of protein essentiality at present, they still have certain limitations. Before the DeepCellEss model proposed by Li (2023), most computational methods do not consider cell line specificity, which leads to the inability to accurately identify essential proteins in different cell lines. Although the model DeepCellEss can accurately identify essential proteins in different cell lines, it is difficult to mine more relevant information to assist in identifying essential proteins relying on sequence information.

[0006] The prior art patent application document with the publication number CN114242168A, a method for identifying essential proteins, the prior method includes: supplementing the missing data to improve robustness. Then respectively adopt constructing PPI network topology, Pearson correlation coefficient, homology correlation coefficient to reduce the convergence speed of deep neural network. Finally, the best correlation between the degree of the node, the Pearson correlation coefficient and the homology correlation coefficient is found through the deep neural network. However, the foregoing prior art does not consider cell specificity. From the existing various data, it can be known that the essentiality of proteins is closely related to the cell environment, and the activity of the same protein gene in different cell lines is different, and the essentiality thereof also has great difference.

[0007] Secondly, for the existing protein essentiality prediction technology in different cell lines, only sequence information is relied on for prediction, and sequence information alone is not enough to obtain better prediction results, and other feature information should be supplemented for prediction.

[0008] Finally, the traditional embedding encoding is composed of word embedding and position embedding. Although the traditional embedding encoding can better associate the context relationship in training, it lacks the introduction of sequence information.

[0009] In summary, the prior art has the technical problems of not fully considering cell line specificity and insufficient information mining, resulting in low prediction accuracy and efficiency for essential proteins. SUMMARY

[0010] The technical problem to be solved by the present application is how to solve the technical problems of not fully considering cell line specificity and insufficient information mining in the prior art, resulting in low prediction accuracy and efficiency for essential proteins.

[0011] The present application solves the above technical problems by adopting the following technical solutions: a method for predicting protein essentiality in different cell lines includes:

[0012] S1, collecting protein sequence data;

[0013] S2, according to the protein sequence data, the exon information is collected, the labeled exon position data set is obtained by the labeling operation processing, and the exon information is used to assist in predicting the necessity of different cell lines of protein;

[0014] S3, a DeepExon network structure is set, wherein the DeepExon network structure comprises: an encoding module, a CNN network, a multi-head self-attention module and a prediction module; wherein in the encoding operation of the encoding module, the amino acid position weight and the dipeptide combination probability weight are fused; the CNN network is used to extract local features from the protein data and the exon information; the sequence information features and the exon features in the local features are spliced to obtain spliced features; the Bi-LSTM is used to capture and store long-term information from the spliced features;

[0015] S4, the prediction module is used to predict the protein necessity prediction result according to the long-term information.

[0016] The protein exon position is collected and combined with the embedded coding and other characteristics to predict the necessity of the protein in different cell lines, so that the auxiliary features can be fully utilized to help the model to better predict the protein necessity in different environments.

[0017] In a more specific technical solution, in S1, the data set is obtained by removing redundant sequences at the protein level and nucleotide level using CD-HIT and CD-HIT-est, wherein the sequence data set comprises: cell lines, binary essentiality labels of proteins, and sequence information.

[0018] In a more specific technical solution, in S2, according to the consensus coding provided in the applicable data set, the exon information corresponding to each protein sequence data is found on NCBI;

[0019] According to the exon position in the exon information, the labeled exon position data set is generated in combination with the applicable sequence data set.

[0020] The present application collects corresponding exon information according to original data, adopts embedded coding mode, fuses sequence position information, amino acid position information and other characteristics to perform prediction, and meanwhile, the data set used by the present application can be used to predict protein necessity for different cell lines. The present application proposes to use exon position information as an additional feature, because the exon directly encodes the functional domain of the protein, and the functional domain determines the specific function of the protein in different cells and biological environments, so introducing the exon position information can enable the neural network to better analyze the relationship between the sequence and the exon position. And the present application has carried out a large number of experiments on DeepExon, compared with the DeepCellEss model, the present application has improved 2.3% in AUC and 2.1% in AUPRC, to prove that the exon has an important role for the prediction of essential proteins in different cell lines. In addition, the present application also carries out feature ablation experiment to prove that the feature way proposed by us has outstanding contribution.

[0021] In a more specific technical solution, in the DeepExon network structure in S3, the coding module includes: adaptive embedding coding, one_hot coding;

[0022] In the coding operation, the non-exon part is shielded, the exon part is one_hot coded, the amino acid fixed feature matrix is obtained, and the exon information is highlighted;

[0023] For protein sequence data, adaptive embedding coding is adopted, including: a learnable embedding layer and a position embedding, the position embedding coding is used to obtain the amino acid sequence in the protein sequence data, so as to understand the sequence information and perform model enhancement and optimization operation;

[0024] In the adaptive embedding coding, the position weight of the amino acid and the dipeptide combination probability are fused to understand the sequence information. For the position weight of the amino acid, the occurrence frequency of the amino acid X in each position in the sequence data set is traversed, and then the total amino acid occurrence frequency in the position is compared with the occurrence frequency of the amino acid X in the position to obtain the position weight of the amino acid X in the position;

[0025] Local features are extracted from protein sequence data and exon information by using a CNN network. The CNN network has strong feature extraction capability, and CNN has strong performance in many biological information prediction. However, the limitation of CNN is obvious, because CNN is only suitable for extracting local information of sequence;

[0026] For the sequence information extraction part, in the CNN network, a double-path CNN is combined, different convolution kernels are used to extract sequence information from the protein sequence data, and element-by-element multiplication is performed.

[0027] The application proposes an improved embedding coding mode, which uses two weights to solve the problem of missing sequence information in embedding coding, and can also speed up the convergence speed.

[0028] The application proposes a DeepExon model, which is a deep learning framework improved from DeepCellEss, and is used for predicting cell line-specific essential proteins. The application collects 16408 pieces of protein exon position information, and combines sequence information, amino acid position information and other features to assist in prediction. Through a large number of experiments and comparisons, the DeepExon model proposed in the application has a significant improvement compared with the DeepCellEss model.

[0029] In a more specific technical solution, in adaptive embedding coding, the position embedding of an amino acid at the pth position in the protein sequence data is expressed as a d k vector PE(p), and the i th element of the vector PE(p) is expressed by using the following logic:

[0030]

[0031] In the formula, 2i and 2i+1 represent odd dimension and even dimension respectively.

[0032] In a more specific technical solution, the position weight of the amino acid X is determined by using the following logic:

[0033]

[0034] In the formula, L X represents the weight distribution of the amino acid X, X i represents the number of occurrences of the amino acid X at the i th position, N represents the sequence length, and L i represents the number of occurrences of the total amino acid at the i th position.

[0035] In a more specific technical solution, the combination probability of two amino acids is expressed according to the following logic by using the dipeptide combination probability:

[0036]

[0037] In the formula, (X,Y) represents the combination probability of the amino acid X and the amino acid Y, A1 represents the total number of combinations of the amino acid X and the amino acid Y, and A2 represents the total number of combinations of the amino acid X and the remaining amino acids.

[0038] The present application improves the traditional embedding coding and proposes to use exon information to assist in predicting the necessity of proteins under different cell lines.

[0039] In a more specific technical solution, the multi-head attention mechanism in the multi-head attention module includes: not less than 2 self-attention:

[0040]

[0041] In the formula, Q, K, V represent query, key, value respectively, W Q , W K , W V represent weight matrix, is a dot product scaling factor.

[0042] In a more specific technical solution, in S3, the Bi-LSTM includes: forward LSTM, backward LSTM; wherein, the forward LSTM and the backward LSTM are expressed by using the following logic:

[0043]

[0044] In a more specific technical solution, a system for predicting the necessity of proteins under different cell lines includes:

[0045] The sequence acquisition module is used to collect protein sequence data;

[0046] The exon information collection module is used to collect exon information according to the protein sequence data, obtain labeled exon position data set through labeling operation processing, and use the exon information to assist in predicting the necessity of proteins under different cell lines, and the exon information collection module is connected with the sequence acquisition module;

[0047] The DeepExon network feature processing module is used to set and utilize the DeepExon network structure, wherein the DeepExon network structure comprises an encoding module, a CNN network, a multi-head self-attention module and a prediction module; wherein in the encoding operation of the encoding module, the amino acid position weight and the dipeptide combination probability weight are integrated; the CNN network is used to extract local features from protein data and exon information; sequence information features and exon features in the local features are subjected to splicing operation to obtain splicing features; the Bi-LSTM is used to capture and store long-term information from the splicing features, and the DeepExon network feature processing module is connected with the exon information collection module and the sequence acquisition module.

[0048] The necessity prediction module is used to utilize the prediction module to obtain protein necessity prediction results according to long-term information, and the necessity prediction module is connected with the DeepExon network feature processing module.

[0049] Compared with the prior art, the present application has the following advantages:

[0050] The present application collects the exon positions of proteins and combines embedded coding and other features to predict the necessity of proteins in different cell lines, which can fully utilize auxiliary features to help the model better predict the necessity of proteins in different environments.

[0051] According to the original data, the present application collects corresponding exon information, adopts embedded coding mode, and fuses sequence position information, amino acid position information and other features to perform prediction. And the present application has carried out a large number of experiments on DeepExon, compared with the DeepCellEss model, the present application has improved 2.3% in AUC and 2.1% in AUPRC, to prove that the exon has an important role in the prediction of necessary proteins in different cell lines. In addition, the present application also carries out feature ablation experiment to prove that the feature mode proposed by us has outstanding contribution.

[0052] The present application proposes a DeepExon model, which is a deep learning framework improved from DeepCellEss, for predicting cell line-specific essential proteins. The present application collects 16408 protein exon position information, combines sequence information, amino acid position information and other features to assist prediction. After a large number of experiments, the DeepExon model proposed by the present application has a significant improvement compared with the DeepCellEss model.

[0053] The present application improves the traditional embedding encoding and proposes to use exon information to assist in predicting the necessity of proteins under different cell lines. For the traditional embedding encoding, the present application proposes to combine the amino acid position weight and the dipeptide combination probability weight into the encoding. The present application respectively makes different ablation experiments to verify the effectiveness of the improved embedding encoding. In addition, the present application proposes to use exon information as another part of features to assist in predicting the necessity. The present application respectively makes different experiments to verify that the exon has excellent effect on the protein necessity prediction. Furthermore, the present application independently trains on the 323 cell line benchmark data set to verify that the model of the present application can better perform the prediction of the necessity of proteins under different cell lines.

[0054] The present application solves the technical problems of not fully considering the cell line specificity and insufficient information mining in the prior art, resulting in low prediction accuracy and efficiency for essential proteins. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 FIG. 1 is a basic step schematic diagram of a protein necessity prediction method under different cell lines according to an embodiment of the present application.

[0056] Figure 2 FIG. 2 is a DeepExon model structure schematic diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in combination with the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0058] Embodiment 1

[0059] As shown in FIG. 1, the present application provides a protein necessity prediction method under different cell lines, which includes the following basic steps: Figure 1

[0060] S1, collecting protein sequence data;

[0061] ​The dataset used in this embodiment is a large-scale protein essentiality dataset for cell lines provided by Li (2023). This data was identified through a large-scale, genome-wide CRISPR-Cas9 exit screening process, including different binary essentiality scores for 17,485 human protein-coding genes from 328 different human cell lines. Specifically, 1 indicates importance and 0 indicates insignificance. Furthermore, the dataset underwent CD-HIT and CD-HIT-est to remove redundant sequences at both the protein and nucleotide levels. The final dataset contains binary essentiality tags and sequence information for 328 cell lines and 16,408 proteins.

[0062] S2. Collect exon information and use it to help predict the necessity of proteins in different cell lines;

[0063] In this embodiment, exons, as part of eukaryotic genes, are fragments retained after mRNA splicing and can be expressed as proteins during protein biosynthesis. Since exons directly participate in protein coding, exon information plays a crucial role in predicting and understanding protein necessity. In this embodiment, based on the consensus coding provided in the original dataset, the corresponding exon information for each protein sequence was found on NCBI. Subsequently, based on the exon positions, a new dataset with labeled exon positions was generated by combining the original dataset.

[0064] S3. Set up a DeepExon network structure, which includes: an encoding module, a CNN, a multi-head self-attention module, a Bi-LSTM, and a prediction module. In the encoding module, amino acid position weights and dipeptide binding probability weights are incorporated into the encoding operation. A CNN network is used to extract local features from the original sequence and exon information. The acquired sequence information features and exon features are concatenated and then fed into the multi-head attention module. The LSTM captures and stores long-term information from the concatenated features by introducing memory cells and gating mechanisms. The Bi-LSTM includes: a forward LSTM and a backward LSTM. The prediction module is used to predict the protein necessity prediction result based on the long-term information.

[0065] like Figure 2 As shown, in this embodiment, the DeepExon network structure combines adaptive embedding learning and one-hot encoding. During the encoding process in this embodiment, to highlight exon information, non-exon parts are masked, and to ensure the uniqueness of each amino acid, one-hot encoding is used for the exon parts, which essentially assigns a fixed feature matrix to each amino acid.

[0066] In this embodiment, we use embedding encoding for sequence information. The traditional embedding encoding consists of a learnable embedding layer and position embedding, where the position encoding can provide the order of amino acids in the protein sequence, and the position encoding can help the model understand the sequence information, and also enhance the performance and flexibility of the model.

[0067] In this embodiment, for the amino acid at the pth position in the sequence, the position embedding is represented as a d k dimensional vector. The i-th element of the vector PE(p) can be represented as:

[0068]

[0069] where 2i and 2i+1 represent the odd dimension and the even dimension respectively. Although the introduction of position embedding can enable the model to capture the order information of amino acids in the sequence, it lacks other information about amino acids. Therefore, in this embodiment, the position weight of amino acid and the dipeptide combination probability are integrated into the embedding encoding to enable the model to better understand the sequence information. The principle of the position weight of amino acid is that for a certain amino acid X, the number of occurrences of amino acid X at each position in the data set is counted, and then the ratio of the total number of amino acids at the position to the number of occurrences of the amino acid at the position is used to obtain the weight of the amino acid at the position. The formula for calculating the position weight of amino acid is as follows:

[0070]

[0071] where L X represents the weight distribution of amino acid X, X i represents the number of occurrences of amino acid X at the i-th position, N represents the length of the sequence, and L i represents the total number of amino acid occurrences at the i-th position.

[0072] In this embodiment, the dipeptide combination probability represents the combination probability of two amino acids, which is represented as:

[0073]

[0074] where (X, Y) represents the combination probability of amino acid X and amino acid Y, A1 represents the total number of combinations of amino acid X and amino acid Y, i.e. the total number of dipeptides XY, and A2 represents the total number of combinations of amino acid X with other amino acids, i.e. the total number of dipeptides combined with amino acid X.

[0075] In this embodiment, the DeepExon network structure further includes: CNN, multi-head self-attention, Bi-LSTM, and a prediction module.

[0076] In the CNN module of the embodiment, a CNN network is used to extract local features from the original sequence and exon information. The CNN network has strong feature extraction capability, and the CNN has strong performance in many biological information prediction. However, the CNN has obvious limitations because the CNN is only suitable for extracting local information of the sequence.

[0077] In the embodiment, for the sequence information extraction part, a two-way CNN combination form is used, different convolution kernels are used to extract features of the sequence information respectively, and then element-by-element multiplication is performed. Using different convolution kernels can enable the CNN to capture features of different scales, and multi-scale features can better understand the structure and pattern of the data; element-by-element multiplication is used to realize a kind of gating mechanism. When performing element-by-element multiplication, if one value is very large, the other corresponding value will also be amplified, and vice versa. The mechanism can more flexibly adjust the intensity of the features, and help the model better process high-dimensional and complex nonlinear data.

[0078] In the multi-head self-attention module of the embodiment, the obtained sequence information features and exon features are spliced and sent to the multi-head attention module. The multi-head attention mechanism is composed of multiple self-attention, and the mathematical formula is as follows:

[0079]

[0080] Wherein, Q, K, V represent query, key, value respectively, which are obtained by linear transformation of X and weight matrix W Q , W K , W V . is a dot product scaling factor.

[0081] In the Bi-LSTM module of the embodiment, LSTM is a special recursive neural network (RNN) designed to handle long-term dependencies in sequence data. LSTM effectively captures and stores long-term information by introducing memory cells and gating mechanisms (input gate, forget gate, and output gate). Bi-LSTM is composed of two LSTM layers, a forward LSTM and a backward LSTM. The two LSTM layers process the input sequence in different directions. The mathematical expressions of the two LSTM layers are as follows:

[0082] For the forward LSTM:

[0083]

[0084]

[0085] The backward LSTM is the same as the forward LSTM, but the input sequence is reversed.

[0086] Example 2

[0087] In this example, since the sample number of essential and non-essential proteins in the dataset is extremely unbalanced, the traditional stratified split strategy will lead to an unbalanced test set. Therefore, 20% of essential proteins and an equal number of non-essential proteins were randomly selected as the independent test set, and the remaining samples were used as the training set. Five-fold cross-validation was used to train the model. In the training process, in order to solve the problem of unbalanced training data, the binary entropy loss function of the benchmark paper was used. All training processes use Nvidia GeForce RTX 4090 GPU.

[0088] The model indicators in this example use the same evaluation indicators as the benchmark paper: area under the curve (AUPROC) and area under the precision-recall curve (AUPRC). AUROC and AUPRC values are between 0 and 1, and the larger the value, the better the model performance. It is worth noting that AUPRC is more sensitive to positive samples, and for protein essentiality prediction under different cell lines, the number of essential and non-essential amino acids in the dataset is inherently unbalanced, so the higher the AUPRC value, the better the model's performance in handling unbalanced datasets.

[0089] In this example, the experimental results are discussed in combination with the prediction performance on large-scale datasets in different cell lines. The ultimate goal of the research in this example is to predict protein essentiality in different cell lines, so we use DeepExon to independently train the 323 cell line benchmark dataset with the same parameter settings. Although DeepExon performs differently on the 328 cell lines, the experimental results show that the indicators obtained for more than ninety percent of the datasets are better than the corresponding indicators of the benchmark paper. Secondly, it can be observed that in the EFO-21 cell line, the improvement of the present application is the largest, with an increase of 4.2% and 4.8% in AUROC and AUPRC, respectively, compared to the benchmark paper. Looking at the datasets in all cell lines, DeepExon can well predict essential proteins in different cell lines. In summary, DeepExon can better complete protein essentiality prediction, and the performance is better than the DeepCellEss model of the benchmark paper.

[0090] In the comparison process of the present embodiment with the benchmark paper, the performance comparison of the model and the benchmark paper model is mainly carried out. First, the AUROC obtained by the benchmark paper DeepCellEss under the HCT-116 cell line is 0.782, the AUPRC is 0.795, and the AUROC obtained by our DeepExon model under the HCT-116 cell line is 0.805, the AUPRC is 0.816, that is, the AUROC and AUPRC are increased by 2.3% and 2.1% respectively. It can be observed that the performance of DeepExon is better than DeepCellEss, first of all, the sequence acquisition feature is not enough, the position information of the exon is introduced into the model as a feature, which can greatly improve this; secondly, we notice that the benchmark paper adopts one-hot encoding, but the vector after one-hot encoding will be very sparse, which will cause a lot of calculation cost, and the continuity of the sequence and other features one-hot encoding cannot perceive, while the embedding encoding adopted in the present application can well solve the above problems. And the embedding layer in the embedding encoding can adjust and optimize according to the training condition, which makes the embedding encoding essentially better than the one-hot encoding method, and the improved embedding encoding of the present application can integrate more sequence features, which can enhance the understanding of protein sequence and improve the prediction performance of the model.

[0091] In the present embodiment, an embedding encoding ablation experiment is carried out, in order to prove that our improved embedding encoding has great contribution, we do the embedding encoding ablation experiment. First of all, our improved embedding encoding is fused with amino acid position information weight and dipeptide fusion probability weight, in order to prove the effectiveness of this method, we will remove the features one by one. According to the experimental results, we found that the performance indicators after removing the amino acid position information weight and the dipeptide fusion weight are as shown in Table 1.

[0092] Table 1. Performance indicators after removing amino acid position information weight and dipeptide fusion weight

[0093]

[0094] From Table 1 above, it can be seen that after removing the two parts of weight respectively, there is not small decrease in AUROC and AUPRC. In the present embodiment, the same model as the original paper is experimented, but the traditional embedding encoding is used instead of one-hot encoding, and the improved embedding encoding is used for experiment, and the experimental results are shown in Table 2:

[0095] Table 2. Experimental results of improved embedding encoding

[0096]

[0097]

[0098] As can be seen from Table 2 above, the embedding encoding is better than the one-hot encoding for the traditional embedding encoding, the embedding encoding can learn and reflect the similarity between the categories, especially in the protein research, the embedding encoding can capture the connection between the amino acids and the hidden information between the sequences. Secondly, by integrating the amino acid position weight feature, the model can better capture the unique role and function of the amino acids at different positions between the sequences, and the introduction of the dipeptide binding probability weight can help the model better understand the strength of the amino acid interaction.

[0099] In this embodiment, for the exon feature, since the exon is directly involved in and synthesizes the protein, the introduction of the exon feature is very effective for protein essentiality prediction. In order to prove this, we did an exon feature ablation experiment, which aims to prove the contribution of the exon. We respectively did the following experiments:

[0100] 1) We fixed the network structure of the sequence information, and stripped the exon feature extraction network.

[0101] 2) We masked the exon information in the sequence, and only kept the non-exon information.

[0102] Table 3. Exon feature ablation experiment results

[0103]

[0104] From the experimental results in Table 3 above, the present application can find that the model performance of the network with the exon feature is the best. The exon directly encodes the functional domain of the protein, and the functional domain determines the specific function of the protein in different cells and biological environments. By analyzing the exon coding region, the functional requirements of the protein in different cell types can be better understood.

[0105] In this embodiment, the present application is compared with the existing method. Specifically, the DeepExon is compared with other three protein essentiality prediction methods, which are EP-GBDT, EP-EDL, Pheg, and DeepCellEss. It can be seen that the DeepExon model adopted by the present application is better than the other methods in terms of AUROC and AUPRC. The contrast nucleotide sequence feature proposed in DeepCellEss, the present application can find that for this we continue to compare the DeepCellEss-nc based on the nucleotide sequence feature. It can be seen that the DeepExon is still better than the model based on the nucleotide sequence. See Table 4 below for details:

[0106] Table 4. Comparison with existing methods

[0107]

[0108]

[0109] In summary, the present application collects the exon position of the protein and combines the embedding coding and other features to predict the necessity of the protein in different cell lines, which can make full use of auxiliary features to help the model better predict the protein necessity in different environments.

[0110] According to the original data, the present application collects the corresponding exon information, adopts the embedding coding mode, and fuses sequence position information, amino acid position information and other features to perform prediction. Compared with the DeepCellEss model, the present application improves the AUC by 2.3% and the AUPRC by 2.1%, which proves that the exon has an important effect on the prediction of necessary proteins in different cell lines. In addition, the present application also performs feature ablation experiments to prove that the feature mode proposed by the present application has outstanding contribution.

[0111] The present application proposes a DeepExon model, which is a deep learning framework improved from DeepCellEss, and is used for predicting cell line-specific essential proteins. The present application collects 16408 protein exon position information, and combines sequence information, amino acid position information and other features to assist in prediction. Through a large number of experiments, the DeepExon model proposed by the present application has a significant improvement compared with the DeepCellEss model.

[0112] The present application improves the traditional embedding coding and proposes to use exon information to assist in predicting the necessity of proteins in different cell lines. For the traditional embedding coding, the present application proposes to combine amino acid position weight and dipeptide combination probability weight into the coding. We have made different ablation experiments to verify the effectiveness of the improved embedding coding. Secondly, we propose to use exon information as another part of the feature to assist in predicting the necessity. We have made different experiments to verify that the exon has excellent effect on the prediction of protein necessity. In addition, the present application also performs independent training on 323 cell line benchmark data sets to verify that the model of the present application can better predict the necessity of proteins in different cell lines.

[0113] The present application solves the technical problems of not fully considering cell line specificity and insufficient information mining in the prior art, which leads to low prediction accuracy and efficiency of essential proteins.

[0114] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting protein essentiality under different cell lines, characterized in that, The method comprises: S1, collecting protein sequence data; S2, according to the protein sequence data, collecting exon information, processing to obtain labeled exon position data set through labeling operation, using the exon information to assist in predicting the necessity of different cell line proteins; S3, setting a DeepExon network structure, wherein the DeepExon network structure comprises: an encoding module, a CNN network, a multi-head self-attention module and a prediction module; wherein in the encoding operation of the encoding module, the amino acid position weight and the dipeptide combination probability weight are fused; the CNN network is used to extract local features from the protein data and the exon information; the sequence information features and exon features in the local features are spliced to obtain spliced features; Bi-LSTM is used to capture and store long-term information from the spliced features; S4, using the prediction module, predicting the protein necessity prediction result according to the long-term information.

2. The method for predicting protein essentiality in different cell lines according to claim 1, wherein, In S1, the dataset is obtained by removing redundant sequences at the protein level and nucleotide level using CD-HIT and CD-HIT-est, wherein the sequence dataset comprises: cell line, protein binary essentiality label, sequence information.

3. The method of claim 2, wherein the method is for predicting essentiality of a protein in different cell lines. In S2, according to the consensus encoding provided in the applicable sequence dataset, the exon information corresponding to each protein sequence data is found on NCBI; According to the exon position in the exon information, the applicable sequence dataset is combined to generate the labeled exon position dataset.

4. The method of claim 1, wherein the method is for predicting essentiality of a protein in different cell lines. In the DeepExon network structure in S3, the encoding module comprises adaptive embedding encoding and one_hot encoding; In the encoding operation, the non-exon part is shielded, and the one_hot encoding is used for the exon part to obtain an amino acid fixed feature matrix for highlighting the exon information; For the protein sequence data, the adaptive embedding encoding is used, including a learnable embedding layer and a position embedding, and the position embedding is used to obtain the amino acid sequence in the protein sequence data, so as to understand the sequence information and perform model enhancement optimization operation; In the adaptive embedding encoding, the position weight of the amino acid and the dipeptide combination probability are fused to understand the sequence information; for the position weight of the amino acid, the occurrence frequency of the amino acid X at each position in the applicable sequence dataset is counted, and then the total amino acid occurrence frequency at the position is compared with the occurrence frequency of the amino acid X at the position to obtain the position weight of the amino acid X at the position; The CNN network is used to extract the local features from the protein sequence data and the exon information; CNN has strong feature extraction capability, and CNN has strong performance in many biological information prediction; however, the limitation of CNN is obvious, because CNN is only suitable for extracting local information of sequences; For extracting sequence information part, in the CNN network, different convolution kernels are used to extract sequence information from the protein sequence data in combination with a two-way CNN, and element-by-element multiplication is performed.

5. The method for predicting protein essentiality in different cell lines according to claim 4, wherein, In the adaptive embedding encoding, for the amino acid at the p-th position in the protein sequence data, the position embedding is represented as d. k The vector PE(p) is expressed using the following logic, where the i-th element of the vector PE(p) is: In the formula, 2i and 2i+1 represent odd dimension and even dimension respectively.

6. The method of predicting essentiality of a protein in different cell lines according to claim 4, wherein, The position weight of the amino acid X is determined by using the following logic: where L X represents the weight distribution of amino acid X, X i represents the number of occurrences of amino acid X at the i-th position, N represents the sequence length, L i represents the total number of amino acid occurrences at the i-th position.

7. The method of claim 4, wherein the method is for predicting essentiality of a protein in different cell lines. The combination probability of two amino acids is represented by using the dipeptide combination probability according to the following logic: In the formula, (X, Y) represents the combination probability of amino acid X and amino acid Y, A1 represents the total number of combinations of amino acid X and amino acid Y, and A2 represents the total number of combinations of amino acid X and other amino acids.

8. The method of predicting essentiality of proteins in different cell lines according to claim 4, wherein, The multi-head attention mechanism in the multi-head attention module includes no less than two self-attention mechanisms: In the formula, Q, K, V respectively represent query, key, value, W Q , W K , W V represent weight matrix, is dot product scale factor.

9. The method of claim 1, wherein the method is for predicting essentiality of a protein in different cell lines. In the S3, the Bi-LSTM includes a forward LSTM and a backward LSTM, and the forward LSTM and the backward LSTM are expressed by using the following logic:

10. A system for predicting protein necessity in different cell lines, characterized in that, The system includes: A sequence acquisition module is configured to collect protein sequence data; An exon information collection module is configured to collect exon information from the protein sequence data, obtain labeled exon position data sets through labeling operation processing, and use the exon information to assist in predicting protein essentiality of different cell lines. A DeepExon network feature processing module is configured to set and use a DeepExon network structure, wherein the DeepExon network structure includes an encoding module, a CNN network, a multi-head self-attention module, and a prediction module. An essentiality prediction module is configured to use the prediction module to predict protein essentiality prediction results based on the long-term information, and the essentiality prediction module is connected to the DeepExon network feature processing module.

Images