A method and device for predicting relative solvent accessibility of proteins based on sequence
By stacking a two-channel model combining a bidirectional long-term short-term memory network and a Transformer encoder, using the feature fusion of dot-product attention mechanism, the long-distance dependence and feature loss problems of bidirectional recurrent neural networks in the prediction of protein relative solvent accessibility are solved. More accurate predictions and wider adaptability to biological applications are achieved.
Patent Information
- Application Number
- CN202411114819.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-14
AI Technical Summary
In the prior art, there are long-distance dependence and feature loss problems when using bidirectional recurrent neural network to predict protein relative solvent accessibility, resulting in inaccurate prediction of prediction results. In the prior art, the prior art designs the calculation of protein relative solvent accessibility into classification problems. It cannot meet the needs of different biological application scenarios.
The stacked bidirectional long and short-term memory network module and the Transformer encoder are used to build a dual-channel model. The global semantic features and context features are fused through the dot-product attention mechanism feature fusion module to avoid information loss, and design prediction problems as regression tasks to adapt to more Biological application scenarios.
It improves the accuracy and prediction accuracy of protein relative solvent accessibility, can better characterize amino acid residues, adapt to more biological application scenarios, and achieve higher prediction accuracy.
Smart Images

Figure CN118629515B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of protein technology, and in particular to a method and device for predicting relative solvent accessibility of proteins based on sequences. Background Art
[0002] The accessible surface area (ASA) of a protein refers to the surface area of each amino acid residue that can be contacted by solvent molecules when the protein is immersed in a solution. The larger the surface area that the solvent molecules can contact, the more the residue is exposed on the protein surface. Conversely, the less exposed, which is an important feature to explain the effect of protein structure on its biological function. The calculation method of the solvent accessible surface area is to roll a probe ball with a set radius along the protein surface to obtain the surface area swept by the center of the probe ball. The specific value of the accessible surface area of each amino acid residue can be calculated using DSSP software.
[0003] However, since the sizes of the 20 natural amino acids are not consistent, their surface areas are also different, and the absolute value of the accessible surface area cannot compare the exposure of different types of amino acids. To solve this problem, Kabsch et al. proposed the concept of relative solvent accessibility (RSA) in 1983. RSA is a ratio value between 0 and 1 (including 0 and 1), which is calculated by dividing the ASA of the amino acid residue at each position by the maximum ASA of the 20 amino acids corresponding to the amino acid residue at that position. The calculation formula is: , , , where i represents the ith amino acid residue in a protein sequence of length N, j represents that the ith amino acid residue belongs to the jth class of amino acids among 20 amino acids, RSA(i) represents the relative solvent accessibility of the ith amino acid residue, ASA(i) represents the accessible surface area of the ith amino acid residue, and maxASA(j) represents the maximum accessible surface area of the jth class of amino acids.
[0004] The calculation of protein relative solvent accessibility needs to rely on its three-dimensional structural information. For example, for the protein sequence TEAEDSIEMYEW, the specific RSA values of each amino acid residue can be calculated based on the spatial structure of the sequence and the DSSP software: T (0.924), E (0.529), A (0.589), E (0.233), D (0.482), S (0.555), I (0.721), E (0.507), M (0.478), Y (0.673), E (0.202), W (0.211). At present, the three-dimensional structure of proteins is mainly determined by experimental methods, such as X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy and electron microscopy (EM). These experimental methods can obtain high-precision protein three-dimensional structures, but there are problems such as long time consumption and high cost. Therefore, the prediction of protein relative solvent accessibility based on computational methods has attracted widespread attention from researchers.
[0005] In the study of using computational methods to predict the relative solvent accessibility of proteins, two major strategies can usually be summarized. The first strategy is to set the prediction task as a classification problem; this method divides the amino acid residues at each site of the protein into two categories: "exposed" and "hidden" based on a preset threshold. This method reveals the exposure status of the amino acid residues through binary classification results. The second strategy is to model the prediction task as a regression problem; under this framework, the trained model predicts a continuous value between 0 and 1 for each amino acid residue in the protein sequence; this value directly reflects the exposure level of the corresponding residue, where larger values indicate higher exposure levels.
[0006] Among them, in the classification strategy, the SANN model proposed by Joo uses the K-nearest neighbor algorithm to complete the prediction of relative solvent accessibility binary classification, ternary classification and regression problems. The DeepCNF model proposed by Wang uses deep convolutional neural networks and conditional random fields to classify the relative solvent accessibility of residues. With the rise of recurrent neural networks in the field of natural language processing, researchers have tried to apply them to protein research, which also has sequence information. The SDBRN model proposed by Zhang et al. uses a bidirectional recurrent neural network, so that the residues at each position can obtain contextual information, thereby improving the prediction performance of the model. Kaleel et al. used a method combining bidirectional recurrent neural networks and convolutional neural networks to predict the relative solvent accessibility of residues in four categories and achieved good performance. Manfredi et al. proposed DeepREx, which also uses a bidirectional recurrent neural network and performs well in binary classification prediction. In the regression strategy, the existing technology mainly uses the architecture of a bidirectional recurrent neural network to capture the contextual information of the protein sequence, so that the amino acid residues at each position can fuse the information of the previous amino acid residues, so as to obtain a vector with contextual information to better characterize each amino acid residue.
[0007] However, the bidirectional recurrent neural network architecture mainly used in the prior art has the problem of long-distance dependence; when the length of the protein sequence is too long, the amino acid residues at the end of the protein chain often find it difficult to fully capture the information of the residues at the beginning of the chain during the encoding process. The contextual information that the amino acids at each position in the sequence can receive during the learning process becomes local and fragmented, and the integrity and continuity of the information are destroyed, making it impossible to accurately characterize the amino acid residues, which in turn leads to inaccurate classification results. In addition, the prior art has the problem of scattered attention when fusing the information of each module; the mainstream approach tends to adopt a simple strategy of directly adding or splicing two feature vectors. Although these methods are simple and fast in implementation, they inadvertently ignore the fine capture and enhancement of the key features of the amino acid residues. The addition operation may inadvertently smooth out some unique and subtle feature differences while merging the information; and although the splicing method retains the integrity of the information, it fails to effectively improve the model's ability to focus on the core features, resulting in the dilution of the amino acid residue information that is crucial to understanding the structure and function in the analysis of complex protein sequences in a large amount of background noise, which also makes the characterization of the amino acid residues inaccurate, and thus leads to inaccurate classification results. In addition, most existing technologies design the calculation of protein relative solvent accessibility as a classification problem. In different biological problems, biologists have different requirements for the exposure levels of amino acid residues, and classification problems cannot meet actual needs. This leads to limited generalization capabilities of existing technologies in different biological application scenarios.
[0008] In summary, the existing regression and classification-based strategies are all based on a model architecture based on a bidirectional recurrent neural network. Due to the existence of long-distance dependency problems and the fact that feature fusion does not consider the key features of amino acid residues, it is impossible to accurately characterize amino acid residues, which in turn leads to inaccurate classification or regression results, resulting in low accuracy in the relative solvent accessibility of proteins. Summary of the invention
[0009] To this end, the technical problem to be solved by the present invention is to overcome the problem of inaccurate prediction results caused by long-distance dependence and feature loss when predicting the relative solvent accessibility of proteins using a model architecture based on a bidirectional recurrent neural network in the prior art.
[0010] In order to solve the above technical problems, the present invention provides a sequence-based protein relative solvent accessibility prediction method, comprising:
[0011] Obtaining a protein sequence of a protein to be predicted, and extracting a feature vector of the protein to be predicted based on the protein sequence;
[0012] Input the feature vector of the protein to be predicted into the pre-trained deep learning model and obtain the output regression value as the relative solvent accessibility of the protein to be predicted, including:
[0013] Inputting the feature vector of the protein to be predicted into a stacked bidirectional long short-term memory network module, and outputting the global semantic feature; the stacked bidirectional long short-term memory network module comprises a plurality of layers of bidirectional long short-term memory network units connected in series in a forward propagation direction;
[0014] Inputting the feature vector of the protein to be predicted into a Transformer encoder and outputting context features; the Transformer encoder includes a multi-head attention mechanism, a first residual connection and layer normalization, a feedforward network, a second residual connection and layer normalization, which are sequentially connected in series along the forward propagation direction;
[0015] The global semantic features and the context features are input into a dot product attention mechanism feature fusion module, and a fused feature vector is output; the dot product attention mechanism feature fusion module comprises a first MatMul unit, a Scale unit, a SoftMax unit and a second MatMul unit which are sequentially connected in series along the forward propagation direction;
[0016] The fused feature vector is input into the output layer, and a regression value is output; the output layer comprises a linear transformation layer and an activation function layer which are sequentially connected in series along a forward propagation direction.
[0017] Preferably, the stacked bidirectional long short-term memory network module comprises: a first bidirectional long short-term memory network unit, a second bidirectional long short-term memory network unit and a third bidirectional long short-term memory network unit;
[0018] Inputting the feature vector of the protein to be predicted into the first bidirectional long short-term memory network unit, obtaining the forward output and the reverse output, and then splicing them to obtain the first output;
[0019] Input the first output into the second bidirectional long short-term memory network unit, obtain the forward output and the reverse output, and then add them bit by bit to obtain the second output;
[0020] The second output is input into the third bidirectional long short-term memory network unit, and the forward output and the reverse output are obtained and then added bit by bit to obtain the global semantic feature.
[0021] Preferably, the step of inputting the feature vector of the protein to be predicted into a Transformer encoder and outputting context features comprises:
[0022] The feature vector of the protein to be predicted is mapped to multiple different spaces. After performing self-attention operation in each space, the outputs of all spaces are concatenated and linearly transformed to obtain multi-head attention output features.
[0023] After performing residual connection on the feature vector of the protein to be predicted and the multi-head attention output feature, normalization is performed to obtain the normalized output;
[0024] Perform nonlinear transformation on the normalized output to obtain the feedforward output;
[0025] The normalized output is residually connected to the feedforward output and then normalized to obtain context features.
[0026] Preferably, the step of inputting the global semantic features and the contextual features into a dot product attention mechanism feature fusion module and outputting a fused feature vector comprises:
[0027] The global semantic features are used as the key vector K and the value vector V, and the context features are used as the query vector Q; the key vector K and the query vector Q are input into the first MatMul unit to perform matrix multiplication, and the similarity matrix is output;
[0028] Input the similarity matrix into the Scale unit, scale each value in the similarity matrix based on a preset scaling factor, and output the scaled matrix;
[0029] Convert the scaling matrix input into the SoftMax unit into a probability distribution and output a weight matrix;
[0030] The weight matrix and the value vector V are input into the second MatMul unit for multiplication, and then weighted summation is performed to output the fused feature vector.
[0031] Preferably, based on the information text of the protein to be predicted including header information and sequence content, the protein sequence of the protein to be predicted in fasta format is obtained.
[0032] Preferably, obtaining the feature vector of the protein to be predicted includes:
[0033] Psi-Blast was used to calculate the position-specific matrix of the protein to be predicted;
[0034] Based on the AAindex database, the physicochemical characteristics of the amino acid residues at each site in the protein sequence of the protein to be predicted are obtained, and the length and relative position information of the sequence of the amino acid residues at each site are calculated;
[0035] Obtaining the number of amino acid species and the number of amino acid types within a preset window at each site in the protein sequence of the protein to be predicted;
[0036] The position-specific matrix of the protein to be predicted, the physicochemical characteristics of the amino acid residues at each site, the length and relative position information of the sequence of the amino acid residues at each site, and the number of amino acid species and the number of amino acid types within the preset window of each site are horizontally spliced to obtain the feature vector of the protein to be predicted.
[0037] Preferably, the size of the position-specific matrix of the protein to be predicted is n×20, where n represents the length of the protein sequence of the protein to be predicted, and a vector of length 20 is generated for the amino acid residues at each site of the protein sequence.
[0038] Preferably, the total number of amino acid types is 5, including polar, non-polar, positively charged, negatively charged and uncharged.
[0039] Preferably, the physicochemical characteristics include physical properties, chemical properties and kinetic properties;
[0040] The physical properties include hydrogen bonding ability, polarity, solubility and molecular weight;
[0041] The chemical properties include acidity, stability, reactivity and lipophilicity;
[0042] The dynamic properties include vibrational frequencies, free energy changes, and thermodynamic stability.
[0043] This embodiment also provides a sequence-based protein relative solvent accessibility prediction device, comprising:
[0044] A feature vector construction module is used to obtain the protein sequence of the protein to be predicted and extract the feature vector of the protein to be predicted based on the protein sequence;
[0045] The prediction module is used to input the feature vector of the protein to be predicted into a pre-trained deep learning model to obtain the output regression value as the relative solvent accessibility of the protein to be predicted, including: inputting the feature vector of the protein to be predicted into a stacked bidirectional long short-term memory network module to output global semantic features; the stacked bidirectional long short-term memory network module includes multiple layers of bidirectional long short-term memory network units connected in series in a forward propagation direction; inputting the feature vector of the protein to be predicted into a Transformer encoder to output context features; the Transformer encoder includes a multi-head attention mechanism, a first residual connection and layer normalization, a feedforward network, a second residual connection and layer normalization connected in series in a forward propagation direction; inputting the global semantic features and the context features into a dot product attention mechanism feature fusion module to output a fused feature vector; the dot product attention mechanism feature fusion module includes a first MatMul unit, a Scale unit, a SoftMax unit and a second MatMul unit connected in series in a forward propagation direction; inputting the fused feature vector into an output layer to output a regression value; the output layer includes a linear transformation layer and an activation function layer connected in series in a forward propagation direction.
[0046] The above technical solution of the present invention has the following beneficial effects compared with the prior art:
[0047] The sequence-based protein relative solvent accessibility prediction method described in the present invention is based on a stacked bidirectional long short-term memory network module and a Transformer encoder to construct a dual-channel model; the stacked bidirectional long short-term memory network module is used to capture the temporal and contextual information of the feature vector of the protein to be predicted, and the global features of the feature vector of the protein to be predicted by the Transformer encoder are used. The outputs of the two channels are fused using a dot product attention mechanism feature fusion module, combining globality and temporality to avoid the problem of information loss caused by long-distance dependence, and obtaining a fused feature vector, so that the deep learning model is more focused on the effective feature part, and can better characterize the amino acid residues and the protein to be predicted, thereby ensuring the accuracy of the relative solvent accessibility of the protein to be predicted obtained by regression. In addition, the present invention designs the protein relative solvent accessibility prediction problem into a regression task, which is more adaptable, so that the present invention can adapt to more biological application scenarios and achieve higher prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein
[0049] Figure 1 is a flowchart of the steps of the method for predicting relative solvent accessibility of proteins based on sequences provided by the present invention;
[0050] Figure 2 It is a flow chart of the method for predicting relative solvent accessibility of proteins based on the prediction system provided by the present invention;
[0051] Figure 3 It is a network structure diagram of the deep learning model SolAcc provided by the present invention. DETAILED DESCRIPTION
[0052] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.
[0053] Reference Figure 1 As shown, the flowchart of the method for predicting relative solvent accessibility of proteins based on sequences of the present invention is shown; the specific steps include:
[0054] S101: obtaining a protein sequence of a protein to be predicted, and extracting a feature vector of the protein to be predicted based on the protein sequence;
[0055] S102: Inputting the feature vector of the protein to be predicted into a pre-trained deep learning model, obtaining an output regression value as the relative solvent accessibility of the protein to be predicted, including:
[0056] S102-1: inputting the feature vector of the protein to be predicted into a stacked bidirectional long short-term memory network module, and outputting a global semantic feature; the stacked bidirectional long short-term memory network module comprises a plurality of layers of bidirectional long short-term memory network units connected in series in a forward propagation direction;
[0057] S102-2: Inputting the feature vector of the protein to be predicted into a Transformer encoder and outputting context features; the Transformer encoder includes a multi-head attention mechanism, a first residual connection and layer normalization, a feedforward network, a second residual connection and layer normalization, which are sequentially connected in series along the forward propagation direction;
[0058] S102-3: inputting the global semantic features and the context features into a dot product attention mechanism feature fusion module, and outputting a fused feature vector; the dot product attention mechanism feature fusion module comprises a first MatMul unit, a Scale unit, a SoftMax unit and a second MatMul unit which are sequentially connected in series along the forward propagation direction;
[0059] S102-4: Input the fused feature vector into the output layer and output the regression value; the output layer includes a linear transformation layer and an activation function layer which are connected in series in the forward propagation direction.
[0060] Specifically, the stacked bidirectional long short-term memory network module includes: a first bidirectional long short-term memory network unit, a second bidirectional long short-term memory network unit and a third bidirectional long short-term memory network unit; wherein, after inputting the feature vector of the protein to be predicted, it includes:
[0061] Inputting the feature vector of the protein to be predicted into the first bidirectional long short-term memory network unit, obtaining the forward output and the reverse output, and then splicing them to obtain the first output;
[0062] Input the first output into the second bidirectional long short-term memory network unit, obtain the forward output and the reverse output, and then add them bit by bit to obtain the second output;
[0063] The second output is input into the third bidirectional long short-term memory network unit, and the forward output and the reverse output are obtained and then added bit by bit to obtain the global semantic feature.
[0064] Specifically, the feature vector of the protein to be predicted is input into the Transformer encoder, and the context features are output, including:
[0065] The feature vector of the protein to be predicted is mapped to multiple different spaces. After performing self-attention operation in each space, the outputs of all spaces are concatenated and linearly transformed to obtain multi-head attention output features.
[0066] After performing residual connection on the feature vector of the protein to be predicted and the multi-head attention output feature, normalization is performed to obtain the normalized output;
[0067] Perform nonlinear transformation on the normalized output to obtain the feedforward output;
[0068] The normalized output is residually connected to the feedforward output and then normalized to obtain context features.
[0069] Specifically, the global semantic features and context features are input into the dot product attention mechanism feature fusion module, and the fused feature vector is output, including:
[0070] The global semantic features are used as the key vector K and the value vector V, and the context features are used as the query vector Q; the key vector K and the query vector Q are input into the first MatMul unit to perform matrix multiplication, and the similarity matrix is output;
[0071] Input the similarity matrix into the Scale unit, scale each value in the similarity matrix based on a preset scaling factor, and output the scaled matrix;
[0072] Convert the scaling matrix input into the SoftMax unit into a probability distribution and output a weight matrix;
[0073] The weight matrix and the value vector V are input into the second MatMul unit for multiplication, and then weighted summation is performed to output the fused feature vector.
[0074] Specifically, in an embodiment of the present invention, based on the information text of the protein to be predicted including the header information and the sequence content, the protein sequence of the protein to be predicted in the fasta format is obtained. Based on the protein sequence of the protein to be predicted, the corresponding feature vector is obtained, including:
[0075] Psi-Blast is used to calculate the position-specific matrix of the protein to be predicted; the size of the position-specific matrix of the protein to be predicted is n×20, n represents the length of the protein sequence of the protein to be predicted, and the amino acid residues at each site of the protein sequence generate a vector with a length of 20;
[0076] Based on the AAindex database, the physicochemical characteristics of the amino acid residues at each site in the protein sequence of the protein to be predicted are obtained, and the length and relative position information of the sequence of the amino acid residues at each site are calculated;
[0077] Obtain the number of amino acid species and the number of amino acid types in the preset window for each site in the protein sequence of the protein to be predicted; the total number of amino acid types is 5, including polar, non-polar, positively charged, negatively charged and uncharged; the number of amino acid types is 20;
[0078] The position-specific matrix of the protein to be predicted, the physicochemical characteristics of the amino acid residues at each site, the length and relative position information of the sequence of the amino acid residues at each site, and the number of amino acid species and the number of amino acid types within the preset window of each site are horizontally spliced to obtain the feature vector of the protein to be predicted.
[0079] Among them, the physicochemical characteristics include physical properties, chemical properties and kinetic properties; the physical properties include hydrogen bonding ability, polarity, solubility and molecular weight; the chemical properties include acidity and alkalinity, stability, reactivity and lipophilicity; the kinetic properties include vibration frequency, free energy change and thermodynamic stability.
[0080] The sequence-based protein relative solvent accessibility prediction method described in the present invention is based on a stacked bidirectional long short-term memory network module and a Transformer encoder to construct a dual-channel model; the stacked bidirectional long short-term memory network module is used to capture the temporal and contextual information of the feature vector of the protein to be predicted, and the global features of the feature vector of the protein to be predicted by the Transformer encoder are used. The outputs of the two channels are fused using a dot product attention mechanism feature fusion module, combining globality and temporality to avoid the problem of information loss caused by long-distance dependence, and obtaining a fused feature vector, so that the deep learning model is more focused on the effective feature part, and can better characterize the amino acid residues and the protein to be predicted, thereby ensuring the accuracy of the relative solvent accessibility of the protein to be predicted obtained by regression. In addition, the present invention designs the protein relative solvent accessibility prediction problem into a regression task, which is more adaptable, so that the present invention can adapt to more biological application scenarios and achieve higher prediction accuracy.
[0081] Based on the above embodiments, in an embodiment of the present invention, a prediction system including a protein information input module, a protein sequence feature calculation module, a protein relative solvent accessibility prediction module and a prediction result output module is used to predict the relative solvent accessibility of a protein. Figure 2 As shown, it is a flowchart of the method for predicting relative solvent accessibility of proteins based on the prediction system; the prediction steps are as follows:
[0082] S201: The protein information input module first obtains the protein sequence according to the protein sequence information input by the user, and then the protein information input module transmits the obtained protein sequence data to the protein sequence feature calculation module. At the same time, all the input data and the user information of the submitted data will be stored by the system;
[0083] S202: After receiving the protein sequence data, the protein sequence feature calculation module calculates the protein position specific matrix (PSSM) through Psi-Blast software, obtains the physicochemical characteristics of the amino acid residues at each site through the AAindex database, and counts the number of 20 amino acids and five types of amino acids (polar, non-polar, positively charged, negatively charged, and uncharged) in a window of size 21 at each site; then, the protein sequence feature calculation module transmits the extracted protein sequence attribute feature value to the protein relative solvent accessibility prediction module;
[0084] The PSSM reflects the frequency of amino acids at each position being replaced by other amino acids; the AAindex database is a database containing various amino acid indices (such as hydrophobicity, charge, etc.);
[0085] S203: After receiving the protein sequence attribute feature value, the module for predicting relative solvent accessibility of proteins will extract 50 features that have been proven to be excellent in experiments from the calculated sequence attribute feature values. The specific values of the amino acid residues at each site will be calculated by the trained deep learning model SolAcc, and the calculated prediction results will be transmitted to the prediction result output module.
[0086] S204: After receiving the test results, the prediction result output module first stores the prediction results, then displays the prediction results on the browser page, generates a PDF file and draws the prediction results into a line graph, and sends an email to the corresponding user who submitted the data according to the task.
[0087] Reference Figure 3 The figure shows the network structure of the deep learning model SolAcc provided by the present invention. SolAcc uses a BiLSTM neural network module consisting of three stacked layers and a Transformer encoder module. This dual-channel strategy includes a global semantic channel and a context information channel, which respectively maps the 50-dimensional input to a 128-dimensional high-level semantic space, generating a global semantic feature transformer_o (128-dimensional) and a context feature bilstm_o (128-dimensional).
[0088] To further integrate these two sets of abstract features, the dot product attention mechanism is introduced, using transformer_o as the key vector (key) and value vector (value), and bilstm_o as the query vector (query). Through the dynamic weighting process, the two sets of features are effectively fused, and finally a comprehensive 128-dimensional feature vector (attention_o) is obtained. In the final output stage of the model, the above fused features (attention_o) are dimensionally mapped through a linear transformation layer, and the Sigmoid activation function is applied to make the output value between 0 and 1 (including 0 and 1), indicating the specific value of the relative solvent accessibility of the amino acid residue.
[0089] The BiLSTM model used in the embodiment of the present invention can capture the timing and contextual information of the input data, but due to the long-distance dependency problem, when the input sequence is too long, the amino acid residues at the tail end will lack the information of the amino acid residues at the front end. The Transformer encoder can capture the global features of the input data, but there is disorder and the relative position relationship between the amino acid residues is not taken into account. We combine the features captured by these two deep learning models to make up for their respective shortcomings and obtain ordered global context features. The dot product attention mechanism feature fusion is different from the traditional additive and splicing feature fusion. This embodiment uses the dot product attention mechanism to fuse the abstract semantic features extracted by the Stacked BiLSTM Block and the Transformer encoder, so that the model pays more attention to information that is positively correlated with the prediction results. In addition, the present invention designs the prediction of relative solvent accessibility as a regression task, so that it can cope with more biological application scenarios.
[0090] Based on the above embodiments, in the embodiments of the present invention, in order to verify the performance of the deep learning model SolAcc provided by the present invention, it is compared with other prediction tools on the test set SATS500, including regression predictors: SPIDER, NetSufP, and classification predictors: DeepREx and PaleAle5.0. The regression performance comparison results of SolAcc, SPIDER, and NetSufP are shown in Table 1:
[0091] Table 1 Comparison of regression performance of SolAcc, SPIDER3, and NetSufP-2.0 on the test set
[0092]
[0093] According to the experimental results, SolAcc has good performance in MAE, MSE, RMSE, MSLE and R 2 It is superior to other existing regression predictors in multiple indicators such as and so on. The value of relative solvent accessibility is between 0 and 1, and there will not be a large deviation between the true value and the predicted value. Therefore, the MAE indicator can better reflect the performance of each prediction tool. On SATS500, SolAcc's MAE indicator is 10.29%, which is 6.44% higher than SPIDER's performance and 2.43% higher than NetSufP, indicating that SolAcc's prediction results are closer to the actual values. The regression prediction results of SolAcc and other regression prediction tools are converted into classification prediction results, and compared with DeepREx and PaleAle5.0. The results are shown in Tables 2 to 4. Under different thresholds, SolAcc is superior to existing predictors in multiple classification indicators, further verifying the effectiveness of SolAcc.
[0094] Table 2 Comparison of SolAcc with other prediction tools at a classification threshold of 20%
[0095]
[0096] According to the results shown in the table, when the threshold is set to 20%, SolAcc's classification performance on the SATS500 dataset is better than or close to other existing prediction tools. Specifically, SolAcc's accuracy (ACC) is improved by 0.9% compared to other existing prediction tools; the recall rate (Recall) is improved by 1.1% compared to other existing prediction tools; and the F1 score (F1) is improved by 2.4% compared to other existing prediction tools. In addition, the Matthews correlation coefficient (MCC) also shows a trend of being better than similar tools, which further verifies SolAcc's overall performance and robustness in solving this problem.
[0097] Table 3 Comparison of SolAcc with other prediction tools at a classification threshold of 25%
[0098]
[0099] When the threshold is set to 25%, SolAcc's recall rate (Recall) is 2.5% higher than other existing prediction tools; the F1 score (F1) is 0.7% higher than other existing prediction tools. In addition, ACC and NPV are also close to other existing prediction tools, ranking second.
[0100] Table 4 Comparison of SolAcc with other prediction tools at a classification threshold of 50%
[0101]
[0102] When the threshold is set at 50%, SolAcc's accuracy (ACC) is improved by 6.7% compared with other existing prediction tools; the negative predictive value (NPV) is improved by 2.2% compared with other existing prediction tools; the recall rate (Recall) is improved by 0.3% compared with other existing prediction tools; and the F1 score (F1) is improved by 3.5% compared with other existing prediction tools.
[0103] Based on the above embodiments, the present invention also provides a sequence-based protein relative solvent accessibility prediction device, which specifically includes:
[0104] A feature vector construction module is used to obtain the protein sequence of the protein to be predicted and extract the feature vector of the protein to be predicted based on the protein sequence;
[0105] The prediction module is used to input the feature vector of the protein to be predicted into a pre-trained deep learning model to obtain the output regression value as the relative solvent accessibility of the protein to be predicted, including: inputting the feature vector of the protein to be predicted into a stacked bidirectional long short-term memory network module to output global semantic features; the stacked bidirectional long short-term memory network module includes multiple layers of bidirectional long short-term memory network units connected in series in a forward propagation direction; inputting the feature vector of the protein to be predicted into a Transformer encoder to output context features; the Transformer encoder includes a multi-head attention mechanism, a first residual connection and layer normalization, a feedforward network, a second residual connection and layer normalization connected in series in a forward propagation direction; inputting the global semantic features and the context features into a dot product attention mechanism feature fusion module to output a fused feature vector; the dot product attention mechanism feature fusion module includes a first MatMul unit, a Scale unit, a SoftMax unit and a second MatMul unit connected in series in a forward propagation direction; inputting the fused feature vector into an output layer to output a regression value; the output layer includes a linear transformation layer and an activation function layer connected in series in a forward propagation direction.
[0106] The sequence-based protein relative solvent accessibility prediction device of the present embodiment is used to implement the aforementioned sequence-based protein relative solvent accessibility prediction method. Therefore, the specific implementation of the sequence-based protein relative solvent accessibility prediction device can be seen in the embodiment section of the sequence-based protein relative solvent accessibility prediction method in the previous text. For example, the feature vector construction module 100 and the prediction module 200 are respectively used to implement steps S101 and S102 in the aforementioned sequence-based protein relative solvent accessibility prediction method. Therefore, its specific implementation can refer to the description of the corresponding embodiments of each part, which will not be repeated here.
[0107] The sequence-based protein relative solvent accessibility prediction method described in the present invention is based on a stacked bidirectional long short-term memory network module and a Transformer encoder to construct a dual-channel model; the stacked bidirectional long short-term memory network module is used to capture the temporal and contextual information of the feature vector of the protein to be predicted, and the global features of the feature vector of the protein to be predicted by the Transformer encoder are used. The outputs of the two channels are fused using a dot product attention mechanism feature fusion module, combining globality and temporality to avoid the problem of information loss caused by long-distance dependence, and obtaining a fused feature vector, so that the deep learning model is more focused on the effective feature part, and can better characterize the amino acid residues and the protein to be predicted, thereby ensuring the accuracy of the relative solvent accessibility of the protein to be predicted obtained by regression. In addition, the present invention designs the protein relative solvent accessibility prediction problem into a regression task, which is more adaptable, so that the present invention can adapt to more biological application scenarios and achieve higher prediction accuracy.
[0108] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0109] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0110] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0111] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0112] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.
Claims
1. A sequence-based method for predicting relative solvent accessibility of proteins, characterized in that: include: Obtain the protein sequence of the protein to be predicted, and use Psi-Blast to calculate the position-specific matrix of the protein to be predicted; Based on the AAindex database, the physicochemical characteristics of the amino acid residues at each site in the protein sequence of the protein to be predicted are obtained, and the length and relative position information of the sequence of the amino acid residues at each site are calculated; Obtaining the number of amino acid species and the number of amino acid types at each site in a preset window in the protein sequence of the protein to be predicted; horizontally splicing the position-specific matrix of the protein to be predicted, the physicochemical characteristics of the amino acid residues at each site, the length and relative position information of the sequence where the amino acid residues at each site are located, and the number of amino acid species and the number of amino acid types at each site in the preset window to obtain a feature vector of the protein to be predicted; Input the feature vector of the protein to be predicted into the pre-trained deep learning model and obtain the output regression value as the relative solvent accessibility of the protein to be predicted, including: Inputting the feature vector of the protein to be predicted into a stacked bidirectional long short-term memory network module to output a global semantic feature; the stacked bidirectional long short-term memory network module includes multiple layers of bidirectional long short-term memory network units sequentially connected in series along the forward propagation direction, including: a first bidirectional long short-term memory network unit, a second bidirectional long short-term memory network unit, and a third bidirectional long short-term memory network unit; inputting the feature vector of the protein to be predicted into the first bidirectional long short-term memory network unit, obtaining a forward output and a reverse output, and then splicing them to obtain a first output; inputting the first output into the second bidirectional long short-term memory network unit, obtaining a forward output and a reverse output, and then bitwise adding them to obtain a second output; inputting the second output into the third bidirectional long short-term memory network unit, obtaining a forward output and a reverse output, and then bitwise adding them to obtain a global semantic feature; Input the feature vector of the protein to be predicted into the Transformer encoder and output context features; the Transformer encoder includes a multi-head attention mechanism, a first residual connection and layer normalization, a feedforward network, and a second residual connection and layer normalization that are sequentially connected in series along the forward propagation direction; map the feature vector of the protein to be predicted to multiple different spaces, perform self-attention operations in each space, concatenate the outputs of all spaces and perform linear transformation to obtain multi-head attention output features; perform residual connection on the feature vector of the protein to be predicted and the multi-head attention output features, and then perform normalization to obtain normalized output; perform nonlinear transformation on the normalized output to obtain feedforward output; perform residual connection on the normalized output and the feedforward output, and then perform normalization to obtain context features; The global semantic features and the contextual features are input into the dot product attention mechanism feature fusion module, and a fused feature vector is output; the dot product attention mechanism feature fusion module includes a first MatMul unit, a Scale unit, a SoftMax unit and a second MatMul unit connected in series along the forward propagation direction; the global semantic features are used as the key vector K and the value vector V, and the contextual features are used as the query vector Q; the key vector K and the query vector Q are input into the first MatMul unit to perform matrix multiplication, and a similarity matrix is output; the similarity matrix is input into the Scale unit, and each value in the similarity matrix is scaled based on a preset scaling factor, and a scaling matrix is output; the scaling matrix is input into the SoftMax unit to convert it into a probability distribution, and a weight matrix is output; the weight matrix and the value vector V are input into the second MatMul unit for multiplication, and then a weighted sum is performed to output a fused feature vector; The fused feature vector is input into the output layer, and a regression value is output; the output layer includes a linear transformation layer and an activation function layer which are sequentially connected in series along the forward propagation direction.
2. The method for predicting relative solvent accessibility of proteins based on sequence according to claim 1, characterized in that: Based on the information text of the protein to be predicted including header information and sequence content, the protein sequence of the protein to be predicted in the fasta format is obtained.
3. The method for predicting relative solvent accessibility of proteins based on sequence according to claim 1, characterized in that: The size of the position-specific matrix of the protein to be predicted is n×20, where n represents the length of the protein sequence of the protein to be predicted, and a vector of length 20 is generated for the amino acid residues at each site in the protein sequence.
4. The method for predicting relative solvent accessibility of proteins based on sequence according to claim 1, characterized in that: There are five types of amino acids in total, including polar, nonpolar, positively charged, negatively charged, and uncharged.
5. The method for predicting relative solvent accessibility of proteins based on sequence according to claim 1, characterized in that: The physicochemical characteristics include physical properties, chemical properties and kinetic properties; The physical properties include hydrogen bonding ability, polarity, solubility and molecular weight; The chemical properties include acidity, stability, reactivity and lipophilicity; The dynamic properties include vibrational frequency, free energy change and thermodynamic stability.
6. A sequence-based protein relative solvent accessibility prediction device, characterized in that: include: The feature vector construction module is used to obtain the protein sequence of the protein to be predicted and calculate the position-specific matrix of the protein to be predicted using Psi-Blast; Based on the AAindex database, the physicochemical characteristics of the amino acid residues at each site in the protein sequence of the protein to be predicted are obtained, and the length and relative position information of the sequence of the amino acid residues at each site are calculated; Obtaining the number of amino acid species and the number of amino acid types at each site in a preset window in the protein sequence of the protein to be predicted; horizontally splicing the position-specific matrix of the protein to be predicted, the physicochemical characteristics of the amino acid residues at each site, the length and relative position information of the sequence where the amino acid residues at each site are located, and the number of amino acid species and the number of amino acid types at each site in the preset window to obtain a feature vector of the protein to be predicted; The prediction module is used to input the feature vector of the protein to be predicted into a pre-trained deep learning model, and obtain the output regression value as the relative solvent accessibility of the protein to be predicted, including: inputting the feature vector of the protein to be predicted into a stacked bidirectional long short-term memory network module, and outputting global semantic features; the stacked bidirectional long short-term memory network module includes a multi-layer bidirectional long short-term memory network unit connected in series along the forward propagation direction, including: a first bidirectional long short-term memory network unit, a second bidirectional long short-term memory network unit and a third bidirectional long short-term memory network unit; inputting the feature vector of the protein to be predicted into the first bidirectional long short-term memory network unit, obtaining the forward output and the reverse output and then splicing them to obtain the first Output; input the first output into the second bidirectional long short-term memory network unit, obtain the forward output and the reverse output and add them bit by bit to obtain the second output; input the second output into the third bidirectional long short-term memory network unit, obtain the forward output and the reverse output and add them bit by bit to obtain the global semantic feature; input the feature vector of the protein to be predicted into the Transformer encoder and output the context feature; the Transformer encoder includes a multi-head attention mechanism, a first residual connection and layer normalization, a feedforward network, a second residual connection and layer normalization, which are sequentially connected in series along the forward propagation direction; map the feature vector of the protein to be predicted to multiple different spaces, perform self-attention operations in each space, and then The spatial outputs are concatenated and then linearly transformed to obtain multi-head attention output features; the feature vector of the protein to be predicted is residually connected with the multi-head attention output features, and then normalized to obtain normalized output; the normalized output is nonlinearly transformed to obtain feedforward output; the normalized output is residually connected with the feedforward output, and then normalized to obtain context features; the global semantic features and the context features are input into the dot product attention mechanism feature fusion module, and a fused feature vector is output; the dot product attention mechanism feature fusion module includes a first MatMul unit, a Scale unit, a SoftMax unit and a second MatMul unit that are sequentially connected in series along the forward propagation direction; The global semantic features are the key vector K and the value vector V, and the context feature is the query vector Q; the key vector K and the query vector Q are input into the first MatMul unit to perform matrix multiplication and output a similarity matrix; the similarity matrix is input into the Scale unit, and each value in the similarity matrix is scaled based on a preset scaling factor, and the scaling matrix is output; the scaling matrix is input into the SoftMax unit to convert it into a probability distribution and output a weight matrix; the weight matrix and the value vector V are input into the second MatMul unit for multiplication, and then weighted summation is performed to output a fused feature vector; the fused feature vector is input into the output layer and a regression value is output; the output layer includes a linear transformation layer and an activation function layer connected in series along the forward propagation direction.
