A method for predicting intrinsically disordered proteins

Abstract

A method for predicting intrinsic disordered proteins, belonging to the field of bioinformatics, is a deep neural network structure for identifying intrinsic disordered protein regions. The method uses three sequence features and other existing protein sequence features as feature matrices, and uses VGG199 variants and two MLP networks to construct a deep neural network to achieve relatively accurate prediction of intrinsic disordered proteins. In this structure, the VGG19 variant is located between the two MLP networks. Three novel sequence features, namely persistent entropy, correlation probability features of two consecutive amino acids, and correlation probability features of three consecutive amino acids, are used for the first time to identify intrinsic disordered protein regions. The method uses VGG199 variants and two MLP networks to construct a deep neural network MLP‑VGG19‑MLP to achieve relatively accurate prediction of intrinsic disordered proteins.

Description

Technical Field

[0001] The invention belongs to the field of bioinformatics and relates to a prediction scheme for intrinsic disordered proteins using novel features. Background Art

[0002] The region without a stable three-dimensional structure in a protein is called a disordered region of the protein, and a protein containing at least one disordered region is called an intrinsically disordered protein. Due to its flexible structure, it is widely involved in important physiological processes, such as nucleic acid folding, DNA transcription and translation, molecular recognition and molecular assembly, and protein-protein interactions. In addition, many human diseases, such as cancer, Alzheimer's disease, and some genetic diseases are related to it. Therefore, the rapid and accurate prediction of intrinsically disordered proteins is of great significance for the understanding of protein function, drug development, and the treatment of some diseases. In the past, there were many experimental methods used to determine the disordered regions of proteins, but the experimental methods were often costly and time-consuming, and could not be measured on a large scale. Therefore, predicting intrinsically disordered proteins by computational methods is a simple and efficient method. Summary of the invention

[0003] The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art and design a method for predicting intrinsic disordered proteins. The method uses three novel sequence features, namely persistent entropy, the correlation probability features of two consecutive amino acids and the correlation probability features of three consecutive amino acids, as well as other existing protein sequence features as feature matrices, and utilizes VGG199 variants and two MLP networks to construct a deep neural network to achieve more accurate prediction of intrinsic disordered proteins.

[0004] The specific steps of the intrinsic disordered protein prediction method provided by the present invention are as follows:

[0005] First, for the learning sample, let w represent a protein sequence of length L, and use a sliding window of length N to intercept a continuous amino acid fragment of corresponding length β j =w j …w j+(N-1) Perform calculations;

[0006] Second, first calculate the persistent entropy, the formula is as follows:

[0007]

[0008] in

[0009]

[0010] Where δ(·) is the impulse function, and the amino acid symbol set is represented by a one-to-one correspondence sequence:

[0011]

[0012] The persistent entropy is:

[0013]

[0014] in Representation and barcode diagram (assumed For all 1≤i≤n) related filter flows, Its k-dimensional barcode [ε s ,ε e ) represents the filtered flow at time ε s Appears at time ε e Finish.

[0015] Third, calculate the correlation probability characteristics of two and three consecutive amino acids. The formula is as follows:

[0016] Define two collections:

[0017]

[0018]

[0019] represents all possibilities of two and three consecutive amino acids in the protein sequence. Then the two new features are:

[0020] H2=[H2(1),…,H2(L)] (7)

[0021] H3=[H3(1),...,H3(L)] (8)

[0022] in

[0023]

[0024]

[0025] The functions I2(j) and I3(j) are defined as:

[0026]

[0027]

[0028] where w j and Respectively represent w j w j+1 (1≤j≤L-1) and w j w j+1 w j+2 (1≤j≤L-2);

[0029]

[0030]

[0031] in, and

[0032]

[0033]

[0034] In the above formula

[0035]

[0036]

[0037] in,

[0038]

[0039]

[0040] where w j and Respectively represent w j w j+1 (1≤j≤L-1) and w j w j+1 w j+2 (1≤j≤L-2).

[0041] Fourth, for a protein sequence w of length L, 20 kinds of evolutionary information, 7 kinds of physical and chemical properties, 3 kinds of tendencies, Shannon entropy and topological entropy of the sequence are calculated. The final feature matrix is ​​obtained by combining the above three novel feature sequences.

[0042] Fifth, use ten-fold cross validation to train the deep neural network and obtain the final network parameters.

[0043] Step 6. For the protein sequence to be predicted, calculate the feature vector of each residue according to steps 1 to 4, and use the neural network parameters obtained in step 5 to determine the amino acid.

[0044] Advantages and positive effects of the present invention:

[0045] The scheme uses three novel sequence features, namely persistent entropy, the correlation probability feature of two consecutive amino acids and the correlation probability feature of three consecutive amino acids, as well as other existing protein sequence features as feature matrices. The three novel sequence features are proposed for the prediction of intrinsic disordered proteins for the first time; secondly, the method uses VGG199 variants and two MLP networks to construct a deep neural network MLP-VGG19-MLP to achieve more accurate prediction of intrinsic disordered proteins. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The present invention is a flowchart for implementing the method for predicting intrinsically disordered proteins. DETAILED DESCRIPTION

[0047] See attached Figure 1 The present invention uses three novel sequence features, namely, persistent entropy, correlation probability features of two consecutive amino acids, and correlation probability features of three consecutive amino acids, as well as other existing protein sequence features as feature matrices, and uses VGG199 variants and two MLP networks to construct a deep neural network to achieve more accurate prediction of intrinsic disordered proteins;

[0048] The method includes the steps of using a sliding window of length N to intercept continuous amino acid fragments of corresponding length for a learning sample protein sequence;

[0049] The method comprises the steps of calculating the persistent entropy, the related probability features of two consecutive amino acids, the related probability features of three consecutive amino acids and other existing protein sequence features as a feature matrix for each sliding window;

[0050] Input the training sample feature vector into the deep neural network MLP-VGG19-MLP for training to obtain the final network model;

[0051] The test sample features are input into the trained deep neural network MLP-VGG19-MLP to give a judgment step for each amino acid.

[0052] The implementation of the present invention is further described below in conjunction with embodiments.

[0053] Embodiment 1:

[0054] The specific steps of the method for predicting intrinsic disordered proteins provided by the present invention are as follows:

[0055] For a protein sequence w with an undetermined disordered region (a protein sequence labeled 1lon in the R80 dataset),

[0056] The specific steps of using the intrinsic disordered protein prediction scheme provided by the present invention to perform prediction are as follows:

[0057] Step 1: The length of the sequence is 457, and the sequence is intercepted using a sliding window of N = 33. The values ​​of thirty-five features are calculated for each window interval.

[0058] Sequence w=GTRASNDRPPGTGGVKRGRLQQEAAATGSRVTV

[0059] For the first sliding window of length N, the persistent entropy, the correlation probability characteristics of two consecutive amino acids and the correlation probability characteristic values ​​of three consecutive amino acids are calculated by formulas (4), (7), and (8). At the same time, 20 kinds of evolutionary information, 7 kinds of physicochemical properties, 3 kinds of tendencies, as well as Shannon entropy and topological entropy characteristic values ​​are calculated to obtain a total of 35 eigenvalues. Then, the sliding window calculates the corresponding eigenvalues ​​to obtain the final feature matrix.

[0060] The calculated feature matrix of the protein sequence w is as follows, where each column corresponds to the feature vector of the amino acid at that position:

[0061]

[0062] Step 2: Use the learning samples to train the deep neural network MLP-VGG19-MLP to obtain the final network parameters to determine X.

[0063] In order to verify the effectiveness of the prediction method, the method was used to predict intrinsic disordered proteins using the DIS166 dataset, the R80 dataset, and the MXD494 dataset. Among them, the DIS166 dataset contains 166 protein sequences; the R80 dataset contains 80 protein sequences, and the MXD494 dataset contains 494 protein sequences. Table 1 lists the prediction accuracy comparison between the prediction method for intrinsic disordered proteins designed by the present invention and existing prediction methods of the same type for the DIS166 dataset. Table 2 lists the prediction accuracy comparison between the prediction method for natural disordered proteins designed by the present invention and existing prediction methods of the same type for the R80 dataset. Table 3 lists the prediction accuracy comparison between the prediction method for natural disordered proteins designed by the present invention and existing prediction methods of the same type for the R494 dataset. Table 4 lists the definitions of various prediction accuracy parameters, where TP represents the number of true positives (disordered amino acids are correctly predicted); TN represents the number of true negatives (ordered amino acids are correctly predicted); FP represents the number of false negatives (disordered amino acids are incorrectly identified as ordered amino acids); and FN represents the number of false positives (ordered amino acids are incorrectly identified as disordered amino acids).

[0064] Table 1

[0065]

[0066] Table 2

[0067]

[0068] Table 3

[0069]

[0070] Table 4

[0071]

[0072] References:

[0073] 1. Dyson H J, Wright P E. Intrinsically unstructured proteins and their functions[J]. Nature reviews Molecular cell biology, 2005, 6(3): 197 - 208.

[0074] 2. Uversky V N. Functional roles of transiently and intrinsically disordered regions within proteins[J]. The FEBS journal, 2015, 282(7): 1182 - 1189.

[0075] 3. Sun X, Jones W T, Rikkerink E H A. GRAS proteins: the versatile roles of intrinsically disordered proteins in plant signaling[J]. Biochemical Journal, 2012, 442(1): 1 - 12.

[0076] 4. Yang Z R, Thomson R, McNeil P, et al. RONN: the bio - basis function neural network technique applied to the detection of natively disordered regions in proteins[J]. Bioinformatics, 2005, 21(16): 3369 - 3376.

[0077] 5、Comprehensive comparative assessment of in-silico predictors ofdisordered regions[J].Current Protein and Peptide Science,2012,13(1):6-18.

[0078] 6、Hatos A,Hajdu-Soltész B,Monzon A M,et al.DisProt:intrinsic proteindisorder annotation in 2020[J].Nucleic acids research,2020,48(D1):D269-D276.

[0079] 7、He H,Zhao J.A low computational complexity scheme for theprediction of intrinsically disordered protein regions[J].MathematicalProblems in Engineering,2018,2018.

[0080] 8、Meiler J,Müller M,Zeidler A,et al.Generation and evaluation ofdimension-reduced amino acid parameter representations by artificial neuralnetworks[J].Molecular modeling annual,2001,7(9):360-369.

[0081] 9、Xu P,Zhao J,Zhang J.Identification of intrinsically disorderedprotein regions based on deep neural network-VGG16[J].Algorithms,2021,14(4):107.

[0082] 10、Hanson J,Paliwal K K,Litfin T,et al.SPOT-Disorder2:improvedprotein intrinsic disorder prediction by ensembled deep learning[J].Genomics,proteomics&bioinformatics,2019,17(6):645-656.

[0083] 11、Liu Y,Wang X,Liu B.RFPR-IDP:reduce the false positive rates forintrinsically disordered protein and region prediction by incorporating bothfully ordered proteins and disordered proteins[J].Briefings inbioinformatics,2021,22(2):2000-2011.

[0084] 12、Tang Y J,Pang Y H,Liu B.IDP-Seq2Seq:identification ofintrinsically disordered regions based on sequence to sequence learning[J].Bioinformatics,2020,36(21):5177-5186.

[0085] 13、He H,Zhao J,Sun G.The Prediction of Intrinsically DisorderedProteins Based on Feature Selection[J].Algorithms,2019,12(2):46。

Claims

1. A method for predicting intrinsically disordered proteins, characterized by: This method uses three sequence features, namely persistent entropy, correlation probability features of two consecutive amino acids, correlation probability features of three consecutive amino acids, and other existing protein sequence features as feature matrices, and uses VGG199 variants and two MLP networks to build a deep neural network to achieve more accurate prediction of intrinsic disordered proteins. The method includes the steps of using a sliding window of length N to intercept continuous amino acid fragments of corresponding length for a learning sample protein sequence; The method comprises the steps of calculating the persistent entropy, the related probability features of two consecutive amino acids, the related probability features of three consecutive amino acids and other existing protein sequence features as a feature matrix for each sliding window; The method includes inputting the training sample feature vector into the deep neural network MLP-VGG19-MLP for training to obtain a final network model; The method includes inputting the test sample features into the trained deep neural network MLP-VGG19-MLP and making a judgment on each amino acid; The calculation formula for the correlation probability characteristics of two consecutive amino acids and the correlation probability characteristics of three consecutive amino acids is as follows: Define two collections: Representing all possibilities of two and three consecutive amino acids in the protein sequence, the two new features are: H2=[H2(1),...,H2(L)] (7) H3=[H3(1),...,H3(L)] (8) in The functions I2(j) and I3(j) are defined as: where w j and Respectively represent w j w j+1 (1≤j≤L-1) and w j w j+1 w j+2 (1≤j≤L-2), in, and In the above formula in, where w j and Respectively represent w j w j+1 (1≤j≤L-1) and w j w j+1 w j+2 (1≤j≤L-2).

2. The method for predicting intrinsically disordered proteins according to claim 1, wherein: Specific process as follows: First, for the learning sample, let w represent a protein sequence of length L, and use a sliding window of length N to intercept a continuous amino acid fragment of corresponding length β j =w j L j+(N-1) Perform calculations; Second, first calculate the persistent entropy, the formula is as follows: in Where δ(·) is the impulse function, and the amino acid symbol set is represented by a one-to-one correspondence sequence: The persistent entropy is: Third, calculate the correlation probability characteristics of two and three consecutive amino acids; Fourth, for a protein sequence w of length L, 20 kinds of evolutionary information, 7 kinds of physical and chemical properties, 3 kinds of tendencies, Shannon entropy and topological entropy of the sequence are calculated, and the final feature matrix is ​​obtained by combining the feature sequence of persistent entropy, the correlation probability characteristics of two consecutive amino acids and the correlation probability characteristics of three consecutive amino acids; Fifth, use ten-fold cross validation to train the deep neural network MLP-VGG19-MLP to obtain the final network parameters; Step 6. For the protein sequence to be predicted, calculate the feature vector of each residue according to steps 1 to 4, and use the neural network parameters obtained in step 5 to determine the amino acid.

Images