A method and apparatus for predicting and analyzing RNA-binding residues

Abstract

A method and apparatus for predicting and analyzing RNA-binding residues are disclosed. The method constructs a GDRBind model using PyTorch and the DGL framework; inputs a protein sequence of a predetermined length into the GDRBind model and obtains the protein structure data of the protein sequence; generates a multiple sequence alignment (MSA) file using the protein sequence search software HHblits; collects several RNA-binding protein sequences from the UniProtKB protein database, performs clustering processing to obtain a pre-training dataset, and uses it to train a general protein language model to obtain an ESM-RBP representation model; obtains a first embedding matrix and a second embedding matrix through processing, and concatenates them to obtain a residue node feature representation matrix; calculates the edge features of all residue pairs in the protein sequence to obtain an edge set; and predicts the RNA-binding residues of the protein sequence using an equivariant graph neural network (EGNN) prediction model, outputting the predicted results. This invention can solve problems such as insufficient domain feature mining, low segmentation accuracy, and lack of interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics, artificial intelligence and computer application, in particular to a prediction analysis method and device for RNA binding residues. BACKGROUND

[0002] The interaction between proteins and RNA plays a fundamental and key role in various life activities such as gene regulation and expression of organisms. Understanding the interaction between proteins and RNA helps to study protein function and drug development. Accurate identification of RNA binding residues is an important step to understand the intrinsic mechanism of interaction. Therefore, accurate RNA binding residue prediction is helpful for the study of protein function and the design of drug target proteins, and has very important significance.

[0003] However, the existing RNA binding residue prediction method is convenient and easy to expand, but the sequence-based method has inherent difficulties in directly mining useful discriminant information from the primary sequence, so its accuracy needs to be improved. The structure-based method takes advantage of the high structural conservation and specificity of nucleic acid binding residues, and is considered to have better performance. However, the long-term lack of high-quality protein 3D structure leads to poor scalability of this method.

[0004] Therefore, how to invent an RNA binding residue prediction method to solve the problems of insufficient field feature mining, low division accuracy and lack of interpretability has become a problem to be solved. SUMMARY

[0005] Therefore, the present application provides a prediction analysis method and device for RNA binding residues, which can overcome the shortcomings of the existing RNA binding residue prediction method in insufficient field feature mining, low division accuracy and lack of interpretability.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: a prediction analysis method for RNA binding residues, comprising:

[0007] building a GDRBind model through Pytorch and DGL framework;

[0008] inputting a protein sequence of a set length into the GDRBind model, and acquiring protein structure data of the protein sequence through a protein data warehouse according to the protein sequence;

[0009] performing multi-sequence comparison on the protein sequence through the fast iterative protein sequence search software HHblits based on HMM-HMM comparison to generate a multi-sequence alignment MSA file;

[0010] Collecting a plurality of RNA binding protein sequences in the UniProtKB protein database, and performing clustering processing, taking the plurality of RNA binding protein sequences processed as a pre-training data set;

[0011] Training a general protein language model using the pre-training data set to obtain an ESM-RBP representation model;

[0012] Inputting the protein sequence into the ESM-RBP representation model to obtain a first embedding matrix through processing;

[0013] Inputting the multiple sequence alignment MSA file into a transformer-based protein multiple sequence alignment neural network ESM-MSA to obtain a second embedding matrix through processing;

[0014] Splicing the first embedding matrix and the second embedding matrix to obtain a residue node feature representation matrix;

[0015] Defining residue pairs with close spatial distances as edges of a protein graph, and calculating edge features of residue pairs in the protein sequence; calculating edge features of all residue pairs in the protein sequence to obtain an edge set;

[0016] Inputting the residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue into an equivariant graph neural network EGNN prediction model, and outputting the RNA binding residue prediction result of the protein sequence through prediction processing of the equivariant graph neural network EGNN prediction model.

[0017] As a preferred scheme of the RNA binding residue prediction analysis method, in the process of clustering the RNA binding protein sequences to obtain the pre-training data set, the CD-HIT tool for quickly clustering protein sequences is used to remove highly similar RNA binding protein sequences with a preset clustering threshold, and the remaining RNA binding protein sequences are taken as the pre-training data set.

[0018] As a preferred scheme of the RNA binding residue prediction analysis method, in the process of training the general protein language model using the pre-training data set, the transformer block set in the general protein language model is trained and updated by randomly masking and predicting 15% of the residues to obtain the ESM-RBP representation model.

[0019] As a preferred scheme of the RNA binding residue prediction analysis method, in the process of calculating the edge features of the residue pairs in the protein sequence, the expression of the edge features is:

[0020] e ij=|R i ,R j | / D max

[0021] In the formula, e ij For residue pairs (R i ,R j Edge features of |R i ,R j |For residue R i and R j The Euclidean distance between Ca atoms; D max This represents the maximum residue distance in the protein sequence.

[0022] As a preferred embodiment of a method for predicting and analyzing RNA-binding residues, the equivariant graph neural network (EGNN) prediction model includes three equilateral graph convolutional layers (EGCL). The residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue are input into the first equilateral graph convolutional layer (EGCL). After transformation processing, the first node embedding set and the first coordinate set are output. The first node embedding set and the first coordinate set are input into the second equilateral graph convolutional layer (EGCL) for transformation processing, and the process is iterated sequentially. The result output by the third equilateral graph convolutional layer (EGCL) is the prediction result of the RNA-binding residues of the protein sequence.

[0023] The present invention also provides an RNA-binding residue prediction and analysis device, based on the above-mentioned RNA-binding residue prediction and analysis method, comprising:

[0024] The GDRBind model building module is used to build GDRBind models using the PyTorch and DGL frameworks.

[0025] The protein sequence input module is used to input a protein sequence of a set length into the GDRBind model, and to obtain the protein structure data of the protein sequence from the protein data warehouse based on the protein sequence.

[0026] The multiple sequence alignment (MSA) file generation module is used to perform multiple sequence alignment on the protein sequence using HHblits, a fast iterative protein sequence search software based on HMM-HMM alignment, and generate a multiple sequence alignment (MSA) file.

[0027] The pre-training dataset acquisition module is used to collect several RNA-binding protein sequences from the UniProtKB protein database, perform clustering processing, and use the processed RNA-binding protein sequences as a pre-training dataset.

[0028] The ESM-RBP representation model acquisition module is used to train the general protein language model using the pre-trained dataset to obtain the ESM-RBP representation model.

[0029] The first embedding matrix acquisition module is used to input the protein sequence into the ESM-RBP characterization model and obtain the first embedding matrix through processing.

[0030] The second embedding matrix acquisition module is used to input the multiple sequence alignment MSA file into the transformer-based protein multiple sequence alignment neural network ESM-MSA, and obtain the second embedding matrix through processing.

[0031] The residue node feature representation matrix acquisition module is used to concatenate the first embedding matrix and the second embedding matrix to obtain the residue node feature representation matrix;

[0032] The edge set acquisition module is used to define residue pairs that are spatially close as edges of the protein graph, and to obtain the edge features of residue pairs in the protein sequence by calculation; the edge features of all residue pairs in the protein sequence are calculated to obtain the edge set.

[0033] The RNA binding residue prediction module is used to input the residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue into the equivariant graph neural network (EGNN) prediction model, and output the RNA binding residue prediction results of the protein sequence through the prediction processing of the EGNN prediction model.

[0034] As a preferred embodiment of a predictive analysis device for RNA-binding residues, in the pre-training dataset acquisition module, during the process of clustering the RNA-binding protein sequences to obtain the pre-training dataset, the tool CD-HIT for fast clustering of protein sequences is used to remove highly similar RNA-binding protein sequences with a preset clustering threshold, and the remaining RNA-binding protein sequences are used as the pre-training dataset.

[0035] As a preferred embodiment of a predictive analysis device for RNA-binding residues, in the ESM-RBP characterization model acquisition module, during the training of the general protein language model using the pre-training dataset, the transformer block set in the general protein language model is trained and updated by randomly masking and predicting 15% of the residues to obtain the ESM-RBP characterization model.

[0036] In a preferred embodiment of a predictive analysis device for RNA-binding residues, the edge set acquisition module, during the calculation of edge features of residue pairs in the protein sequence, uses the following expression for the edge features:

[0037] e ij =|R i ,R j | / D max

[0038] In the formula, e ij For residue pairs (R i ,R j Edge features of |R i ,R j |For residue R i and R j The Euclidean distance between Ca atoms; D max This represents the maximum residue distance in the protein sequence.

[0039] As a preferred embodiment of an RNA-binding residue prediction and analysis device, the RNA-binding residue prediction module includes an isovariant graph neural network (EGNN) prediction model comprising three equilateral graph convolutional layers (EGCLs). The residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue are input into the first EGCL, and after transformation processing, a first node embedding set and a first coordinate set are output. The first node embedding set and the first coordinate set are input into the second EGCL for transformation processing, and this process is iterated sequentially. The result output by the third EGCL is the RNA-binding residue prediction result of the protein sequence.

[0040] This invention has the following advantages: A GDRBind model is constructed using PyTorch and the DGL framework; a protein sequence of a set length is input into the GDRBind model, and protein structure data of the protein sequence is obtained from a protein data warehouse based on the protein sequence. Multiple sequence alignment (MSA) is performed on the protein sequence using HHblits, a fast iterative protein sequence search software based on HMM-HMM alignment, generating an MSA file; several RNA-binding protein sequences are collected from the UniProtKB protein database and clustered, with the processed RNA-binding protein sequences used as a pre-training dataset; a general protein language model is trained using the pre-training dataset to obtain an ESM-RBP representation model; the protein sequence is input into the ESM-RBP representation model, and a first embedding matrix is ​​obtained through processing; the MSA file is input into a transformer-based... The protein multiple sequence alignment neural network (ESM-MSA) is used to obtain a second embedding matrix. The first and second embedding matrices are then concatenated to obtain a residue node feature representation matrix. Spatially close residue pairs are defined as edges in the protein graph, and edge features of residue pairs in the protein sequence are calculated. Edge features of all residue pairs in the protein sequence are calculated to obtain an edge set. The residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue are input into an isovariant graph neural network (EGNN) prediction model. The EGNN prediction model then performs prediction processing and outputs the prediction results of RNA-binding residues in the protein sequence. This invention further enhances the ability of ESM2 to characterize RNA-binding residues based on the common characteristics of all protein families described by ESM2, obtaining more specific and useful information, and further improving the characterization ability of RNA-binding residues in protein sequences. Simultaneously, using EGNN effectively obtains contextual information about spatial structure, fully utilizing the characterization information, and possessing a certain degree of visibility and interpretability. This improves the accuracy and interpretability of RNA-binding residue prediction. It solves problems such as insufficient domain feature mining, low segmentation accuracy, and lack of interpretability. Attached Figure Description

[0041] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0042] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0043] Figure 1 This is a schematic diagram of the predictive analysis method for RNA binding residues provided in Embodiment 1 of the present invention;

[0044] Figure 2 This is a schematic diagram of the specific implementation process of the RNA binding residue prediction and analysis method provided in Embodiment 1 of the present invention; wherein, A is a schematic diagram of the training of a large protein characterization model in the ESM-RBP field; B is a schematic diagram of the RNA binding residue prediction process;

[0045] Figure 3 This is a schematic diagram showing the results of predicting RNA-binding residues of protein 1AIL in one possible embodiment of the present invention, as provided in Embodiment 1 of the present invention.

[0046] Figure 4 This is a schematic diagram of the architecture of a predictive analysis device for RNA binding residues provided in Embodiment 2 of the present invention. Detailed Implementation

[0047] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] See Figure 1 and Figure 2 Embodiment 1 of the present invention provides a method for predicting and analyzing RNA-binding residues, comprising the following steps:

[0050] S1. Construct the GDRBind model using PyTorch and the DGL framework;

[0051] S2. Input the protein sequence of a set length into the GDRBind model, and obtain the protein structure data of the protein sequence from the protein data warehouse based on the protein sequence.

[0052] S3. Perform multiple sequence alignment on the protein sequence using HHblits, a fast iterative protein sequence search software based on HMM-HMM alignment, and generate a multiple sequence alignment MSA file;

[0053] S4. Collect several RNA-binding protein sequences from the UniProtKB protein database and perform clustering processing. Use the processed RNA-binding protein sequences as a pre-training dataset.

[0054] S5. Use the pre-trained dataset to train the general protein language model to obtain the ESM-RBP representation model.

[0055] S6. Input the protein sequence into the ESM-RBP characterization model and obtain the first embedding matrix through processing;

[0056] S7. Input the multiple sequence alignment MSA file into the transformer-based protein multiple sequence alignment neural network ESM-MSA, and obtain the second embedding matrix through processing;

[0057] S8. Concatenate the first embedding matrix and the second embedding matrix to obtain the residue node feature representation matrix;

[0058] S9. Define spatially close residue pairs as edges of the protein graph, and calculate the edge features of residue pairs in the protein sequence; calculate the edge features of all residue pairs in the protein sequence to obtain the edge set;

[0059] S10. Input the residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue into the equivariant graph neural network (EGNN) prediction model, and output the prediction results of RNA binding residues of the protein sequence through the EGNN prediction model prediction process.

[0060] In this embodiment, in step S1, the GDRBind model is constructed using PyTorch and the DGL framework.

[0061] Specifically, the GDRBind model is built using PyTorch and DGL, with a batch size of 1, meaning each batch uses one protein map for both forward and backward propagation. Cross-entropy is used as the loss function. The optimization was performed using the AdamW optimizer with a learning rate of 1e-04; the loss weights for positive and negative samples were 0.7 and 0.3, respectively; and regularization with a coefficient of 1e-04 was used to limit parameter overfitting.

[0062] In this embodiment, in step S2, a protein sequence of a set length is input into the GDRBind model, and protein structure data of the protein sequence is obtained from the protein data warehouse based on the protein sequence.

[0063] Specifically, in the GDRBind model, the input is a protein sequence of length s to be used for RNA binding residue prediction, denoted as S; the protein structure is obtained from the PDB (Protein Data Bank) protein database based on the input protein sequence S, and the protein structure is represented by a graph G = (V, E). Where V = {v...} 0 ,...,v i ,...,v s-1} represents the set of all residue nodes; e ij ∈E represents the edge set of interacting residues.

[0064] In this embodiment, in step S3, the protein sequence is subjected to multiple sequence alignment using HHblits, a fast iterative protein sequence search software based on HMM-HMM alignment, to generate a multiple sequence alignment MSA file.

[0065] Specifically, the protein sequence S was subjected to multiple sequence alignment using HHblits, a fast iterative protein sequence search software based on HMM-HMM alignment. After multiple comparisons in the Uniclust30 protein database, the e-value threshold was set to 1e-04, and a Multiple Sequence Alignment (MSA) file was generated.

[0066] In this embodiment, in step S4, several RNA-binding protein sequences are collected from the UniProtKB protein database and clustered. The processed RNA-binding protein sequences are used as a pre-training dataset.

[0067] Specifically, 9,743,473 redundant RNA-binding protein sequences (RBPs) were collected from the UniProtKB protein database. To prevent the model from overfitting to highly redundant RBPs, the CD-HIT tool for fast clustering of protein sequences was used to remove highly similar sequences with a clustering threshold of 40%. The remaining 459,656 non-redundant RBPs were used as a pre-training dataset and named UniRBP40.

[0068] In this embodiment, in step S5, the pre-trained dataset is used to train the general protein language model to obtain the ESM-RBP representation model.

[0069] Specifically, based on the general protein language model ESM2, the ESM2 model is further trained by randomly masking and predicting 15% of the residues, using the pre-training dataset UniRBP40 obtained in step S4 as the training data. ESM2 has a total of 33 transformer blocks; the first 28 blocks are frozen, and the remaining 5 blocks are used as update modules. After further training, the ESM-RBP representation model is obtained, as shown below. Figure 2 As shown in B.

[0070] In this embodiment, in step S6, the protein sequence is input into the ESM-RBP characterization model, and the first embedding matrix is ​​obtained through processing;

[0071] Specifically, the protein sequence S of length s input in step S1 is input into the ESM-RBP characterization model obtained in step S5 to obtain the first embedding matrix of size s×1280, denoted as E1.

[0072] In this embodiment, in step S7, the multiple sequence alignment MSA file is input into the transformer-based protein multiple sequence alignment neural network ESM-MSA, and the second embedding matrix is ​​obtained through processing;

[0073] Specifically, the MSA file obtained in step S3 is input into the ESM-MSA protein multiple sequence alignment neural network based on transformers to obtain a second embedding matrix of size s×768, denoted as E2, which is used to depict the evolutionary information of the protein.

[0074] In this embodiment, in step S8, the first embedding matrix and the second embedding matrix are concatenated to obtain the residue node feature representation matrix;

[0075] Specifically, the first embedding matrix E1 and the second embedding matrix E2 obtained in steps S6 and S7 are concatenated to obtain the final residue node feature representation matrix f = {f0, ..., f...} i ,…,f s-1}, with a size of s×2048, and each residue node v i The features are all composed of a vector f of length 2048. i express.

[0076] In this embodiment, in step S9, residue pairs that are spatially close are defined as edges of the protein graph, and the edge features of residue pairs in the protein sequence are obtained by calculation; the edge features of all residue pairs in the protein sequence are calculated to obtain the edge set.

[0077] Specifically, residue pairs that are spatially close are defined as edges in the protein graph, meaning the Euclidean distance between the Ca atoms of two residues is less than 1. Defined as having contact.

[0078] Target residue pair (R) i ,R j The edge characteristics of ) are:

[0079] e ij =|R i ,R j | / D max

[0080] Where |R i ,R j | indicates residue R i and R j The Euclidean distance between Ca atoms, D max The maximum residue distance of the target protein is represented by the edge set E, which is obtained by calculating all the edge features within protein S.

[0081] In this embodiment, in step S10, the residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue are input into the equivariant graph neural network (EGNN) prediction model. The EGNN prediction model is then used for prediction processing to output the prediction results of RNA binding residues of the protein sequence.

[0082] Specifically, the equilateral graph neural network (EGNN) prediction model includes three equilateral graph convolutional layers (EGCL); the residue node feature representation matrix f l The edge set E and the Ca coordinate set x of each residue l Input the Graph Equivariate Neural Network (EGNN) prediction model; for EGCL(f l ,x l E) undergoes the following transformation:

[0083]

[0084] In the formula, and These represent edge operations and node operations based on MLP (Multilayer Perception), respectively; C = 1 / (s-1) represents taking the average value; φ x m ij Convert to scalar values.

[0085] Output node embedding set f l+1 and coordinate set x l+1 This will be used as the input for the next EGCL; the process will be iterated sequentially; the output of the third equilateral graph convolutional layer EGCL will be the prediction result of the RNA binding residues of the protein sequence.

[0086] In one possible embodiment, an example of providing RNA-binding residue predictions for protein 1AIL is as follows:

[0087] T1. Construct the GDRBind model using PyTorch and DGL, where the batch size is set to 1, meaning each batch uses one protein map for forward and backward propagation; use cross-entropy as the loss function. The AdamW optimizer with a learning rate of 1e-04 was used for optimization; the loss weights for positive and negative samples were 0.7 and 0.3 respectively; regularization with a coefficient of 1e-04 was used to limit parameter overfitting.

[0088] T2. Input the 1AIL sequence information of the protein to be predicted for RNA binding residues, with a length of 73, denoted as S; obtain the protein structure from the PDB (Protein Data Bank) protein database based on the input protein sequence S, and represent the protein structure using a graph G = (V, E); where V = {v 0 ,...,v i ,...,v 72} represents the set of all residue nodes; e ij ∈E represents the edge set of interacting residues;

[0089] T3. The protein sequence S was subjected to multiple sequence alignment using HHblits, a fast iterative protein sequence search software based on HMM-HMM alignment. After multiple comparisons in the Uniclust30 protein database, the e-value threshold was set to 1e-04, and a Multiple Sequence Alignment (MSA) file was generated.

[0090] T4. Collect 9,743,473 redundant RNA-binding protein sequences (RBPs) from the UniProtKB protein database. To prevent the model from overfitting to highly redundant RBPs, use the fast protein sequence clustering tool CD-HIT to remove highly similar sequences with a clustering threshold of 40%. The remaining 459,656 non-redundant RBPs are used as a pre-training dataset and named UniRBP40.

[0091] T5. Based on the general protein language model ESM2, the ESM2 model is trained by randomly masking and predicting 15% of the residues. The pre-training dataset UniRBP40 obtained in step T4 is used as the training data. ESM2 has a total of 33 transformer blocks. The first 28 blocks are frozen and the remaining 5 blocks are used as update modules. After further training, the ESM-RBP representation model is obtained.

[0092] T6. Input the protein sequence S with a length of 73 from step T1 into the ESM-RBP characterization model obtained in step T5 to obtain the first embedding matrix with a size of 73×1280, denoted as E1.

[0093] T7. Input the MSA file obtained in step T3 into the ESM-MSA protein multiple sequence alignment neural network based on transformer to obtain a second embedding matrix of size 73×768, denoted as E2, which is used to depict the evolutionary information of the protein.

[0094] T8. Concatenate the first embedding matrix E1 and the second embedding matrix E2 obtained in steps T6 and T7 to obtain the final residue node feature representation matrix f = {f0, ..., f...} i ,…,f s-1}, with a size of 73×2048, each residue node v i The features are all composed of a vector f of length 2048. i express;

[0095] T9. Define spatially close residue pairs as edges in the protein graph, i.e., the Euclidean distance between the Ca atoms of two residues is less than 1 / 2. Defined as having contact.

[0096] Target residue pair (R) i ,R j The edge characteristics of ) are:

[0097] e ij =|R j ,R j | / D max

[0098] Where |R i ,R j | indicates residue R i and R j The Euclidean distance between Ca atoms, D max The maximum residue distance of the target protein is represented by the edge set E, which is obtained by calculating all the edge features within protein S.

[0099] The T10 equilateral graph neural network (EGNN) prediction model includes three equilateral graph convolutional layers (EGCL); the residue node feature representation matrix f... l The edge set E and the Ca coordinate set x of each residue l Input the Graph Equivariate Neural Network (EGNN) prediction model; for EGCL(f l ,x l E) undergoes the following transformation:

[0100]

[0101] In the formula, and These represent edge operations and node operations based on MLP (Multilayer Perception), respectively; C = 1 / (s-1) represents taking the average value; φ x m ij Convert to scalar value;

[0102] Output node embedding set f l+1 and coordinate set x l+1 This will be used as the input for the next EGCL; the process will be iterated sequentially; the output of the third equilateral graph convolutional layer EGCL is the prediction result of the RNA binding residues of protein 1AIL, as shown below. Figure 3 As shown.

[0103] In summary, this invention constructs a GDRBind model using PyTorch and the DGL framework; inputs a protein sequence of a set length into the GDRBind model, and obtains the protein structure data of the protein sequence from a protein data warehouse based on the protein sequence. Multiple sequence alignment (MSA) is performed on the protein sequence using HHblits, a fast iterative protein sequence search software based on HMM-HMM alignment, generating an MSA file; several RNA-binding protein sequences are collected from the UniProtKB protein database and clustered, with the processed RNA-binding protein sequences used as a pre-training dataset; a general protein language model is trained using the pre-training dataset to obtain an ESM-RBP representation model; the protein sequence is input into the ESM-RBP representation model, and a first embedding matrix is ​​obtained through processing; the MSA file is then input into a transformer-based... The protein multiple sequence alignment neural network (ESM-MSA) is used to obtain a second embedding matrix. The first and second embedding matrices are then concatenated to obtain a residue node feature representation matrix. Spatially close residue pairs are defined as edges in the protein graph, and edge features of residue pairs in the protein sequence are calculated. Edge features of all residue pairs in the protein sequence are calculated to obtain an edge set. The residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue are input into an isovariant graph neural network (EGNN) prediction model. The EGNN prediction model then performs prediction processing and outputs the prediction results of RNA-binding residues in the protein sequence. This invention further enhances the ability of ESM2 to characterize RNA-binding residues based on the common characteristics of all protein families described by ESM2, obtaining more specific and useful information, and further improving the characterization ability of RNA-binding residues in protein sequences. Simultaneously, using EGNN effectively obtains contextual information about spatial structure, fully utilizing the characterization information, and possessing a certain degree of visibility and interpretability. This improves the accuracy and interpretability of RNA-binding residue prediction. It solves problems such as insufficient domain feature mining, low segmentation accuracy, and lack of interpretability.

[0104] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0105] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0106] Example 2

[0107] See Figure 4 Embodiment 2 of the present invention also provides a predictive analysis device for RNA-binding residues, comprising:

[0108] GDRBind model building module 001 is used to build GDRBind models using the PyTorch and DGL frameworks.

[0109] The protein sequence input module 002 is used to input a protein sequence of a set length into the GDRBind model, and to obtain the protein structure data of the protein sequence through a protein data warehouse based on the protein sequence.

[0110] The multiple sequence alignment MSA file generation module 003 is used to perform multiple sequence alignment on the protein sequence using HHblits, a fast iterative protein sequence search software based on HMM-HMM alignment, and generate a multiple sequence alignment MSA file.

[0111] The pre-training dataset acquisition module 004 is used to collect several RNA-binding protein sequences from the UniProtKB protein database, perform clustering processing, and use the processed RNA-binding protein sequences as a pre-training dataset.

[0112] ESM-RBP representation model acquisition module 005 is used to train a general protein language model using the pre-trained dataset to obtain an ESM-RBP representation model.

[0113] The first embedding matrix acquisition module 006 is used to input the protein sequence into the ESM-RBP characterization model and obtain the first embedding matrix through processing.

[0114] The second embedding matrix acquisition module 007 is used to input the multiple sequence alignment MSA file into the transformer-based protein multiple sequence alignment neural network ESM-MSA, and obtain the second embedding matrix through processing;

[0115] The residue node feature representation matrix acquisition module 008 is used to concatenate the first embedding matrix and the second embedding matrix to obtain the residue node feature representation matrix;

[0116] Edge set acquisition module 009 is used to define spatially close residue pairs as edges of the protein graph, and to obtain the edge features of residue pairs in the protein sequence by calculation; the edge features of all residue pairs in the protein sequence are calculated to obtain the edge set;

[0117] The RNA binding residue prediction module 010 is used to input the residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue into the equivariant graph neural network (EGNN) prediction model, and output the RNA binding residue prediction result of the protein sequence through the prediction processing of the EGNN prediction model.

[0118] In this embodiment, in the pre-training dataset acquisition module 004, during the process of clustering the RNA-binding protein sequences to obtain the pre-training dataset, the tool CD-HIT for fast clustering of protein sequences is used to remove highly similar RNA-binding protein sequences with a preset clustering threshold, and the remaining RNA-binding protein sequences are used as the pre-training dataset.

[0119] In this embodiment, in the ESM-RBP representation model acquisition module 005, during the training of the general protein language model using the pre-training dataset, the transformer block set in the general protein language model is trained and updated by randomly masking and predicting 15% of the residues to obtain the ESM-RBP representation model.

[0120] In this embodiment, in the edge set acquisition module 009, during the calculation of the edge features of residue pairs in the protein sequence, the expression of the edge features is:

[0121] e ij =|R i ,R j | / D max

[0122] In the formula, e ij For residue pairs (R i ,R j Edge features of |R i ,R j |For residue R i and R j The Euclidean distance between Ca atoms; D max This represents the maximum residue distance in the protein sequence.

[0123] In this embodiment, the RNA binding residue prediction module 010 includes an isomorphic graph neural network (EGNN) prediction model comprising three isomorphic graph convolutional layers (EGCLs). The residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue are input into the first isomorphic graph convolutional layer (EGCL). After transformation processing, a first node embedding set and a first coordinate set are output. The first node embedding set and the first coordinate set are input into the second isomorphic graph convolutional layer (EGCL) for transformation processing, and this process is iterated sequentially. The result output by the third isomorphic graph convolutional layer (EGCL) is the RNA binding residue prediction result of the protein sequence.

[0124] It should be noted that the information interaction and execution process between the modules of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0125] Example 3

[0126] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code of a method for predicting and analyzing RNA-binding residues. The program code includes instructions for executing the method for predicting and analyzing RNA-binding residues of Embodiment 1 or any possible implementation thereof.

[0127] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives, SSDs).

[0128] Example 4

[0129] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0130] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute a method for predicting and analyzing RNA binding residues according to Embodiment 1 or any possible implementation thereof by calling the program instructions.

[0131] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0132] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0133] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0134] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for predicting analysis of RNA binding residues, characterized by, The method comprises the following steps: The GDRBind model is built through the Pytorch and DGL frameworks, in which one protein graph is used for forward and backward propagation in each batch; the cross-entropy is used as the loss function: The AdamW optimizer with a learning rate of 1e-04 is used for optimization; the loss weight of positive and negative samples is divided into 0.7 and 0.3; the regularization with a coefficient of 1e-04 is used to limit parameter overfitting; The protein sequence of a set length is input into the GDRBind model, in which protein sequence information of a length of s to be subjected to RNA binding residue prediction is denoted as S , and protein structure data of the protein sequence is obtained from a protein data warehouse according to the protein sequence, and the protein structure is represented by a graph ; wherein represents a set of all residue nodes; represents a set of interacting residues. performing multiple sequence alignment on the protein sequence by using a fast iteration protein sequence search software to generate a multiple sequence alignment (MSA) file; collecting a plurality of RNA-binding protein sequences in a protein database and performing clustering processing, and taking the plurality of RNA-binding protein sequences after processing as a pre-training data set; training a general protein language model using the pre-training data set to obtain an ESM-RBP representation model; inputting the protein sequence into the ESM-RBP representation model to obtain a first embedding matrix through processing; inputting the multiple sequence alignment (MSA) file into a transformer-based protein multiple sequence alignment neural network (ESM-MSA) to obtain a second embedding matrix through processing; splicing the first embedding matrix and the second embedding matrix to obtain a residue node feature representation matrix; defining residue pairs with close spatial distances as edges of a protein graph, and calculating edge features of residue pairs in the protein sequence; calculating edge features of all residue pairs in the protein sequence to obtain an edge set; inputting the residue node feature representation matrix, the edge set, and a Ca coordinate set of each residue into an equivariant graph neural network (EGNN) prediction model, and outputting an RNA-binding residue prediction result of the protein sequence through processing of the equivariant graph neural network (EGNN) prediction model.

2. The method for predicting and analyzing RNA-binding residues according to claim 1, characterized in that, In the process of clustering the RNA-binding protein sequences to obtain the pre-training data set, a CD-HIT tool for quickly clustering protein sequences is used to remove highly similar RNA-binding protein sequences with a preset clustering threshold, and the remaining RNA-binding protein sequences are taken as the pre-training data set.

3. The method for predicting and analyzing RNA-binding residues according to claim 2, characterized in that, In the process of training the general protein language model using the pre-training data set, the transformer blocks in the general protein language model are trained and updated by randomly masking and predicting 15% of the residues to obtain the ESM-RBP representation model.

4. The method for predicting and analyzing RNA-binding residues according to claim 3, characterized in that, In the process of calculating the edge features of the residue pairs in the protein sequence, the expression of the edge features is: = wherein is the edge feature of the residue pair ; is the Euclidean distance between the Ca atoms of the residues and ; is the maximum residue distance of the protein sequence.

5. The method for predicting and analyzing RNA-binding residues according to claim 4, characterized in that, The equivariant graph neural network (EGNN) prediction model comprises three EGCLs; the residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue are input into the first EGCL, and a first node embedding set and a first coordinate set are output through transformation processing; the first node embedding set and the first coordinate set are input into the second EGCL for transformation processing, and the process is iterated in turn; the result output by the third EGCL is the RNA-binding residue prediction result of the protein sequence.

6. An apparatus for predicting RNA-binding residues, using the method for predicting RNA-binding residues according to any one of claims 1-5, wherein, The method comprises the following steps: a GDRBind model construction module for constructing a GDRBind model through a Pytorch and DGL framework; a protein sequence input module, configured to input a protein sequence of a set length into the GDRBind model and obtain protein structure data of the protein sequence from a protein data warehouse according to the protein sequence; a multiple sequence alignment (MSA) file generation module, configured to perform multiple sequence alignment on the protein sequence by using a fast iterative protein sequence search software to generate an MSA file; a pre-training data set acquisition module, configured to collect a plurality of RNA-binding protein sequences in a protein database and perform clustering processing on the plurality of RNA-binding protein sequences, and use the plurality of RNA-binding protein sequences after the clustering processing as a pre-training data set; an ESM-RBP representation model acquisition module, configured to train a general protein language model using the pre-training data set to obtain an ESM-RBP representation model; a first embedding matrix acquisition module, configured to input the protein sequence into the ESM-RBP representation model to obtain a first embedding matrix through processing; a second embedding matrix acquisition module, configured to input the MSA file into an ESM-MSA neural network based on a transformer to obtain a second embedding matrix through processing; a residue node feature representation matrix acquisition module, configured to splice the first embedding matrix and the second embedding matrix to obtain a residue node feature representation matrix; an edge set acquisition module, configured to define residue pairs with close spatial distances as edges of a protein graph, and obtain edge features of residue pairs in the protein sequence through calculation; the edge set is obtained by calculating edge features of all residue pairs in the protein sequence; an RNA-binding residue prediction module, configured to input the residue node feature representation matrix, the edge set, and a Ca coordinate set of each residue into an equivariant graph neural network (EGNN) prediction model, and output an RNA-binding residue prediction result of the protein sequence through prediction processing of the EGNN prediction model.

7. The device for predicting analysis of RNA binding residues according to claim 6, wherein, In the pre-training data set acquisition module, in the process of clustering the RNA-binding protein sequences to obtain the pre-training data set, a CD-HIT tool for quickly clustering protein sequences is used to remove highly similar RNA-binding protein sequences with a preset clustering threshold, and the remaining RNA-binding protein sequences are used as the pre-training data set.

8. The device for predicting analysis of RNA binding residues according to claim 7, wherein, In the ESM-RBP representation model acquisition module, in the process of training the general protein language model using the pre-training data set, the transformer blocks in the general protein language model are trained and updated by randomly masking and predicting 15% of the residues to obtain the ESM-RBP representation model.

9. The device for predicting analysis of RNA binding residues according to claim 8, wherein, In the edge set acquisition module, in the process of calculating the edge features of the residue pairs in the protein sequence, the expression of the edge features is: = wherein is the edge feature of residue pair ; is the Euclidean distance between the Ca atoms of residues and ; is the maximum residue distance of the protein sequence.

10. The device for predicting analysis of RNA binding residues according to claim 9, wherein, In the RNA-binding residue prediction module, the EGNN prediction model includes three EGCLs; the residue node feature representation matrix, the edge set, and the Ca coordinate set of each residue are input into the first EGCL, and the first node embedding set and the first coordinate set are output through transformation processing; the first node embedding set and the first coordinate set are input into the second EGCL for transformation processing, and the iteration is sequentially performed; and the result output by the third EGCL is the RNA-binding residue prediction result of the protein sequence.

Images