DNA binding residue prediction method based on multi-modal protein language model

Abstract

The invention discloses a DNA binding residue prediction method based on a multi-modal protein language model in the technical field of bioinformatics. The DNA binding residue prediction method based on the multi-modal protein language model comprises the following steps: sequentially using tools to respectively obtain feature files of a protein sequence P to be subjected to DNA binding residue prediction; inputting all the feature files into the ESM3 to obtain an embedded matrix M; the method comprises the following steps: processing a protein sequence into a residue sample, constructing an isovariant graph neural network model according to a protein structure diagram, constructing a data set by utilizing the protein sequence of known DNA binding residues, and training the constructed network; and inputting a residue sample of the protein sequence into the trained model to obtain a DNA binding residue of the protein sequence. According to the DNA binding residue prediction method based on the multi-modal protein language model, the efficiency and accuracy of DNA binding residue prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for predicting DNA binding residues based on a multimodal protein language model. Background Art

[0002] DNA-protein interactions run through many life processes and play a fundamental regulatory role. In addition, mutations in DNA-bound proteins are the root cause of many difficult genetic diseases. Accurate identification of DNA-binding residues helps to deeply understand the intrinsic mechanism of DNA-protein interactions and clarify the pathogenic mechanism of diseases. Therefore, accurate prediction of DNA-binding residues is helpful for the study of protein regulatory functions and the development of targeted drugs, which has very important biological significance.

[0003] At present, the literature methods specifically used for DNA binding residue prediction include: ESM-DBP (Zeng W, Dou Y, Pan L, et al. Improving prediction performance of general protein language model by domain-adaptive pretraining on DNA-binding protein [J]. Nature Communications, 2024, 15 (1): 7838. Zeng W et al. Improving prediction performance of general protein language model by domain-adaptive pretraining on DNA-binding protein [J]. Nature Communications, 2024, 15 (1): 7838); GraphBind (Xia Y, Xia CQ, Pan X, et al. GraphBind: protein structural context embedded rules learned by hierarchical graph neural networks for recognizing nucleic-acid-binding residues [J]. Nucleic acids research, 2021, 49 (9): e51-e51. Y et al. GraphBind: Protein structure context embedding rules learned by hierarchical graph neural network for identifying nucleic acid binding residues[J]. Nucleic Acids Research, 2021, 49(9): e51-e51); EquiPNAS (Roche R, Moussad B, Shuvo MH, et al. EquiPNAS: improved protein-nucleic acid binding site prediction using protein-language-model-informed equivariant deep graph neural networks[J]. Nucleic Acids Research, 2024, 52(5): e27-e27. That is: Roche R et al. EquiPNAS: Improved protein-nucleic acid binding site prediction using equivariant deep graph neural networks based on protein language model[J]. Nucleic Acids Research, 2024, 52(5): e27-e27), ESM-NBR (Zeng W, Lv D, Liu X, et al.ESM-NBR: fast and accurate nucleic acid-binding residue prediction via protein language model feature representation and multi-task learning [C]. 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2023: 76-81. Zeng W et al. Fast and accurate prediction of nucleic acid-binding residues based on protein language model feature representation and multi-task learning [C]. 2023 IEEE International Conference on Bioinformatics and Biomedicine, 2023: 76-81) etc. .

[0004] ESM-DBP collects a large number of DNA-bound proteins for improving the large universal protein language model ESM2. GraphBind constructs a local protein graph centered on the target residue to learn spatial context embedding information, and uses a hierarchical graph neural network as a classifier to learn its discriminative features. EquiPNAS uses a variety of protein sequence and structural features, as well as the feature embedding of the protein language model ESM2, and then uses an equivariant graph neural network as a model to identify DNA-bound residues. ESM-NBR uses the representation of the universal protein language model ESM2 and combines it with multi-task learning technology to simultaneously predict DNA and RNA binding residues. In summary, the existing DNA binding residue prediction methods are still far from the requirements of practical applications in terms of computational cost and partitioning accuracy, and are in urgent need of improvement. Summary of the invention

[0005] In order to solve the above problems, the purpose of the present invention is to provide a DNA binding residue prediction method based on a multimodal protein language model, which overcomes the shortcomings of existing DNA binding residue prediction methods in terms of computational cost and division accuracy, has low computational cost and high division accuracy, and improves the efficiency and accuracy of DNA binding residue prediction.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for predicting DNA binding residues based on a multimodal protein language model comprises the following steps:

[0008] S1. The protein sequence P to be used for DNA binding residue prediction is used to obtain the protein structure PDB feature file, protein secondary structure DSSP feature file, protein solvent accessibility SS feature file and InterPro function annotation feature file respectively using AlphaFold2, DSSP, Shrake-Rupley and InterProScan tools;

[0009] S2. Input all feature files obtained in S1 into the multimodal protein language model ESM3 to obtain an embedding matrix M as the feature representation of protein P;

[0010] S3. Process the protein sequence into residue samples, build an isovariant graph neural network model based on the protein structure graph, use the protein sequence of known DNA binding residues to construct a data set and train the built network;

[0011] S4. Input the residue samples of the protein sequence for which DNA binding residues are to be predicted into the trained model to obtain the DNA binding residues of the protein sequence.

[0012] Furthermore, the S1 comprises the following steps:

[0013] S11. Input a protein sequence information with a residue number of L for which DNA binding residues are to be predicted, denoted as P;

[0014] S12. For protein P, use the AlphaFold2 tool to predict its three-dimensional structure information to obtain a PDB feature file, and represent its atomic coordinates as a matrix denoted as Coord;

[0015] S13. For the PDB feature file of protein P obtained in S12, the DSSP tool was used to calculate the eight-state secondary structure of protein P: SS = {s1, s2, ..., s i , ..., s L}, where s i ∈(H, B, E, G, I, T, S, L) represents the secondary structure state of the i-th residue of protein P;

[0016] S14. Use the Shrake-Rupley algorithm to calculate the solvent accessible surface area SA of the protein P obtained in S12 using the PDB feature file. i , ..., a L}, non-negative real value a i represents the solvent accessible surface area of ​​the ith residue of protein P;

[0017] S15. Use InterProScan tool to search Pfam, PROSITE and CATH databases for protein P sequence to obtain a protein function annotation information feature file

[0018] Furthermore, the matrix denoted as Coord in S12 is a matrix of size L×NNN×3, wherein NNN is the number of atoms of each residue, and 3 is the three-dimensional spatial coordinate of each atom.

[0019] Furthermore, the protein function annotation information file in S15 in, represents the i-th functional annotation information in FUN, NNN represents the total number of functional annotation information in FUN, and each functional annotation information Contains a protein function annotation text STR and the corresponding function interval R = [start, end], where 1≤start≤end≤L.

[0020] Furthermore, the size of the embedding matrix M in S2 is L×1534, where L is the number of residues in the protein P and 1534 is the feature dimension of each residue.

[0021] Further, the S3 comprises the following steps:

[0022] S31. The sample corresponding to any residue in the protein sequence P is represented by (M i ,lab i ), i = 1, 2, ..., L, M i and lab i They represent the feature vector and DNA-bound residue label information of the ith residue, respectively, with 0 representing non-DNA-bound residues and 1 representing DNA-bound residues;

[0023] S32. Collect all DNA-protein complex structures, define DNA binding residues, extract protein sequence sets with known DNA binding residue labels from all DNA-protein complexes, use MMseqs2 tool to cluster protein sequences and select representative sequences as training sets;

[0024] S33. Generate all proteins P i The residue sample set Ω={fea i , label i}, the feature matrix and label vector of each protein sample are denoted as fea i and label i , i=1,2,...,C;

[0025] S34. According to the Coord in S12 and the feature matrix fea in S24 i , construct a protein graph G = (V, E, X), where V = {v 0 , ..., v i , ..., v L-1} represents a set of residue nodes, any node v i The feature is a feature vector of length 1534 The edges are defined by the contact residue pairs of the i-th and j-th residues, and the Euclidean distance is less than The residue pairs are considered to be in contact, X = {x0, x1, ..., x i , ..., x L} is the central carbon atom C of each residue a The coordinate set of ;

[0026] S35. Construct an equivariant graph neural network model, denoted as EGNN, which consists of three equivariant graph convolutional layers, each of which is denoted as EGCL. Input the protein graph G in S34 into EGCL and perform the following transformation:

[0027]

[0028] in, and represents edge and node operations based on multi-layer perceptron, C = 1 / (L-1) represents averaging, Y ij Converted into a scalar, the node feature set M output by each layer of EGCL l+1 and coordinate set X l+1 Used as input for the next layer of EGCL, the node feature set of the last layer of EGCL is a vector of length L, in which the i-th element represents the probability that the i-th node residue is predicted to be a DNA binding residue.

[0029] Further, the S23 comprises the following steps:

[0030] S231. Collect all DNA-protein complex structures from the PDB library and select those structures whose Euclidean distance to any atom of the DNA molecule is less than The residues of are defined as DNA binding residues;

[0031] S232. Extracting a set of protein sequences with known DNA-binding residue signatures from all DNA-protein complexes

[0032] S233. Use the MMseqs2 tool with a clustering threshold of 0.4 to cluster Seq into C clusters and take out all C representative sequences as the training set, denoted as D = {P i , Y i}, i = 1, 2, ..., C, P i represents the i-th protein sequence in the training set D, Y i Indicates P i The label information of , C is the total number of protein sequences.

[0033] Further, the S4 comprises the following steps:

[0034] S41. Input the protein residue sample set Ω in S33 into the EGNN model constructed in S35 for training, use the cross entropy function as the loss, AdamW as the parameter optimizer of the model, and use the Dropout algorithm with a loss rate of 0.5 and L2 regularization with a λ parameter of 0.1 to prevent the model from overfitting;

[0035] S42. The output p of each residue node is used as the probability that the target residue is predicted as a positive sample. p ≥ 0.5 indicates a DNA-binding residue, and p < 0.5 indicates a non-DNA-binding residue.

[0036] By adopting the above technical solution, the present invention has the following advantages:

[0037] A method for predicting DNA binding residues based on a multimodal protein language model provides a method for predicting DNA binding residues based on a multimodal protein language model with low computational cost and high partitioning accuracy. On the one hand, the multimodal information of protein sequence, structure and functional annotation is extracted, and more useful constraint and evolutionary information is obtained, which is ready for improving the accuracy of DNA binding residue prediction; on the other hand, the rotational invariance of the protein graph is maintained by using an equivariant graph neural network, and the spatial constraint information of neighboring residues is learned, which further improves the efficiency and accuracy of DNA binding residue prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of the process of the method for predicting DNA binding residues based on a multimodal protein language model of the present invention;

[0039] Figure 2 The results of predicting DNA binding residues of protein P0ACF8 using the method of the present invention in the examples are shown. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is described in detail below in conjunction with the drawings in the specification. It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0041] A method for predicting DNA binding residues based on a multimodal protein language model comprises the following steps:

[0042] S1. The protein sequence P to be used for DNA binding residue prediction is used to obtain the protein structure PDB feature file, protein secondary structure DSSP feature file, protein solvent accessibility SS feature file and InterPro function annotation feature file respectively using AlphaFold2, DSSP, Shrake-Rupley and InterProScan tools;

[0043] S2. Input all feature files obtained in S1 into the multimodal protein language model ESM3 to obtain an embedding matrix M as the feature representation of protein P;

[0044] S3. Process the protein sequence into residue samples, build an isovariant graph neural network model based on the protein structure graph, use the protein sequence of known DNA binding residues to construct a data set and train the built network;

[0045] S4. Input the residue samples of the protein sequence for which DNA binding residues are to be predicted into the trained model to obtain the DNA binding residues of the protein sequence.

[0046] S1 includes the following steps:

[0047] S11. Input a protein sequence information with a residue number of which DNA binding residues are to be predicted, denoted as P;

[0048] S12. For protein P, the AlphaFold2 tool is used to predict its three-dimensional structure information to obtain the PDB feature file, and its atomic coordinates are represented as a matrix denoted as Coord; the matrix denoted as Coord in S12 is a matrix of size L×NNN×3, where NNN is the number of atoms in each residue and 3 is the three-dimensional space coordinate of each atom.

[0049] S13. For the PDB feature file of protein P obtained in S12, the DSSP tool was used to calculate the eight-state secondary structure of protein P: SS = {s1, s2, ..., s i , ..., s L}, where s i ∈(H, B, E, G, I, T, S, L) represents the secondary structure state of the i-th residue of protein P;

[0050] S14. Use the Shrake-Rupley algorithm to calculate the solvent accessible surface area SA of the protein P obtained in S12 using the PDB feature file. i , ..., a L}, non-negative real value a i represents the solvent accessible surface area of ​​the ith residue of protein P;

[0051] S15. Use InterProScan tool to search Pfam, PROSITE and CATH databases for protein P sequence to obtain a protein function annotation information feature file Protein function annotation information file in, represents the i-th functional annotation information in FUN, NNN represents the total number of functional annotation information in FUN, and each functional annotation information Contains a protein function annotation text STR and the corresponding function interval R = [start, end], where 1≤start≤end≤L.

[0052] The size of the embedding matrix M in S2 is L×1534, where L is the number of residues in protein P and 1534 is the feature dimension of each residue.

[0053] S3 includes the following steps:

[0054] S31. The sample corresponding to any residue in the protein sequence P is represented by (M i ,lab i ), i = 1, 2, ..., L, M i and lab i They represent the feature vector and DNA-bound residue label information of the ith residue, respectively, with 0 representing non-DNA-bound residues and 1 representing DNA-bound residues;

[0055] S32. Collect all DNA-protein complex structures, define DNA binding residues, extract protein sequence sets with known DNA binding residue labels from all DNA-protein complexes, use MMseqs2 tool to cluster protein sequences and select representative sequences as training sets;

[0056] S33. Generate all proteins P i The residue sample set Ω={fea i , label i}, the feature matrix and label vector of each protein sample are denoted as fea i and label i , i=1,2,...,C;

[0057] S34. According to the Coord in S12 and the feature matrix fea in S24 i , construct a protein graph G = (V, E, X), where V = {v 0 , ..., v i , ..., v L-1} represents a set of residue nodes, any node v i The feature is a feature vector of length 1534 The edges are defined by the contact residue pairs of the i-th and j-th residues, and the Euclidean distance is less than The residue pairs are considered to be in contact, X = {x0, x1, ..., x i , ..., x L} is the central carbon atom C of each residue a The coordinate set of ;

[0058] S35. Construct an equivariant graph neural network model, denoted as EGNN, which consists of three equivariant graph convolutional layers, each of which is denoted as EGCL. Input the protein graph G in S34 into EGCL and perform the following transformation:

[0059]

[0060] in, and represents edge and node operations based on multi-layer perceptron, C = 1 / (L-1) represents averaging, Y ij Converted into a scalar, the node feature set M output by each layer of EGCL l+1 and coordinate set X l+1 Used as input for the next layer of EGCL, the node feature set of the last layer of EGCL is a vector of length L, in which the i-th element represents the probability that the i-th node residue is predicted to be a DNA binding residue.

[0061] S32 includes the following steps:

[0062] S321. Collect all DNA-protein complex structures from the PDB library and select those structures whose Euclidean distance to any atom of the DNA molecule is less than The residues of are defined as DNA binding residues;

[0063] S322. Extracting a set of protein sequences with known DNA-binding residue signatures from all DNA-protein complexes

[0064] S323. Use the MMseqs2 tool with a clustering threshold of 0.4 to cluster Seq into C clusters and take out all C representative sequences as the training set, denoted as D = {P i , Y i}, i = 1, 2, ..., C, P i represents the i-th protein sequence in the training set D, Y i Indicates P i The label information of , C is the total number of protein sequences.

[0065] S4 includes the following steps:

[0066] S41. Input the protein residue sample set Ω in S33 into the EGNN model constructed in S35 for training, use the cross entropy function as the loss, AdamW as the parameter optimizer of the model, and use the Dropout algorithm with a loss rate of 0.5 and L2 regularization with a λ parameter of 0.1 to prevent the model from overfitting;

[0067] S42. The output p of each residue node is used as the probability that the target residue is predicted as a positive sample. p ≥ 0.5 indicates a DNA-binding residue, and p < 0.5 indicates a non-DNA-binding residue.

[0068] In a specific embodiment, taking the prediction of DNA binding residues of protein P0ACF8 as an example, a method for predicting DNA binding residues based on a multimodal protein language model is as follows: Figure 1 As shown, the following steps are included:

[0069] S1. The protein sequence P to be used for DNA binding residue prediction is used to obtain the protein structure PDB feature file, protein secondary structure DSSP feature file, protein solvent accessibility SS feature file and InterPro function annotation feature file respectively using AlphaFold2, DSSP, Shrake-Rupley and InterProScan tools;

[0070] S11. Input the sequence information of a protein P0ACF8 with 137 residues for which DNA binding residues prediction is to be performed, denoted as P;

[0071] S12. For protein P, use the AlphaFold2 tool to predict its three-dimensional structure information to obtain the PDB file, and represent its atomic coordinates as a matrix of size 137×NNN×3, denoted as Coord, where NNN is the number of atoms in each residue and 3 is the three-dimensional spatial coordinate of each atom;

[0072] S13. For the PDB file of protein P obtained in S12, the DSSP tool was used to calculate the eight-state secondary structure of protein P: SS = {s1, s2, ..., s i , ..., s 137}, where s i ∈(H, B, E, G, I, T, S, L) represents the secondary structure state of the i-th residue of protein P;

[0073] S14. The solvent accessible surface area SA = {a1, a2, ..., a i , ..., a 137}, non-negative real value a i represents the solvent accessible surface area of ​​the ith residue of protein P;

[0074] S15. Use InterProScan tool to search Pfam, PROSITE and CATH databases for protein P sequence to obtain a protein function annotation information file in represents the i-th functional annotation information in FUN, NNN represents the total number of functional annotation information in FUN, and each functional annotation information Contains a protein function annotation text STR and the corresponding function interval R = [start, end], where 1≤start≤end≤137.

[0075] S2. Input all the feature inputs obtained in S1, namely P, Coord, SS, SA and FUN, into the multimodal protein language model ESM3 to obtain an embedding matrix M of size 137×1534 as the feature representation of protein P, where 137 is the number of residues in protein P and 1534 is the feature dimension of each residue.

[0076] S3. Process the protein sequence into residue samples, build an isovariant graph neural network model based on the protein structure graph, use the protein sequence of known DNA binding residues to construct a data set and train the built network;

[0077] S31. The sample corresponding to any residue in the protein sequence P is represented by (M i ,lab i ), i = 1, 2, ..., 137, M i and lab i They represent the feature vector and DNA-bound residue label information of the ith residue, respectively, with 0 representing non-DNA-bound residues and 1 representing DNA-bound residues;

[0078] S32. Collect all DNA-protein complex structures, define DNA binding residues, extract protein sequence sets with known DNA binding residue labels from all DNA-protein complexes, use MMseqs2 tool to cluster protein sequences and select representative sequences as training sets;

[0079] S32 includes the following specific steps:

[0080] S321. Collect all DNA-protein complex structures from the PDB library and select those structures whose Euclidean distance to any atom of the DNA molecule is less than The residues of are defined as DNA binding residues;

[0081] S322. Extracting a set of protein sequences with known DNA-binding residue signatures from all DNA-protein complexes

[0082] S323. Use the MMseqs2 tool with a clustering threshold of 0.4 to cluster Seq into C clusters and take out all C representative sequences as the training set, denoted as D = {P i , Y i}, i = 1, 2, ..., C, P i represents the i-th protein sequence in the training set D, Y i Indicates P i The label information of , C is the total number of protein sequences.

[0083] S33. Generate all proteins P i The residue sample set Ω={fea i , label i}, the feature matrix and label vector of each protein sample are denoted as fea i and label i , i=1,2,...,C;

[0084] S34. Protein structure atomic coordinate matrix Coord generated in S12 and feature matrix fea generated in S24 i, construct a protein graph G = (V, E, X), where V = {v 0 , ..., v i , ..., v 136} represents a set of residue nodes, any node v i The feature is a feature vector of length 1534 The edges are defined by the contact residue pairs of the i-th and j-th residues, and the Euclidean distance is less than The residue pairs are considered to be in contact, X = {x0, x1, ..., x i , ..., x 136} is the central carbon atom C of each residue a The coordinate set of ;

[0085] S35. Construct an equivariant graph neural network model, denoted as EGNN, which consists of three equivariant graph convolutional layers, each denoted as EGCL. Input the protein graph G obtained in S34 into EGCL and perform the following transformation:

[0086]

[0087] in, and represents edge and node operations based on multi-layer perceptron, C=1 / 136 represents averaging, Y ij Converted into a scalar, the node feature set M output by each layer of EGCL l+1 and coordinate set X l+1 Used as input for the next layer of EGCL, the node feature set of the last layer of EGCL is a vector of length L, in which the i-th element represents the probability that the i-th node residue is predicted to be a DNA binding residue.

[0088] S4. Input the residue samples of the protein sequence for which DNA binding residues are to be predicted into the trained model to obtain the DNA binding residues of the protein sequence.

[0089] S41. Input the protein residue sample set Ω in S33 into the EGNN model constructed in S35 for training, use the cross entropy function as the loss, AdamW as the parameter optimizer of the model, and use the Dropout algorithm with a loss rate of 0.5 and L2 regularization with a λ parameter of 0.1 to prevent the model from overfitting;

[0090] S42. Generate a protein graph and a residue feature vector for protein P, and input them into the EGNN model trained in step S41. The output p of each residue node is used as the probability that the target residue is predicted as a positive sample. p ≥ 0.5 indicates a DNA-bound residue, and p < 0.5 indicates a non-DNA-bound residue.

[0091] Taking the prediction of DNA binding residues of protein P0ACF8 as an example, the DNA binding residues of protein P0ACF8 were obtained by using the above method. Figure 2 shown.

[0092] Finally, it should be pointed out that although the present invention has been described with reference to the current specific embodiments, ordinary technicians in this technical field should realize that the above embodiments are only used to illustrate the present invention and are not used as a limitation of the present invention. Various equivalent changes or substitutions can be made without departing from the concept of the present invention. Therefore, as long as the changes and modifications to the above embodiments are within the essential spirit of the present invention, they will fall within the scope of the claims of the present invention.

Claims

1. A method for predicting DNA binding residues based on a multimodal protein language model, characterized in that: The following steps are involved: S1. The protein sequence P to be used for DNA binding residue prediction is used to obtain the protein structure PDB feature file, protein secondary structure DSSP feature file, protein solvent accessibility SS feature file and InterPro function annotation feature file respectively using AlphaFold2, DSSP, Shrake-Rupley and InterProScan tools; S2. Input all feature files obtained in S1 into the multimodal protein language model ESM3 to obtain an embedding matrix M as the feature representation of protein P; S3. Process the protein sequence into residue samples, build an isovariant graph neural network model based on the protein structure graph, use the protein sequence of known DNA binding residues to construct a data set and train the built network; S4. Input the residue samples of the protein sequence for which DNA binding residues are to be predicted into the trained model to obtain the DNA binding residues of the protein sequence.

2. The method for predicting DNA binding residues based on a multimodal protein language model according to claim 1, characterized in that: The S1 comprises the following steps: S11. Input a protein sequence information with a residue number of which DNA binding residues are to be predicted, denoted as P; S12. For protein P, use the AlphaFold2 tool to predict its three-dimensional structure information to obtain a PDB feature file, and represent its atomic coordinates as a matrix denoted as Coord; S13. For the PDB feature file of protein P obtained in S12, the DSSP tool was used to calculate the eight-state secondary structure of protein P: SS = {s1, s2, …, s i ,…,s L }, where s i ∈(H,B,E,G,I,T,S,L) represents the secondary structure state of the i-th residue of protein P; S14. Use the Shrake-Rupley algorithm to calculate the solvent accessible surface area SA of the protein P obtained in S12 using the PDB feature file. i ,…,a L }, non-negative real value a i represents the solvent accessible surface area of ​​the ith residue of protein P; S15. Use InterProScan tool to search Pfam, PROSITE and CATH databases for protein P sequence to obtain a protein function annotation information feature file 3. The method for predicting DNA binding residues based on a multimodal protein language model according to claim 2, characterized in that: The matrix denoted as Coord in S12 is a matrix of size L×N×3, wherein N is the number of atoms in each residue and 3 is the three-dimensional spatial coordinate of each atom.

4. The method for predicting DNA binding residues based on a multimodal protein language model according to claim 2, characterized in that: The protein function annotation information file in S15 in, represents the i-th functional annotation information in FUN, N represents the total number of functional annotation information in FUN, and each functional annotation information Contains a protein function annotation text STR and the corresponding function interval R = [start, end], where 1≤start≤end≤L.

5. The method for predicting DNA binding residues based on a multimodal protein language model according to claim 2, characterized in that: The size of the embedding matrix M in S2 is L×1534, where L is the number of residues in protein P and 1534 is the feature dimension of each residue.

6. The method for predicting DNA binding residues based on a multimodal protein language model according to claim 2, characterized in that: The S3 comprises the following steps: S31. The sample corresponding to any residue in the protein sequence P is represented by (M i ,lab i ), i=1,2,…,L,M i and lab i They represent the feature vector and DNA-bound residue label information of the ith residue, respectively, with 0 representing non-DNA-bound residues and 1 representing DNA-bound residues; S32. Collect all DNA-protein complex structures, define DNA binding residues, extract protein sequence sets with known DNA binding residue labels from all DNA-protein complexes, use MMseqs2 tool to cluster protein sequences and select representative sequences as training sets; S33. Generate all proteins P i The residue sample set Ω={fea i ,label i }, the feature matrix and label vector of each protein sample are denoted as fea i and label i , i=1,2,…,C; S34. According to Coored in S12 and the feature matrix fea in S24 i , construct a protein graph G = (V, E, X), where V = {v 0 ,…,v i ,…,v L-1 } represents a set of residue nodes, any node v i The feature is a feature vector of length 1534 e ij ∈E represents the edge defined by the contact residue pair of the i-th and j-th, and the Euclidean distance is less than The residue pairs are considered to be in contact, X = {x0, x1, …, x i ,…,x L } is the central carbon atom C of each residue a The coordinate set of ; S35. Construct an equivariant graph neural network model, denoted as EGNN, which consists of three equivariant graph convolutional layers, each of which is denoted as EGCL. Input the protein graph G in S34 into EGCL and perform the following transformation: in, and represents edge and node operations based on multi-layer perceptron, C = 1 / (L-1) represents averaging, Y ij Converted into a scalar, the node feature set M output by each layer of EGCL l+1 and coordinate set X l+1 Used as input for the next layer of EGCL, the node feature set of the last layer of EGCL is a vector of length L, in which the i-th element represents the probability that the i-th node residue is predicted to be a DNA binding residue.

7. The method for predicting DNA binding residues based on a multimodal protein language model according to claim 6, characterized in that: The S32 comprises the following steps: S321. Collect all DNA-protein complex structures from the PDB library and select those structures whose Euclidean distance to any atom of the DNA molecule is less than The residues of are defined as DNA binding residues; S322. Extracting a set of protein sequences with known DNA-binding residue signatures from all DNA-protein complexes S323. Use the MMseqs2 tool with a clustering threshold of 0.4 to cluster Seq into C clusters and take out all C representative sequences as the training set, denoted as D = {P i ,Y i }, i = 1, 2, ..., C, P i represents the i-th protein sequence in the training set D, Y i Indicates P i The label information of , C is the total number of protein sequences.

8. The method for predicting DNA binding residues based on a multimodal protein language model according to claim 7, characterized in that: The S4 comprises the following steps: S41. Input the protein residue sample set Ω in S33 into the EGNN model constructed in S35 for training, use the cross entropy function as the loss, AdamW as the parameter optimizer of the model, and use the Dropout algorithm with a loss rate of 0.5 and L2 regularization with a λ parameter of 0.1 to prevent the model from overfitting; S42. The output p of each residue node is used as the probability that the target residue is predicted as a positive sample. p ≥ 0.5 means it is a DNA-binding residue, and p < 0.5 means it is a non-DNA-binding residue.

Images