A deep learning algorithm for predicting the monomer and complex structures of cyclic and linear peptides containing non-standard amino acids.

Abstract

This invention discloses a deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids. The specific steps include: constructing training datasets containing monomers and complexes of head-to-tail cyclic peptides, disulfide-bonded cyclic peptides, and linear peptides containing non-standard amino acids; training and optimizing a peptide structure prediction model containing basic protein knowledge; constructing test datasets containing linear peptide monomers, linear peptide complexes, head-to-tail cyclic peptide monomers, and disulfide-bonded cyclic peptide monomers; and testing the trained peptide structure prediction model on the test datasets. Compared with existing peptide structure prediction models, this invention proposes for the first time a high-precision structure prediction method for monomers and complexes of cyclic peptides and linear peptides containing non-standard amino acids, without requiring any target protein structural information or pocket information. It only requires the one-dimensional sequence of the protein and non-standard amino acid side chain groups as input to accurately predict the three-dimensional structure of the peptide.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided drug design and development technology, specifically involving a deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids. Technical Background

[0002] In recent years, peptides have been continuously explored as potential drugs. Compared with small molecule drugs, peptides possess unique physicochemical properties, including larger size and more flexible backbones, giving them an inherent advantage as inhibitors or activators of protein-protein interaction (PPIs). Compared with linear peptides, cyclic peptides have advantages in stability and membrane permeability. Considering the significant potential of peptides in the therapeutic market, understanding the structural information of peptides is crucial before further designing novel peptides. In recent years, the rise of deep learning has led to significant progress in biomolecular modeling. AlphaFold2 (AF2) and RoseTTAFold have achieved unprecedented accuracy in protein structure prediction. Studies have shown that even when AF2's training set excludes peptides shorter than 16, it still learns the protein folding mechanism from tens of millions of datasets. Since peptide binding is essentially an extension of folds, it can still make relatively accurate predictions of peptide-protein complexes. Based on this understanding, AfCycDesign adds cyclic constraints to the AF2 network to embed cyclic information for the structure prediction of cyclic peptides. HighFold then further considered the structural information of the cyclic beginning and end and the disulfide bond bridge, which improved the accuracy of disulfide bond prediction while ensuring the cyclic beginning and end.

[0003] Despite these findings, the idealistic-realistic limitations persist in real-world drug development. Most importantly, the aforementioned models can only predict peptide structures containing natural amino acids. While these peptides have the potential to be therapeutic agents, their practical application is limited, primarily due to their low stability to proteolytic enzymes and low bioavailability. Introducing non-natural amino acids can significantly improve the pharmacokinetic characteristics of peptides. However, to date, there is no dedicated model for predicting the structure of peptides containing non-standard amino acids. Considering the current need for predicting the structure of peptides containing non-standard amino acids, there is an urgent need for a high-precision peptide structure calculation method that requires no three-dimensional structural information. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a deep learning algorithm for predicting the monomer and complex structures of cyclic and linear peptides containing non-standard amino acids, thereby helping researchers quickly understand the peptide structures containing non-standard amino acids and promoting the research progress of peptide drugs.

[0005] The specific technical solution is as follows:

[0006] A deep learning algorithm for predicting the monomer and complex structures of cyclic and linear peptides containing non-standard amino acids includes the following steps:

[0007] Step 1: Obtain a training dataset from the RCSB Protein Data Bank database containing monomers and complexes of cyclic peptides with heads and tails, monomers and complexes of cyclic peptides with disulfide bonds, and linear peptide monomers and complexes containing non-standard amino acids.

[0008] Step two involves further processing the obtained training dataset and organizing it into a data format usable by the model: a txt file containing feature information of peptide chains and their atoms in all data samples for training the RHT module; and a YAML file for training peptide structure prediction.

[0009] Step 3: The model incorporates cyclic information by defining the relative position matrix of the model feature construction module, and enhances the model's ability to resolve different levels of amino acid details by introducing a Transformer-based RHT module in the model feature construction.

[0010] Step 4: Use the training dataset prepared in Step 1 to pre-train the RHT module, and then merge the trained RHT module with the main model.

[0011] Step 5: Input the file prepared in Step 2 into the model for training. After obtaining the encoder output of the RHT module and the predicted 3D coordinates of the atoms, calculate the loss function value and perform gradient backpropagation. Iterate the above process until the model converges.

[0012] Step 6: Obtain the test dataset containing non-standard amino acids, including cyclic peptide monomers, disulfide bond cyclic peptide monomers, linear peptide monomers, and linear peptide complexes.

[0013] Step 7: Repeat the steps in Step 2 to configure the YAML file;

[0014] Step 8: Input the YAML file into the model to obtain the atomic 3D coordinates of the peptide monomer or complex, and save it in PDB format; after aligning the predicted structure with the true structure, calculate the RMSD. For complexes, also calculate the DockQ value and F. nat The value is used to evaluate the model's performance in protein-peptide docking.

[0015] Furthermore, in step one, apart from the monomer structures of cyclic peptides, the rest of the training dataset samples were obtained from the literature. Then, PDB data were downloaded in batches from the PDB database according to PDB_ID. For the monomer structures of cyclic peptides, the sequences of cyclic peptides were extracted and input into the HighFold model to predict their three-dimensional structures. The predicted three-dimensional structures were used as a supplement to the cyclic peptide data as part of the training dataset.

[0016] Furthermore, the txt file in step two includes residue sequence, atom name, three-dimensional coordinates of atom, number of atoms per residue, peptide chain identifier, atom identifier, atom index, number of non-standard amino acids, peptide chain length, number of atoms per residue, and complex identifier. The construction process of the txt file is as follows: extract residue sequence, atom name, three-dimensional coordinates of atom, number of atoms per residue, peptide chain identifier, atom identifier, atom index, number of non-standard amino acids, peptide chain length, number of atoms per residue, and complex identifier from the PDB file obtained in step one.

[0017] Furthermore, the YAML file in step two contains the file path of the input sequence, the SDF file path of the side chain small molecule groups required for non-standard amino acids, the covalent binding position information of the small molecules and peptide chains, and information on whether head-to-tail cyclization is required. For linear peptide monomers and their complex files containing non-standard amino acids, the non-standard amino acids are split into standard amino acids and side chain small molecule residues. The FASTA file of the one-dimensional amino acid sequence is placed in the proteininputs position of the configuration file, and the side chain small molecule residues are placed in the sminputs position of the configuration file.

[0018] Furthermore, the cyclization information in step three is achieved by modifying the relative position matrix. The relative position matrix of NCPepFold is divided into two parts. One part is the relative position matrix of standard amino acid residues, with a size of NX×NX, where N is the peptide chain length and X is the number of non-standard amino acids contained in the peptide chain. The relative positions of the N-terminus and C-terminus of the peptide are set to 1, and the relative positions of other residues are changed accordingly, thereby achieving cyclization constraint. The basic idea is:

[0019]

[0020] Among them, c ij d represents the relative positions of the i-th and j-th amino acids in the relative position matrix of the cyclic peptide. ij It represents the relative positions of the i-th and j-th amino acids in the relative position matrix of the linear peptide, L SIt represents the number of residues in the peptide chain, including non-standard amino acids. |·| represents the absolute value sign. signal(x) represents the directional relationship of relative positions. signal(x) = 1 indicates that amino acid i is in the positive direction relative to amino acid j. signal(x) = -1 indicates that amino acid i is in the negative direction relative to amino acid j. signal(x) = 0 indicates that the relative position between amino acid i and amino acid i itself is 0.

[0021] Furthermore, in step three, the RHT module extracts the encoder output and merges it with the feature embeddings of the two relative position matrices, providing the model with information at different levels and enhancing its understanding of peptide information. The dimension of the encoder output is... Where N AT It represents the total number of atoms in the peptide chain, while the dimension of the feature embedding of the two relative position matrices is... Where N RA It is the sum of the number of standard amino acids and the total number of atoms of non-standard amino acids. To match the two, a converter is used to summarize the information of the standard amino acid portion in the encoder output from atoms to the amino acid level. The atomic characterization of the non-standard amino acid portion remains unchanged. Then, the pairwise features between residues and atoms are calculated by Einstein summation. Finally, the amino acid pairwise features and the atom pairwise features are merged by diagonal splicing of blocks to obtain features with the same dimension as the feature embedding of the two relative position matrices.

[0022] Furthermore, the loss function used in step four when training the RHT module is the cross-entropy loss term, and the specific calculation formula is as follows:

[0023]

[0024] Where N is the number of samples, R is the number of classes, and y ji This represents the true label of the j-th sample in the i-th class. This represents the predicted probability of the j-th sample in the i-th class.

[0025] Furthermore, the loss function during training in step five is the frame alignment error, which is used for the structural loss of the atomic 3D coordinate prediction module. The specific calculation formula is shown below:

[0026]

[0027] Where, N frames N represents the number of frames. atoms This represents the number of atoms, where p represents the actual structural coordinates of the atoms. Represents the predicted structure coordinates of the atom.

[0028] Furthermore, the data samples in Step 6 are all obtained from the literature, and then the PDB data are downloaded in batches from the PDB database according to the PDB_ID; the YAML file in Step 7 contains the file path of the input sequence, the file path of the side-chain small molecule group SDF file required for non-standard amino acids, the covalent binding position information of the small molecule and the peptide chain, and the information on whether ring closure is required at the beginning and end.

[0029] Furthermore, the RMSD in Step 8 is used to measure the average distance deviation of the corresponding atoms between the native structure and the predicted structure. RMSD includes peptide chain RMSD and non-standard residue RMSD, which respectively represent the deviation of the overall peptide chain and the accuracy of the local structure of non-standard amino acids. The RMSD calculation formula is as follows:

[0030]

[0031] where p1 is the true structure of the peptide chain, p2 is the predicted structure of the peptide chain, and i is the serial number of the C α atom, x i , y i, z i respectively represent the three-dimensional coordinates of the structure. The chain RMSD is to first extract the common C α atoms of the native structure and the predicted structure. For monomeric structures, the Kabsch algorithm is used to align the coordinates of the entire peptide chain, and the RMSD of the C α atoms of the entire peptide chain is calculated. For complex structures, the Kabsch algorithm is used to align the coordinates of the protein part, and the RMSD of the C α atoms of the entire peptide chain; the non-standard residue RMSD is to align the heavy atoms on the peptide chain within the non-standard residue and calculate the all-atom RMSD of the non-standard residue. If there are multiple non-standard residues on a peptide chain, the RMSD of each non-standard residue is calculated and then averaged; the DockQ in Step 8 is used to evaluate the docking quality of the peptide-protein complex. DockQ is a comprehensive score of three indicators, namely F nat , LRMS and iRMS. DockQ divides the structural quality of the complex into four grades. DockQ < 0.23 is unacceptable quality, 0.23 < DockQ < 0.49 is acceptable quality, 0.49 < DockQ < 0.8 is medium quality, and DockQ > 0.8 is high-quality prediction. The DockQ calculation formula is as follows:

[0032] DockQ(F nat , LRMS, iRMS, d1, d2) = (F nat +RMS scaled (LRMS, d1)+RMS scaled (iRMS, d2)) / 3 where, F natDefined as heavy atoms in a crystal complex Using protein-peptide residue pairs from mainland China, LRMS represents the local root mean square error, iRMS refers to the global root mean square error, and d1 and d2 are scaling factors used to adjust the weights of the local RMSD.

[0033] F nat Used to represent the success rate of grounding; for peptide-protein complexes, F nat A composite material with a density ≥0.5 is considered a high-quality docking. nat The calculation formula is as follows:

[0034]

[0035] Among them, S N S represents the set of amino acid pairs in the real structure where the distance between the ligand sequence and the receptor sequence is less than a threshold θ. M This represents the set of amino acid pairs in the predicted structure where the distance between the ligand sequence and the receptor sequence is less than the threshold θ, and ∩ represents the intersection.

[0036] The beneficial effects of this invention are as follows: Firstly, by introducing cyclization constraints, the model's understanding of cyclic structures is achieved, overcoming the limitation of previous techniques being unable to predict cyclic peptide structures and realizing high-precision prediction of cyclic peptide structures containing non-standard amino acids. Secondly, an RHT module based on the Transformer model is introduced, expanding the residue atomization range to the entire polypeptide chain, achieving comprehensive characterization of peptides with different particle sizes at the residue and atomic levels, and effectively integrating the inherent hierarchical relationships of the peptide chain. This integration method significantly improves the model's ability to resolve multi-level details of amino acids, especially non-standard amino acids, further enhancing the model's prediction accuracy. The structural prediction algorithms for monomers and complexes of cyclic peptides and linear peptides containing non-standard amino acids involved in this invention will provide valuable insights into the influence of non-standard amino acids on the local structure of peptides, which is of particular importance for the discovery and design of peptide drugs. Attached Figure Description

[0037] Figure 1 The flowchart shows the structure prediction algorithm for monomers and complexes of cyclic peptides and linear peptides containing non-standard amino acids.

[0038] Figure 2 This is a flowchart of the training process for a model based on transfer learning. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings, providing a clear and complete description of the technical solutions in the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0040] Example 1

[0041] like Figure 1 As shown, a deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids includes the following steps:

[0042] Step 1: Obtain training datasets from the PDB database containing monomers and complexes of cyclic peptides with anterior and posterior ends, monomers and complexes of cyclic peptides with disulfide bonds, and monomers and complexes of linear peptides containing non-standard amino acids. Except for the monomer structures of cyclic peptides with anterior and posterior ends, the other data samples in the training dataset are obtained from literature. Then, download PDB data in batches from the PDB database according to PDB_ID. For the monomer structures of cyclic peptides with anterior and posterior ends, extract the sequences of the cyclic peptides with anterior and posterior ends, input them into the HighFold model to predict their three-dimensional structures, and use the predicted three-dimensional structures as a supplement to the cyclic peptide data with anterior and posterior ends as part of the training dataset.

[0043] Step two involves further processing the obtained training dataset, organizing it into a data format usable by the model: a txt file containing feature information of peptide chains and their atoms in all data samples, used for training the RHT module; and a YAML file for training peptide structure prediction. The txt file includes residue sequences, atom names, three-dimensional coordinates of atoms, number of atoms per residue, peptide chain identifier, atom identifier, atom index, number of non-standard amino acids, peptide chain length, number of atoms per residue, and complex identifier. The txt file is constructed as follows: extracting residue sequences, atom names, three-dimensional coordinates of atoms, number of atoms per residue, peptide chain identifier, atom identifier, atom index, number of non-standard amino acids, peptide chain length, number of atoms per residue, and complex identifier from the PDB file obtained in step one. The YAML file contains the file path of the input sequence, the SDF file path of the side chain small molecule groups required for non-standard amino acids, information on the covalent binding positions of small molecules and peptide chains, and information on whether head-to-tail looping is required. For linear peptide monomers and their complex files containing non-standard amino acids, the non-standard amino acids are split into standard amino acids and side chain small molecule residues. The FASTA file of the one-dimensional amino acid sequence is placed into the protein configuration file. The inputs position places side chain small molecule residues into the sminputs position of the configuration file.

[0044] Step three: The model incorporates cyclization information by defining the relative position matrix of the model feature construction module, and enhances its ability to resolve different levels of amino acid details by introducing a Transformer-based RHT module into the model feature construction. The cyclization information in step three is achieved by modifying the relative position matrix. The relative position matrix of NCPepFold is divided into two parts: one part is the relative position matrix of standard amino acid residues, with a size of NX×NX, where N is the peptide chain length and X is the number of non-standard amino acids contained in the peptide chain. The relative positions of the N-terminus and C-terminus of the peptide are set to 1, and the relative positions of other residues are changed accordingly, thereby achieving cyclization constraint. The basic idea is:

[0045]

[0046] Among them, c ij d represents the relative positions of the i-th and j-th amino acids in the relative position matrix of the cyclic peptide. ij It represents the relative positions of the i-th and j-th amino acids in the relative position matrix of the linear peptide, L s It represents the number of residues in the peptide chain, including non-standard amino acids. |·| represents the absolute value sign. signal(x) represents the directional relationship of relative positions. signal(x) = 1 indicates that amino acid i is in the positive direction relative to amino acid j. signal(x) = -1 indicates that amino acid i is in the negative direction relative to amino acid j. signal(x) = 0 indicates that the relative position between amino acid i and amino acid i itself is 0.

[0047] Step three's RHT module extracts the encoder output and merges it with the feature embeddings of the two relative position matrices, providing the model with information at different levels and enhancing its understanding of peptide information. The dimension of the encoder output is... Where N AT It represents the total number of atoms in the peptide chain, while the dimension of the feature embedding of the two relative position matrices is... Where N RA It is the sum of the number of standard amino acids and the total number of atoms of non-standard amino acids. To match the two, a converter is used to summarize the information of the standard amino acid portion in the encoder output from atoms to the amino acid level. The atomic characterization of the non-standard amino acid portion remains unchanged. Then, the pairwise features between residues and atoms are calculated by Einstein summation. Finally, the amino acid pairwise features and the atom pairwise features are merged by diagonal splicing of blocks to obtain features with the same dimension as the feature embedding of the two relative position matrices.

[0048] Step four: Using the training dataset prepared in step one, pre-train the RHT module. Then, merge the trained RHT module with the main model. The loss function used when training the RHT module in step four is the cross-entropy loss term, and the specific calculation formula is as follows:

[0049]

[0050] Where N is the number of samples, R is the number of classes, and y ji This represents the true label of the j-th sample in the i-th class. This represents the predicted probability of the j-th sample in the i-th class.

[0051] Step 5: Input the file prepared in Step 2 into the model for training. After obtaining the encoder output of the RHT module and the predicted 3D coordinates of the atoms, calculate the loss function value and perform gradient backpropagation. Iterate the above process until the model converges. The loss function during training in Step 5 is the frame alignment error, which is used for the structural loss of the atom 3D coordinate prediction module. The specific calculation formula is as follows:

[0052]

[0053] Where, N frames N represents the number of frames. atoms This represents the number of atoms, where p represents the actual structural coordinates of the atoms. Represents the predicted structure coordinates of the atom.

[0054] Step six involves obtaining test datasets containing non-standard amino acids for linear peptide monomers, linear peptide complexes, first- and last-chain cyclic peptide monomers, and disulfide bond cyclic peptide monomers from the PDB database. The data samples in step six are obtained from literature, and then PDB data are downloaded in batches from the PDB database according to PDB_ID.

[0055] Step 7: Repeat the steps in Step 2 to configure the YAML file. The YAML file in Step 7 includes the file path of the input sequence, the SDF file path of the side chain small molecule groups required for non-standard amino acids, the covalent binding position information of the small molecule and peptide chain, and information on whether head-to-tail cyclization is required.

[0056] Step 8: Input the YAML file into the model to obtain the atomic 3D coordinates of the peptide monomer or complex, and save it in PDB format; after aligning the predicted structure with the true structure, calculate the RMSD. For complexes, also calculate the DockQ value and F. natValues are used to evaluate the performance of the model in protein-peptide docking. The RMSD in Step 8 is used to measure the average distance deviation of corresponding atoms between the native structure and the predicted structure. RMSD includes peptide chain RMSD and non-standard residue RMSD, which respectively represent the deviation of the overall peptide chain and the accuracy of the local structure of non-standard amino acids. The RMSD calculation formula is as follows:

[0057]

[0058] where p1 is the true structure of the peptide chain, p2 is the predicted structure of the peptide chain, and i is the serial number of the C α atom, x i , y i, z i respectively represent the three-dimensional coordinates of the structure. Chain RMSD is to first extract the C α atoms shared by the native structure and the predicted structure. For monomer structures, the Kabsch algorithm is used to align the coordinates of the entire peptide chain, and the RMSD of the C α atoms of the entire peptide chain is calculated. For complex structures, the Kabsch algorithm is used to align the coordinates of the protein part, and the RMSD of the C α atoms of the entire peptide chain; Non-standard residue RMSD is to align the heavy atoms on the peptide chain within the non-standard residue , and calculate the all-atom RMSD of the non-standard residue. If there are multiple non-standard residues on a peptide chain, the RMSD of each non-standard residue is calculated and then averaged; DockQ in Step 8 is used to evaluate the docking quality of the peptide-protein complex. DockQ is a comprehensive score of three indicators, namely F nat , LRMS, and iRMS. DockQ divides the structural quality of the complex into four grades. DockQ < 0.23 represents unacceptable quality, 0.23 < DockQ < 0.49 represents acceptable quality, 0.49 < DockQ < 0.8 represents medium quality, and DockQ > 0.8 represents high-quality prediction. The DockQ calculation formula is as follows:

[0059] DockQ(F nat , LRMS, iRMS, d1, d2) = (F nat + RMS scaled (LRMS, d1) + RMS scaled (iRMS, d2)) / 3 where F nat is defined as the heavy atoms in the crystal complex within the protein-peptide residue pairs, LRMS is the local root mean square deviation, iRMS is the global root mean square deviation, and d1, d2 are scaling factors used to adjust the weight of the local RMSD,

[0060] F natUsed to represent the success rate of grounding; for peptide-protein complexes, F nat A composite material with a density ≥0.5 is considered a high-quality docking. nat The calculation formula is as follows:

[0061]

[0062] Among them, S N S represents the set of amino acid pairs in the real structure where the distance between the ligand sequence and the receptor sequence is less than a threshold θ. M This represents the set of amino acid pairs in the predicted structure where the distance between the ligand and receptor sequences is less than a threshold θ, and ∩ represents the intersection. The data in Tables 1 to 3 were calculated using the above method.

[0063] Table 1. RMSD results of the first and last cyclic peptide monomers.

[0064] PDB ID peptide RMSD Modify RMSD 2N8E 0.274 1.209 2MSQ 0.458 0.895 2M1P 1.239 1.207 1T9E 0.229 1.019 2M2G 2.397 1.463 4M6E 4.063 2.09 2J15 1.45 0.845 2KNN 2.108 1.741 2M2S 1.767 1.753 2M2X 2.418 3.402 average 1.640 1.562 median 1.503 1.336

[0065] Table 2. RMSD results of disulfide-bonded cyclic peptide monomers

[0066] PDB ID peptide RMSD Modify RMSD PDB ID peptide RMSD Modify RMSD 1P9G 2.104 2.301 2MFX 1.824 3.118 2CRD 1.203 1.313 2MG6 1.637 3.283 1BIG 1.252 1.49 3L09 0.474 1.218 1K64 1.74 1.132 4E86 0.763 0.907 2EW4 3.587 1.558 4E83 1.075 0.963 6MY3 1.684 1.484 3HJD 0.632 1.995 6MY2 1.399 1.126 3LO6 0.697 0.717 1OMC 1.032 1.864 7N21 1.513 1.751 6MY1 0.688 1.388 7N24 1.396 1.892 2M62 2.646 2.295 7N25 1.026 2.831 5UG3 2.116 2.714 7N20 1.728 2.76 1KFP 1.241 1.703 average value 1.455 1.818 median 1.396 1.751

[0067] The average RMSD of the cyclic peptide monomer was The median is The average value of disulfide bond-linked cyclic peptide monomers is The median is The RMSDs of all modifications were slightly smaller than those of peptides, with the RMSD of cyclic peptide monomer modifications being [value missing]. Modification of disulfide bond cyclic peptide monomers with RMSD: The performance improvements over RFAA predictions were 81.9% and 21.1%, respectively. This indicates high-quality prediction results. Therefore, these results confirm that NCPepFold can accurately predict structures containing modified residues.

[0068] Table 3. RMSD, DockQ, and F of line peptide monomers and complexes nat result

[0069]

[0070] Table 3 shows the significant similarity between the linear peptide monomers and complexes and the experimental structures, with the average RMSD of the monomeric peptides being [value missing]. The average of modified RMSD is

[0071] In online peptide-protein docking, NCPepFold demonstrated impressive effectiveness. The average RMSD of the peptide in the complex was [value missing]. The average of modified RMSD is Compared to monomer structure prediction, the complexity of complexes lies in the fact that incorrect selection of binding sites can significantly affect the overall spatial configuration of the complex. Even if the peptide chain structure itself is accurately reconstructed, if the binding site deviation is large, the final complex structure will still not closely approximate the experimental values. NCPepFold demonstrates its accuracy in capturing binding sites. DockQ is used to evaluate the accuracy of complex docking; DockQ > 0.8 indicates a high-quality prediction. The median DockQ of the complex structure predicted by NCPepFold is 0.911, further confirming the good performance of NCPepFold.

[0072] It should be noted that the FASTA of the linear peptide monomer and the FASTA of the linear peptide complex are mainly from the PDB database, but the FASTA of the linear peptide monomer and the FASTA of the linear peptide complex that can be used in this invention are not limited to the above data sources. The evaluation of the results of the linear peptide monomer and the linear peptide complex requires the actual experimental structures of the linear peptide monomer and the linear peptide complex as a reference; this is an optional but not mandatory step.

[0073] The above description is only a preferred embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent substitutions or modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A deep learning algorithm for predicting the monomer and complex structures of cyclic and linear peptides containing non-standard amino acids, characterized in that, Includes the following steps: Step 1: Obtain a training dataset from the RCSB Protein Data Bank database containing monomers and complexes of cyclic peptides with heads and tails, monomers and complexes of cyclic peptides with disulfide bonds, and linear peptide monomers and complexes containing non-standard amino acids. Step two, further process the obtained training dataset and organize it into a data format for the model: a txt file containing feature information of peptide chains and their atoms in all data samples, used for training the RHT module; A YAML file used for training peptide structure prediction; Step 3: The model incorporates cyclic information by defining the relative position matrix of the model feature construction module, and enhances the model's ability to resolve different levels of amino acid details by introducing a Transformer-based RHT module in the model feature construction. Step 4: Use the training dataset prepared in Step 1 to pre-train the RHT module, and then merge the trained RHT module with the main model. Step 5: Input the file prepared in Step 2 into the model for training. After obtaining the encoder output of the RHT module and the predicted 3D coordinates of the atoms, calculate the loss function value and perform gradient backpropagation. Iterate the above process until the model converges. Step 6: Obtain test datasets containing non-standard amino acids, including cyclic peptide monomers with head and tail, disulfide bond cyclic peptide monomers, linear peptide monomers, and linear peptide complexes. Step 7: Repeat the steps in Step 2 to configure the YAML file; Step 8: Input the YAML file into the model to obtain the atomic 3D coordinates of the peptide monomer or complex, and save it in PDB format; after aligning the predicted structure with the true structure, calculate the RMSD. For complexes, also calculate the DockQ value and F. nat The value is used to evaluate the model's performance in protein-peptide docking.

2. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, In Step 1, apart from the monomer structures of cyclic peptides, the rest of the training dataset samples were obtained from the literature. Then, PDB data were downloaded in batches from the PDB database according to PDB_ID. For the monomer structures of cyclic peptides, the sequences of cyclic peptides were extracted and input into the HighFold model to predict their three-dimensional structures. The predicted three-dimensional structures were used as a supplement to the cyclic peptide data and as part of the training dataset.

3. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, The txt file in step two includes the residue sequence, atom name, three-dimensional coordinates of the atom, number of atoms per residue, peptide chain identifier, atom identifier, atom index, number of non-standard amino acids, peptide chain length, number of atoms per residue, and complex identifier. The construction process of the txt file is as follows: extract the residue sequence, atom name, three-dimensional coordinates of the atom, number of atoms per residue, peptide chain identifier, atom identifier, atom index, number of non-standard amino acids, peptide chain length, number of atoms per residue, and complex identifier from the PDB file obtained in step one.

4. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, The YAML file in step two contains the file path of the input sequence, the SDF file path of the side chain small molecule groups required for non-standard amino acids, the covalent binding position information of the small molecules and peptide chains, and information on whether head-to-tail cyclization is required. For linear peptide monomers and their complex files containing non-standard amino acids, the non-standard amino acids are split into standard amino acids and side chain small molecule residues. The FASTA file of the one-dimensional amino acid sequence is placed in the protein inputs position of the configuration file, and the side chain small molecule residues are placed in the sminputs position of the configuration file.

5. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, The cyclization information in step three is achieved by modifying the relative position matrix. The relative position matrix of NCPepFold consists of two parts. One part is the relative position matrix of standard amino acid residues, with a size of NX×NX, where N is the peptide chain length and X is the number of non-standard amino acids contained in the peptide chain. The relative positions of the N-terminus and C-terminus of the peptide are set to 1, and the relative positions of other residues are changed accordingly, thereby achieving cyclization constraint. The idea is: ， in, This represents the relative positions of the i-th and j-th amino acids in the relative position matrix of the cyclic peptide. It represents the relative positions of the i-th and j-th amino acids in the relative position matrix of the linear peptide. It refers to the number of residues contained in the peptide chain, including non-standard amino acids. The absolute value symbol is represented by signal(x). The direction relationship between the relative positions is represented by signal(x) = 1, which means that amino acid i is in the positive direction relative to amino acid j. The direction of signal(x) = -1 means that amino acid i is in the negative direction relative to amino acid j. The direction of signal(x) = 0 means that the relative position between amino acid i and amino acid j is 0.

6. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, Step three's RHT module extracts the encoder output and merges it with the feature embeddings of the two relative position matrices, providing the model with information at different levels and enhancing its understanding of peptide information. The dimension of the encoder output is... ,in It represents the total number of atoms in the peptide chain, while the dimension of the feature embedding of the two relative position matrices is... ,in It is the sum of the number of standard amino acids and the total number of atoms of non-standard amino acids. To match the two, a converter is used to summarize the information of the standard amino acid portion in the encoder output from atoms to the amino acid level. The atomic characterization of the non-standard amino acid portion remains unchanged. Then, the pairwise features between residues and atoms are calculated by Einstein summation. Finally, the amino acid pairwise features and the atom pairwise features are merged by diagonal splicing of blocks to obtain features with the same dimension as the feature embedding of the two relative position matrices.

7. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, In step four, the loss function used when training the RHT module is the cross-entropy loss term, and the specific calculation formula is as follows: ， Where N is the number of samples and R is the number of categories. This represents the true label of the j-th sample in the i-th class. This represents the predicted probability of the j-th sample in the i-th class.

8. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, The loss function used during training in step five is the frame alignment error, which is applied to the structural loss of the atomic 3D coordinate prediction module. The specific calculation formula is shown below: ， in, Indicates the number of frames. Indicates the number of atoms. Represents the actual structural coordinates of the atom. Represents the predicted structure coordinates of the atom.

9. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, The data samples in step six are all obtained from the literature, and then PDB data are downloaded in batches from the PDB database according to PDB_ID; the YAML file in step seven contains the file path of the input sequence, the SDF file path of the side chain small molecule groups required for non-standard amino acids, the covalent binding position information of small molecules and peptide chains, and information on whether head-to-tail cyclization is required.

10. The deep learning algorithm for predicting the monomer and complex structures of cyclic peptides and linear peptides containing non-standard amino acids as described in claim 1, characterized in that, The RMSD in step eight is used to measure the average distance deviation between corresponding atoms in the native and predicted structures. RMSD includes peptide chain RMSD and non-standard residue RMSD, which respectively represent the overall deviation of the peptide chain and the accuracy of the local structure of non-standard amino acids. The RMSD calculation formula is as follows: ， Among them is the true structure of the peptide chain, is the predicted structure of the peptide chain, is C α atomic serial number, x i , y i, z i respectively represent the three-dimensional coordinates of the structure. The chain RMSD is to first extract the C α atoms shared by the native structure and the predicted structure. For monomer structures, the Kabsch algorithm is used to align the coordinates of the entire peptide chain, and the RMSD of the C α atoms of the entire peptide chain is calculated. For complex structures, the Kabsch algorithm is used to align the coordinates of the protein part, and the RMSD of the C α atoms of the entire peptide chain; the non-standard residue RMSD is to align the heavy atoms on the peptide chain within 10 Å of the non-standard residue, and calculate the all-atom RMSD of the non-standard residue. If there are multiple non-standard residues on a peptide chain, the RMSD of each non-standard residue is calculated and then averaged; DockQ in step eight is used to evaluate the docking quality of the peptide-protein complex. DockQ is a comprehensive score of three indicators, namely F nat , LRMS and iRMS. DockQ divides the structural quality of the complex into four grades. DockQ < 0.23 is unacceptable quality, 0.23 < DockQ < 0.49 is acceptable quality, 0.49 < DockQ < 0.8 is medium quality, and DockQ > 0.8 is high-quality prediction. The calculation formula of DockQ is as follows: ， Among them, F nat Defined as protein-peptide residue pairs with heavy atoms within 5 Å in a crystal complex, LRMS is the local root mean square error, iRMS is the global root mean square error, and d1 and d2 are scaling factors used to adjust the weights of the local RMSD. F nat Used to represent the success rate of grounding; for peptide-protein complexes, F nat A composite material with a density ≥0.5 is considered a high-quality docking. nat The calculation formula is as follows: ， in, This represents the set of amino acid pairs in the real structure where the distance between the ligand sequence and the receptor sequence is less than a threshold θ. This represents the set of amino acid pairs in the predicted structure where the distance between the ligand and receptor sequences is less than a threshold θ. Indicates intersection.

Images