A deep learning algorithm for predicting cyclic peptide and its complex structure

Abstract

The application discloses a kind of deep learning algorithm for predicting cyclic peptide and its complex structure, which is composed of HighFold model and HighFold_Multimer model, specific steps include: obtaining cyclic peptide monomer benchmark dataset;HighFold model is with AlphaFold as framework, by defining the relative position matrix state of the input of model to add the head-to-tail loop and disulfide bridge information of cyclic peptide, realize reasonable cyclic peptide monomer structure prediction;Obtain cyclic peptide complex benchmark dataset and construct external dataset;HighFold_Multimer model is with AlphaFold_Multimer as framework, by defining the relative position matrix state of the input of model to effectively distinguish the input difference between cyclic peptide ligand structure and target protein structure, ensure the learning of the head-to-tail loop and disulfide bridge information of cyclic peptide, while realizing the reasonable structure prediction of target protein, complete the structure prediction of cyclic peptide complex.The algorithm provides effective cyclic peptide structure theoretical support for researchers, promotes cyclic peptide function exploration, cyclic peptide-protein interaction, cyclic peptide design and the development of related research.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided drug development technology, specifically relating to a deep learning algorithm for predicting the structure of cyclic peptides and their complexes. Background Technology

[0002] Cyclic peptides are cyclic peptide compounds composed of amino acids. Due to their unique cyclic structure, these peptides have the following advantages over linear peptides: (1) Hydrogen bonds form within the cyclic peptide molecule, greatly reducing its molecular polarity, resulting in significantly higher metabolic stability and bioavailability compared to linear peptides; (2) Cyclic peptides have a clear structural pre-organization, enabling them to better adhere to target proteins; (3) Cyclic peptides have a large surface area, thus they typically possess high affinity and recognition specificity. In recent years, cyclic peptides have attracted increasing attention due to their broad biological activities, bringing new developments to drug research and development. In the research and development of cyclic peptides, it is necessary to observe the structure of cyclic peptides and their complexes to understand the biological mechanisms of this class of drugs.

[0003] While traditional experimental observations offer high precision, they are time-consuming and labor-intensive. With technological advancements, the prediction of cyclic peptide monomers can be achieved using CADD (Conformance-Based Derivatives) technology. This technique typically combines physics-based energy functions and machine learning algorithms such as Monte Carlo simulations to predict cyclic peptide structures using conformational sampling. However, the accuracy of these algorithms still needs improvement. As for the prediction of cyclic peptide complex structures, current tools still rely on experimentally observed structural information of the target protein. This limits the application of this approach to predicting cyclic peptide complexes involving unverified protein targets, significantly impacting its real-world application. With the emergence of novel deep learning models such as AlphaFold and AlphaFold_Multimer, research on cyclic peptides and their complexes has begun to see new breakthroughs. The implementation of AfCycDesign has confirmed the feasibility of predicting cyclic peptide monomers using the AlphaFold framework. However, when predicting cyclic peptide monomer structures, this model only considers the overall ring structure with loops formed at both ends, ignoring the disulfide bridge structures that may exist within the ring structure. Disulfide bridge structures can significantly influence the structure and activity of cyclic peptides. Considering the current practical application problems in predicting the structure of cyclic peptides and their complexes, there is an urgent need for a high-precision, reasonable computational method for predicting the structure of cyclic peptides and their complexes that does not require experimental structural information of the target site. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a deep learning algorithm for predicting the structure of cyclic peptides and their complexes, which can help researchers quickly understand the structure of cyclic peptides and promote the study of mechanisms such as cyclic peptide-protein interactions.

[0005] The specific technical solution is as follows:

[0006] A deep learning algorithm for predicting the structure of cyclic peptides and their complexes includes the following steps:

[0007] Step 1: Obtain the cyclic peptide monomer benchmark dataset;

[0008] Step 2: The HighFold model uses AlphaFold as a framework and incorporates information on the head-to-tail ring formation and disulfide bridge structure of cyclic peptides by defining the feature state of the relative position matrix of the model's input.

[0009] Step 3: Input the cyclic peptide monomer information from the data in Step 1 into the HighFold model obtained in Step 2 in FASTA text format. The FASTA text information will be combined with MSA (Multiple Sequence Alignment) and template information to be converted into a matrix. The model will infer based on the features of this matrix to achieve reasonable prediction of the cyclic peptide monomer structure. The output of the model is the atomic 3D coordinates of the cyclic peptide monomer, which is saved in PDB (Protein Data Bank) format. There are 5 output results, which are sorted according to the pLDDT (predicted Local-Distance Difference Test) index.

[0010] Step four involves aligning the results obtained in step three with the actual cyclic peptide structures in the data from step one to calculate the Cα-RMSD (Cα-Root Mean Square Deviation) value. This value is used to evaluate the model's performance in the cyclic peptide monomer prediction task. The evaluation of cyclic peptide monomer results requires the actual experimental structure of the cyclic peptide monomer as a reference; this is an optional, not mandatory, step.

[0011] Step 5: Obtain the benchmark dataset of cyclic peptide complexes and construct the external dataset;

[0012] Step 6: The HighFold_Multimer model uses AlphaFold_Multimer as a framework. By defining the state of the input matrix of the model, it effectively distinguishes the differences between the cyclic peptide ligand structure and the target protein structure, ensuring the learning of the cyclic peptide's head-to-tail ring formation and disulfide bridge information, and maintaining the reasonable structure of the target protein.

[0013] Step 7: Input the cyclic peptide complex information from the data in Step 5 into the HighFold_Multimer model obtained in Step 6 in FASTA text format. The FASTA text information will be combined with MSA and template information to be converted into a matrix. The model will perform inference based on the features of this matrix to achieve reasonable prediction of the cyclic peptide complex structure. The output of the model is the 3D coordinates of the atoms of the cyclic peptide complex, which is saved in PDB format. There are 5 output results in total, which are sorted according to the confidence index.

[0014] Step 8: Calculate the F value by combining the results obtained in Step 7 with the actual cyclic peptide complex structure from the data in Step 5. nat The value is used to evaluate the model's performance in the cyclic peptide complex prediction task.

[0015] Furthermore, the PDB_ID of the cyclic peptide monomer benchmark dataset from step one is obtained from the literature, and then PDB data is downloaded in batches from the PDB database based on the PDB_ID.

[0016] Further, step two specifically includes: the relative position matrix of the HighFold model is defined as an N×N cyclic offset matrix, where the sequence distance between the terminal residues of the cyclic peptide of length N is changed to 1, thereby realizing the writing of cyclic structure information. The model then obtains the position information of potentially bonded cysteine ​​residues in the sequence from the input, and uses the Floyd algorithm to calculate the minimum distance between two cysteine ​​residues. The sequence spacing between the corresponding cysteine ​​residues in the matrix is ​​also changed to 1. The Floyd algorithm uses dynamic programming to solve for the minimum distance between any two residues in the cyclic peptide sequence topology. Its basic idea is:

[0017] d[k][i][j]=min(d[k-1][i][j],d[k-1][i][k]+d[k-1][k][j])

[0018] Where d[k][i][j] represents the minimum distance between residue i and residue j that can be reached through residues 1…k.

[0019] Further, step three of the HighFold model's monomer prediction process includes: inputting the cyclic peptide monomer information to be predicted into the HighFold model as FASTA text information. After the FASTA information is entered into the relative offset matrix, it undergoes one-hot encoding and linear projection, and is then input into the paired features in the Evoformer module of the model. The model uses these features to infer and achieve a reasonable prediction of the cyclic peptide monomer structure. The model output is the atomic 3D coordinates of the cyclic peptide monomer, saved in PDB format. There are five output results, sorted according to the pLDDT index. pLDDT measures how close the model's prediction result is to the true LDDT (Local-Distance Difference Test), and is a type of model confidence. Taking a single amino acid residue as an example, its calculation formula is:

[0020] r pLDDT =p pLDDT ·v bins

[0021] Where, p pLDDT =softmax(f(s)) means that the feature s of the residue after training the network is mapped to a value of |v. bins |The feature space is then used to take the softmax value, v bins =[1,3,…,99] T This represents a bucket vector of length 50. Of course, the bucket length can also be any natural number.

[0022] Further, step four specifically includes: using the RMSD calculation formula, aligning the amino acid atomic spatial coordinates of the actual cyclic peptide monomer structure and the predicted monomer structure, and finally calculating the distance between the two structures. The specific RMSD formula is as follows:

[0023]

[0024] Where N represents the number of amino acids in the cyclic peptide. These represent the 3D coordinates of the actual structure and the model-predicted structure, respectively.

[0025] Furthermore, step five specifically includes: obtaining the PDB_ID of the cyclic peptide complex benchmark dataset from the literature, and then downloading PDB data in batches from the PDB database according to the PDB_ID; and downloading the cyclic peptide complex external dataset after manual screening from the PDB database.

[0026] Further, step six specifically includes: The relative position matrix of the HighFold_Multimer model contains information on both the cyclic peptide ligand and the target protein. This algorithm first separates the matrix range containing the cyclic peptide ligand and defines it as an N×N cyclic offset matrix. Simultaneously, the sequence distance between the terminal residues of the cyclic peptide of length N in this ligand matrix is ​​changed to 1, while the matrix range of the target protein remains unchanged, ensuring the model can correctly predict the true target protein structure. The model then obtains the position information of potentially pairwise bonded cysteine ​​residues in the cyclic peptide sequence matrix from the cyclic peptide ligand input, and then uses the Floyd algorithm to calculate the minimum distance between two cysteine ​​residues. Furthermore, the sequence interval between corresponding cysteine ​​residues in the cyclic peptide ligand matrix is ​​changed to 1. The Floyd algorithm uses dynamic programming to solve for the minimum distance between any two residues in the cyclic peptide sequence topology. Its basic idea is:

[0027] d[k][i][j]=min(d[k-1][i][j],d[k-1][i][k]+d[k-1][k][j])

[0028] Where d[k][i][j] represents the minimum distance between residue i and residue j that can be reached by residues 1…k.

[0029] Furthermore, step seven, the AlphaFold_Multimer prediction of the cyclic peptide complex, includes: inputting the cyclic peptide complex as FASTA text information into the modified AlphaFold_Multimer model obtained in step six. The input FASTA text format is as follows:

[0030] Target protein: cyclic peptide

[0031] The FASTA information, combined with MSA and template information, is transformed into a matrix, then one-hot encoded and linearly projected before being incorporated into the pairwise features of the Evoformer module in the model. The model uses these features to infer reasonable predictions of the cyclic peptide complex structure. The model output is the 3D atomic coordinates of the cyclic peptide complex, stored in PDB format. There are five output results, ordered according to confidence scores. The model confidence score is calculated using the following formula:

[0032] S conf = 0.8·ipTM + 0.2·pTM

[0033] Here, pTM represents the predicted TM-score, which is a confidence level of the TM-score of the model prediction result, and its calculation method is similar to pLDDT. ipTM represents the pTM of the surface portion of the complex protein interaction.

[0034] The TM-score formula is:

[0035]

[0036] Among them, L N N represents the number of sequence residues in the native conformation. T d represents the number of residue pairs aligned to the template in the alignment results. i This represents the spatial distance between a pair of aligned residues.

[0037] Furthermore, step eight specifically includes: aligning the spatial coordinates of amino acid atoms between the actual cyclic peptide complex structure and the predicted complex structure, and calculating F... nat F nat The specific formula is:

[0038] S native ={(i,j)|i∈NChainA,j∈NChainB,d ij ≤θ}

[0039] S model ={(i,j)|i∈MChainA,j∈MChainB,d ij ≤θ}

[0040]

[0041]

[0042] Where N is the number of amino acids in the cyclic peptide and the target protein, i is the 3D coordinate of the cyclic peptide structure, j is the 3D coordinate of the target protein structure, and S... native S is the set of amino acid pairs in the native conformation where the distance between the ligand and receptor sequences is less than a threshold θ. model F represents the set of amino acid pairs in the model-predicted conformation where the distance between the ligand and receptor sequences is less than a threshold θ. ∩ represents the intersection of these sets. |.| represents the size of the set, i.e., the number of elements in the set. nat This indicates the recall of the model's predicted conformation at the binding site relative to the native conformation. Generally, F... nat A value >0.2 indicates that the predicted complex result is acceptable. nat A value >0.5 indicates that the predicted complex result is generally poor. nat A value >0.8 indicates that the predicted complex result is excellent.

[0043] The beneficial effects of this invention are as follows: Firstly, by using the input relative position matrix of the HighFold model, the input of the cyclic peptide structure and the learning of potential disulfide bridge information within the ring are completed, overcoming the computational drawbacks of the original technology and achieving high-precision cyclic peptide structure prediction. Secondly, by using the input relative position matrix of the HighFold_Multimer model, while ensuring the input of the ligand cyclic peptide structure and the learning of potential disulfide bridge information within the ring, normal prediction of the target protein can be achieved. This solves the problem that experimental structural information of the target protein is required for cyclic peptide complex structure prediction in reality, achieving high-precision cyclic peptide complex structure prediction without experimental structural information of the target protein. The algorithm for predicting the structure of cyclic peptides and their complexes designed in this invention will greatly promote the development of cyclic peptide function research, cyclic peptide-protein interactions, and other fields, further accelerating the research speed of cyclic peptide drugs and reducing the development cost of such drugs. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the prediction of cyclic peptide monomers and complexes in a specific implementation of the present invention. Detailed Implementation

[0045] 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.

[0046] Example 1

[0047] like Figure 1 As shown, a deep learning algorithm for predicting the structure of cyclic peptides and their complexes includes the following steps:

[0048] Step 1: Obtain the cyclic peptide monomer benchmark dataset. The PDB_ID of the cyclic peptide monomer benchmark dataset is obtained from the literature. Then, download PDB data in batches from the PDB database according to the PDB_ID.

[0049] Step two: The HighFold model, based on the AlphaFold framework, incorporates information about the beginning and end ring formation and disulfide bridge structure of the cyclic peptide by defining the feature state of the relative position matrix of the input model. The relative position matrix of the HighFold model is defined as an N×N cyclic offset matrix, where the sequence distance between the terminal residues of the cyclic peptide of length N is changed to 1, thereby realizing the writing of cyclic structure information. The model then obtains the position information of potentially pairwise bonded cysteine ​​residues in the sequence from the input and uses the Floyd algorithm to calculate the minimum distance between two cysteine ​​residues. The sequence spacing between the corresponding cysteine ​​residues in the matrix is ​​also changed to 1. The Floyd algorithm uses dynamic programming to solve for the minimum distance between any two residues in the cyclic peptide sequence topology. Its basic idea is:

[0050] d[k][i][j]=min(d[k-1][i][j],d[k-1][i][k]+d[k-1][k][j])

[0051] Where d[k][i][j] represents the minimum distance between residue i and residue j that can be reached through residues 1…k.

[0052] Step 3: Input the cyclic peptide monomer information from Step 1 into the HighFold model obtained in Step 2 in FASTA text format. The FASTA text information is combined with MSA and template information to form a matrix. The model uses the features of this matrix to infer and achieve reasonable cyclic peptide monomer structure prediction. The model output is the atomic 3D coordinates of the cyclic peptide monomer, saved in PDB format. There are 5 output results, sorted according to the pLDDT index. Input the cyclic peptide monomer information to be predicted into the HighFold model in FASTA text format. After the FASTA information enters the relative offset matrix, it undergoes one-hot encoding and linear projection, and then enters the paired features in the Evoformer module of the model. The model uses these features to infer and achieve reasonable cyclic peptide monomer structure prediction. The model output is the atomic 3D coordinates of the cyclic peptide monomer, saved in PDB format. There are 5 output results, sorted according to the pLDDT index. pLDDT is used to measure how close the model prediction result is to the true LDDT, and is a type of model confidence. Taking an amino acid residue as an example, its calculation formula is:

[0053] r pLDDT =p pLDDT ·v bins

[0054] Where, p pLDDT =softmax(f(s)) means that the feature s of the residue after training the network is mapped to a value of |v. bins|The feature space is then used to take the softmax value, v bins = [1, 3, ..., 99] T This represents a bucket vector of length 50. Of course, the bucket length can also be any natural number.

[0055] Step four involves aligning the results obtained in step three with the actual cyclic peptide structures in the data from step one to calculate the Cα-RMSD value. This value is used to evaluate the model's performance in the cyclic peptide monomer prediction task. The RMSD calculation formula is used to align the amino acid atomic spatial coordinates of the actual and predicted cyclic peptide monomer structures and finally calculate the distance between these two structures. The specific formula for RMSD is as follows:

[0056]

[0057] Where N represents the number of amino acids in the cyclic peptide. The values ​​represent the 3D coordinates of the actual structure and the model-predicted structure, respectively. The results are shown in Table 1.

[0058] Table 1. Cα-RMSD results of the HighFold model on cyclic peptide monomer baseline data.

[0059]

[0060]

[0061]

[0062] It should be noted that the cyclic peptide monomer FASTA mainly comes from the cyclic peptide monomer benchmark dataset, but the cyclic peptide monomer FASTA used in this invention is not limited to the above data sources. Evaluation of cyclic peptide monomer results requires the actual experimental structure of the cyclic peptide monomer as a reference; this is an optional but not mandatory step.

[0063] Step 5: Obtain the cyclic peptide complex benchmark dataset and construct the external dataset. The PDB_ID of the cyclic peptide complex benchmark dataset is obtained from the literature. Then, PDB data is downloaded in batches from the PDB database according to the PDB_ID. The cyclic peptide complex external dataset is downloaded after manual screening from the PDB database.

[0064] Step six: The HighFold_Multimer model, based on the AlphaFold_Multimer framework, effectively distinguishes the differences between the cyclic peptide ligand structure and the target protein structure by defining the input matrix state of the model. This ensures the learning of the beginning and end ring formation and disulfide bridge information of the cyclic peptide, while maintaining the reasonable structure of the target protein. The relative position matrix of the HighFold_Multimer model contains information on both the cyclic peptide ligand and the target protein. This algorithm first divides the matrix range where the cyclic peptide ligand is located and defines it as an N×N cyclic offset matrix. Simultaneously, this ligand matrix... The sequence distance between the terminal residues of a cyclic peptide of length N in the matrix is ​​changed to 1, while the matrix content of the target protein remains unchanged, ensuring that the model can correctly predict the actual target protein structure. The model then obtains the position information of potentially pairwise bonded cysteine ​​residues in the cyclic peptide sequence matrix from the cyclic peptide ligand input, and then uses the Floyd algorithm to calculate the minimum distance between two cysteine ​​residues. In addition, the sequence spacing between corresponding cysteine ​​residues in the cyclic peptide ligand matrix is ​​changed to 1. The Floyd algorithm uses dynamic programming to solve for the minimum distance between any two residues in the cyclic peptide sequence topology graph. Its basic idea is:

[0065] d[k][i][j]=min(d[k-1][i][j], d[k-1][i][k]+d[k-1][k][j])

[0066] Where d[k][i][j] represents the minimum distance between residue i and residue j that can be reached through residues 1...k;

[0067] Step 7: Input the cyclic peptide complex information from the data in Step 5 into the HighFold_Multimer model obtained in Step 6 in FASTA text format. The input FASTA text format is as follows:

[0068] Target protein: cyclic peptide

[0069] The FASTA information, combined with MSA and template information, is transformed into a matrix, then one-hot encoded and linearly projected before being incorporated into the pairwise features of the Evoformer module in the model. The model uses these features to infer reasonable predictions of the cyclic peptide complex structure. The model output is the 3D atomic coordinates of the cyclic peptide complex, stored in PDB format. There are five output results, ordered according to confidence scores. The model confidence score is calculated using the following formula:

[0070] S conf = 0.8·ipTM + 0.2·pTM

[0071] Here, pTM represents the predicted TM-score, which is a confidence level of the TM-score of the model prediction result, and its calculation method is similar to pLDDT. ipTM represents the pTM of the surface portion of the complex protein interaction.

[0072] The TM-score formula is:

[0073]

[0074] Among them, L N N represents the number of sequence residues in the native conformation. T d represents the number of residue pairs aligned to the template in the alignment results. i This represents the spatial distance between a pair of aligned residues.

[0075] Step 8: Calculate the F value by combining the results obtained in Step 7 with the actual cyclic peptide complex structure from the data in Step 5. nat The F-value is used to evaluate the model's performance in the cyclic peptide complex prediction task. It aligns the spatial coordinates of amino acid atoms between the actual and predicted cyclic peptide complex structures and calculates the F-value. nat F nat The specific formula is:

[0076] S native ={(i,j)|i∈NChainA,j∈NChainB,d ii ≤θ}

[0077] S model ={(i,j)|i∈MChainA,j∈MChainB,d ij ≤θ}

[0078]

[0079]

[0080] Where N is the number of amino acids in the cyclic peptide and the target protein, i is the 3D coordinate of the cyclic peptide structure, j is the 3D coordinate of the target protein structure, and S... native S: The set of amino acid pairs in the native conformation where the distance between the ligand and receptor sequences is less than a threshold θ. model F: The set of amino acid pairs in the model-predicted conformation where the distance between the ligand and receptor sequences is less than a threshold θ. ∩ represents the intersection of the sets. |.|: The size of the set, i.e., the number of elements in the set. nat This indicates the recall of the model's predicted conformation at the binding site relative to the native conformation. Generally, F... nat A value >0.2 indicates that the predicted complex result is acceptable. natA value >0.5 indicates that the predicted complex result is generally poor. nat A value >0.8 indicates excellent results for the predicted complex. The results are shown in Tables 2 and 3.

[0081] Table 2. F-values ​​of the HighFold_Multimer model on baseline data of cyclic peptide complexes. nat result

[0082]

[0083]

[0084] Table 3. F-values ​​of the HighFold_Multimer model on external data of cyclic peptide complexes. nat result

[0085]

[0086] It should be noted that the cyclic peptide complex FASTA datasets primarily originate from cyclic peptide complex benchmark datasets and external datasets; however, the cyclic peptide complex FASTA datasets used in this invention are not limited to these sources. Evaluation of cyclic peptide complex results requires actual experimental structures of the cyclic peptide complex as a reference; this is an optional, not mandatory, step.

[0087] 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 structure of cyclic peptides and their complexes, characterized in that... Includes the following steps: Step 1: Obtain the cyclic peptide monomer benchmark dataset; Step 2: The HighFold model uses AlphaFold as a framework and incorporates information on the head-to-tail ring formation and disulfide bridge structure of cyclic peptides by defining the feature state of the relative position matrix of the model's input. Step 3: Input the cyclic peptide monomer information from the data in Step 1 into the HighFold model obtained in Step 2 in FASTA text format. The FASTA text information will be combined with MSA and template information to be converted into a matrix. The model will infer based on the features of this matrix to achieve reasonable prediction of the cyclic peptide monomer structure. The output of the model is the atomic 3D coordinates of the cyclic peptide monomer, which is saved in PDB format. There are 5 output results in total, which are sorted according to the pLDDT index. Step four: Align the results obtained in step three with the actual cyclic peptide structures in the data from step one to calculate the Cα-RMSD value, which is used to evaluate the model's performance in the cyclic peptide monomer prediction task. Step 5: Obtain the benchmark dataset of cyclic peptide complexes and construct the external dataset; Step 6: The HighFold_Multimer model uses AlphaFold_Multimer as a framework. By defining the state of the input matrix of the model, it effectively distinguishes the differences between the cyclic peptide ligand structure and the target protein structure, ensuring the learning of the cyclic peptide's head-to-tail ring formation and disulfide bridge information, and maintaining the reasonable structure of the target protein. Step 7: Input the cyclic peptide complex information from the data in Step 5 into the HighFold_Multimer model obtained in Step 6 in FASTA text format. The FASTA text information will be combined with MSA and template information to be converted into a matrix. The model will perform inference based on the features of this matrix to achieve reasonable prediction of the cyclic peptide complex structure. The output of the model is the 3D coordinates of the atoms of the cyclic peptide complex, which is saved in PDB format. There are 5 output results in total, which are sorted according to the confidence index. Step 8: Calculate the F value by combining the results obtained in Step 7 with the actual cyclic peptide complex structure from the data in Step 5. nat The value is used to evaluate the model's performance in the cyclic peptide complex prediction task.

2. The deep learning algorithm for predicting the structure of cyclic peptides and their complexes as described in claim 1, characterized in that... In step one, the PDB_ID of the cyclic peptide monomer benchmark dataset was obtained from the literature, and then PDB data was downloaded in batches from the PDB database based on the PDB_ID.

3. The deep learning algorithm for predicting the structure of cyclic peptides and their complexes as described in claim 1, characterized in that... Step two specifically includes: The relative position matrix of the HighFold model is defined as an N×N cyclic offset matrix, where the sequence distance between the terminal residues of the cyclic peptide of length N is changed to 1, thereby realizing the writing of cyclic structure information. The model then obtains the position information of potentially bonded cysteine ​​residues in the sequence from the input, and uses the Floyd algorithm to calculate the minimum distance between two cysteine ​​residues. The sequence spacing between the corresponding cysteine ​​residues in the matrix is ​​also changed to 1. The Floyd algorithm uses dynamic programming to solve for the minimum distance between any two residues in the cyclic peptide sequence topology. Its basic idea is: d[k][i][j]=min(d[k-1][i][j],d[k-1][i][k]+d[k-1][k][j]) Where d[k][i][j] represents the minimum distance between residue i and residue j that can be reached through residues 1…k.

4. The deep learning algorithm for predicting the structure of cyclic peptides and their complexes as described in claim 3, characterized in that... Step four specifically includes: using the RMSD calculation formula, aligning the amino acid atomic spatial coordinates of the actual cyclic peptide monomer structure with the predicted monomer structure, and finally calculating the distance between the two structures. The specific RMSD formula is as follows: Where N represents the number of amino acids in the cyclic peptide. These represent the 3D coordinates of the actual structure and the model-predicted structure, respectively.

5. The deep learning algorithm for predicting the structure of cyclic peptides and their complexes as described in claim 4, characterized in that... Step five specifically includes: obtaining the PDB_ID of the cyclic peptide complex benchmark dataset from the literature, then downloading PDB data in batches from the PDB database according to the PDB_ID, and downloading the cyclic peptide complex external dataset after manual screening from the PDB database.

6. The deep learning algorithm for predicting the structure of cyclic peptides and their complexes as described in claim 5, characterized in that... Step six specifically includes: The relative position matrix of the HighFold_Multimer model contains information on both the cyclic peptide ligand and the target protein. This algorithm first separates the matrix range containing the cyclic peptide ligand and defines it as an N×N cyclic offset matrix. Simultaneously, the sequence distance between the terminal residues of the cyclic peptide of length N in this ligand matrix is ​​changed to 1, while the matrix range of the target protein remains unchanged, ensuring the model can correctly predict the true target protein structure. The model then obtains the position information of potentially pairwise bonded cysteine ​​residues in the cyclic peptide sequence matrix from the cyclic peptide ligand input, and then uses the Floyd algorithm to calculate the minimum distance between two cysteine ​​residues. Furthermore, the sequence spacing between corresponding cysteine ​​residues in the cyclic peptide ligand matrix is ​​changed to 1. The Floyd algorithm uses dynamic programming to solve for the minimum distance between any two residues in the cyclic peptide sequence topology. Its basic idea is: d[k][i][j]=min(d[k-1][i][j],d[k-1][i][k]+d[k-1][k][j]) Where d[k][i][j] represents the minimum distance between residue i and residue j that can be reached by residues 1…k.

7. The deep learning algorithm for predicting the structure of cyclic peptides and their complexes as described in claim 6, characterized in that... Step eight specifically includes: aligning the spatial coordinates of amino acid atoms between the actual cyclic peptide complex structure and the predicted complex structure, and calculating F. nat F nat The specific formula is: S native ={(i,j)|i∈NChainA,j∈NChainB,d ij ≤θ} S model ={(i,j)|i∈MChainA,j∈MChainB,d ij ≤θ} Where N is the number of amino acids in the cyclic peptide and the target protein, i is the 3D coordinate of the cyclic peptide structure, j is the 3D coordinate of the target protein structure, and S... native S is the set of amino acid pairs in the native conformation where the distance between the ligand and receptor sequences is less than a threshold θ. model F represents the set of amino acid pairs in the model-predicted conformation where the distance between the ligand and receptor sequences is less than a threshold θ. ∩ represents the intersection of these sets. |.| represents the size of the set, i.e., the number of elements in the set. nat This indicates the recall of the model's predicted conformation at the binding site relative to the native conformation. Generally, F... nat A value >0.2 indicates that the predicted complex result is acceptable. nat A value >0.5 indicates that the predicted complex result is generally poor. nat A value >0.8 indicates that the predicted complex result is excellent.

Images