Deep learning algorithm for predicting the three-dimensional structures of non-natural cyclic peptide monomers and their complexes based on AlphaFold

Abstract

The present invention discloses a deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes based on AlphaFold. The specific steps include: obtaining a non-natural linear peptide dataset; dividing the dataset into training, validation, and test sets; adding a neural network to AlphaFold for capturing the atomic-scale representation of peptides; pooling the obtained atomic-scale representation and combining it with AlphaFold's original residue-scale representation; predicting the structure of non-natural amino acids by defining rigid groups of non-natural amino acids and initializing their atomic coordinates; training the modified AlphaFold using the non-natural linear peptide training and validation sets; testing the trained model on a non-natural linear peptide test set; obtaining a non-natural cyclic peptide test dataset; modifying the relative position encoding matrix in the model; and testing the modified model on a non-natural cyclic peptide test set. Compared with previous structure prediction models, the present invention can accurately predict the structure of non-natural cyclic peptide monomers and their complexes.

Description

Technical Field

[0001] The present invention relates to the technical field of cyclic peptide drug research and development, and specifically to a deep learning algorithm based on AlphaFold to predict the three-dimensional structure of non-natural cyclic peptide monomers and their complexes. Background Art

[0002] Peptides are compounds composed of amino acids linked by peptide bonds. Their molecular weight lies between small and large molecules, and they can selectively regulate a variety of protein-protein interactions. In drug development, peptide-based therapies have demonstrated unique advantages: compared with small molecule drugs, they have higher specificity and lower toxicity; compared with protein drugs, they exhibit lower immunogenicity and lower production costs. These characteristics make peptides attractive drug candidates for treating various diseases. However, linear peptides are susceptible to proteolysis. By cyclizing linear peptides and introducing non-natural amino acids, not only can their binding affinity and specificity be significantly improved, but also their stability and membrane permeability can be enhanced. Therefore, the development of cyclic peptides containing non-natural amino acids has become a hot topic in peptide drug research.

[0003] Predicting the three-dimensional structure of unnatural cyclic peptide monomers and their complexes with protein receptors is crucial for structure-based drug design. As one of the core approaches in computer-aided drug design, structure-based drug design has been widely applied in areas such as lead compound discovery and optimization, ADMET property prediction, and drug repurposing, becoming an effective tool for developing novel drug candidates. The rapid development of artificial intelligence algorithms has significantly improved structure prediction models. Tools such as AlphaFold2, RoseTTAFold, and ESMFold have achieved accurate structure prediction for protein monomers and protein-protein complexes. Building on these models, subsequent research has extended structure prediction to protein-nucleic acid, protein-peptide, protein-small molecule, and protein-cyclic peptide complexes. Recently, models such as RoseTTAFold All-Atom and AlphaFold3 have further expanded the scope of deep learning-based spatial structure prediction to complex complexes involving proteins, nucleic acids, small molecules, ions, and modified residues. However, none of these approaches have addressed the problem of structure prediction for unnatural cyclic peptide monomers and their complexes. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention successfully achieved accurate spatial structure prediction of non-natural cyclic peptide monomers and their complexes based on the AlphaFold model, atomic feature extraction and non-natural amino acid coding information expansion. Taking into account that polypeptides have certain small molecule characteristics, the present invention integrates the neural network module that characterizes the atomic features of polypeptides into the AlphaFold model, realizing multi-scale modeling of polypeptide molecules. In addition, the present invention has also successfully expanded the coding information of non-natural amino acids, so that the model has the ability to predict non-natural amino acids. Although the model can be trained directly on cyclic peptide structures with non-natural amino acids, such structures are relatively scarce in the RCSB ProteinDataBank database. Therefore, the present invention first uses linear peptide data containing non-natural amino acids to fine-tune the model, and then modifies the relative position coding matrix in the model so that it can accurately predict the cyclic peptide structure. Experimental results show that the model can accurately predict the three-dimensional structure of non-natural cyclic peptide monomers and their complexes.

[0005] The specific technical solutions are as follows:

[0006] A deep learning algorithm based on AlphaFold for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes includes the following steps:

[0007] Step 1: Obtain the spatial structure file containing non-natural linear peptides and perform preprocessing. The linear peptide data containing non-natural amino acids comes from the ModPep dataset, which contains a total of 501 linear peptide samples containing non-natural amino acid residues, most of which are complex structures formed by non-natural linear peptides and proteins. Using the PDB ID provided, the corresponding crystal structure was downloaded from the RCSBProteinDataBank database, duplicates were deleted, and finally 419 unique three-dimensional structures were obtained. The solvent, hydrogen atoms and heteroatoms in the crystal structure were then deleted, and only the first state of each structure was retained. In order to reduce the redundancy of the crystal structure and reduce the computational requirements during the training process, certain chains in the spatial structure were deleted according to the following rules: chains with less than 50 amino acid residues in the three-dimensional structure are classified as polypeptide chains, while other chains are classified as protein chains. If a structure contains at least one protein chain, the sequences of all polypeptide chains are aligned, and for those with the same or partial overlap and a distance greater than The peptide sequence was then sorted and only the longest peptide chain was retained, while the remaining peptide chains were deleted. In addition, if there were at least two protein chains in the structure at this time, only the protein chain with the most contact with the peptide chain was retained. Then, the types of unnatural amino acids in the crystal structure were analyzed, and any structure containing unnatural amino acids that were only found in this crystal structure was also deleted. After this pre-processing process, a total of 382 unnatural linear peptide samples were obtained, containing 23 different unnatural amino acid residues;

[0008] Step 2: Randomly split the dataset obtained in Step 1 into training, validation, and test sets in a ratio of 7.5:1.5:1. Ensure that each unnatural amino acid appears at least once in both the validation and test sets. Use ColabFold to generate the features required for training and validation.

[0009] Step three: In the original AlphaFold model, an additional neural network module is added to the polypeptide molecule to capture its atomic scale characterization. The atomic scale features come from the atomic elements and the bond information between them, which is extracted using RDKit after converting the FASTA of the polypeptide molecule into SMILES. The atomic element features are encoded using one-hot encoding. In order to reduce the sparsity of the encoding matrix, the present invention only encodes common atoms (B, C, F, I, N, O, P, S, Br, Cl), and rare atoms share the encoding. The generated one-hot matrix is ​​input into the multi-head attention and multi-layer perceptron. The bond information between atoms is also encoded using the one-hot method, which includes five bond types (single bond, double bond, triple bond, aromatic bond and ionic bond). The obtained encoding matrix is ​​transformed by a linear layer;

[0010] Step 4: Pool the atomic-scale representation obtained in step 3 to facilitate its combination with the original residue-scale representation of the AlphaFold model, so as to achieve multi-scale modeling of polypeptide molecules. Both the atomic element representation and the representation of the bond information between atoms use an attention-based pooling mechanism to map them into embeddings at the amino acid residue scale. The pooled atomic element embeddings will be added to the unique-hot matrix encoding the polypeptide type in the AlphaFold model, and then input into the Evoformer module to generate a single representation. The pooled bond information embeddings will be added to the paired activation matrix in the AlphaFold model to generate paired representations in the Evoformer module. The following is the calculation formula for the attention-based pooling mechanism:

[0011]

[0012] A score =Softmax(x atom +F),

[0013]

[0014] Among them, M is a mask matrix used to indicate which amino acid residue the atom belongs to, ∈ is a minimum positive value to prevent division by zero error, and x atom represents the atomic scale representation, x residueThis represents the pooled amino acid residue scale representation. Step 5: Predict the spatial structure of non-natural amino acids by defining rigidity groups and initializing the atomic coordinates of the non-natural amino acids. The Structure module in the AlphaFold model uses the single and paired representations obtained from the Evoformer module to predict the torsion angles of each amino acid residue in the protein. These predicted torsion angles, together with the predefined rigidity groups and initialized atomic coordinates, construct the coordinates of all atoms in the protein's spatial structure. Based on the dependence of each atom in the non-natural amino acid residue on a specific torsion angle, they are divided into five rigidity groups: backbone rigidity group, ω rigidity group, Φ rigidity group, ψ rigidity group, and χ rigidity group. The backbone rigidity group includes the four atoms C, Cα, Cβ, and N. The hydrogen atoms attached to Cα are classified as the ω rigidity group. The Φ rigidity group includes only the hydrogen atoms of the amino group, the ψ rigidity group includes only the oxygen atoms of the carboxyl group, and the χ rigidity group includes all side chain atoms. Since hydrogen atoms are not predicted in this invention, the ω rigidity group and the Φ rigidity group can be ignored. After determining the rigidity groups, the initial coordinates of each atom are determined based on the crystal structure of the non-natural amino acid. For the backbone rigidity group, a coordinate system is established with Cα as the origin, C placed on the positive x-axis, and N placed in the xy plane. The initial coordinates of the Cα, Cβ, C, and N atoms are calculated. In the ψ rigidity group, a coordinate system is established with C as the origin, Cα on the negative x-axis, and the N of the next amino acid residue in the xy plane to determine the initial coordinates of O. The χ rigidity group contains all atoms that depend on the torsion angle of the side chain. There are four groups in total. Based on the true coordinates of the four atoms in each group, the third atom is set as the origin, the second atom is placed on the negative x-axis, and the first atom is placed in the xy plane. This establishes a coordinate system to determine the relative coordinates of the fourth atom in this coordinate system. The fourth atom is then rotated around the x-axis to the xy plane to determine the initial coordinates of the fourth atom. The calculation formula is as follows:

[0015]

[0016] Coordinate init =R x Coordinate relative ,

[0017] Among them, Coordinate relative is the relative coordinate of the fourth atom, and θ is the angle between it and the xy plane, R x is the rotation matrix, Coordinate init represents the initialization coordinates of the fourth atom;

[0018] Step 6: Use the training set and validation set obtained in step 2 to train the above-mentioned non-natural line peptide monomer and its complex structure prediction model. During training, the amino acid residues in each training sample are trimmed to 220, while the validation set remains uncropped. For samples with breakpoints in the training set, the loss value is halved. In order to prevent overfitting, an early stopping strategy is implemented based on the loss of the validation set. The five sets of parameters of the AlphaFold model are fine-tuned using the non-natural line peptide structure;

[0019] Step 7: Use the five trained models to test on the test set obtained in step 2;

[0020] Step 8: Obtain a test data set of non-natural cyclic peptide monomers and their complexes. Part of the cyclic peptide data containing non-natural amino acids comes from cPEPmatch. In order to effectively evaluate the accuracy of the model on the cyclic peptide data, we filtered the data set, excluded samples containing uncoded non-natural amino acids, and limited the cyclization form to head-to-tail cyclization and disulfide bond cyclization. In addition, a certain search was conducted on the RCSB ProteinDataBank database to determine the cyclic peptide structures that meet the standards. After downloading the corresponding non-natural cyclic peptide crystal structures from the RCSB ProteinDataBank database, the solvent molecules, hydrogen atoms and heteroatoms were removed, and only the first state of each structure was retained, and samples containing nucleic acids and repetitions were deleted. Finally, a total of 34 cyclic peptide test samples containing non-natural amino acids were obtained;

[0021] Step 9: Modify the relative position encoding matrix in the structure prediction model of the non-natural linear peptide monomer and its complex so that the relative positions of the two loop residues in the model can be accurately represented and adjacent to each other;

[0022] Step 10: Use the model obtained in step 9 to test on the non-natural cyclic peptide test set obtained in step 8. As a technical solution of the present invention, in the process of constructing the multiple sequence alignment feature, all non-natural amino acids are regarded as unknown amino acids. By integrating the atomic scale embedding information of the polypeptide into the AlphaFold model, the model can distinguish various non-natural amino acid residues and realize multi-scale modeling of polypeptide molecules. As a technical solution of the present invention, after defining the corresponding rigid groups and initialized atomic coordinates for non-natural amino acids, the modeling of each protein atom can be completed by predicting the seven torsion angles of each amino acid, which greatly reduces the complexity of the prediction and enables the model to predict their atomic coordinates in an end-to-end manner. As a technical solution of the present invention, five sets of AlphaFold model parameter sets are fine-tuned using non-natural linear peptide data, and by modifying the relative position encoding matrix in the model, it has the ability to predict the structure of non-natural cyclic peptides.

[0023] The present invention provides a deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes based on the AlphaFold model, atomic feature extraction, and non-natural amino acid encoding information expansion, which has the following beneficial effects:

[0024] (1) The present invention proposes a multi-scale cyclic peptide molecular characterization method, which has better feature extraction capabilities. Compared with previous methods, the present invention can better characterize cyclic peptide molecules, and this method is also applicable to the atomic characterization of linear peptide molecules.

[0025] (2) The present invention proposes a method for predicting the structure of non-natural amino acids. By defining the corresponding rigid groups and initializing the atomic coordinates of non-natural amino acids, the prediction range of the structure prediction model is successfully extended to non-natural linear peptides and non-natural cyclic peptide structures.

[0026] (3) The present invention designs a zero-shot learning method for predicting the structure of non-natural cyclic peptides. The model is first trained on non-natural linear peptide data, and then the relative position encoding matrix is ​​modified to make it suitable for the prediction of non-natural cyclic peptide structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flowchart of the non-natural peptide three-dimensional structure prediction algorithm based on AlphaFold;

[0028] Figure 2 Diagram of the training and inference framework of the non-natural cyclic peptide three-dimensional structure prediction model. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present invention are described clearly and completely below. The embodiments described are only a part of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work are within the scope of protection of the present invention.

[0030] like Figure 1 and Figure 2 As shown, this embodiment provides a deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes. The method includes the following steps:

[0031] Step 1: Obtain the spatial structure files containing non-natural linear peptides from the RCSB ProteinDataBank database and preprocess them. The linear peptide data containing non-natural amino acids comes from the ModPep dataset, which contains a total of 501 linear peptide samples containing non-natural amino acid residues, most of which are complex structures formed by non-natural linear peptides and proteins. Using the PDB ID provided, download the corresponding crystal structure from the RCSB ProteinDataBank database, delete duplicates, and finally obtain 419 unique three-dimensional structures. Then delete the solvent, hydrogen atoms and heteroatoms in the crystal structure, and only retain the first state of each structure. In order to reduce the redundancy of the crystal structure and reduce the computational requirements during the training process, certain chains in the spatial structure are deleted according to the following rules: chains with less than 50 amino acid residues in the three-dimensional structure are classified as polypeptide chains, while other chains are classified as protein chains. If a structure contains at least one protein chain, the sequences of all polypeptide chains are aligned, and for those with the same or partial overlap and a distance greater than The peptide sequence is then sorted, and only the longest peptide chain is retained, while the remaining peptide chains are deleted. In addition, if there are at least two protein chains in the structure at this time, only the protein chain with the most contact with the peptide chain is retained. Then, the types of non-natural amino acids in the crystal structure are analyzed, and any structure containing non-natural amino acids that are only found in this crystal structure will also be deleted. After this pre-processing process, a total of 382 non-natural linear peptide samples were obtained, which contained 23 different non-natural amino acid residues.

[0032] Step 2: Randomly divide the dataset obtained in step 1 into training, validation, and test sets in a ratio of 7.5:1.5:1, ensuring that each unnatural amino acid appears at least once in both the validation and test sets. Use ColabFold to generate the various features required for training and validation.

[0033] Step three: In the original AlphaFold model, an additional neural network is added to the polypeptide molecule to capture its atomic-scale characterization. The atomic-scale features come from the atomic elements and the bond information between them. This information is extracted using RDKit after converting the FASTA of the polypeptide molecule into SMILES. The atomic element features are encoded using one-hot encoding. In order to reduce the sparsity of the encoding matrix, the present invention only encodes common atoms (B, C, F, I, N, O, P, S, Br, Cl). Rare atoms share the encoding, and the generated one-hot matrix is ​​input into the multi-head attention and multi-layer perceptron. The bond information between atoms is also encoded using the one-hot method, which includes a total of five bond types (single bond, double bond, triple bond, aromatic bond and ionic bond). The obtained encoding matrix is ​​transformed through a linear layer;

[0034] Step 4: Pool the atomic-scale representation obtained in step 3 and combine it with the original residue-scale representation of the AlphaFold model to achieve multi-scale modeling of polypeptide molecules. Both the atomic element representation and the representation of the bond information between atoms use an attention-based pooling mechanism to map them to amino acid residue-scale embeddings. The pooled atomic element embeddings will be added to the unique-hot matrix encoding the polypeptide type in the AlphaFold model, and then input into the Evoformer module to generate a single representation. The pooled bond information embeddings will be added to the paired activation matrix in the AlphaFold model to generate paired representations in the Evoformer module. The following is the calculation formula for the attention-based pooling mechanism:

[0035]

[0036] A score =Softmax(x atom +F),

[0037]

[0038] Among them, M is a mask matrix used to indicate which amino acid residue the atom belongs to, ∈ is a minimum positive value to prevent division by zero error, and x atom represents the atomic scale representation, x residue=This represents the pooled amino acid residue scale representation. Step 5: Predicting the structure of non-natural peptides is achieved by defining rigidity groups for non-natural amino acids and initializing their atomic coordinates. The Structure module in the AlphaFold model uses the single and paired representations obtained from the Evoformer module to predict the torsion angles of each amino acid residue in the protein. These predicted torsion angles, combined with the predefined rigidity groups and initialized atomic coordinates, construct the coordinates of all atoms in the protein's spatial structure. Based on the dependence of each atom in the non-natural amino acid residue on a specific torsion angle, they are divided into five rigidity groups: backbone rigidity group, ω rigidity group, Φ rigidity group, ψ rigidity group, and χ rigidity group. The backbone rigidity group includes the four atoms C, Cα, Cβ, and N. The hydrogen atoms attached to Cα are classified as the ω rigidity group. The Φ rigidity group includes only the hydrogen atoms of the amino group, the ψ rigidity group includes only the oxygen atoms of the carboxyl group, and the χ rigidity group includes all side chain atoms. Since hydrogen atoms are not predicted in this invention, the ω rigidity group and the Φ rigidity group can be ignored. After determining the rigidity groups, the initial coordinates of each atom are determined based on the crystal structure of the non-natural amino acid. For the backbone rigidity group, a coordinate system is established with Cα as the origin, C placed on the positive x-axis, and N placed in the xy plane. The initial coordinates of the Cα, Cβ, C, and N atoms are calculated. In the ψ rigidity group, a coordinate system is established with C as the origin, Cα on the negative x-axis, and the N of the next amino acid residue in the xy plane to determine the initial coordinates of O. The χ rigidity group contains all atoms that depend on the torsion angle of the side chain. There are four groups in total. Based on the true coordinates of the four atoms in each group, the third atom is set as the origin, the second atom is placed on the negative x-axis, and the first atom is placed in the xy plane. This establishes a coordinate system to determine the relative coordinates of the fourth atom in this coordinate system. The fourth atom is then rotated around the x-axis to the xy plane to determine the initial coordinates of the fourth atom. The calculation formula is as follows:

[0039]

[0040] Coordinate init =R x Coordinate relative ,

[0041] Among them, Coordinate relative is the relative coordinate of the fourth atom, and θ is the angle between it and the xy plane, R x is the rotation matrix, Coordinate init represents the initialization coordinates of the fourth atom;

[0042] Step 6: Use the obtained non-natural peptide training set and validation set to train the non-natural peptide monomer and complex structure prediction model. During training, the amino acid residues in each training sample are trimmed to 220, while the validation set remains uncropped. For samples with breakpoints in the training set, the loss value is halved. To prevent overfitting, an early stopping strategy is implemented based on the loss of the validation set. The five sets of parameters of the AlphaFold model are fine-tuned using the non-natural peptide structure;

[0043] Step 7: Use the trained model to test on the non-natural peptide test set;

[0044] Step 8: Obtain the test data set of non-natural cyclic peptide monomers and their complexes from the RCSB ProteinDataBank database, and obtain a total of 34 cyclic peptide test samples containing non-natural amino acids;

[0045] Step 9: Modify the relative position encoding matrix algorithm in the structure prediction model of the non-natural linear peptide monomer and its complex, so that the relative positions of the two loop residues in the model can be accurately represented and adjacent to each other;

[0046] Step 10: Use the modified structure prediction model to test on the non-natural cyclic peptide test set.

[0047] The present invention evaluates the predictive performance of the model by calculating the root mean square deviation (RMSD) between the predicted structure and the true structure. Specifically, the RMSD of the pocket-aligned peptide is all atom , pocket-aligned peptide RMSD Cα RMSD of peptide aligned with pocket unaa Used to evaluate the structural accuracy of non-natural peptide complexes. RMSD all atom , RMSD Cα and RMSD unaa It is used to evaluate the structural accuracy of non-natural polypeptide monomers and is specifically defined as follows:

[0048] For each crystal structure, the chain with less than 25 residues is defined as the polypeptide chain, and the remaining chains are defined as the protein chain. The protein chain backbone atoms (N, C, Cα) within the pocket are then aligned using the Kabsch algorithm to align the backbone atoms in the predicted structure and the crystal structure. The polypeptide chains are then aligned using the same rotation and translation matrix. The RMSD of the pocket-aligned polypeptides is all atom is the RMSD between all heavy atoms in the peptide molecule, while the RMSD of the pocket-aligned peptide is unaaOnly the RMSD between heavy atoms of unnatural amino acid residues is calculated. RMSD of pocket-aligned peptides Cα A similar procedure is followed, with the main difference being that only the distance to the Cα atom of the polypeptide chain is selected. The protein chain backbone atoms within are aligned, and the RMSD is calculated using only the Cα atoms in the polypeptide chain. all atom It is determined by aligning all heavy atoms in the predicted structure and the crystal structure using the Kabsch algorithm and then calculating the RMSD between all heavy atoms. When the calculation is limited to unnatural amino acid residues, the RMSD is obtained. unaa RMSD Cα It is obtained by aligning the Cα atoms in the predicted structure and the crystal structure and calculating the RMSD between these atoms.

[0049] In one embodiment of the present invention, the structure prediction model of the non-natural line peptide monomer and its complex obtained by training according to the present invention was tested on the non-natural line peptide test set. The results are shown in Table 1. RMSD Cα The median is 0.994 (RMSD for the complex Cα Refers to the RMSD of the peptide aligned with the pocket Cα , the same below), RMSD all atom The median is 1.906 (RMSD for the complex all atom Refers to the RMSD of the peptide aligned with the pocket all atom , the same below), RMSD unaa The median is 1.971 (RMSD for the complex unaa Refers to the RMSD of the peptide aligned with the pocket unaa , the same below), confirming the powerful ability of the model in predicting the three-dimensional structures of non-natural linear peptide monomers and their complexes.

[0050] Table 1 Performance of the structure prediction model of non-natural peptide monomers and their complexes on the non-natural peptide test set

[0051]

[0052]

[0053] In one embodiment of the present invention, the trained non-natural linear peptide monomer and its complex structure prediction model was applied to the prediction of non-natural cyclic peptide structure by modifying the relative position encoding matrix. The results are shown in Table 2. RMSD Cα The median is 1.891, and the RMSD of most test samples is Cα Less than This indicates that the model is highly accurate in predicting the backbone of non-natural cyclic peptides. all atom The median is 2.872, and there are 6 samples with RMSD all atom Less than This shows that the model also has excellent side chain structure prediction capabilities. RMSD unaa The median score is 2.579, which shows that the model can also accurately predict the three-dimensional structures of most unnatural amino acid residues.

[0054] Table 2 Performance of the non-natural linear peptide monomer and complex structure prediction model with modified relative position encoding matrix on the non-natural cyclic peptide test set

[0055]

[0056]

[0057] The foregoing description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution or modification to the technical solution and technical content disclosed in the present invention, or directly or indirectly applies such changes to other related technical fields, all of which fall within the scope of the present invention and are still within the scope of protection of the present invention.

Claims

1. A deep learning algorithm based on AlphaFold to predict the three-dimensional structure of non-natural cyclic peptide monomers and their complexes, characterized by: The following steps are involved: Step 1: Obtain linear peptide structures containing unnatural amino acids from the RCSB Protein DataBank database, delete solvents, hydrogen atoms, and heteroatoms from the crystal structures, and only retain the first state of each structure; Step 2: Divide the non-natural peptide dataset obtained in step 1 into training set, validation set, and test set and construct features using ColabFold; Step 3: Add a neural network module to the original AlphaFold model to capture the atomic-scale representation of peptide molecules. The atomic-scale features come from the atomic elements and the bond information between them. Step 4: Pool the atomic-scale representation obtained in step 3 to facilitate its combination with the original residue-scale representation of the AlphaFold model to achieve multi-scale modeling of peptide molecules. The atomic element representation and the representation of the bond information between atoms use an attention-based pooling mechanism to map them into amino acid residue-scale embeddings. Step 5: Predict the spatial structure of non-natural amino acids by defining rigid groups and initializing the atomic coordinates of non-natural amino acids. The Structure module in the AlphaFold model uses the single representation and pairwise representation obtained from the Evoformer module to predict the torsion angle of each amino acid residue in the protein. These predicted torsion angles, together with the predefined rigid groups and initialized atomic coordinates, construct the coordinates of all atoms in the protein spatial structure. Step 6: Use the training set and validation set obtained in step 2 to train the structure prediction model of the non-natural linear peptide monomer and its complex. During training, the amino acid residues in each training sample are trimmed to 220, while the validation set remains untrimmed. For samples with breakpoints in the training set, the loss value is halved. To prevent overfitting, an early stopping strategy is implemented based on the loss of the validation set. Step 7: Use the trained model to test on the test set obtained in step 2; Step 8: Obtain a test data set of non-natural cyclic peptide monomers and their complexes, remove solvent molecules, hydrogen atoms, and heteroatoms, retain only the first state of each structure, and delete samples containing nucleic acids and repeats; Step 9: Modify the relative position encoding matrix in the structure prediction model of the non-natural linear peptide monomer and its complex so that the relative positions of the two loop residues in the model can be accurately represented and adjacent to each other; Step 10: Use the model obtained in step 9 to test on the non-natural cyclic peptide test set obtained in step 8.

2. A deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes based on AlphaFold according to claim 1, characterized in that: The specific process of obtaining the linear peptide structure containing unnatural amino acids from the RCSB Protein DataBank database in step 1 is as follows: using the PDB ID provided by the ModPep dataset, download the corresponding crystal structure from the RCSB Protein DataBank database and delete duplicates.

3. A deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes based on AlphaFold according to claim 2, characterized in that: After retaining the first state of each structure in step 1, in order to reduce the redundancy of crystal structures and reduce the computational requirements during training, some chains in the spatial structure are deleted according to the following rules: chains with less than 50 amino acid residues in the three-dimensional structure are classified as polypeptide chains, while other chains are classified as protein chains. If a structure contains at least one protein chain, the sequences of all polypeptide chains are aligned, and sequences with the same or partially overlapping residues and a distance greater than The polypeptide sequence of the structure is analyzed, and only the longest polypeptide chain is retained. The remaining polypeptide chains are deleted. If there are at least two protein chains in the structure at this time, only the protein chain with the most contact with the polypeptide chain is retained. The types of unnatural amino acids in the crystal structure are analyzed, and any structure containing unnatural amino acids that are only found in this crystal structure will also be deleted.

4. A deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes based on AlphaFold according to claim 3, characterized in that: The atomic element information and bond information in step 3 were obtained by converting the FASTA of the peptide molecule into SMILES and then extracting it using RDKit.

5. A deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes based on AlphaFold according to claim 4, characterized in that: The specific operation process of step 4 is that the pooled atomic element embedding will be added to the one-hot matrix encoding the peptide type in the AlphaFold model, and then input into the Evoformer module to generate a single representation. The pooled key information embedding will be added to the pairwise activation matrix in the AlphaFold model to generate the pairwise representation in the Evoformer module. The following is the calculation formula of the attention-based pooling mechanism: A score =Softmax(x atom +F), Among them, M is a mask matrix used to indicate which amino acid residue the atom belongs to, ∈ is a minimum positive value to prevent division by zero error, and x atom represents the atomic scale representation, x residue is the scale representation of the amino acid residues after pooling.

6. A deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes based on AlphaFold according to claim 5, characterized in that: The specific process of defining the rigidity group and initializing the atomic coordinates of the non-natural amino acid in step 5 is to divide them into five rigidity groups according to the dependence of each atom in the non-natural amino acid residue on a specific torsion angle, namely the backbone rigidity group, ω rigidity group, Φ rigidity group, ψ rigidity group and χ rigidity group. The four atoms C, Cα, Cβ and N on the main chain are the backbone rigidity group, the hydrogen atoms connected to Cα are classified as the ω rigidity group, the Φ rigidity group only contains the hydrogen atoms on the amino group, the ψ rigidity group only contains the oxygen atoms in the carboxyl group, and the χ rigidity group contains all the side chain atoms. After determining the rigidity group, the initialization coordinates of each atom of the non-natural amino acid are determined according to the crystal structure of the non-natural amino acid. For the backbone rigidity group, Cα is taken as the origin, C is placed on the positive x-axis, and N is placed In the xy plane, a coordinate system is established to calculate the initial coordinates of Cα, Cβ, C and N atoms respectively. In the ψ rigid group, C is used as the coordinate origin, Cα is on the negative x-axis, and N in the next amino acid residue is on the xy plane. In this way, a coordinate system is established to determine the initial coordinates of O. The χ rigid group contains all atoms that depend on the torsion angle of the side chain. There are four groups in total. According to the true coordinates of the four atoms in each group, the third atom is set as the origin, the second atom is placed on the negative x-axis, and the first atom is placed on the xy plane to establish a coordinate system, thereby determining the relative coordinates of the fourth atom in the coordinate system. Then, the fourth atom is rotated around the x-axis to the xy plane to determine the initial coordinates of the fourth atom. The calculation formula is as follows: Coordinate init =R x ·Coordinate relative , Among them, Coordinate relative is the relative coordinate of the fourth atom, and θ is the angle between it and the xy plane, R x is the rotation matrix, Coordinate init Represents the initial coordinates of the fourth atom.

7. A deep learning algorithm for predicting the three-dimensional structure of non-natural cyclic peptide monomers and their complexes based on AlphaFold according to claim 6, characterized in that: The specific operation process of step eight is to filter the non-natural cyclic peptide data from cPEPmatch, exclude samples containing uncoded non-natural amino acids, and limit the cyclization form to head-to-tail cyclization and disulfide bond cyclization. At the same time, the RCSB ProteinDataBank database is searched to determine the cyclic peptide structure that meets the standards. After downloading the corresponding non-natural cyclic peptide crystal structure from the RCSB ProteinDataBank database, the solvent molecules, hydrogen atoms and heteroatoms are removed, and only the first state of each structure is retained. Samples containing nucleic acids and repeats are deleted.

Images