A peptide design method based on protein language model and Monte Carlo

Abstract

The present invention discloses a polypeptide design method based on a protein language model and Monte Carlo. On the one hand, the present invention uses a protein language model to extract characteristic representations of receptor proteins and newly designed polypeptides based on sequence information, and builds a deep neural network model to predict the binding probability between receptor proteins and polypeptides, thus preparing for improving polypeptide design performance. On the other hand, the Monte Carlo algorithm is used to design polypeptide sequences of varying lengths, and the above-mentioned deep neural network model is used to evaluate the binding probability of the newly designed polypeptides and the receptor proteins. The polypeptide sequence corresponding to the highest binding probability is the final design result.

Description

Technical Field

[0001] The present invention relates to a polypeptide design method based on a protein language model and Monte Carlo, belonging to the technical field of protein polypeptide design. Background Art

[0002] Proteins do not exist in isolation during life; they must interact with various ligands to fulfill specific biological functions. Most therapeutic drugs also act as ligands, interacting with target proteins (i.e., receptors) and altering their function to achieve therapeutic effects. Common therapeutic drug molecules include small organic molecules, proteins, and peptides. Compared to small organic molecules and proteins, peptides offer inherent advantages, including strong targeting and minimal toxicity, making them one of the most promising candidates for drug development. Although biological experiments are the most precise method for designing peptide drugs, they are time-consuming, labor-intensive, and costly, significantly hindering the development of peptide drugs. Computational peptide drug design methods, primarily structure-based and sequence-based, have garnered increasing attention. A limitation of structure-based methods is that they require the known three-dimensional structure of the receptor protein. However, accurate determination of the structure of all proteins is currently unavailable, significantly hindering the progress of structure-based computational methods. In contrast, sequence-based methods can design peptides with binding capabilities based solely on protein sequence information, significantly accelerating the design of peptide drugs. The main disadvantage of sequence-based computational methods is that the indicators used to evaluate the binding ability of peptides to receptor proteins are imprecise, resulting in lower design performance.

[0003] The Joker method proposed by William F. Porto et al. in 2018 (see the literature Joker: An algorithm to insert patterns into sequences for designing antimicrobial peptides) is one of the few sequence-based computational methods for peptide design. It uses regular matching of sequence patterns to design a peptide sequence of fixed length. Although the Joker method can be used for peptide design, it does not fully consider the specific information of the receptor protein sequence information, thereby limiting the peptide design accuracy of the Joker method. Existing sequence-based computational methods for peptide design are still far from the requirements of practical applications in terms of design accuracy. Summary of the Invention

[0004] In order to overcome the shortcomings of existing sequence-based computational methods in terms of peptide design accuracy, the present invention proposes a peptide design method based on a protein language model with high design accuracy.

[0005] The first object of the present invention is to provide a polypeptide design method based on a protein language model and Monte Carlo, the method comprising the following steps:

[0006] S1. Input a protein sequence S for peptide ligand design;

[0007] S2, randomly generate a polypeptide sequence P with a residue number of l;

[0008] S3. Build a deep neural network to predict the binding probability between protein sequence S and peptide sequence P;

[0009] S4. Collect data on protein and peptide sequence pairs that have been determined to interact in the PDB database, construct a training set to train the deep neural network model built in step S3, and obtain a trained deep neural network model;

[0010] S5. Using the Monte Carlo algorithm, the deep neural network model trained in step S4 is used as the energy function to generate a Monte Carlo trajectory. Any point in the trajectory is a potential peptide design result:

[0011] S501, using the deep neural network model trained in step S4 to predict the binding probability between the protein sequence S and the polypeptide sequence P, denoted as E;

[0012] S502. Randomly generate an integer R between 1 and 5. If R is 1, randomly select an amino acid type from 20 common amino acid types and place an amino acid of this type at the N-terminus of the polypeptide sequence P. If R is 2, randomly select an amino acid type from 20 common amino acid types and place an amino acid of this type at the C-terminus of the polypeptide sequence P. If R is 3, delete an amino acid from the N-terminus of the polypeptide sequence P. If R is 4, delete an amino acid from the C-terminus of the polypeptide sequence P. If R is 5, randomly select an amino acid from the polypeptide sequence P and replace it with an amino acid of a random type. The newly generated polypeptide sequence is denoted as Pnew.

[0013] S503, using the deep neural network model trained in step S4 to predict the binding probability between the protein sequence S and the polypeptide sequence Pnew generated in step S502, denoted as Enew;

[0014] S504. If Enew is greater than E, then Pnew is a point in the Monte Carlo trajectory, and steps S501 to S503 are re-executed with Pnew as the polypeptide sequence P; if Enew is less than or equal to E, a number rd between 0 and 1 is randomly generated. If rd is less than e (Enew-E) / T, T is the temperature parameter in the Monte Carlo algorithm, then Pnew is a point in the Monte Carlo trajectory, and steps S501 to S503 are re-executed with Pnew as the polypeptide sequence P, otherwise steps S501 to S503 are directly executed;

[0015] S505: The above process is repeated until N Monte Carlo trajectory points are obtained. From these N trajectory points, the trajectory point with the highest binding probability to the protein sequence S is selected as the final designed polypeptide sequence P. final .

[0016] Furthermore, the neural network includes a first submodule, a second submodule, a cross-attention layer and a fully connected layer; the protein sequence S is input into the first submodule, the polypeptide sequence P is input into the second submodule, the outputs of the first submodule and the second submodule are directly input into the cross-attention layer, and the output of the cross-attention layer is input into the fully connected layer to obtain the binding probability of S and P.

[0017] Furthermore, the first submodule contains a protein language pre-trained model ESM2 with the first 32 layers of parameters frozen, three serial convolutional modules and a fully connected layer. Each convolutional module contains a convolutional layer, a normalization layer and an excitation layer.

[0018] Furthermore, the second submodule contains a protein language pre-trained model ESM2 with the first 32 layers of parameters frozen, three serial convolutional modules and a fully connected layer. Each convolutional module contains a convolutional layer, a normalization layer and an excitation layer.

[0019] Furthermore, in step S4, when training the deep neural network model built in step S3, the Adam optimizer and the second-class cross entropy loss function are used to adjust the adjustable parameters in the network. The Adam optimizer automatically adjusts the parameters that minimize the loss.

[0020] Furthermore, the second-class cross entropy loss function is

[0021] Among them, y i is the label of the i-th training sample, y i ∈{0,1},p(y i ) is the output of the deep neural network model belonging to y i The probability of the label, N is the total number of training samples.

[0022] Furthermore, the residue number l is 5-30.

[0023] Furthermore, the number N of Monte Carlo trajectory points is at least 3000.

[0024] The second object of the present invention is to provide a memory capable of executing the polypeptide design method based on protein language model and Monte Carlo.

[0025] The third object of the present invention is to provide an electronic device comprising a processor; and a memory; the memory stores executable code, and when the executable code is executed by the processor, the processor executes the polypeptide design method based on protein language model and Monte Carlo.

[0026] The beneficial effects of the present invention are:

[0027] On the one hand, the present invention uses a protein language model to extract the characteristic representations of receptor proteins and newly designed peptides based on sequence information, and builds a deep neural network model to predict the binding probability of receptor proteins and peptides, thus preparing for improving the performance of peptide design. On the other hand, the Monte Carlo algorithm is used to design peptide sequences of varying lengths, and the above-mentioned deep neural network model is used to evaluate the binding probability of the newly designed peptides and receptor proteins. The peptide sequence corresponding to the highest binding probability is the final design result. DETAILED DESCRIPTION

[0028] The present invention will be further described below with reference to specific examples so that those skilled in the art can better understand the present invention and implement it, but the examples are not intended to limit the present invention.

[0029] Example 1:

[0030] This example uses the design of peptides that interact with protein 6CQRE as an example. A peptide design method based on a protein language model and Monte Carlo analysis includes the following steps:

[0031] 1) Input a protein sequence S for peptide ligand design;

[0032] 2) Randomly generate a polypeptide sequence P with 7 residues;

[0033] 3) Build a deep neural network to predict the binding probability of a protein sequence S and a peptide sequence P. The neural network consists of two submodules, a cross-attention layer, and a fully connected layer. Each submodule contains a protein language pre-trained model ESM2 (33 layers in total, with the first 32 layers frozen and not involved in learning) with the parameters of the first 32 layers frozen, three convolutional modules, and a fully connected layer. Each convolutional module contains a convolutional layer, a normalization layer, and an excitation layer. The protein sequence S and the peptide sequence P are input into the two submodules respectively. The outputs of the two submodules are directly input into the cross-attention layer. The output of the cross-attention layer is input into the final fully connected layer to obtain the binding probability of S and P.

[0034] 4) Collecting protein-peptide sequence pairs that have been determined to interact with each other from the PDB database, constructing a training set to train the deep neural network model built in step 3), and using the Adam optimizer and the two-class cross entropy loss function to adjust the adjustable parameters in the network;

[0035] Among them, the second-class cross entropy loss function is

[0036] Among them, y i is the label of the i-th training sample, y i ∈{0,1},p(y i ) is the output of the deep neural network model belonging to y i The probability of the label, N is the total number of training samples; the Adam optimizer automatically adjusts the parameters to minimize the loss;

[0037] 5) Using the Monte Carlo algorithm, the deep neural network model trained in step 4) is used as the energy function to generate a Monte Carlo trajectory. Any point in the trajectory is a potential peptide design result:

[0038] 5.1) Using the deep neural network model trained in step 4), predict the binding probability between protein sequence S and peptide sequence P, denoted as E;

[0039] 5.2) Randomly generate an integer R between 1 and 5. If R is 1, randomly select an amino acid type from 20 common amino acid types and place an amino acid of this type at the N-terminus of the polypeptide sequence P; if R is 2, randomly select an amino acid type from 20 common amino acid types and place an amino acid of this type at the C-terminus of the polypeptide sequence P; if R is 3, delete an amino acid from the N-terminus of the polypeptide sequence P; if R is 4, delete an amino acid from the C-terminus of the polypeptide sequence P; if R is 5, randomly select an amino acid from the polypeptide sequence P and replace it with an amino acid of a random type; the newly generated polypeptide sequence is denoted as Pnew;

[0040] 5.3) Using the deep neural network model trained in step 4), predict the binding probability between the protein sequence S and the peptide sequence Pnew generated in step 5.2), denoted as Enew;

[0041] 5.4) If Enew is greater than E, then Pnew is a point in the Monte Carlo trajectory, and steps 5.1) to 5.3) are re-executed with Pnew as P; if Enew is less than or equal to E, a number rd between 0 and 1 is randomly generated. If rd is less than e(Enew-E) / T, where T=1 is the temperature parameter in the Monte Carlo algorithm, then Pnew is a point in the Monte Carlo trajectory, and steps 5.1) to 5.3) are re-executed with Pnew as the polypeptide sequence P; otherwise, directly execute steps 5.1) to 5.3);

[0042] 5.5) The above process is repeated until 10,000 Monte Carlo trajectory points are obtained. From these 10,000 trajectory points, the trajectory point with the highest binding probability to S is selected as the final designed polypeptide sequence Pfinal.

[0043] Taking the design of interacting peptides for protein 6CQRE as an example, the newly designed peptide sequence obtained by the above method is GFYKTLRGEQGSQ. The sequence similarity between the newly designed peptide calculated by the sequence alignment method NW-align and the natural peptide in the PDB database is 76.9%, proving that the above method is effective.

[0044] The above embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. A peptide design method based on protein language model and Monte Carlo, characterized in that: The method comprises the following steps: S1. Input a protein sequence for peptide ligand design S ; S2, randomly generate a residue number l Peptide sequence P ; S3. Build a deep neural network to predict protein sequences S With peptide sequence P The binding probability of S4. Collect data on protein and peptide sequence pairs that have been determined to interact in the PDB database, construct a training set to train the deep neural network model built in step S3, and obtain a trained deep neural network model; S5. Using the Monte Carlo algorithm, the deep neural network model trained in step S4 is used as the energy function to generate a Monte Carlo trajectory. Any point in the trajectory is a potential peptide design result: S501: Use the deep neural network model trained in step S4 to predict protein sequences S With peptide sequence P The binding probability is denoted as E; S502, randomly generate an integer R between 1 and 5. If R is 1, randomly select an amino acid type from 20 common amino acid types and place an amino acid of this type in the polypeptide sequence. P If R is 2, an amino acid type is randomly selected from 20 common amino acid types and an amino acid of this type is placed in the peptide sequence. P If R is 3, then from the polypeptide sequence P If R is 4, then delete an amino acid from the N-terminus of the polypeptide sequence. P Delete an amino acid from the C-terminus of the peptide sequence; if R is 5, P Select an amino acid from the sequence and replace it with a random type of amino acid; record the newly generated polypeptide sequence as Pnew ; S503: Use the deep neural network model trained in step S4 to predict protein sequences S The polypeptide sequence generated in step S502 Pnew The binding probability is denoted as Enew; S504. If Enew is greater than E, then Pnew is a point in the Monte Carlo trajectory, and Pnew As a polypeptide sequence P Re-execute steps S501 to S503; if Enew is less than or equal to E, randomly generate a number rd between 0 and 1. If rd is less than e (Enew-E) / T , T is the temperature parameter in the Monte Carlo algorithm, then Pnew is a point in the Monte Carlo trajectory, and Pnew As a polypeptide sequence P Re-execute steps S501 to S503, otherwise directly execute steps S501 to S503; S505, the above process is repeated until the N The Monte Carlo trajectory ends at this point. N Among the trajectory points, select the one with the protein sequence S The track point with the highest binding probability is used as the final designed peptide sequence P final ; The neural network comprises a first submodule, a second submodule, a cross attention layer and a fully connected layer; the protein sequence S Input the first submodule, peptide sequence P Input the second submodule, the output of the first and second submodules are directly input to the cross attention layer, and the output of the cross attention layer is input to the fully connected layer to obtain S and P The binding probability.

2. The polypeptide design method according to claim 1, characterized in that The first submodule contains a protein language pre-trained model ESM2 with the first 32 layers of parameters frozen, three serial convolutional modules and a fully connected layer. Each convolutional module contains a convolutional layer, a normalization layer and an excitation layer.

3. The polypeptide design method according to claim 1, characterized in that The second submodule contains a protein language pre-trained model ESM2 with the first 32 layers of parameters frozen, three serial convolutional modules and a fully connected layer. Each convolutional module contains a convolutional layer, a normalization layer and an excitation layer.

4. The polypeptide design method according to claim 1, characterized in that In step S4, when training the deep neural network model built in step S3, the Adam optimizer and the two-class cross entropy loss function are used to adjust the adjustable parameters in the network. The Adam optimizer automatically adjusts the parameters that minimize the loss.

5. The polypeptide design method according to claim 4, characterized in that The two-class cross entropy loss function is Loss= ; in, , The output of the deep neural network model belongs to .

6. The polypeptide design method according to claim 1, characterized in that: Number of residues l 5~30.

7. The polypeptide design method according to claim 1, characterized in that Number of Monte Carlo trajectory points N At least 3000.

8. A memory, characterized in that: The memory stores executable code, and the executable code is used to execute the polypeptide design method based on protein language model and Monte Carlo according to any one of claims 1 to 7.

9. An electronic device, characterized in that: The method comprises a processor and a memory; the memory stores executable code, and when the executable code is executed by the processor, the processor executes the polypeptide design method based on protein language model and Monte Carlo according to any one of claims 1 to 7.

Images