Binding force prediction model training method, polypeptide sequence design method, and device

Abstract

The present application relates to the field of polypeptide design, and provides a training method of a binding force prediction model, a polypeptide sequence design method and device, the training method comprising obtaining sequence features of a plurality of candidate polypeptide sequence samples, the sequence features comprising proportions of various amino acids and occurrence frequencies of each adjacent residue pairing group; selecting a plurality of target polypeptide sequence samples from the plurality of candidate polypeptide sequence samples according to the proportions of various amino acids and the occurrence frequencies of each adjacent residue pairing group contained in each sequence feature; simulating binding forces between each target polypeptide sequence sample and an extracellular vesicle membrane; and using at least part of the target polypeptide sequence samples and the simulated binding forces thereof as binding force labels to train a binding force prediction model, thereby obtaining a final binding force prediction model. The present application can improve the binding force prediction accuracy of the final binding force prediction model while reducing the training cost.

Description

Technical Field

[0001] This invention relates to the field of peptide design technology, and in particular to a training method for a binding force prediction model, a peptide sequence design method, and an apparatus. Background Technology

[0002] Extracellular vesicles possess crucial clinical and research value in intercellular communication and as diagnostic biomarkers for diseases. Developing highly specific and high-affinity targeting ligands is a critical technological need for the effective detection, isolation, and targeted delivery of extracellular vesicles. Peptides, due to their small molecular weight, ease of synthesis, and good biocompatibility, are widely considered ideal candidates for targeting ligands. Therefore, rapidly and accurately evaluating and obtaining peptide sequences capable of efficiently binding to extracellular vesicle membranes is a key research focus in this field.

[0003] In existing technologies, to meet the above requirements, researchers often use machine learning techniques to establish peptide screening models. The specific implementation typically involves: randomly sampling a large amount of peptide sequence data from a public peptide database; simulating the binding ability of the peptide sequences in the database using molecular docking software or molecular dynamics simulation platforms; calculating the binding parameters with the target membrane structure; and using these parameters as labels to input into the machine learning algorithm for model training. Finally, the trained model is used to predict and evaluate the binding activity of unknown peptides.

[0004] However, existing model training mechanisms have significant limitations when dealing with massive amounts of peptide data. Directly performing extensive membrane binding simulations on a large peptide database not only requires extremely high computational resources and time costs, making efficient data iteration difficult, but also the peptide samples selected for simulation and training often lack structural certainty and representativeness. This results in low accuracy of the constructed prediction model when assessing binding affinity during the application phase, leading to low efficiency of the designed peptide sequences in separating extracellular vesicle membranes in scenarios such as plasma. Summary of the Invention

[0005] This invention provides a training method for a binding force prediction model, a peptide sequence design method, and an apparatus to solve the problem that the prediction model trained by randomly sampling a large number of peptide sequences in the prior art has low accuracy in practical applications, thereby improving the accuracy of the trained prediction model in practical applications.

[0006] This invention provides a training method for a binding force prediction model, comprising the following steps: The sequence features of multiple candidate polypeptide sequence samples are obtained, including the proportion of various amino acids and the frequency of occurrence of each adjacent residue pairing. Based on the proportion of various amino acids contained in each sequence feature and the frequency of occurrence of each adjacent residue pairing, multiple target polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples; Simulate the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane; Using at least a portion of the target peptide sequence samples and their respective simulated binding forces as binding force labels, the binding force prediction model is trained to obtain the final binding force prediction model.

[0007] According to the training method of the binding force prediction model provided by the present invention, multiple target polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples based on the proportion of various amino acids contained in each sequence feature and the frequency of occurrence of each adjacent residue pairing group, including: The overall proportion of each amino acid in multiple candidate polypeptide sequence samples is determined based on the frequency of occurrence of each adjacent residue pairing in each candidate polypeptide sequence sample. Based on the difference between the proportion of each amino acid in each candidate polypeptide sequence sample and the corresponding overall proportion, multiple candidate polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples and used as target polypeptide sequence samples.

[0008] According to a training method for a binding affinity prediction model provided by the present invention, the overall proportion of each amino acid in multiple candidate polypeptide sequence samples is determined based on the frequency of occurrence of each adjacent residue pairing group in each candidate polypeptide sequence sample, including: The total frequency of each adjacent residue pairing group was counted in all candidate polypeptide sequence samples. Select the first number of adjacent residue pairs with the highest total frequency of each pair to obtain multiple target adjacent residue pairs. Determine the first frequency of each amino acid in all target adjacent residue pairs and use the sum of the first frequencies of all amino acids as the second frequency; For each amino acid, the ratio between the first frequency and the second frequency is taken as the overall proportion of the amino acid.

[0009] According to the training method of the binding force prediction model provided by the present invention, based on the difference between the proportion of each amino acid in each candidate polypeptide sequence sample and the corresponding overall proportion, multiple candidate polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples as target polypeptide sequence samples, including: For each candidate polypeptide sequence sample, the difference between the proportion of each amino acid in the candidate polypeptide sequence sample and the corresponding overall proportion is obtained and the differences are statistically analyzed to obtain the total difference value of the candidate polypeptide sequence sample. Candidate polypeptide sequence samples with a total difference value less than or equal to a preset difference value are selected as target polypeptide sequence samples.

[0010] According to the training method of the binding force prediction model provided by the present invention, before obtaining the sequence features of multiple candidate polypeptide sequence samples, the method further includes: Multiple initial polypeptide sequence samples were selected from the polypeptide database; Each initial polypeptide sequence sample is screened according to at least one screening dimension to obtain each candidate polypeptide sequence sample. The at least one screening dimension includes sequence length, overall charge and / or secondary structure.

[0011] According to a training method for a binding force prediction model provided by the present invention, multiple initial polypeptide sequence samples are selected from a polypeptide database, including: Based on multiple search keywords, initial polypeptide sequence samples were selected from the polypeptide database. Each search keyword was associated with the binding ability to target extracellular vesicle membranes. In this process, the keyword tag of each initial polypeptide sequence sample is matched with one or more membrane binding search keywords.

[0012] According to the training method of the binding force prediction model provided by the present invention, the method for obtaining the sequence features of each candidate polypeptide sequence sample includes: Determine the category of each amino acid in the candidate polypeptide sequence sample, which can be cationic, hydrophobic, aromatic, acidic, nucleophilic, or amide. Based on the category to which each amino acid belongs, determine the proportion of each amino acid and the frequency of occurrence of each pair of adjacent residues.

[0013] According to a training method for a binding force prediction model provided by the present invention, the binding force prediction model is trained using at least a portion of the target polypeptide sequence samples and their respective simulated binding forces as binding force labels to obtain the final binding force prediction model, comprising: Based on the binding force simulated from each target polypeptide sequence sample, the second-highest number of target polypeptide sequence samples with the highest binding force are selected and added to the training dataset. The binding force prediction model is trained using the target peptide sequence samples and their binding force labels in the training dataset to obtain the final binding force prediction model.

[0014] According to the training method of the binding force prediction model provided by the present invention, there are multiple binding force prediction models to be trained, and the different binding force prediction models are of different types. The binding force prediction models are trained using target peptide sequence samples and their binding force labels in the training dataset to obtain the final binding force prediction model, including: Using the target polypeptide sequence samples and their binding force labels in the training dataset, each binding force prediction model was iteratively trained. After the iteration stopping condition is met, each bonding force prediction model is tested, and based on the test results of each bonding force prediction model, at least one bonding force prediction model with better test results is selected as the final bonding force prediction model.

[0015] This invention also provides a method for designing polypeptide sequences, comprising the following steps: Obtain multiple peptide sequences to be evaluated; Multiple peptide sequences to be evaluated were assessed using a binding force prediction model to obtain the predicted binding force between each peptide sequence and the extracellular vesicle membrane. The binding force prediction model was trained based on any of the above-mentioned binding force prediction model training methods. Based on the predicted binding forces, the target polypeptide sequence is selected from multiple polypeptide sequences to be evaluated.

[0016] According to a peptide sequence design method provided by the present invention, the number of binding force prediction models is at least two, and a target peptide sequence is selected from multiple peptide sequences to be evaluated based on each predicted binding force, including: For each predicted binding force output by the binding force prediction model, based on the magnitude of each predicted binding force, the peptide sequences corresponding to the top three highest predicted binding forces are selected as candidate peptide sequences corresponding to the binding force prediction model. The relative signal intensity between each candidate polypeptide sequence and the extracellular vesicle membrane was verified to obtain the relative signal intensity corresponding to each candidate polypeptide sequence. Based on the relative signal intensity corresponding to each candidate polypeptide sequence, at least one candidate polypeptide sequence is selected as the target polypeptide sequence.

[0017] According to a polypeptide sequence design method provided by the present invention, based on the relative signal intensities corresponding to each candidate polypeptide sequence, at least one candidate polypeptide sequence is selected as the target polypeptide sequence, including: The relative signal intensities were compared with the relative signal intensities between known polypeptide sequences and extracellular vesicle membranes; Based on the alignment results, the target polypeptide sequence was obtained.

[0018] According to a polypeptide sequence design method provided by the present invention, before obtaining multiple polypeptide sequences to be evaluated, the method further includes: Multiple reference sequence features are used as input to the sequence generation model, and the multiple peptide sequences output by the sequence generation model are used as the peptide sequences to be evaluated. The sequence generation model is trained based on sequence feature samples and their corresponding polypeptide sequences as labels.

[0019] The present invention also provides a training device for a binding force prediction model, comprising the following modules: The first feature acquisition module is used to acquire the sequence features of multiple candidate polypeptide sequence samples. The sequence features include the proportion of various amino acids and the frequency of occurrence of each pair of adjacent residues. The sample screening module is used to select multiple target polypeptide sequence samples from multiple candidate polypeptide sequence samples based on the proportion of various amino acids contained in each sequence feature and the frequency of occurrence of each adjacent residue pairing. The binding force label determination module is used to simulate the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane; The training module is used to train the binding force prediction model using at least a portion of the target peptide sequence samples and their respective simulated binding forces as binding force labels, so as to obtain the final binding force prediction model.

[0020] The present invention also provides a polypeptide sequence design device, comprising the following modules: The second feature acquisition module is used to acquire multiple polypeptide sequences to be evaluated; The binding force assessment module is used to assess multiple peptide sequences to be evaluated using a binding force prediction model, and to obtain the predicted binding force between each peptide sequence to be evaluated and the extracellular vesicle membrane. The binding force prediction model is trained based on any of the above-mentioned binding force prediction model training methods. The sequence screening module is used to select the target peptide sequence from multiple peptide sequences to be evaluated based on each predicted binding force.

[0021] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a training method for any of the binding force prediction models described above or implements any of the polypeptide sequence design methods.

[0022] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a training method for any of the binding force prediction models described above or implements any of the polypeptide sequence design methods.

[0023] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a training method for any of the binding force prediction models described above or implements any of the polypeptide sequence design methods.

[0024] This invention provides a training method, peptide sequence design method, and apparatus for a binding force prediction model. By acquiring the sequence characteristics of multiple candidate peptide sequence samples and selecting target peptide sequence samples based on amino acid proportions and the frequency of adjacent residue pairings, the binding force between each target peptide sequence sample and the extracellular vesicle membrane is obtained through simulation. Finally, the binding force prediction model is trained using the target peptide sequence samples and their binding forces as training data. Compared to directly performing extensive membrane binding simulations on a large peptide database, this method not only reduces the computational simulation cycle and resource consumption, but also effectively characterizes the sequence composition features of candidate peptide sequence samples because the proportions of various amino acids and the frequency of adjacent residue pairings can effectively characterize the sequence composition features. This makes the selected target peptide sequence samples more representative in terms of sequence composition, helping to incorporate the relationship between sequence composition rules and membrane binding ability into the modeling. Therefore, the trained binding force prediction model can more accurately predict the binding force between the peptide sequence to be evaluated and the extracellular vesicle membrane in practical applications, thereby improving the separation efficiency of extracellular vesicle membranes. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the training method for the bonding force prediction model provided by the present invention.

[0027] Figure 2 This is a schematic diagram showing the frequency statistics of adjacent residue pairing groups provided by the present invention.

[0028] Figure 3 This is a schematic diagram showing the overall percentage distribution of various amino acids provided by the present invention.

[0029] Figure 4 This is a schematic diagram of the two-dimensional distribution of multiple target polypeptide sequence samples after dimensionality reduction provided by the present invention.

[0030] Figure 5 This is a schematic diagram of the binding force distribution of multiple target polypeptide sequence samples provided by the present invention.

[0031] Figure 6 This is a flowchart illustrating the polypeptide sequence design method provided by the present invention.

[0032] Figure 7 This is a schematic diagram of the structure of the training device for the bonding force prediction model provided by the present invention.

[0033] Figure 8 This is a schematic diagram of the polypeptide sequence design device provided by the present invention.

[0034] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0036] The following is combined Figures 1 to 5 This invention describes a training method for the bonding force prediction model provided by the present invention. This training method for the bonding force prediction model can be executed by an electronic device with data processing capabilities, such as a server, workstation, or personal computer; the present invention does not limit this method. Figure 1 This is a flowchart illustrating the training method for the bonding force prediction model provided by the present invention, as shown below. Figure 1 As shown, the training method includes the following steps: Step 101: Obtain the sequence characteristics of multiple candidate polypeptide sequence samples.

[0037] In this invention, candidate polypeptide sequence samples can be considered as polypeptide sequences used as training data to construct a binding affinity prediction model. Each candidate polypeptide sequence sample can consist of several amino acid residues arranged in sequence. The number of candidate polypeptide sequence samples can be multiple, such as hundreds, thousands, or even tens of thousands. This invention does not limit the specific number of candidate polypeptide sequence samples.

[0038] Sequence features can be considered as characteristic information used to characterize the compositional regularity of a candidate polypeptide sequence sample. In this embodiment, sequence features include the proportion of various amino acids and the frequency of occurrence of each adjacent residue pair. The proportion of each amino acid can be considered as the percentage of that amino acid in a candidate polypeptide sequence sample. Here, amino acids can refer to specific amino acid types among the 20 standard amino acids, or to amino acid categories obtained after classifying the 20 amino acids according to certain rules; this embodiment does not limit this. An adjacent residue pair is a pairing combination consisting of two adjacent amino acids in a candidate polypeptide sequence sample. The frequency of occurrence of one type of adjacent residue pair can be considered as the number of times that type of adjacent residue pair appears in a candidate polypeptide sequence sample. For example, for a candidate polypeptide sequence sample, starting from the beginning of the sequence, every two adjacent amino acids can be sequentially divided into groups, and the frequency of occurrence of each type of adjacent residue pair can be obtained by counting the number of occurrences of each type of adjacent residue pair.

[0039] In some application scenarios, the method for obtaining sequence features in step 101 can be: analyzing each candidate polypeptide sequence sample position by position, counting the number of various amino acids and calculating the corresponding proportion of amino acids, and pairing adjacent amino acids in the sequence and counting the frequency of occurrence of various adjacent residue pairing groups, thereby obtaining the sequence features of the candidate polypeptide sequence sample.

[0040] Step 102: Select multiple target polypeptide sequence samples from multiple candidate polypeptide sequence samples based on the proportion of various amino acids contained in each sequence feature and the frequency of occurrence of each adjacent residue pairing group.

[0041] In some application scenarios, step 102 can be implemented as follows: Based on the proportion of various amino acids and the frequency of occurrence of each adjacent residue pair in each candidate polypeptide sequence sample, determine the overall sequence characteristics of each candidate polypeptide sequence sample in terms of sequence composition. Then, select candidate polypeptide sequence samples that have minimal differences from these overall sequence characteristics as the target polypeptide sequence samples. This eliminates candidate polypeptide sequence samples whose sequence composition deviates from the overall sequence characteristics, resulting in a more concentrated distribution of amino acid proportions and adjacent residue pairs in the obtained target polypeptide sequence samples.

[0042] Step 103: Simulate the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane.

[0043] The extracellular vesicle membrane can be considered as a lipid bilayer structure on the outer layer of extracellular vesicles, and is the target for peptide ligands to bind to. Extracellular vesicles can include, but are not limited to, exosomes, microvesicles, and other nanoparticles encapsulated by a lipid bilayer. Binding force is a quantitative indicator used to characterize the strength of the binding ability between a peptide sequence and the extracellular vesicle membrane, such as the binding constant.

[0044] In some application scenarios, step 103 can employ molecular docking simulation to model the binding force between the target polypeptide sequence sample and the extracellular vesicle membrane. For example, molecular docking simulation tools (such as AutoDockVina 2) can be used to simulate the docking of each target polypeptide sequence sample with the lipid bilayer of the extracellular vesicle, calculate the binding constant between each target polypeptide sequence sample and the extracellular vesicle membrane, and use this binding constant as the binding force between the target polypeptide sequence sample and the extracellular vesicle membrane.

[0045] Step 104: Using at least a portion of the target polypeptide sequence samples and their respective simulated binding forces as binding force labels, the binding force prediction model is trained to obtain the final binding force prediction model.

[0046] Among them, the types of predictive models for associative force can be Extra Trees Regressor (ETR), Random Forest Regressor (RFR), Extreme Gradient Boosting Regression (XGB), Support Vector Regression (SVR), Kernel Ridge Regression (KRR), Ridge Regression (hereinafter referred to as Ridge), Least Absolute Shrinkage and Selection Operator Regression (hereinafter referred to as Lasso), Gaussian Process Regression (GPR), and K-Nearest Neighbors Regression (KNN).

[0047] At least part of this can be either using all target polypeptide sequence samples and their simulated binding forces for training, or selecting a portion of the target polypeptide sequence samples and their simulated binding forces from all target polypeptide sequence samples for training. This embodiment does not limit this.

[0048] For example, at least a portion of the sequence features of the target polypeptide sequence sample used in the training are used as input to the binding force prediction model, and the simulated binding force of the target polypeptide sequence sample is used as the binding force label to construct training data. Then, the binding force prediction model is iteratively trained using this training data. During the iterative training process, the model parameters of the binding force prediction model are continuously adjusted so that the difference between the predicted binding force output by the binding force prediction model and the binding force label gradually decreases until the training stopping condition is met (e.g., the number of iterations reaches a preset number or the loss function converges), thus obtaining the final binding force prediction model. The final binding force prediction model uses at least a portion of the sequence features of the polypeptide sequence as input and the binding force between the polypeptide sequence and the extracellular vesicle membrane as output.

[0049] In the above scheme, sequence characteristics of multiple candidate polypeptide sequence samples are obtained, and target polypeptide sequence samples are selected based on amino acid proportions and the frequency of occurrence of adjacent residue pairings. The binding force between each target polypeptide sequence sample and the extracellular vesicle membrane is then simulated. Finally, the target polypeptide sequence samples and their binding forces are used as training data to train the binding force prediction model. Compared to directly performing extensive membrane binding simulations on a large polypeptide database, this scheme not only reduces the computational simulation cycle and resource consumption, but also effectively characterizes the sequence composition features of candidate polypeptide sequence samples because the proportions of various amino acids and the frequency of occurrence of adjacent residue pairings can effectively characterize the sequence composition features of the candidate polypeptide sequence samples. This makes the selected target polypeptide sequence samples more representative in terms of sequence composition, helping to incorporate the relationship between sequence composition rules and membrane binding ability into the modeling. Therefore, the trained binding force prediction model can more accurately predict the binding force between the polypeptide sequence to be evaluated and the extracellular vesicle membrane in practical applications, thereby improving the separation efficiency of extracellular vesicle membranes.

[0050] In some embodiments, before performing step 101 above, the method further includes: First, select multiple initial polypeptide sequence samples from the polypeptide database.

[0051] The peptide database can be considered as a database containing a large number of peptide sequences and related information, such as including but not limited to the Peptipedia database, SATPdb database, APD6 database, UniProtKB database, etc. The number of peptide databases can be one or multiple. For example, in this embodiment, multiple initial peptide sequence samples can be selected from the Peptipedia database and the SATPdb database.

[0052] In some application scenarios, selecting multiple initial peptide sequence samples from a peptide database can be achieved by sampling peptide sequences from a peptide database (such as Peptipedia and SATPdb) and using the sampled peptide sequences as initial peptide sequence samples. Alternatively, in some embodiments, selecting multiple initial peptide sequence samples from a peptide database can also be achieved by selecting each initial peptide sequence sample from the peptide database based on multiple search keywords, where each search keyword is related to the binding ability to the extracellular vesicle membrane; wherein the keyword tag of each initial peptide sequence sample matches one or more membrane binding search keywords.

[0053] The number of search keywords is multiple. Keyword tags can be considered as tag information carried by peptide sequences in a peptide database, used to identify their functional attributes.

[0054] For example, multiple search keywords related to the binding ability to target extracellular vesicle membranes are pre-defined. These keywords are then used to search a peptide database. If a peptide sequence's keyword tag matches one or more of these search keywords, that peptide sequence is selected as the initial peptide sequence sample. In other words, a peptide sequence is retained as long as it carries any keyword tag corresponding to any one of the multiple search keywords, thereby preserving as many membrane-binding-related peptide sequences as possible and ensuring the sufficiency of the initial dataset.

[0055] In some application scenarios, six search keywords can be set, such as general membrane binding, fusion lipid binding, pore forming, hemolytic / cytolytic toxicity, cell penetrating delivery, and antimicrobial membrane activity.

[0056] Among these, general membrane binding refers to the ability of a polypeptide sequence to non-specifically adhere to the surface of a cell membrane or lipid layer; for example, a positively charged polypeptide sequence being electrostatically attracted to a negatively charged cell membrane falls into this category. Fusion-lipid binding refers to the ability of a polypeptide sequence not only to bind to a lipid layer but also to promote the adhesion and eventual fusion of two membranes. Pore forming refers to the ability of a polypeptide sequence to assemble on a membrane to form channels or pores, allowing ions, small molecules, and even water molecules to enter or exit. Hemolytic / cytotoxic refers to the properties of a polypeptide sequence that can lyse red blood cells, lyse various types of cells, or exhibit cytotoxicity. Cell penetrating delivery refers to the ability of a polypeptide sequence to carry itself or other cargo across the cell membrane into the cell. Antimicrobial membrane activity refers to the ability of a polypeptide sequence to exert its antimicrobial effect primarily by acting on the membranes of microorganisms (such as bacteria and fungi); common mechanisms include disrupting membrane integrity, forming pores, and interfering with membrane potential.

[0057] After retrieving peptide sequences from the peptide database using the aforementioned search keywords, duplicate sequences can be further removed, and the retained peptide sequences after deduplication can be used as the initial peptide sequence sample. The six search keywords proposed in this embodiment are all designed around the binding ability of peptide sequences to extracellular vesicle membranes, thus providing a relatively comprehensive coverage of membrane-binding peptide sequences.

[0058] In the above scheme, by setting multiple search keywords that are all related to the binding ability of the target extracellular vesicle membrane, and selecting peptide sequences whose keyword tags match one or more of the search keywords as initial peptide sequence samples, it is possible to obtain peptide sequences related to membrane binding in a targeted manner while ensuring the sufficiency of the dataset, thus ensuring the membrane binding ability of the final designed peptide.

[0059] Then, each initial polypeptide sequence sample is screened according to at least one screening dimension to obtain each candidate polypeptide sequence sample.

[0060] The screening dimension can be one or more. At least one screening dimension includes sequence length, overall charge, and / or secondary structure.

[0061] The sequence length screening dimension can be considered as using the number of amino acid residues contained in the initial polypeptide sequence sample as the screening criterion. The overall charge screening dimension can be considered as using the overall net charge of the initial polypeptide sequence sample as the screening criterion. The secondary structure screening dimension can be considered as using the types of secondary structures that the initial polypeptide sequence sample can form as the screening criterion.

[0062] In some application scenarios, the method of screening each initial polypeptide sequence sample according to at least one screening dimension to obtain each candidate polypeptide sequence sample can be as follows: for each initial polypeptide sequence sample, examine it one by one according to at least one screening dimension, and retain the initial polypeptide sequence samples that simultaneously meet the screening conditions corresponding to each screening dimension as candidate polypeptide sequence samples.

[0063] For example, the following three screening dimensions can be used to screen each initial polypeptide sequence sample, and only the initial polypeptide sequence sample that meets the following conditions can be retained as candidate polypeptide sequence samples: (1) the sequence length is between 10 and 25; (2) the overall charge is positive; (3) the secondary structure is α-helix.

[0064] After the above screening, multiple (e.g., over ten thousand) candidate polypeptide sequence samples can be obtained from tens of thousands of initial polypeptide sequence samples. In this embodiment, the α-helix is ​​used as a screening condition for the secondary structure dimension because polypeptide sequences capable of forming α-helices typically have their hydrophobic and hydrophilic surfaces arranged on different sides of the helix, facilitating insertion into the lipid bilayer on one side and interaction with the aqueous environment on the other, thereby stably binding to the extracellular vesicle membrane. Using a positive overall charge as a screening condition is based on the consideration that positively charged polypeptide sequences are easily adsorbed onto the surface of the extracellular vesicle membrane through electrostatic interactions. Through the above screening dimensions, it can be preliminarily ensured that the candidate polypeptide sequence samples have good membrane-binding potential.

[0065] In the above scheme, by first selecting initial peptide sequence samples from the peptide database, and then screening the initial peptide sequence samples according to screening dimensions such as sequence length, overall charge and / or secondary structure to obtain candidate peptide sequence samples, peptide sequences that obviously do not have membrane binding potential can be eliminated before obtaining sequence features. This is beneficial to improving the training efficiency of the binding force prediction model and the prediction accuracy of the final binding force prediction model.

[0066] In some embodiments, the method of obtaining the sequence characteristics of each candidate polypeptide sequence sample includes: first, determining the category to which each amino acid in the candidate polypeptide sequence sample belongs, such as cationic, hydrophobic, aromatic, acidic, nucleophilic, or amide; and then, determining the proportion of each amino acid and the frequency of occurrence of each adjacent residue pairing group according to the category to which each amino acid belongs.

[0067] In order to classify amino acids based on the interaction mechanism between polypeptide sequences and extracellular vesicle membranes, this embodiment divides 20 standard amino acids into six functional categories: cationic, hydrophobic, aromatic, acidic, nucleophilic, and amide / flexible. That is, each standard amino acid is classified into one of these six functional categories.

[0068] This embodiment mainly considers that these six functional categories can better correspond to the different interaction modes involved in the interaction between peptides and membranes. For example, cationic amino acids mainly interact with negatively charged membrane surfaces, hydrophobic amino acids are mainly responsible for inserting into the lipid bilayer, and aromatic amino acids are often distributed in the interfacial region of the membrane. Therefore, amino acids are classified according to the above six functional categories.

[0069] In some application scenarios, determining the category of each amino acid in a candidate polypeptide sequence sample can be achieved by: analyzing the candidate polypeptide sequence sample position by position, and determining the functional category of each amino acid in the candidate polypeptide sequence sample based on the pre-established correspondence between amino acids and the six major functional categories. That is, in this embodiment, each amino acid in the aforementioned steps can be understood as each of the six major functional categories, i.e., the proportion of amino acids and the pairing of adjacent residues are statistically analyzed on a functional category basis.

[0070] In some application scenarios, determining the proportion of each amino acid and the frequency of each adjacent residue pairing based on the category to which each amino acid belongs can be achieved as follows: After determining the functional category to which each amino acid in the candidate polypeptide sequence sample belongs, the number of amino acids belonging to each functional category in the candidate polypeptide sequence sample is counted, and their proportion in the candidate polypeptide sequence sample is calculated to obtain the proportion of each amino acid (i.e., the proportion of amino acids in each functional category). In addition, adjacent amino acids in the candidate polypeptide sequence sample are paired according to their respective functional categories, and the frequency of each adjacent residue pairing composed of functional categories is counted.

[0071] For example, since amino acids are divided into six major functional categories, there are a total of 21 types of adjacent residue pairings that can be formed by pairing them according to functional categories. For example, there are 21 types of adjacent residue pairings such as cationic-hydrophobic and cationic-aromatic. The frequency of occurrence of these 21 types of adjacent residue pairings in candidate polypeptide sequence samples can be counted.

[0072] In addition, the frequency of amino acid letters at each position in each candidate polypeptide sequence sample can be statistically analyzed, and a sequence frequency stacked plot can be drawn accordingly. The horizontal axis of the sequence frequency stacked plot represents the sequence position in the candidate polypeptide sequence sample, and the vertical axis represents the information content, measured in bits (character height). A higher value indicates a more homogeneous and conserved amino acid composition at that position in different candidate polypeptide sequence samples; a lower value indicates a more random and diverse amino acid composition at that position. Different letters in the sequence frequency stacked plot represent different amino acids, and colors represent different categories, such as green for hydrophobic amino acids.

[0073] In the above scheme, by classifying amino acids into six functional categories—cationic, hydrophobic, aromatic, acidic, nucleophilic, and amide—and determining the proportion of amino acids and the frequency of adjacent residue pairings by functional category, the scheme can reduce the dimensionality of the peptide sequence by analyzing the compositional rules of the peptide sequence, based on the mechanism of peptide-membrane interaction. This not only reduces the dimensionality and statistical complexity of sequence features but also highlights the functional attributes closely related to membrane binding capacity, thus providing more targeted input features for the binding force prediction model.

[0074] In some embodiments, step 102 above may include the following steps: First, based on the frequency of occurrence of each adjacent residue pair in each candidate polypeptide sequence sample, the overall proportion of each amino acid in multiple candidate polypeptide sequence samples is determined.

[0075] Among them, the frequency of occurrence of each adjacent residue pair in each candidate polypeptide sequence sample can be used to summarize and count the adjacent residue pairs, and then based on the amino acid composition of the adjacent residue pairs, the overall proportion of each amino acid in the entire candidate polypeptide sequence sample can be deduced.

[0076] In some embodiments, determining the overall proportion of each amino acid in multiple candidate polypeptide sequence samples based on the frequency of occurrence of each adjacent residue pair in each candidate polypeptide sequence sample can be achieved in the following ways: First, the total frequency of each adjacent residue pairing group was counted in all candidate polypeptide sequence samples.

[0077] The total frequency of occurrence can be considered as the sum of the number of times a certain adjacent residue pair appears in all candidate polypeptide sequence samples. For example, the occurrence frequencies of the same adjacent residue pair in each candidate polypeptide sequence sample can be summed to obtain the total frequency of occurrence of that adjacent residue pair in all candidate polypeptide sequence samples.

[0078] Secondly, the top number of adjacent residue pairs with the highest total frequency of occurrence are selected to obtain multiple target adjacent residue pairs.

[0079] The first quantity can be considered as a threshold for screening adjacent residue pairs, and its value can be set according to actual needs. The target adjacent residue pair can be considered as the adjacent residue pair with the highest total frequency selected from all adjacent residue pairs, reflecting the dominant adjacent residue pairing pattern in all candidate polypeptide sequence samples.

[0080] In some application scenarios, selecting the top number of adjacent residue pairs with the highest total frequency to obtain multiple target adjacent residue pairs can be achieved by: arranging all adjacent residue pairs in descending order of total frequency, and then selecting the top number of adjacent residue pairs from the top of the list. For example, after arranging all adjacent residue pairs in descending order of total frequency, the top few pairs (i.e., the top number of pairs) whose total frequency accounts for approximately 50% of the total frequency of all adjacent residue pairs can be selected as target adjacent residue pairs. In other words, the selected target adjacent residue pairs are the dominant pairings. Testing has shown that the top number of pairs can be the top 5.

[0081] Then, the first frequency of each amino acid in all target adjacent residue pairs is determined, and the sum of the first frequencies of all amino acids is taken as the second frequency.

[0082] The first frequency can be considered as the frequency of a certain amino acid appearing in all target adjacent residue pairs. For example, for a certain amino acid, the total frequency of that amino acid appearing as a member of all target adjacent residue pairs can be counted as the first frequency of that amino acid. Taking a certain amino acid as an example, its first frequency can be expressed as the sum of the occurrence frequencies of the target adjacent residue pairs formed by that amino acid and various other amino acids, that is, the frequencies of that amino acid appearing in the position before and after the pair are added together.

[0083] The second frequency can be considered as the sum of the first frequencies of all types of amino acids, which reflects the total amount of amino acids contained in all target adjacent residue pairs.

[0084] The total frequency of occurrence for each adjacent residue pairing can be found in [reference needed]. Figure 2 The statistics are presented in the format shown. Figure 2 The chart shows the total frequency of various adjacent residue pairs. The horizontal axis represents the frequency of occurrence, and the vertical axis represents various adjacent residue pairs, for example... Figure 2 The middle vertical axis shows 11 pairs of adjacent residues: Cationic-Hydrophobic, Hydrophobic-Nucleophilic, Cationic-Nucleophilic, Aromatic-Cationic, Acidic-Hydrophobic, Acidic-Cationic, Nucleophilic-Nucleophilic, Aromatic-Aromatic, Acidic-Nucleophilic, Acidic-Acidic, and Acidic-Amide.

[0085] For each amino acid, the ratio between the first frequency and the second frequency is taken as the overall proportion of the amino acid.

[0086] In other words, for each amino acid, the first frequency of that amino acid can be divided by the second frequency, and the resulting ratio is the overall proportion of that amino acid in multiple candidate polypeptide sequence samples.

[0087] For example, the overall proportions of the various amino acids obtained can be found in [reference needed]. Figure 3 The statistics are presented in the format shown. Figure 3 The statistics show the percentage of various amino acids in each candidate polypeptide sequence sample, such as Cationic (16.87%), Hydrophobic (58.55%), Aromatic (8.28%), Acidic (2.34%), Nucleophilic (10.49%), and Amide (3.48%).

[0088] In the above scheme, by statistically analyzing the total frequency of various adjacent residue pairings and selecting the top number of dominant adjacent residue pairings as target adjacent residue pairings, and then determining the overall proportion of each amino acid based on the ratio between the first and second frequencies in the target adjacent residue pairings, it is possible to focus on the dominant pairing combinations in all candidate polypeptide sequence samples to characterize the overall distribution pattern of amino acids. This reduces the statistical complexity and highlights the compositional characteristics of amino acids that play a dominant role in membrane binding.

[0089] Secondly, based on the difference between the proportion of each amino acid in each candidate polypeptide sequence sample and the corresponding overall proportion, multiple candidate polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples and used as target polypeptide sequence samples respectively.

[0090] Here, the difference can be considered as the degree of deviation between the proportion of a certain amino acid in a single candidate polypeptide sequence sample and the overall proportion of that amino acid. For example, it can be measured by the difference between the two, the absolute difference, or the relative deviation.

[0091] For example, for each candidate polypeptide sequence sample, the difference between the proportion of each amino acid in the candidate polypeptide sequence sample and the overall proportion of that amino acid is compared one by one. If the overall difference between the proportion of each amino acid in the candidate polypeptide sequence sample and the corresponding overall proportion is small, it indicates that the amino acid composition of the candidate polypeptide sequence sample is relatively close to the overall distribution of all candidate polypeptide sequences and is highly representative, so it can be selected as the target polypeptide sequence sample. Conversely, if the overall difference between the proportion of each amino acid in a candidate polypeptide sequence sample and the corresponding overall proportion is large, it indicates that the amino acid composition of the candidate polypeptide sequence sample deviates far from the overall distribution and can be eliminated. Alternatively, a criterion can be set to measure whether the difference is acceptable, selecting candidate polypeptide sequence samples where the differences between the proportions of each amino acid and the corresponding overall proportion are all within an acceptable range (e.g., the difference does not exceed 20%) as the target polypeptide sequence samples.

[0092] In some embodiments, the above-mentioned method of selecting multiple candidate polypeptide sequence samples from multiple candidate polypeptide sequence samples based on the difference between the proportion of each amino acid in each candidate polypeptide sequence sample and the corresponding overall proportion, and using them as target polypeptide sequence samples, can be implemented as follows: First, for each candidate polypeptide sequence sample, the difference between the proportion of each amino acid in the candidate polypeptide sequence sample and the corresponding overall proportion is obtained and the differences are statistically analyzed to obtain the total difference of the candidate polypeptide sequence samples.

[0093] The total difference value can be considered as the sum of the differences between the proportion of all types of amino acids in a candidate polypeptide sequence sample and the corresponding overall proportion. It is used to measure the degree to which the amino acid composition of the candidate polypeptide sequence sample deviates from the overall distribution of all candidate polypeptide sequence samples.

[0094] For example, for each candidate polypeptide sequence sample, the difference between the proportion of each amino acid in the candidate polypeptide sequence sample and the overall proportion corresponding to that amino acid is calculated one by one. Then, the differences corresponding to each amino acid in the candidate polypeptide sequence sample are statistically analyzed to obtain the total difference value of the candidate polypeptide sequence sample. The statistical method can be to sum the differences corresponding to each amino acid, or to calculate the average or weighted sum, etc., and the present invention does not limit this.

[0095] For example, if a candidate polypeptide sequence sample involves six amino acids, the difference between the proportion of each of the six amino acids and its corresponding overall proportion can be calculated. The differences corresponding to the six amino acids are then summed, and the sum is taken as the total difference value of the candidate polypeptide sequence sample. Alternatively, the maximum value, average value, or other statistical values ​​of the differences corresponding to the six amino acids can be used as the total difference value.

[0096] Then, candidate polypeptide sequence samples with a total difference value less than or equal to a preset difference value are selected as target polypeptide sequence samples.

[0097] The preset difference value can be a difference threshold used to determine whether a candidate polypeptide sequence sample is selected as the target polypeptide sequence sample. Its value can be set according to actual needs, such as 20%.

[0098] For example, the total difference value of each candidate polypeptide sequence sample is compared with a preset difference value. If the total difference value of a candidate polypeptide sequence sample is less than or equal to the preset difference value, it indicates that the amino acid composition of the candidate polypeptide sequence sample is relatively close to the overall distribution of all candidate polypeptide sequence samples, and it is selected as the target polypeptide sequence sample. Conversely, if the total difference value of a candidate polypeptide sequence sample is greater than the preset difference value, it is eliminated. That is, candidate polypeptide sequence samples whose differences between the proportion of each amino acid and the corresponding overall proportion do not exceed 20% can be selected as target polypeptide sequence samples.

[0099] In the above scheme, the total difference value is obtained by statistically analyzing the differences between the proportion of various amino acids in a single candidate polypeptide sequence sample and the corresponding overall proportion. The target polypeptide sequence sample is selected by using the relationship between the total difference value and the preset difference value as the screening criterion. This allows for the screening of candidate polypeptide sequence samples in a quantitative and controllable manner, thereby efficiently and accurately eliminating candidate polypeptide sequence samples whose composition deviates significantly from the overall distribution. This further ensures the consistency of the selected target polypeptide sequence samples in terms of amino acid composition.

[0100] In the above scheme, the overall proportion of each amino acid is determined by the frequency of occurrence of adjacent residue pairing groups, and the target polypeptide sequence sample is selected based on the difference between the proportion of amino acids in a single candidate polypeptide sequence sample and the corresponding overall proportion. This can eliminate candidate polypeptide sequence samples whose amino acid composition deviates significantly from the overall distribution, making the selected target polypeptide sequence samples closer to the overall distribution pattern in terms of amino acid composition. This ensures the consistency and representativeness of the standard dataset used for subsequent training in terms of sequence composition, which is beneficial to improving the training effect of the binding force prediction model.

[0101] In some embodiments, step 104 may include the following steps: First, based on the binding forces simulated for each target polypeptide sequence sample, select the second-highest number of target polypeptide sequence samples with the highest binding forces and add them to the training dataset. Then, use each target polypeptide sequence sample in the training dataset and its binding force label to train the binding force prediction model to obtain the final binding force prediction model.

[0102] The second quantity is a threshold for selecting the number of target polypeptide sequence samples, and its value can be set according to actual needs. The training dataset is a set of samples used to train the binding force prediction model.

[0103] For example, the target peptide sequence samples are sorted in descending order of their simulated binding strength from highest to lowest. Then, the second-highest number of target peptide sequences with the highest binding strength are selected, and these selected sequences and their simulated binding strengths are added to the training dataset. For instance, the training dataset can be constructed by selecting the top 900 target peptide sequences with the highest binding strength from all target peptide sequence samples. These 900 sequences can be split into a training dataset and a validation dataset in a 9:1 ratio. The binding strength prediction model is then trained using the split training dataset, while the validation dataset is used to evaluate the model's generalization performance and prediction accuracy in real time. The sequence features of these 900 target peptide sequences can be presented using dimensionality reduction analysis. Figure 4 This shows the two-dimensional distribution results obtained from dimensionality reduction analysis of 900 target polypeptide sequence samples. Figure 4 It can intuitively display the distribution of the selected target polypeptide sequence sample in the sequence composition space, specifically, Figure 4 The horizontal and vertical axes represent the coordinates of each target polypeptide sequence sample in two-dimensional space (e.g., t-SNE 1 and t-SNE 2). Each point represents a target polypeptide sequence sample, and the color of the point indicates the binding affinity of that target polypeptide sequence sample, with units of kcal / mol. For example... Figure 4 The closer the color is to yellow, the lower the binding force; the closer it is to dark blue, the higher the binding force.

[0104] In addition, the distribution of binding forces obtained from simulations of the 900 selected target polypeptide sequence samples can be found in [reference needed]. Figure 5 As shown. Figure 5 The horizontal axis represents binding force, measured in kcal / mol, and the vertical axis represents the number of target polypeptide sequence samples. A more negative binding force indicates a stronger binding force. For example... Figure 5 As shown, the peak values ​​are approximately -8.0 to -7.5 kcal / mol, and the overall distribution is relatively concentrated, without obvious multi-peak or long-tailed dispersion. This is consistent with... Figure 4 The target polypeptide sequence samples form relatively compact clusters in two-dimensional space, and the high-binding samples (dark blue dots) show a consistent trend.

[0105] The method of training the binding force prediction model using the target peptide sequence samples and their binding force labels in the training dataset to obtain the final binding force prediction model can be as follows: take the sequence features of each target peptide sequence sample in the training dataset as the input of the binding force prediction model, take the binding force simulated by the target peptide sequence sample as the binding force label, and iteratively train the binding force prediction model until the training stopping condition is met to obtain the final binding force prediction model.

[0106] In the above scheme, by simulating the binding force based on each target polypeptide sequence sample, the second-highest number of target polypeptide sequence samples with the highest binding force are selected and added to the training dataset. The binding force prediction model is then trained using this training dataset. This allows the target polypeptide sequence samples with strong membrane binding ability to be the focus of training, enabling the binding force prediction model to learn more fully the sequence determinants of high-affinity polypeptide sequences, thereby improving the prediction accuracy of the binding force prediction model for high-affinity polypeptide sequences.

[0107] In some embodiments, the number of cohesion prediction models to be trained is multiple, and the different cohesion prediction models are of different types. For example, the cohesion prediction models to be trained may include at least two of the following: Extreme Random Tree Regression (ETR), Random Forest Regression (RFR), Extreme Gradient Boosting (XGB), Support Vector Regression (SVR), Kernel Ridge Regression (KRR), Ridge Regression, Lasso Regression with Minimum Absolute Shrinkage and Selection Operator (Lasso), Gaussian Process Regression (GPR), and K Nearest Neighbor Regression (KNN).

[0108] The above-mentioned method of training the binding force prediction model using target peptide sequence samples and their binding force labels in the training dataset can include the following steps: Using the target peptide sequence samples and their binding force labels in the training dataset, each binding force prediction model was iteratively trained.

[0109] For example, at least a portion of the sequence features of each target peptide sequence sample in the training dataset are used as input to each binding force prediction model, and the simulated binding force of the target peptide sequence sample is used as the binding force label. The nine binding force prediction models are iteratively trained respectively. The at least a portion of the sequence features can be feature vectors mapped from the proportions of various amino acids in the target peptide sequence sample. That is, the input to the binding force prediction model is the feature vector mapped from the proportions of various amino acids in each target peptide sequence sample, and the output is the predicted binding force of each target peptide sequence sample. Then, in each round of iterative training, the target loss is determined by the difference between the predicted binding force and the binding force label. The parameters in the binding force prediction model to be trained in this round are then adjusted according to the target loss to obtain the binding force prediction model to be trained in the next round of iterative training. For example, the type of target loss can be regression loss such as mean squared error or mean absolute error.

[0110] After the iteration stopping condition is met, each bonding force prediction model is tested, and based on the test results of each bonding force prediction model, at least one bonding force prediction model with better test results is selected as the final bonding force prediction model.

[0111] The iteration stopping condition is used to determine whether the iterative training of the binding force prediction model should terminate. This condition may include, but is not limited to, reaching a preset number of iterations or the loss function converging. In each round of iterative training, a validation set can be used to validate the parameter-adjusted binding force prediction model.

[0112] In some application scenarios, after each binding force prediction model satisfies the iteration stopping condition, the test data is used to test each binding force prediction model to obtain the test results of each binding force prediction model. Then, based on the test results of each binding force prediction model, at least one binding force prediction model with better test performance is selected from multiple binding force prediction models as the final binding force prediction model. For example, the test results may include evaluation indicators such as prediction accuracy and goodness of fit of each binding force prediction model.

[0113] Selecting at least one of the better-performing binding force prediction models as the final binding force prediction model can be done by starting with at least one preset evaluation dimension and selecting one or more binding force prediction models with better test results as the final binding force prediction model. For example, the evaluation dimension can be the coefficient of determination dimension, the root mean square error dimension, or the prediction bias shown by the scatter distribution of the predicted and actual values. For example, first determine the top n optimal binding force prediction models under each evaluation dimension, and then select the final binding force prediction model based on the weight of each evaluation dimension.

[0114] For example, the nine associative force prediction models can be tested and their test results compared. For instance, the prediction bias shown by the coefficient of determination, root mean square error, and scatter distribution of predicted and true values ​​of each associative force prediction model in the validation set can be compared to determine that the extreme gradient boosting (XGB) and random forest regression (RFR) as the two associative force prediction models have the best performance among the nine associative force prediction models. Therefore, the extreme gradient boosting (XGB) and random forest regression (RFR) as the two associative force prediction models can be selected as the final associative force prediction models.

[0115] In the above scheme, by preparing multiple different types of binding force prediction models and iteratively training them, and then selecting at least one of them as the final binding force prediction model based on the test results of each binding force prediction model, it is possible to fully compare the prediction capabilities of different types of binding force prediction models for the binding force between peptide sequences and extracellular vesicle membranes, thereby selecting the best binding force prediction model and further improving the accuracy and reliability of peptide membrane binding ability prediction.

[0116] Please see Figure 6 The present invention provides a polypeptide sequence design method comprising the following steps: Step 201: Obtain multiple peptide sequences to be evaluated.

[0117] Among them, the polypeptide sequence to be evaluated is the polypeptide sequence whose binding force with the extracellular vesicle membrane needs to be assessed.

[0118] Step 202: Use the binding force prediction model to evaluate multiple peptide sequences to be evaluated, and obtain the predicted binding force between each peptide sequence to be evaluated and the extracellular vesicle membrane.

[0119] Among them, the bonding force prediction model is the final bonding force prediction model trained based on the above bonding force prediction model training method.

[0120] For example, the proportions of various amino acids in the sequence features of each peptide sequence to be evaluated are used as input to the binding affinity prediction model. The meaning, composition, and acquisition method of the sequence features of the peptide sequence to be evaluated can be found in the relevant descriptions of sequence features in the training method of the aforementioned binding affinity prediction model. For instance, the proportions of various amino acids included in the sequence features can be obtained by statistically analyzing them in units of six functional categories: cationic, hydrophobic, aromatic, acidic, nucleophilic, and amide. This will not be elaborated further here. Then, the binding affinity prediction model processes the proportions of various amino acids in each peptide sequence to be evaluated and outputs the predicted binding affinity between each peptide sequence and the extracellular vesicle membrane.

[0121] Step 203: Based on each predicted binding affinity, select the target polypeptide sequence from multiple polypeptide sequences to be evaluated.

[0122] The number of target peptide sequences can be one or more. For example, the predicted binding forces of each peptide sequence to be evaluated are sorted by magnitude, and one or more peptide sequences with higher predicted binding forces are selected as target peptide sequences.

[0123] In the above scheme, the binding force prediction model trained by the above binding force prediction model is more accurate in predicting the binding force of each peptide sequence to be evaluated, thereby improving the separation efficiency of extracellular vesicle membrane.

[0124] In some embodiments, before obtaining multiple peptide sequences to be evaluated, the method further includes: using multiple reference sequence features as input to a sequence generation model, and using the multiple peptide sequences output by the sequence generation model as each peptide sequence to be evaluated; wherein, the sequence generation model is trained based on sequence feature samples and their corresponding peptide sequences as labels.

[0125] The sequence generation model can be considered a machine learning model capable of generating corresponding peptide sequences based on input sequence features. For example, the architecture of the sequence generation model can be a generative model based on recurrent neural networks (such as Long Short-Term Memory networks (LSTM) or gated recurrent units (GRUs), a generative model based on autoregressive transformers (such as conditional Transformers), or a variational autoencoder (such as conditional VAEs). Reference sequence features can be considered as sequence features used as input to the sequence generation model to guide the generation of peptide sequences. Sequence feature samples can be considered as sequence features used to train the sequence generation model. In some applications, reference sequence features can be variant features obtained by perturbation, interpolation, or derivation based on the sequence features of the aforementioned target peptide sequence samples (e.g., the sequence features of the Top900 targets), to explore diverse sequences while maintaining compositional rationality. In other scenarios, reference sequence features can also be obtained by extracting sequence features from peptide sequences sampled from the aforementioned peptide database, determining features based on desired improvements to the acquisition function, or constructing features according to design requirements. The sources of reference sequence features are not elaborated here. In addition, the reference sequence features and sequence feature samples may include the proportions of various amino acids, or may further include the frequency of occurrence of adjacent residue pairings.

[0126] The sequence generation model is trained based on sequence feature samples and their corresponding peptide sequences as labels. That is, when training the sequence generation model, the sequence feature samples are used as input and the peptide sequences corresponding to the sequence feature samples are used as labels to train the sequence generation model. This allows the trained sequence generation model to learn the mapping relationship between sequence features and peptide sequences, thereby generating the corresponding peptide sequences based on the input sequence features.

[0127] In some application scenarios, the above steps can be implemented as follows: input multiple reference sequence features into the trained sequence generation model, generate corresponding peptide sequences based on each reference sequence feature, and use the multiple peptide sequences output by the sequence generation model as each peptide sequence to be evaluated.

[0128] For example, corresponding sequence generation models can be constructed for the two binding affinity prediction models selected above: Extreme Gradient Boosting (XGB) and Random Forest Regression (RFR). Each sequence generation model can then generate a large number of peptide sequences to be evaluated (e.g., 5000 sequences each), providing sufficient peptide sequences for subsequent binding affinity assessment and target peptide sequence screening. In this approach, the sequence generation model is pre-trained based on sequence feature samples and their corresponding peptide sequences. Multiple reference sequence features are input into the sequence generation model to generate multiple peptide sequences as peptide sequences to be evaluated. This allows for the direct and targeted generation of novel peptide sequences based on sequence features, thus overcoming the constraints of existing peptide libraries. This provides a rich and diverse range of peptide sequences to be evaluated for subsequent binding affinity prediction and target peptide sequence screening, further improving the design efficiency of membrane-bound peptides and the membrane binding ability of the designed peptide sequences.

[0129] In some embodiments, the number of binding force prediction models is at least two, and a target polypeptide sequence is selected from multiple polypeptide sequences to be evaluated based on each predicted binding force, including: First, for each predicted binding force output by the binding force prediction model, based on the magnitude of each predicted binding force, the peptide sequences corresponding to the top three highest predicted binding forces are selected as candidate peptide sequences corresponding to the binding force prediction model.

[0130] For example, for each binding force prediction model, the predicted binding forces of each peptide sequence to be evaluated output by the binding force prediction model are sorted in descending order, and the third highest number of peptide sequences to be evaluated are selected as the candidate peptide sequences corresponding to the binding force prediction model.

[0131] The third quantity is a threshold number used to select the polypeptide sequence with the highest predicted binding affinity, and its value can be set according to actual needs.

[0132] For example, if the binding affinity prediction models are the aforementioned Extreme Gradient Boosting (XGB) and Random Forest Regression (RFR) models, the selection operation described above can be performed on each of these two models. For instance, the two models can be used to evaluate more than 5,000 peptide sequences to be evaluated, and the predicted binding affinity of each sequence can be output. Then, for each model, the top 6 peptide sequences with the highest predicted binding affinity are selected as candidate sequences for that model. In this way, a total of 12 candidate peptide sequences are obtained from the two models.

[0133] In some embodiments, the binding force prediction model can also output the uncertainty of each predicted binding force. When determining candidate peptide sequences, the expected improvement (EI) acquisition function can be used to determine the expected improvement value based on the predicted binding force and the corresponding uncertainty of each peptide sequence to be evaluated. Then, based on the level of the expected improvement value, the top three orders of peptide sequences to be evaluated with the highest value are selected as candidate peptide sequences.

[0134] Secondly, the relative signal intensity between each candidate polypeptide sequence and the extracellular vesicle membrane was verified to obtain the relative signal intensity corresponding to each candidate polypeptide sequence.

[0135] For example, candidate polypeptide sequences of the entity are synthesized separately, and then the relative signal intensity between each synthesized candidate polypeptide sequence and the extracellular vesicle is examined using a dot blot experiment, thereby obtaining the relative signal intensity between each candidate polypeptide sequence and the extracellular vesicle membrane. The relative signal intensity can reflect the actual binding ability between each candidate polypeptide sequence and the extracellular vesicle membrane.

[0136] Then, based on the relative signal intensity corresponding to each candidate polypeptide sequence, at least one candidate polypeptide sequence is selected as the target polypeptide sequence.

[0137] In some embodiments, selecting at least one candidate polypeptide sequence as the target polypeptide sequence based on the relative signal intensity corresponding to each candidate polypeptide sequence may include the following steps: comparing each relative signal intensity with the relative signal intensity between a known polypeptide sequence and an extracellular vesicle membrane; and obtaining the target polypeptide sequence based on the comparison results.

[0138] Among them, the known polypeptide sequence can be considered as a polypeptide sequence that has been reported and can target the extracellular vesicle membrane, such as a lipid-targeting polypeptide ligand reported in the literature.

[0139] For example, multiple experiments can be conducted on candidate peptide sequences and reported peptide sequences for each entity, and the results can be statistically analyzed. For instance, the relative signal intensity generated on a DotBlot membrane after each peptide sequence binds to the extracellular vesicle membrane can be calculated, and then the average value and the standard deviation of the difference between the relative signal intensity obtained in each experiment and the average value can be calculated. To further verify the statistical significance of the differences, statistical hypothesis testing can be used to calculate the p-value, and the significance level can be marked with an asterisk in the results. This allows for a quantitative assessment of the differences in the binding ability of each candidate peptide sequence to the extracellular vesicle membrane, screening out the optimal peptide with significantly superior binding affinity compared to other sequences.

[0140] In the above scheme, known peptide sequences are used as a reference to screen target peptide sequences, thereby ensuring that the selected target peptide sequences have better membrane binding ability than conventional random screening or reported peptide sequences, thus achieving rapid, unbiased identification and efficient separation of extracellular vesicles.

[0141] The training apparatus for the bonding force prediction model provided by this invention will be described below. The training apparatus for the bonding force prediction model described below can be referred to in correspondence with the training method for the bonding force prediction model described above. Figure 7 As shown, the training device 300 for the combined force prediction model may include the following modules: The first feature acquisition module 301 is used to acquire the sequence features of multiple candidate polypeptide sequence samples. The sequence features include the proportion of various amino acids and the frequency of occurrence of each adjacent residue pairing. The sample screening module 302 is used to select multiple target polypeptide sequence samples from multiple candidate polypeptide sequence samples based on the proportion of various amino acids contained in each sequence feature and the frequency of occurrence of each adjacent residue pairing group. The binding force label determination module 303 is used to simulate the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane; Training module 304 is used to train the binding force prediction model using at least a portion of the target polypeptide sequence samples and the binding forces obtained by simulation of each as binding force labels, so as to obtain the final binding force prediction model.

[0142] According to the training device 300 of the binding force prediction model provided by the present invention, the sample screening module 302 selects multiple target polypeptide sequence samples from multiple candidate polypeptide sequence samples based on the proportion of various amino acids contained in each sequence feature and the frequency of occurrence of each adjacent residue pairing group, including: The overall proportion of each amino acid in multiple candidate polypeptide sequence samples is determined based on the frequency of occurrence of each adjacent residue pairing in each candidate polypeptide sequence sample. Based on the difference between the proportion of each amino acid in each candidate polypeptide sequence sample and the corresponding overall proportion, multiple candidate polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples and used as target polypeptide sequence samples.

[0143] According to the training apparatus 300 of the binding force prediction model provided by the present invention, the sample screening module 302 determines the overall proportion of each amino acid in multiple candidate polypeptide sequence samples based on the occurrence frequency of each adjacent residue pairing group in each candidate polypeptide sequence sample, including: The total frequency of each adjacent residue pairing group was counted in all candidate polypeptide sequence samples. Select the first number of adjacent residue pairs with the highest total frequency of each pair to obtain multiple target adjacent residue pairs. Determine the first frequency of each amino acid in all target adjacent residue pairs and use the sum of the first frequencies of all amino acids as the second frequency; For each amino acid, the ratio between the first frequency and the second frequency is taken as the overall proportion of the amino acid.

[0144] According to the training device 300 of the binding force prediction model provided by the present invention, the sample screening module 302 selects multiple candidate polypeptide sequence samples from multiple candidate polypeptide sequence samples based on the difference between the proportion of each amino acid in each candidate polypeptide sequence sample and the corresponding overall proportion, and uses them as target polypeptide sequence samples respectively, including: For each candidate polypeptide sequence sample, the difference between the proportion of each amino acid in the candidate polypeptide sequence sample and the corresponding overall proportion is obtained and the differences are statistically analyzed to obtain the total difference value of the candidate polypeptide sequence sample. Candidate polypeptide sequence samples with a total difference value less than or equal to a preset difference value are selected as target polypeptide sequence samples.

[0145] Before acquiring the sequence features of multiple candidate polypeptide sequence samples, the first feature acquisition module 301 of the training apparatus 300 for a binding force prediction model provided by the present invention is further configured to: Multiple initial polypeptide sequence samples were selected from the polypeptide database; Each initial polypeptide sequence sample is screened according to at least one screening dimension to obtain each candidate polypeptide sequence sample. The at least one screening dimension includes sequence length, overall charge and / or secondary structure.

[0146] According to the training apparatus 300 of the binding force prediction model provided by the present invention, the first feature acquisition module 301 selects multiple initial polypeptide sequence samples from a polypeptide database, including: Based on multiple search keywords, initial polypeptide sequence samples were selected from the polypeptide database. Each search keyword was associated with the binding ability to target extracellular vesicle membranes. In this process, the keyword tag of each initial polypeptide sequence sample is matched with one or more membrane binding search keywords.

[0147] According to the training apparatus 300 for a binding force prediction model provided by the present invention, the first feature acquisition module 301 acquires the sequence features of each candidate polypeptide sequence sample in the following manner: Determine the category of each amino acid in the candidate polypeptide sequence sample, which can be cationic, hydrophobic, aromatic, acidic, nucleophilic, or amide. Based on the category to which each amino acid belongs, determine the proportion of each amino acid and the frequency of occurrence of each pair of adjacent residues.

[0148] According to the present invention, a training device 300 for a binding force prediction model includes a training module 304 that uses at least a portion of target polypeptide sequence samples and the respective simulated binding forces as binding force labels to train the binding force prediction model, thereby obtaining a final binding force prediction model, comprising: Based on the binding force simulated from each target polypeptide sequence sample, the second-highest number of target polypeptide sequence samples with the highest binding force are selected and added to the training dataset. The binding force prediction model is trained using the target peptide sequence samples and their binding force labels in the training dataset to obtain the final binding force prediction model.

[0149] According to the training apparatus 300 of the binding force prediction model provided by the present invention, there are multiple binding force prediction models to be trained, and the different binding force prediction models are of different types. The training module 304 uses the target polypeptide sequence samples and their binding force tags in the training dataset to train the binding force prediction model to obtain the final binding force prediction model, including: Using the target polypeptide sequence samples and their binding force labels in the training dataset, each binding force prediction model was iteratively trained. After the iteration stopping condition is met, each bonding force prediction model is tested, and based on the test results of each bonding force prediction model, at least one bonding force prediction model with better test results is selected as the final bonding force prediction model.

[0150] The polypeptide sequence design apparatus provided by this invention is described below. The polypeptide sequence design apparatus described below can be referred to in correspondence with the polypeptide sequence design method described above. For example... Figure 8 As shown, the polypeptide sequence design device 400 may include the following modules: The second feature acquisition module 401 is used to acquire multiple polypeptide sequences to be evaluated. The binding force assessment module 402 is used to assess multiple peptide sequences to be assessed using a binding force prediction model to obtain the predicted binding force between each peptide sequence to be assessed and the extracellular vesicle membrane. The binding force prediction model is the final binding force prediction model trained by the training device 400 based on any of the above-mentioned binding force prediction models. The sequence screening module 403 is used to select the target polypeptide sequence from multiple polypeptide sequences to be evaluated based on each predicted binding force.

[0151] According to the present invention, a polypeptide sequence design apparatus 400 includes at least two binding affinity prediction models. A sequence screening module 403 selects a target polypeptide sequence from multiple polypeptide sequences to be evaluated based on each predicted binding affinity, including: For each predicted binding force output by the binding force prediction model, based on the magnitude of each predicted binding force, the peptide sequences corresponding to the top three highest predicted binding forces are selected as candidate peptide sequences corresponding to the binding force prediction model. The relative signal intensity between each candidate polypeptide sequence and the extracellular vesicle membrane was verified to obtain the relative signal intensity corresponding to each candidate polypeptide sequence. Based on the relative signal intensity corresponding to each candidate polypeptide sequence, at least one candidate polypeptide sequence is selected as the target polypeptide sequence.

[0152] According to the present invention, a polypeptide sequence design apparatus 400 includes a sequence screening module 403 that selects at least one candidate polypeptide sequence as a target polypeptide sequence based on the relative signal intensity corresponding to each candidate polypeptide sequence, comprising: The relative signal intensities were compared with the relative signal intensities between known polypeptide sequences and extracellular vesicle membranes; Based on the alignment results, the target polypeptide sequence was obtained.

[0153] According to the polypeptide sequence design apparatus 400 provided by the present invention, before acquiring multiple polypeptide sequences to be evaluated, the second feature acquisition module 401 is further configured to: Multiple reference sequence features are used as input to the sequence generation model, and the multiple peptide sequences output by the sequence generation model are used as the peptide sequences to be evaluated. The sequence generation model is trained based on sequence feature samples and their corresponding polypeptide sequences as labels.

[0154] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a training method for the binding force prediction model, including the following steps: acquiring sequence features of multiple candidate polypeptide sequence samples, the sequence features including the proportion of various amino acids and the frequency of occurrence of each adjacent residue pairing; selecting multiple target polypeptide sequence samples from the multiple candidate polypeptide sequence samples based on the proportion of various amino acids and the frequency of occurrence of each adjacent residue pairing contained in each sequence feature; simulating the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane; using at least some of the target polypeptide sequence samples and the simulated binding force as binding force labels to train the binding force prediction model, thereby obtaining the final binding force prediction model. Alternatively, a peptide sequence design method may be implemented, including the following steps: obtaining multiple peptide sequences to be evaluated; evaluating the multiple peptide sequences to be evaluated using a binding force prediction model to obtain the predicted binding force between each peptide sequence to be evaluated and the extracellular vesicle membrane, wherein the binding force prediction model is trained based on any of the above-mentioned binding force prediction model training methods; and selecting a target peptide sequence from the multiple peptide sequences to be evaluated based on each predicted binding force.

[0155] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0156] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the training method of the binding force prediction model provided by the above methods, including the following steps: acquiring sequence features of multiple candidate polypeptide sequence samples, the sequence features including the proportion of various amino acids and the frequency of occurrence of each adjacent residue pair; selecting multiple target polypeptide sequence samples from the multiple candidate polypeptide sequence samples according to the proportion of various amino acids and the frequency of occurrence of each adjacent residue pair contained in each sequence feature; simulating the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane; using at least some of the target polypeptide sequence samples and the binding force obtained by each simulation as binding force labels to train the binding force prediction model to obtain the final binding force prediction model. Alternatively, a peptide sequence design method may be implemented, including the following steps: obtaining multiple peptide sequences to be evaluated; evaluating the multiple peptide sequences to be evaluated using a binding force prediction model to obtain the predicted binding force between each peptide sequence to be evaluated and the extracellular vesicle membrane, wherein the binding force prediction model is trained based on any of the above-mentioned binding force prediction model training methods; and selecting a target peptide sequence from the multiple peptide sequences to be evaluated based on each predicted binding force.

[0157] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a training method for the above-described binding force prediction model, comprising the following steps: acquiring sequence features of multiple candidate polypeptide sequence samples, the sequence features including the proportion of various amino acids and the frequency of occurrence of each adjacent residue pair; selecting multiple target polypeptide sequence samples from the multiple candidate polypeptide sequence samples based on the proportion of various amino acids and the frequency of occurrence of each adjacent residue pair contained in each sequence feature; simulating the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane; using at least some of the target polypeptide sequence samples and the simulated binding force of each as binding force labels to train the binding force prediction model, thereby obtaining the final binding force prediction model. Alternatively, a peptide sequence design method may be implemented, including the following steps: obtaining multiple peptide sequences to be evaluated; evaluating the multiple peptide sequences to be evaluated using a binding force prediction model to obtain the predicted binding force between each peptide sequence to be evaluated and the extracellular vesicle membrane, wherein the binding force prediction model is trained based on any of the above-mentioned binding force prediction model training methods; and selecting a target peptide sequence from the multiple peptide sequences to be evaluated based on each predicted binding force.

[0158] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0159] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A training method for a combined force prediction model, characterized in that, include: The sequence features of multiple candidate polypeptide sequence samples are obtained, including the proportion of various amino acids and the frequency of occurrence of each adjacent residue pairing; Based on the proportion of various amino acids contained in each of the aforementioned sequence features and the frequency of occurrence of each of the aforementioned adjacent residue pairings, multiple target polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples; Simulate the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane; Using at least a portion of the target polypeptide sequence samples and the respective simulated binding forces as binding force labels, the binding force prediction model is trained to obtain the final binding force prediction model.

2. The training method for the binding force prediction model according to claim 1, characterized in that, Based on the proportions of various amino acids contained in each of the aforementioned sequence features and the frequency of occurrence of each of the aforementioned adjacent residue pairings, multiple target polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples, including: Based on the frequency of occurrence of each adjacent residue pairing group in each candidate polypeptide sequence sample, the overall proportion of each amino acid in multiple candidate polypeptide sequence samples is determined; Based on the difference between the proportion of each amino acid in each candidate polypeptide sequence sample and the corresponding overall proportion, multiple candidate polypeptide sequence samples are selected from multiple candidate polypeptide sequence samples and used as the target polypeptide sequence samples respectively.

3. The training method for the binding force prediction model according to claim 2, characterized in that, Based on the frequency of occurrence of each adjacent residue pair in each candidate polypeptide sequence sample, the overall proportion of each amino acid in the multiple candidate polypeptide sequence samples is determined, including: The total frequency of each adjacent residue pairing group in all candidate polypeptide sequence samples was calculated. Select the first number of adjacent residue pairs with the highest total frequency of each adjacent residue pairing group to obtain multiple target adjacent residue pairs; Determine the first frequency of each amino acid in all target adjacent residue pairs and use the sum of the first frequencies of all amino acids as the second frequency; For each amino acid, the ratio between the first frequency and the second frequency is taken as the overall percentage of the amino acid.

4. The training method for the binding force prediction model according to claim 2, characterized in that, Based on the difference between the proportion of each amino acid in each candidate polypeptide sequence sample and the corresponding overall proportion, multiple candidate polypeptide sequence samples are selected from the multiple candidate polypeptide sequence samples as the target polypeptide sequence samples, including: For each candidate polypeptide sequence sample, the difference between the proportion of each amino acid in the candidate polypeptide sequence sample and the corresponding overall proportion is obtained, and the differences are statistically analyzed to obtain the total difference value of the candidate polypeptide sequence sample. The candidate polypeptide sequence samples with a total difference value less than or equal to a preset difference value are selected as the target polypeptide sequence samples.

5. The training method for the binding force prediction model according to any one of claims 1 to 4, characterized in that, Before obtaining the sequence features of multiple candidate polypeptide sequence samples, the method further includes: Multiple initial polypeptide sequence samples were selected from the polypeptide database; Each of the initial polypeptide sequence samples is screened according to at least one screening dimension to obtain each of the candidate polypeptide sequence samples, wherein the at least one screening dimension includes sequence length, overall charge and / or secondary structure.

6. The training method for the binding force prediction model according to claim 5, characterized in that, Multiple initial peptide sequence samples were selected from the peptide database, including: Based on multiple search keywords, each initial polypeptide sequence sample is selected from the polypeptide database, and each search keyword is related to the binding ability to target extracellular vesicle membranes; In this context, the keyword tag of each initial polypeptide sequence sample is matched with one or more membrane binding search keywords.

7. The training method for the binding force prediction model according to any one of claims 1 to 4, characterized in that, Methods for obtaining sequence features for each candidate polypeptide sequence sample include: The category to which each amino acid in the candidate polypeptide sequence sample belongs is determined, wherein the category is cationic, hydrophobic, aromatic, acidic, nucleophilic, or amide. Based on the category to which each amino acid belongs, determine the proportion of each amino acid and the frequency of occurrence of each adjacent residue pairing.

8. The training method for the binding force prediction model according to any one of claims 1 to 4, characterized in that, Using at least a portion of the target polypeptide sequence samples and the respective simulated binding forces as binding force labels, the binding force prediction model is trained to obtain the final binding force prediction model, including: Based on the binding force simulated from each target polypeptide sequence sample, the second-highest number of target polypeptide sequence samples with the highest binding force are selected and added to the training dataset. The binding force prediction model is trained using the target polypeptide sequence samples and their binding force labels in the training dataset to obtain the final binding force prediction model.

9. The training method for the binding force prediction model according to claim 8, characterized in that, The number of binding force prediction models to be trained is multiple, and the different binding force prediction models are of different types. Using the target peptide sequence samples and their binding force labels in the training dataset, the binding force prediction models are trained to obtain the final binding force prediction model, including: Using the target polypeptide sequence samples and their binding force labels in the training dataset, each binding force prediction model is iteratively trained. After the iteration stopping condition is met, each of the bonding force prediction models is tested, and based on the test results of each bonding force prediction model, at least one of the bonding force prediction models with better test results is selected as the final bonding force prediction model.

10. A method for designing polypeptide sequences, characterized in that, include: Obtain multiple peptide sequences to be evaluated; Multiple peptide sequences to be evaluated are evaluated using a binding force prediction model to obtain the predicted binding force between each peptide sequence to be evaluated and the extracellular vesicle membrane. The binding force prediction model is trained based on the training method of the binding force prediction model according to any one of claims 1 to 9. Based on the predicted binding forces, a target polypeptide sequence is selected from the plurality of polypeptide sequences to be evaluated.

11. The polypeptide sequence design method according to claim 10, characterized in that, The number of binding force prediction models is at least two. Based on each predicted binding force, a target polypeptide sequence is selected from the plurality of polypeptide sequences to be evaluated, including: For each predicted binding force output by the binding force prediction model, based on the magnitude of each predicted binding force, the peptide sequences corresponding to the third highest number of predicted binding forces are selected as candidate peptide sequences corresponding to the binding force prediction model. The relative signal intensity between each candidate polypeptide sequence and the extracellular vesicle membrane is verified to obtain the relative signal intensity corresponding to each candidate polypeptide sequence. Based on the relative signal intensity corresponding to each of the candidate polypeptide sequences, at least one candidate polypeptide sequence is selected as the target polypeptide sequence.

12. The polypeptide sequence design method according to claim 11, characterized in that, Based on the relative signal intensity corresponding to each of the candidate polypeptide sequences, at least one candidate polypeptide sequence is selected as the target polypeptide sequence, including: The relative signal intensities of each of these parameters are compared with the relative signal intensities between known polypeptide sequences and extracellular vesicle membranes. Based on the alignment results, the target polypeptide sequence was obtained.

13. The polypeptide sequence design method according to any one of claims 10 to 12, characterized in that, Before obtaining multiple peptide sequences to be evaluated, the method further includes: Multiple reference sequence features are used as input to the sequence generation model, and the multiple polypeptide sequences output by the sequence generation model are used as the respective polypeptide sequences to be evaluated. The sequence generation model is trained based on sequence feature samples and their corresponding polypeptide sequences as labels.

14. A training device for a bonding force prediction model, characterized in that, include: The first feature acquisition module is used to acquire the sequence features of multiple candidate polypeptide sequence samples. The sequence features include the proportion of various amino acids and the frequency of occurrence of each adjacent residue pairing. The sample screening module is used to select multiple target polypeptide sequence samples from multiple candidate polypeptide sequence samples based on the proportion of various amino acids contained in each sequence feature and the frequency of occurrence of each adjacent residue pairing. A binding force label determination module is used to simulate the binding force between each target polypeptide sequence sample and the extracellular vesicle membrane; The training module is used to train the binding force prediction model using at least a portion of the target polypeptide sequence samples and the binding forces obtained by simulation of each as binding force labels, so as to obtain the final binding force prediction model.

15. A polypeptide sequence design device, characterized in that, include: The second feature acquisition module is used to acquire multiple polypeptide sequences to be evaluated; A binding force assessment module is used to assess multiple polypeptide sequences to be assessed using a binding force prediction model to obtain the predicted binding force between each polypeptide sequence to be assessed and the extracellular vesicle membrane, wherein the binding force prediction model is trained based on the training method of the binding force prediction model according to any one of claims 1 to 9. A sequence screening module is used to select a target polypeptide sequence from the plurality of polypeptide sequences to be evaluated based on the predicted binding forces.

16. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the training method of the binding force prediction model as described in any one of claims 1 to 9 or the polypeptide sequence design method as described in any one of claims 10 to 13.

17. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method of the binding force prediction model as described in any one of claims 1 to 9 or the polypeptide sequence design method as described in any one of claims 10 to 13.

18. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method of the binding force prediction model as described in any one of claims 1 to 9 or the polypeptide sequence design method as described in any one of claims 10 to 13.

Images