An artificial intelligence-based candidate polypeptide drug generation method and system
This artificial intelligence-based method for generating peptide drugs utilizes graph convolutional neural networks and diffusion models to generate peptide backbones. By combining multi-task learning and defolding techniques, it solves the problem of low efficiency in peptide drug design and achieves efficient generation of candidate drug molecules.
Patent Information
- Application Number
- CN202510274649.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-10
- Publication Date
- 2025-12-12
- Estimated Expiration
- 2045-03-10
AI Technical Summary
Current peptide drug design relies on the identification of natural functional peptides or expert experience, resulting in low efficiency, and peptides are difficult to model effectively in their free state.
An artificial intelligence-based approach is employed to generate peptide backbones through graph convolutional neural networks and diffusion models. By combining multi-task learning and inverse folding techniques, new candidate peptide molecules are generated. The Transformer architecture is used to learn the mapping relationship from structure to sequence, and the Monte Carlo algorithm is used for sequence generation.
It enables accurate, efficient, and low-cost identification and generation of potential drug candidate molecules, solves the problem of difficulty in modeling peptides in their free state, and improves drug development efficiency.
Smart Images
Figure CN120183538B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of drug discovery, and in particular to a candidate polypeptide drug generation method and system based on artificial intelligence. BACKGROUND
[0002] Drug discovery is the process of discovering new candidate drugs. Polypeptides are composed of 20 amino acids through dehydration condensation, and can form a huge chemical space through different combinations. Identifying different combinations through wet experiments will consume a lot of time, financial and material resources. Past polypeptide drug design often relies on identifying natural functional polypeptides or obtaining polypeptide fragments from functional proteins based on expert experience. This discovery and identification method hinders the efficiency of polypeptide drug development. With the development of artificial intelligence, AI technology provides new tools and methods for drug discovery. AI can quickly identify potential drug targets and candidate drug molecules from a large amount of biological data through deep learning and machine learning algorithms, further accelerating the efficiency of drug development. SUMMARY
[0003] In order to solve the technical problems of polypeptide drug toxicity prediction in the prior art, the present application provides a candidate polypeptide drug generation method and system based on artificial intelligence.
[0004] The present application is realized by the following technical solutions:
[0005] A candidate polypeptide drug generation method based on artificial intelligence, comprising:
[0006] S1: Collecting sequence data, structure data and related property labels of polypeptide drug molecules and their corresponding target proteins, wherein the related properties include functions and chemical and physical properties;
[0007] S2: Using the interaction site information of polypeptides and their targets, fixing the target protein structure, and generating the skeleton of the polypeptide using a diffusion model;
[0008] S3: Reverse folding the generated polypeptide skeleton to generate new candidate polypeptide molecules, and analyzing the prediction results.
[0009] Further, the data collected in step S1 includes sequence data and structure data of existing drug molecules and target proteins, sequence data and biological activity data of functionally identified polypeptides.
[0010] Further, the diffusion model construction process in step S2 includes:
[0011] S21, data preprocessing: obtaining relevant protein interaction data, extracting hydrophobicity, surface charge, affinity, activity and spatial geometry information by processing interaction site information;
[0012] S22, modeling the target protein using a graph convolutional neural network;
[0013] S23, constructing a diffusion model: fixing the target protein and the binding protein in a geometric space, and gradually adding noise to them;
[0014] S24, multi-task learning: introducing a multi-task learning framework in the training of the diffusion model, and simultaneously predicting the binding affinity, geometric structure and activity.
[0015] Further, in step S22, the isomorphic graph convolutional neural network is used to model the Cα atoms as nodes and the chemical bonds between adjacent atoms as edges, and the edge features composed of different chemical bonds are transmitted to the nodes through the information propagation mechanism, so as to capture the key features of the protein structure and output the feature vector.
[0016] Further, the diffusion formula of the diffusion model in step S23 is as follows:
[0017]
[0018]
[0019] wherein, is the polypeptide backbone structure diffused in the t-th step, is a time step control parameter, is the polypeptide backbone structure diffused in the t-1-th step, is the noise updated according to the adjacent nodes, is a neural network, is the feature of the i-th node, is the updated feature of the neighbor node of the i-th node.
[0020] Further, the inverse folding in step S3 includes:
[0021] S31, data preprocessing: obtaining the sequence and structure information of the related polypeptide, introducing the polypeptide dynamic structure simulation data, and using molecular dynamics simulation to extract the structural change characteristics of the polypeptide in different environments;
[0022] S32, obtaining the feature vector of the polypeptide sequence, using a word embedding model to map each amino acid to a feature vector, capturing the semantic relationship between amino acids, and mapping it to a fixed-dimensional polypeptide sequence feature vector;
[0023] S33: obtaining a structure feature vector, inputting the dynamic structure features of the polypeptide into an isomorphic graph convolutional neural network, the isomorphic graph convolutional neural network inputting the atom types, positions and chemical bonds of the polypeptide backbone, and adopting information edge transmission;
[0024] S34. Input the feature vector and structural feature vector of the polypeptide sequence into the neural network of the Transformer architecture to predict the three-dimensional coordinates of C, H, O and N of the polypeptide and learn the mapping relationship from structure to sequence.
[0025] S35. Use the Monte Carlo algorithm to decode the latent vector generated from the polypeptide backbone to generate a new polypeptide sequence;
[0026] S36. Use an automatic learning strategy to select sequences with high information entropy for iterative experiments, calculate the feature representation and predicted output distribution for each sequence, and calculate the uncertainty.
[0027] Furthermore, the update formula in step S33 is as follows:
[0028]
[0029] in, For the features of node i in the l-th layer, For the features of node i in the (l+1)th layer, Let be an adjacency matrix, representing the connections between nodes i and j. Let be the trainable parameters of the l-th layer. Activated by Sigmod, where i and j represent nodes i and j respectively. Let i be the set of adjacent nodes of node i.
[0030] Furthermore, the learning process, specifically the mapping relationship from structure to sequence, is described as follows:
[0031] Loss =
[0032] in, Representing the The number of amino acids, N represents the length of the polypeptide chain, and stur represents the structure of the polypeptide.
[0033] This invention also provides an artificial intelligence-based candidate peptide drug generation system, based on the aforementioned artificial intelligence-based peptide drug toxicity prediction method, comprising:
[0034] The data acquisition module is used to acquire sequence data, structural data, and related property tags of peptide drug molecules and their corresponding target proteins;
[0035] The peptide backbone generation module is used to fix the target protein structure using information about the interaction sites between the peptide and its target, and to generate the peptide backbone using a diffusion model.
[0036] The peptide backbone defolding module is used to defold the generated peptide backbone to generate new candidate peptide molecules and analyze the prediction results.
[0037] In addition, to achieve the above object, the application further provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions of an artificial intelligence based candidate polypeptide drug generation method, and the program instructions of the artificial intelligence based candidate polypeptide drug generation method can be executed by one or more processors to implement the steps of the artificial intelligence based candidate polypeptide drug generation method.
[0038] Compared with the prior art, the application has the following beneficial effects:
[0039] The application realizes accurate, efficient and low-cost rapid identification and generation of potential candidate drug molecules, effectively solves the structural information of proteins and polypeptides, solves the problem that polypeptides are difficult to model in a free state, and adjusts the diffusion process in combination with the physical properties of proteins to make it conform to the constraints of the actual chemical environment. BRIEF DESCRIPTION OF DRAWINGS
[0040] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the principles of the application, and do not limit the application in any way. In the drawings:
[0041] Figure 1 is a flowchart of an artificial intelligence based candidate polypeptide drug generation method according to an embodiment of the application;
[0042] Figure 2 is a Monte Carlo simulation sampling schematic diagram according to an embodiment of the application. DETAILED DESCRIPTION
[0043] The embodiments of the application will be described in detail below with reference to the accompanying drawings.
[0044] The embodiments of the application will be described in detail below with reference to the accompanying drawings.
[0045] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and only show the components related to the present application, not the number, shape and size of the components when actually implemented. The shape, number and proportion of each component when actually implemented can be arbitrarily changed, and the component layout form can also be more complex.
[0046] Referring to Figure 1 An artificial intelligence-based candidate polypeptide drug generation method, comprising:
[0047] S1: Collecting sequence data, structure data and related property tags of polypeptide drug molecules and their corresponding target proteins, the related properties including functions and chemical and physical properties;
[0048] Specifically, the collected data includes sequence data and structure data of existing drug molecules and target proteins, sequence data of functionally identified polypeptides, and biological activity data;
[0049] S2: Using the interaction site information of polypeptides and their targets, fixing the target protein structure, and generating the skeleton of the polypeptide using a diffusion model, thereby improving the generation effect of the model;
[0050] The diffusion model construction process is as follows:
[0051] S21, data preprocessing: obtaining relevant protein interaction data, extracting hydrophobicity, surface charge, affinity, activity and spatial geometry information by processing interaction site information;
[0052] S22, modeling the target protein using a graph convolutional neural network: fully considering the differences in the properties of different amino acids and the side chain attributes in the protein structure. By using a heterogeneous graph convolutional neural network, the Cα atom is taken as a node, and the chemical bond between adjacent atoms is taken as an edge for modeling. Through an information propagation mechanism, the edge features composed of different chemical bonds are transmitted to the nodes, thereby capturing the key features of the protein structure and outputting a polypeptide graph feature vector. This method can effectively solve the structural information of proteins and polypeptides, and solve the problem of difficult modeling of polypeptides in the free state;
[0053] S23, constructing a diffusion model: fixing the target protein and the binding protein in a geometric space and gradually adding noise to them. The addition process introduces an edge weight weighted diffusion method, and the physical properties (such as charge distribution and bond energy) of the binding protein are adjusted to adjust the diffusion process to conform to the constraints of the actual chemical environment. The formula of the diffusion process is:
[0054]
[0055]
[0056] wherein, is the polypeptide backbone structure of diffusion step t, is the time step control parameter, is the polypeptide backbone structure of diffusion step t-1, is the noise updated according to the neighboring nodes, is the neural network, is the feature of i node, is the updated feature of i node neighbor node.
[0057] The conditional feature c is introduced into the diffusion model, which guides the generation of polypeptide structure according to the local characteristics of the target protein. In the denoising process, the noise prediction value is calculated by the heterogeneous graph convolutional neural network and the conditional feature c:
[0058]
[0059] wherein, is the graph feature of target protein, is the graph feature of polypeptide, is the polypeptide backbone structure of diffusion step t, denotes the heterogeneous graph convolutional neural network model.
[0060] S24, multi-task learning: in order to predict the binding affinity, geometric structure and activity at the same time, the multi-task learning framework is introduced in the training of diffusion model. The loss function is in the form of multi-objective optimization, and the expression of loss function is as follows:
[0061]
[0062] wherein, , , is the weight of different tasks, which is used to balance the influence of multiple objectives and fine-tune the target, binding site prediction loss, is the loss of binding geometric structure prediction, is the binding energy prediction loss.
[0063] is used to measure the difference between the predicted affinity of the model and the real affinity. Mean square error (MSE) loss function is adopted:
[0064]
[0065] wherein, N represents the number of samples, represents the i th molecule, represents the predicted affinity of the i th molecule, represents the i th molecule, represents the predicted affinity of the i th molecule, represents the i th molecule, Experimental affinity of a molecule, i.e. Kd value.
[0066]
[0067] where M is the number of atoms, and represent predicted and real atomic coordinates, respectively, K is the number of bonds, and are the predicted and real bond lengths, respectively.
[0068] is used to measure the difference between the predicted activity of a model and the real activity. The mean squared error (MSE) loss function is adopted:
[0069]
[0070] where N represents the number of samples, represents the th molecule, represents the predicted activity of the th molecule, represents the experimental activity (IC50 value) of the th molecule.
[0071] After training is completed, the inverse process of the diffusion model starts from the geometric features of the target protein, and gradually denoises the Gaussian noise at the active site of the target protein to generate the binding polypeptide of the target protein. The relevant expressions are as follows. At time t, the model predicts the denoised state and the geometric features of the target protein according to the current state :
[0072]
[0073] where is the trained model.
[0074] S3: The generated polypeptide skeleton is subjected to inverse folding to generate new candidate polypeptide molecules, and the prediction results are analyzed.
[0075] The inverse folding process is as follows:
[0076] S31, data preprocessing: obtain the sequence and structure information of the relevant polypeptide, introduce polypeptide dynamic structure simulation data, and extract the structural change characteristics of the polypeptide in different environments by using molecular dynamics simulation. Assuming that the dynamic conformation of the polypeptide is , the time series structure calculated by molecular dynamics simulation is:
[0077]
[0078] in, is the coordinate of the i-th Cα atom, N is the amino acid length of the polypeptide, and dynamic structural features are extracted through dimensionality reduction or clustering as input for subsequent models.
[0079] S32. Obtain the feature vector of the polypeptide sequence. Using a word embedding model, map each amino acid to a feature vector to capture the semantic relationship between amino acids and map it to a fixed-dimensional polypeptide sequence feature vector.
[0080] S33: Obtain the structural feature vector. Input the dynamic structural features of the peptide into a heterogeneous graph convolutional neural network. The heterogeneous graph convolutional neural network is input with the types, positions, and chemical bonds of the peptide backbone, and uses information edge passing. The specific update formula is as follows:
[0081]
[0082] in, For the features of node i in the l-th layer, For the features of node i in the (l+1)th layer, Let be an adjacency matrix, representing the connections between nodes i and j. Let be the trainable parameters of the l-th layer. Activated by Sigmod, where i and j represent nodes i and j respectively. Let i be the set of adjacent nodes of node i.
[0083] S34. Input the feature vector and structural feature vector of the polypeptide sequence into a neural network with a Transformer architecture to predict the three-dimensional coordinates of C, H, O, and N of the polypeptide, and learn the mapping relationship from structure to sequence. The relevant description is as follows:
[0084] Loss =
[0085] in, Representing the The number of amino acids, N represents the length of the polypeptide chain, and stur represents the structure of the polypeptide.
[0086] S35. The latent vector generated from the polypeptide backbone obtained in S2 is decoded using the Monte Carlo algorithm to generate a new polypeptide sequence.
[0087] S36. Using an automatic learning strategy, select sequences with high information entropy for iterative experiments, and calculate the feature representation of each sequence. and predicted output distribution ( ), Calculate the uncertainty:
[0088]
[0089] wherein, represents a feature vector of a sequence, represents a predicted first class, is a predicted probability of the affinity of the polypeptide predicted by the model.
[0090] The uncertainty of the generated new polypeptide is calculated, the top-ranked polypeptide is selected for biological experiment verification, and the biological activity information obtained by the biological experiment is used for model parameter adjustment.
[0091] Preferably, the data set in step S31 is used for training, parameter adjustment and evaluation of the model, and the training set of polypeptide sequences and structures is used for training to learn the mapping relationship between the sequences and structures of active polypeptides.
[0092] After the training is completed, the model generates potential vector representations by inputting the active polypeptide skeleton, and these potential vectors capture key feature information of the polypeptide sequence. Further, by operating and sampling the potential vectors, new polypeptide sequences are generated, which have similar or better biological activity characteristics to the input polypeptide in theory.
[0093] Preferably, in step S35, a Monte Carlo sampling method is used to sample the potential vectors of the model to generate new polypeptide sequences. By sampling different potential vectors, the potential space of polypeptide sequences can be explored, thereby generating polypeptide sequences with different biological activities.
[0094] In this embodiment, the present application realizes accurate, efficient and low-cost rapid identification and generation of potential candidate drug molecules, effectively solves the problem of structural information of proteins and polypeptides, solves the problem of difficulty in modeling polypeptides in free state, and combines the physical characteristics of proteins to adjust the diffusion process to make it conform to the constraints of the actual chemical environment.
[0095] The present application also provides an artificial intelligence-based candidate polypeptide drug generation system based on the artificial intelligence-based candidate polypeptide drug generation method described above, comprising:
[0096] A data acquisition module is used to acquire sequence data, structure data and related property labels of polypeptide drug molecules and their corresponding target proteins;
[0097] A polypeptide skeleton generation module is used to use the interaction site information of the polypeptide and its target, fix the target protein structure, and generate the polypeptide skeleton using the diffusion model;
[0098] A polypeptide skeleton unfolding module is used to unfold the generated polypeptide skeleton to generate new candidate polypeptide molecules, and analyze the prediction results.
[0099] In addition, the embodiment of the present application also provides a computer readable storage medium, and program instructions of the artificial intelligence-based candidate polypeptide drug generation method are stored on the computer readable storage medium. The program instructions of the artificial intelligence-based candidate polypeptide drug generation method can be executed by one or more processors to implement the steps of the artificial intelligence-based candidate polypeptide drug generation method as described above.
[0100] The above-described embodiments are only used to describe the preferred embodiments of the present application, and are not used to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope of the present application defined by the claims.
Claims
1. An artificial intelligence-based candidate polypeptide drug generation method, characterized in that, The method comprises the following steps: S1: collecting sequence data, structure data and related property tags of polypeptide drug molecules and their corresponding target proteins, wherein the related properties include functions and chemical physicochemical properties; S2: using the interaction site information of the polypeptide and its target to fix the target protein structure, and generating a polypeptide skeleton by using a diffusion model; The diffusion model construction process in step S2 comprises: S21, data preprocessing: obtaining relevant protein interaction data, extracting hydrophobicity, surface charge, affinity, activity and spatial geometry information by processing interaction site information; S22, modeling the target protein by using a graph convolutional neural network; S23, constructing a diffusion model: fixing the target protein and the binding protein in a geometric space, and gradually adding noise to them; S24, multi-task learning: introducing a multi-task learning framework in the training of the diffusion model to simultaneously predict binding affinity, geometric structure and activity; S3: reverse folding the generated polypeptide skeleton to generate new candidate polypeptide molecules, and analyzing the prediction results; The reverse folding in step S3 comprises: S31, data preprocessing: obtaining sequence and structure information of relevant polypeptides, introducing polypeptide dynamic structure simulation data, and extracting structural variation characteristics of polypeptides in different environments by using molecular dynamics simulation; S32, obtaining a feature vector of the polypeptide sequence, mapping each amino acid to a feature vector by using a word embedding model, capturing the semantic relationship between amino acids, and mapping it to a fixed-dimensional polypeptide sequence feature vector; S33: obtaining a structure feature vector, inputting the dynamic structure characteristics of the polypeptide into a heterogeneous graph convolutional neural network, and inputting the atom type, position and chemical bond of the polypeptide skeleton into the heterogeneous graph convolutional neural network, and adopting information edge transmission; S34, inputting the feature vector of the polypeptide sequence and the structure feature vector into a neural network with a Transformer architecture to predict the three-dimensional coordinates of C, H, O and N of the polypeptide, and learning the mapping relationship from structure to sequence; S35, using a Monte Carlo algorithm to decode the potential vector generated from the polypeptide skeleton to generate a new polypeptide sequence; S36, using an automatic learning strategy to select a sequence with high information entropy for experimental iteration, calculating the feature representation and prediction output distribution of each sequence, and calculating the uncertainty, which is calculated as follows: ； in, The feature vector representing the sequence. Indicates the predicted first Categories The predicted probability of peptide affinity predicted by the model; The mapping relationship from structure to sequence is described as follows: Loss = ; in, Representing the The number of amino acids, N represents the length of the polypeptide chain, and stur represents the structure of the polypeptide.
2. The artificial intelligence-based candidate polypeptide drug generation method of claim 1, wherein, The data collected in step S1 includes sequence data and structure data of existing drug molecules and target proteins, sequence data and biological activity data of functionally identified polypeptides.
3. The artificial intelligence-based candidate polypeptide drug generation method of claim 2, wherein, In step S22, the heterogeneous graph convolutional neural network is used to model the Cα atom as a node and the chemical bond between adjacent atoms as an edge, and the edge features composed of different chemical bonds are transmitted to the node by an information propagation mechanism, so as to capture the key features of the protein structure and output a polypeptide graph feature vector.
4. The artificial intelligence-based candidate polypeptide drug generation method of claim 2, wherein, The diffusion formula of the diffusion model in step S23 is as follows: ； ； wherein, is the polypeptide backbone structure of diffusion step t, is a time step control parameter, is the polypeptide backbone structure of diffusion step t-1, is the noise updated according to the neighboring nodes, is a neural network, is a feature of the i-th node, is an updated feature of the i-th node neighbor node.
5. The artificial intelligence-based candidate polypeptide drug generation method of claim 1, wherein, The update formula in step S33 is as follows: ； wherein, is a feature of the i-th node of the l-th layer, is a feature of the i-th node of the l+1-th layer, is an adjacency matrix representing the connection between i, j nodes, is a trainable parameter of the l-th layer, is a Sigmod activation, i, j represent nodes i and j respectively, is a set of adjacent nodes of the i-th node.
6. An artificial intelligence-based candidate polypeptide drug generation system based on the artificial intelligence-based candidate polypeptide drug generation method of any one of claims 1 to 5, comprising: a data collection module for collecting sequence data, structure data and related property labels of polypeptide drug molecules and their corresponding target proteins; a polypeptide skeleton generation module for fixing the target protein structure using the interaction site information of the polypeptide and its target, and generating the polypeptide skeleton using a diffusion model; a polypeptide skeleton inverse folding module for inverse folding the generated polypeptide skeleton to generate new candidate polypeptide molecules, and analyzing the prediction results.
7. A computer readable storage medium characterized in that, The computer readable storage medium stores program instructions of the artificial intelligence-based candidate polypeptide drug generation method, which can be executed by one or more processors to implement the steps of the artificial intelligence-based candidate polypeptide drug generation method of any one of claims 1 to 5.