3D molecule generation method and system based on dynamic conformation guidance

Abstract

The invention relates to a 3D molecule generation method and system based on dynamic conformation guidance. The 3D molecule generation method based on dynamic conformation guidance comprises the following steps: acquiring a dynamic conformation set of protein through molecular dynamics simulation; processing the dynamic conformation set of the protein by adopting a multi-level equilateral graph attention network architecture to obtain a dynamic condition vector; a 3D molecule is generated based on an isotropic diffusion trunk and the dynamic condition vector. The dynamic conformation set of the protein is obtained through molecular dynamics simulation, after the dynamic condition vector is obtained based on the dynamic conformation set, the 3D molecule is generated based on the dynamic condition vector, it can be ensured that the generated 3D molecule not only meets the static binding requirement, but also can adapt to the real dynamic environment of the target spot, and the application prospect is wide. The problem of static-dynamic performance difference in the prior art can be effectively solved, so that the druggability of generated molecules and the success rate of experimental verification are greatly improved, and the later elimination rate of drug research and development is reduced.

Description

Technical Field

[0001] This application relates to the field of molecular generation technology, and in particular to a method and system for generating 3D molecules based on dynamic conformation guidance. Background Technology

[0002] Current diffusion-based molecular generation techniques primarily rely on static protein structures as input. This approach is limited by its inability to reflect the dynamic characteristics of real biological environments. Proteins exhibit dynamic behaviors under physiological conditions, including conformational changes, side-chain rotation, and global domain movement. Water molecule networks also undergo dynamic reorganization following ligand binding. While molecules generated by existing techniques show good binding patterns in static docking, they often dissociate rapidly in molecular dynamics simulations because they fail to adapt to the dynamic characteristics of the target site. This discrepancy between static and dynamic performance leads to the elimination of many candidate molecules in later experimental validation stages, significantly increasing drug development costs. Summary of the Invention

[0003] Therefore, it is necessary to provide a method and system for generating 3D molecules based on dynamic conformation guidance to address the problems in related technologies.

[0004] To achieve the above objectives, in a first aspect, this application provides a method for generating 3D molecules based on dynamic conformation guidance, the method comprising:

[0005] Obtain a set of dynamic conformations of proteins through molecular dynamics simulations;

[0006] A multi-level equilateral graph attention network architecture is used to process the dynamic conformation set of proteins to obtain dynamic condition vectors;

[0007] 3D molecules are generated based on the isovariant diffusion backbone and the dynamic condition vector.

[0008] The aforementioned 3D molecule generation method based on dynamic conformation guidance obtains a set of dynamic conformations of proteins through molecular dynamics simulation. After obtaining a dynamic condition vector based on the set of dynamic conformations, 3D molecules are generated based on the dynamic condition vector. This ensures that the generated 3D molecules not only meet the static binding requirements but also adapt to the real dynamic environment of the target site. It can effectively overcome the problem of static-dynamic performance differences in existing technologies, thereby significantly improving the drug-likeness of the generated molecules and the success rate of experimental verification, and reducing the late-stage rejection rate in drug development.

[0009] In some embodiments, a set of dynamic conformations of the protein is obtained through molecular dynamics simulations, including:

[0010] Trajectory data is obtained through molecular dynamics simulations;

[0011] Time-correlated principal component analysis was performed on the trajectory data to obtain the principal component space;

[0012] K-means clustering analysis was performed on the principal component space to obtain the conformational state;

[0013] Water molecule analysis is performed on the trajectory data to obtain a water molecule density map and a hydrogen bond network; the conformational state, the water molecule density map, and the hydrogen bond network together constitute the dynamic conformation set.

[0014] In some embodiments, a three-dimensional spatial grid method is used to calculate the water molecule density distribution at the binding sites in the trajectory data to obtain the water molecule density map.

[0015] In some embodiments, the dynamic characteristics of the water network are quantified using a hydrogen bond lifetime autocorrelation function to obtain the hydrogen bond network.

[0016] In some embodiments, a multi-level equilateral graph attention network architecture is used to process the dynamic conformation set of proteins to obtain a dynamic conditional vector, including:

[0017] The configuration state is EGAT encoded to obtain the global feature vector of the configuration;

[0018] The water molecule density map is subjected to multi-layer convolution using a 3D convolutional neural network to obtain a solvated global feature vector.

[0019] The hydrogen bond network is temporally encoded using LSTM to obtain the global feature vector of hydrogen bonds;

[0020] The global feature vector of the conformation, the global feature vector of solvation, and the global feature vector of hydrogen bonding are spliced ​​together from multiple sources to obtain a multi-source spliced ​​feature vector.

[0021] The multi-source spliced ​​feature vectors are subjected to multi-scale gating fusion to obtain the dynamic condition vector.

[0022] In some embodiments, the conformational state is EGAT encoded to obtain a global feature vector of the conformation, including:

[0023] An isomorphic map is constructed based on the aforementioned conformational state;

[0024] The isomorphic graph is encoded using a graph attention mechanism to obtain the global feature vector of the conformation.

[0025] In some embodiments, generating 3D molecules based on the isovariant diffusion backbone and the dynamic conditional vector includes:

[0026] Generate random noise molecules;

[0027] The random noise molecules are denoised using an isovariant diffusion backbone to generate a molecular skeleton that conforms to 3D geometric rules; the dynamic condition vector is injected during the denoising process.

[0028] Guide the molecular intermediate structure to adapt to the target dynamic conformation through conformational attention gating.

[0029] The intermediate structure of the molecule is subjected to multi-objective adaptive optimization to balance static binding and dynamic stability in order to generate a 3D molecule.

[0030] In some embodiments, the accompanying noise molecules are denoised multiple times using an isotropic diffusion backbone; the dynamic condition vector is injected during each denoising process.

[0031] In some embodiments, the composite loss function for multi-objective adaptive optimization of the molecular intermediate structure is:

[0032]

[0033] in, This is the total loss function; For static docking loss; λ static This represents the weighting coefficient for static docking loss; For dynamic molecular dynamics loss; λ dyn This represents the weighting coefficient for dynamic molecular dynamics loss.

[0034] Secondly, this application also provides a 3D molecular generation system based on dynamic conformation guidance, the 3D molecular generation system based on dynamic conformation guidance includes:

[0035] The dynamic conformation set acquisition module is used to obtain the dynamic conformation set of proteins through molecular dynamics simulation.

[0036] The dynamic condition vector acquisition module is used to process the dynamic conformation set of proteins using a multi-level equilateral graph attention network architecture to obtain dynamic condition vectors.

[0037] A 3D molecule generation module is used to generate 3D molecules based on an isovariant diffusion backbone and the dynamic condition vector.

[0038] The aforementioned 3D molecule generation system based on dynamic conformation guidance involves a dynamic conformation set acquisition module that obtains the dynamic conformation set of proteins through molecular dynamics simulation. After the dynamic condition vector acquisition module obtains the dynamic condition vector based on the dynamic conformation set, the 3D molecule generation module generates 3D molecules based on the dynamic condition vector. This ensures that the generated 3D molecules not only meet the static binding requirements but also adapt to the real dynamic environment of the target site. It can effectively overcome the problem of static-dynamic performance differences in existing technologies, thereby significantly improving the drug-likeness of the generated molecules and the success rate of experimental verification, and reducing the late-stage rejection rate in drug development. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a 3D molecule generation method based on dynamic conformation guidance provided in one embodiment of this application;

[0041] Figure 2 This is a structural block diagram of a 3D molecular generation system based on dynamic conformation guidance provided in another embodiment of this application.

[0042] Figure labeling: 10, Dynamic configuration set acquisition module; 20, Optimal design scheme acquisition module; 30, Construction monitoring, analysis and optimization module. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] In one embodiment, see Figure 1 This application also provides a method for generating 3D molecules based on dynamic conformation guidance, which includes the following steps: S10 to S30.

[0045] S10: Obtain a set of dynamic conformations of proteins through molecular dynamics simulations.

[0046] S20: A multi-level equilateral graph attention network architecture is used to process the dynamic conformation set of proteins to obtain dynamic condition vectors.

[0047] S30: Generate 3D molecules based on the isovariant diffusion backbone and the dynamic condition vector.

[0048] The dynamic conformation-guided 3D molecule generation method in this embodiment obtains a set of dynamic conformations of proteins through molecular dynamics simulation. After obtaining a dynamic condition vector based on the set of dynamic conformations, 3D molecules are generated based on the dynamic condition vector. This ensures that the generated 3D molecules not only meet the static binding requirements but also adapt to the real dynamic environment of the target site. It can effectively overcome the problem of static-dynamic performance differences in the prior art, thereby significantly improving the drug-likeness of the generated molecules and the success rate of experimental verification, and reducing the late-stage rejection rate of drug development.

[0049] In step S10, please refer to Figure 1 In step S10, the dynamic conformation set of the protein is obtained through molecular dynamics simulation.

[0050] As an example, in step S10, obtaining the set of dynamic conformations of the protein through molecular dynamics simulation may include the following steps: S101 to S104.

[0051] S101: Obtain trajectory data through molecular dynamics simulation.

[0052] S102: Time-correlated principal component analysis is performed on the trajectory data to obtain the principal component space.

[0053] S103: Perform k-means clustering analysis on the principal component space to obtain the conformational state.

[0054] S104: Perform water molecule analysis on the trajectory data to obtain a water molecule density map and a hydrogen bond network. The conformational states, the water molecule density map, and the hydrogen bond network together constitute the dynamic conformation set.

[0055] As an example, in step S101, molecular dynamics simulations of 100–500 ns can be run in an explicit solvent environment, using a temperature copy exchange method to enhance conformational sampling efficiency. During the molecular dynamics simulation, the atomic coordinate trajectories of the protein backbone and side chains are recorded, with a sampling frequency set to 10 ps / frame.

[0056] As an example, in step S102, the process of performing time-correlated principal component analysis (tPCA) on the trajectory data to obtain the principal component space includes: calculating the covariance matrix of the trajectory.

[0057]

[0058] Among them, C ij Let x be the covariance matrix between the i-th atom and the j-th atom; i Let be the coordinates of the i-th atom; x is the average of the coordinates of the i-th atom; j Let be the coordinates of the j-th atom; The average value of the coordinates of the j-th atom; <.> is the time averaging operator.

[0059] By solving the covariance matrix above, we can obtain the eigenvector V and eigenvalue Λ. The first 3 to 5 principal components can usually explain more than 80% of the conformational changes.

[0060] As an example, step S103, in the process of performing k-means clustering analysis on the principal component space (setting the number of clusters k = 3 to 5) to obtain the conformational state, may include the following: calculating the Markov state transition matrix:

[0061]

[0062] Among them, T ij (τ) represents the probability that conformational state i transitions to conformational state j after a lag time τ; C ij (τ) represents the total number of direct transitions from conformational state i to conformation j within the lag time τ (transition count); ∑ j C ij (τ) is the sum of the total number of transitions from conformational state i to all possible conformational states within the lag time τ (total transition count).

[0063] As an example, in step S104, a three-dimensional spatial mesh method is used ( The water molecule density distribution at binding sites in the trajectory data is calculated using a resolution method to obtain the water molecule density map. The formula for calculating the water molecule density distribution is as follows:

[0064]

[0065] Where ρ(x,y,z) represents the water molecule density distribution at spatial location (x,y,z); N is the total number of water molecules; W is the set of water molecules; (x w ,y w ,z w ) represents the three-dimensional coordinates of the water molecule in the current frame; δ(.) represents the Dirac function.

[0066] As an example, in step S104, the dynamic characteristics of the water network are quantified using the hydrogen bond lifetime autocorrelation function to obtain the hydrogen bond network. The formula for quantifying the dynamic characteristics of the water network using the hydrogen bond lifetime autocorrelation function can be:

[0067]

[0068] Among them, C HBh(t) is the hydrogen bond lifetime autocorrelation function; h(t) is the indicator function for the presence of hydrogen bonds at time t.<h(0)h(t)> The correlation between the hydrogen bond states at time t=0 and at time t is measured; h(0) 2 is the normalization factor; <.> is the time averaging operator.

[0069] In step S20, please refer to Figure 1 In step S20, a multi-level equilateral graph attention network architecture is used to process the dynamic conformation set of proteins to obtain dynamic condition vectors.

[0070] As an example, in step S20, a multi-level equilateral graph attention network architecture is used to process the dynamic conformation set of proteins to obtain a dynamic condition vector, which may include the following steps: S201 to S205.

[0071] S201: Encode the configuration state using EGAT to obtain the global feature vector of the configuration.

[0072] S202: Use a 3D convolutional neural network to perform multi-layer convolution on the water molecule density map to obtain a solvated global feature vector.

[0073] S203: Use LSTM to perform temporal encoding on the hydrogen bond network to obtain the global feature vector of hydrogen bonds.

[0074] S204: Perform multi-source concatenation of the conformational global feature vector, the solvation global feature vector, and the hydrogen bond global feature vector to obtain a multi-source concatenated feature vector.

[0075] S205: Perform multi-scale gating fusion on the multi-source spliced ​​feature vector to obtain the dynamic condition vector.

[0076] As an example, in step S201, the conformation state is EGAT encoded to obtain the global feature vector of the conformation, which may include the following steps: S2011 to S2022.

[0077] S2011: Construct an isomorphic graph based on the aforementioned conformational state.

[0078] S2022: Encode the isomorphic graph using a graph attention mechanism to obtain the global feature vector of the conformation.

[0079] As an example, in step S2011, constructing the isomorphic diagram based on the conformational state may include the following:

[0080] Each conformational state is represented as an atomic-level graph G = (V, E), with node v i ∈V contains atomic coordinates and eigenvectors

[0081] Side e ij ∈E based on distance threshold Establish, including the relative displacement vector r ij =r j -r i r i Let r be the absolute coordinate vector of atom i. j Let be the absolute coordinate vector of atom j;

[0082] Coordinate information is processed using an SE(3)-isovariant linear layer:

[0083] r i′ =W r r i +b r

[0084] h i1′ =W h h i1 +b h

[0085] Where, r i′ W represents the coordinates after linear transformation. r r is the weight matrix applied to the coordinates. i h represents the original coordinates before the linear transformation. i1′ W represents the eigenvectors after linear transformation. h h is the weight matrix applied to the eigenvectors. i1 b is the eigenvector before linear transformation; r and b h This is a bias term.

[0086] As an example, in step S2022, the graph attention mechanism is used to encode the isomorphic graph to obtain the global feature vector of the configuration, which may include the following:

[0087] Calculate the rotation-invariant attention coefficient:

[0088] α ij =softmax(σ(a) T [W′h i ||W′h j ||||r ij ||]))

[0089] Where, α ij σ is the attention coefficient (i.e., attention weight); softmax is the normalization function; σ(.) is the sigmoid activation function; W′ is the learnable linear transformation matrix; a T Let h be the weight vector of the attention scoring function, and let h be a learnable parameter.i h is the eigenvector of atom i; j r is the eigenvector of atom j; ij Let be the three-dimensional spatial vector pointing from atom i to atom j;

[0090] For equivariant message passing, the corresponding formula can be as follows:

[0091] m ij =φ m (h i ,h j ,r ij )·r ij

[0092]

[0093] Where, m ij φ is the message vector passed from atom j to atom i. m (.) represents a multilayer perceptron; h i h is the eigenvector of atom i; j r is the eigenvector of atom j; ij Let Δr be the three-dimensional spatial vector pointing from atom i to atom j; i α represents the total displacement of atom i; N(i) represents all neighboring atoms of atom i; ij The attention coefficient (i.e., attention weight); h i′ φ is the updated eigenvector of atom i; h (.) represents another multilayer perceptron.

[0094] As an example, in step S202, a 3D convolutional neural network is used to perform multi-layer convolution on the water molecule density map to obtain a solvated global feature vector. This may include the following: the water molecule network features in the water molecule density map are processed by 3D convolution kernels, and the corresponding formula can be:

[0095]

[0096] Where Ψ(x,y,z) is the global function; N is the number of water molecule network features to be processed; w i ρ(x,y,z) represents the weight of the i-th water molecule network feature to be processed; ρ(x,y,z) represents the water molecule density map; and K(x,y,z) represents the 3D Gaussian kernel.

[0097] As an example, in step S203, LSTM is used to perform temporal encoding on the hydrogen bond network to obtain the global feature vector of hydrogen bonds. The formulas involved in LSTM can be as follows:

[0098] f t =σ(W f [ht ,c t-1 ]+b f )

[0099] i t =σ(W i [h t ,c t-1 ]+b i )

[0100] o t =σ(W o [h t ,c t-1 ]+b o )

[0101] Among them, f t For the Gate of Oblivion; i t For input gate; o t σ is the output gate; σ(.) is the Sigmoid activation function; W f W is the gating weight matrix for the forgetting gate; i W is the gating weight matrix of the input gate; o Here is the gating weight matrix for the output gate; [h t ,c t-1 [] represents the concatenation of inputs, combining the input features h at the current time step. t The cell state c at the previous time step t-1 Concatenate them into a longer vector according to the feature dimension; b f b is the gating bias vector for the forget gate; i b is the gate bias vector of the input gate; o This is the gate bias vector for the output gate.

[0102] As an example, in step S205, the formula for obtaining the dynamic condition vector can be as follows:

[0103] c dyn =MLP([h EGAT ||Ψ||h LSTM ])

[0104] Among them, c dyn For dynamic conditional vectors; MLP stands for Multilayer Perceptron; GAT stands for Conformational Global Feature Vector; Ψ stands for Solvational Global Feature Vector; h LSTM This is the global feature vector of hydrogen bonds.

[0105] In step S30, please refer to Figure 1 In step S30, 3D molecules are generated based on the isovariant diffusion backbone and the dynamic condition vector.

[0106] As an example, step S30, generating a 3D molecule based on the isovariant diffusion backbone and the dynamic condition vector, may include the following steps: S301 to S304.

[0107] S301: Generate random noise molecules.

[0108] S302: Denoise the random noise molecules using an isotropic diffusion backbone to generate a molecular skeleton that conforms to 3D geometric rules; inject the dynamic condition vector during the denoising process.

[0109] S303: Guides the dynamic conformation of molecular intermediate structures to target sites through conformational attention gating.

[0110] S304: Perform multi-objective adaptive optimization on the intermediate structure of the molecule to balance static binding and dynamic stability in order to generate a 3D molecule.

[0111] As an example, in step S301, random noise molecules can be generated by sampling from a standard normal distribution.

[0112] As an example, in step S302, the accompanying noise molecules are denoised multiple times through an isotropic diffusion backbone; the dynamic condition vector is injected in each denoising process.

[0113] As an example, in step S302, the random noise molecules are denoised using an isovariant diffusion backbone to generate a molecular skeleton that conforms to 3D geometric rules; the dynamic condition vector injected during the denoising process may include the following:

[0114] SE(3) - Equivalent diffusion backbone construction, including: defining noise scheduling for the diffusion process and designing an equivalent denoising network;

[0115] Dynamic conditional vector injection.

[0116] Specifically, the formula for defining extended process noise scheduling can be:

[0117] β t =β min +t(β max -β min ) / T

[0118] Where, β t β is the noise intensity at step t; T is the total number of diffusion steps; min For minimum noise intensity, β min =10 -4 ;β max For the maximum noise intensity, β max =0.02.

[0119] Specifically, the expression for the designed isovariant denoising network can be:

[0120] ∈ θ (x t ,t,c dyn ) = EGNN(x t ,t,c dyn )

[0121] Among them, the equivariant noise reduction network ∈ θ With noisy random noise molecule x t Time step t and dynamic condition vector c dyn As input, this information is processed by EGNN (Equivariant Graph Neural Network) to output a denoised molecular skeleton that conforms to 3D geometric rules.

[0122] As an example, dynamic conditional vector injection can be implemented based on a conditional adaptive normalization layer. The expression for the conditional adaptive normalization layer can be:

[0123]

[0124] Among them, h out The output features after injecting dynamic conditional vectors; h in σ is the input feature; μ is the mean of the input feature; σ is the standard deviation of the input feature; γ(c dyn ) represents the dynamic scaling parameter; β(c) dyn ) represents the dynamic offset parameter.

[0125] As an example, in step S303, during the process of guiding the intermediate molecular structure to adapt to the dynamic conformation of the target site through conformational attention gating, the formula for conformational attention gating can be as follows:

[0126] g conf =σ(W g [h mol ||c dyn ])

[0127] h final =g conf ⊙h mol +(1-g conf )⊙c dyn

[0128] Among them, g conf The gate signal has a value range of [0,1]; σ(.) is the Sigmoid activation function; [h mol ||c dyn [H] is a molecular characteristic. mol and dynamic condition vector c dyn splicing; h final The final characteristic after fusion (i.e., it can be an intermediate molecular structure); g conf ⊙hmol Indicates the contribution ratio of molecular characteristics; (1-g) conf )⊙c dyn This represents the contribution ratio of the dynamic feature vector.

[0129] As an example, in step S304, the composite loss function for multi-objective adaptive optimization of the molecular intermediate structure can be:

[0130]

[0131] in, This is the total loss function; For static docking loss; λ static This represents the weighting coefficient for static docking loss; For dynamic molecular dynamics loss; λ dyn This represents the weighting coefficient for dynamic molecular dynamics loss.

[0132] Specifically, the weighting coefficient λ of the static docking loss is obtained based on a dynamic weighting adjustment strategy. static The weighting coefficient λ of dynamic molecular dynamics loss dyn Among them, the weighting coefficient λ for static docking loss static The formula can be:

[0133]

[0134] Weighting coefficient λ of dynamic molecular dynamics loss dyn The formula can be:

[0135] λ dyn =1-λ static

[0136] Where t is the current iteration step; k is the switching rate, k = 0.1; y0 is the switching midpoint, y0 = 0.5t; and T is the total number of iteration steps.

[0137] Specifically, static docking loss The formula can be:

[0138]

[0139] Wherein, PLP(.) is a commonly used scoring function in molecular docking; x is the 3D coordinate of the generated molecule; and p is the 3D structure of the target protein. This is the expected value.

[0140] Specifically, dynamic molecular dynamics loss The formula can be:

[0141]

[0142] Among them, RMSD flex It is a flexible RMSD, a special RSMD, which optimizes the conformation of the molecule and the target protein before calculation; x is the 3D coordinate of the generated molecule; The expected value; {p t} represents the set of conformations of the target protein during molecular dynamics simulations.

[0143] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0144] In another embodiment, please refer to Figure 1 See Figure 2 This application also provides a 3D molecular generation system based on dynamic conformation guidance. The 3D molecular generation system based on dynamic conformation guidance includes: a dynamic conformation set acquisition module 10, a dynamic condition vector acquisition module 20, and a 3D molecular generation module 30. The dynamic conformation set acquisition module 10 is used to acquire the dynamic conformation set of proteins through molecular dynamics simulation. The dynamic condition vector acquisition module 20 is used to process the dynamic conformation set of proteins using a multi-level equilateral graph attention network architecture to obtain a dynamic condition vector. The 3D molecular generation module 30 is used to generate 3D molecules based on the isovariant diffusion backbone and the dynamic condition vector.

[0145] The aforementioned 3D molecule generation system based on dynamic conformation guidance involves a dynamic conformation set acquisition module that obtains the dynamic conformation set of proteins through molecular dynamics simulation. After the dynamic condition vector acquisition module obtains the dynamic condition vector based on the dynamic conformation set, the 3D molecule generation module generates 3D molecules based on the dynamic condition vector. This ensures that the generated 3D molecules not only meet the static binding requirements but also adapt to the real dynamic environment of the target site. It can effectively overcome the problem of static-dynamic performance differences in existing technologies, thereby significantly improving the drug-likeness of the generated molecules and the success rate of experimental verification, and reducing the late-stage rejection rate in drug development.

[0146] As an example, the dynamic conformation-guided 3D molecular generation system of this embodiment can be used to perform tasks such as... Figure 1The method for generating 3D molecules based on dynamic conformation guidance in its corresponding embodiments.

[0147] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features of the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0148] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for generating 3D molecules based on dynamic conformation guidance, characterized in that, include: Obtain a set of dynamic conformations of proteins through molecular dynamics simulations; A multi-level equilateral graph attention network architecture is used to process the dynamic conformation set of proteins to obtain dynamic condition vectors; 3D molecules are generated based on the isovariant diffusion backbone and the dynamic condition vector.

2. The 3D molecular generation method based on dynamic conformation guidance according to claim 1, characterized in that, A set of dynamic conformations of proteins is obtained through molecular dynamics simulations, including: Trajectory data is obtained through molecular dynamics simulations; Time-correlated principal component analysis was performed on the trajectory data to obtain the principal component space; K-means clustering analysis was performed on the principal component space to obtain the conformational state; Water molecule analysis is performed on the trajectory data to obtain a water molecule density map and a hydrogen bond network; the conformational state, the water molecule density map, and the hydrogen bond network together constitute the dynamic conformation set.

3. The 3D molecular generation method based on dynamic conformation guidance according to claim 2, characterized in that, The water molecule density distribution at the binding sites in the trajectory data is calculated using a three-dimensional spatial grid method to obtain the water molecule density map.

4. The 3D molecular generation method based on dynamic conformation guidance according to claim 2, characterized in that, The dynamic characteristics of the water network are quantified using the hydrogen bond lifetime autocorrelation function to obtain the hydrogen bond network.

5. The 3D molecular generation method based on dynamic conformation guidance according to claim 2, characterized in that, A multi-level equilateral graph attention network architecture is used to process the dynamic conformation set of proteins to obtain dynamic conditional vectors, including: The configuration state is EGAT encoded to obtain the global feature vector of the configuration; The water molecule density map is subjected to multi-layer convolution using a 3D convolutional neural network to obtain a solvated global feature vector. The hydrogen bond network is temporally encoded using LSTM to obtain the global feature vector of hydrogen bonds; The global feature vector of the conformation, the global feature vector of solvation, and the global feature vector of hydrogen bonding are spliced ​​together from multiple sources to obtain a multi-source spliced ​​feature vector. The multi-source spliced ​​feature vectors are subjected to multi-scale gating fusion to obtain the dynamic condition vector.

6. The 3D molecular generation method based on dynamic conformation guidance according to claim 5, characterized in that, The configurational state is EGAT encoded to obtain the global feature vector of the configuration, including: An isomorphic map is constructed based on the aforementioned conformational state; The isomorphic graph is encoded using a graph attention mechanism to obtain the global feature vector of the conformation.

7. The method for generating 3D molecules based on dynamic conformation guidance according to any one of claims 1 to 6, characterized in that, Generating 3D molecules based on the isotropic diffusion backbone and the dynamic conditional vector includes: Generate random noise molecules; The random noise molecules are denoised using an isovariant diffusion backbone to generate a molecular skeleton that conforms to 3D geometric rules; the dynamic condition vector is injected during the denoising process. Guide the molecular intermediate structure to adapt to the target dynamic conformation through conformational attention gating. The intermediate structure of the molecule is subjected to multi-objective adaptive optimization to balance static binding and dynamic stability in order to generate a 3D molecule.

8. The 3D molecular generation method based on dynamic conformation guidance according to claim 7, characterized in that, The accompanying noise molecules are denoised multiple times using an isotropic diffusion backbone; the dynamic condition vector is injected in each denoising process.

9. The 3D molecular generation method based on dynamic conformation guidance according to claim 7, characterized in that, The composite loss function for multi-objective adaptive optimization of the aforementioned molecular intermediate structure is: in, This is the total loss function; For static docking loss; λ static This represents the weighting coefficient for static docking loss; For dynamic molecular dynamics loss; λ dyn This represents the weighting coefficient for dynamic molecular dynamics loss.

10. A 3D molecular generation system based on dynamic conformation guidance, characterized in that, include: The dynamic conformation set acquisition module is used to obtain the dynamic conformation set of proteins through molecular dynamics simulation. The dynamic condition vector acquisition module is used to process the dynamic conformation set of proteins using a multi-level equilateral graph attention network architecture to obtain dynamic condition vectors. A 3D molecule generation module is used to generate 3D molecules based on an isovariant diffusion backbone and the dynamic condition vector.

Images