Structure sensing small molecule generation method and device based on pocket condition diffusion
By employing a structure-aware small molecule generation method based on pocket conditional diffusion, combined with deep learning and reinforcement learning, the shortcomings of existing models in target identification and multi-objective optimization in targeted drug design are addressed, resulting in small molecule compounds with high binding activity, good drug diversity, and synthetic potential.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-03-17
AI Technical Summary
Existing generative models lack the ability to identify target structures, cannot achieve multi-objective collaborative optimization, have poor syntheticability of generated molecules, insufficient drug diversity, and lack physical feedback and reinforcement learning mechanisms, resulting in poor actual drugability of generated molecules.
A structure-aware small molecule generation method based on pocket conditional diffusion is adopted, which combines deep learning, molecular dynamics and reinforcement learning. By extracting target protein structure data, a pocket-ligand generator and affinity scorer model are used to generate candidate molecules that meet the binding requirements of the target protein. The binding energy, ADMET performance and synthetic accessibility of the molecules are optimized by a multi-task discriminator.
It significantly improves the binding activity, drug-likeness, and syntheticity of the generated molecules, enabling efficient molecular design and drug generation, avoiding performance bias, and improving the practicality and druggability of the generated molecules.
Smart Images
Figure CN121687285A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to small molecule generation, and in particular to a structure-sensing method and apparatus for generating small molecules based on pocket conditional diffusion. Background Technology
[0002] Drug development, a crucial intersection of life sciences and artificial intelligence, has traditionally relied on high-throughput screening and molecular docking methods, which suffer from high costs and low efficiency. In recent years, deep learning technologies, through variational autoencoders, generative adversarial networks, graph neural networks, and diffusion models, have enabled data-driven molecular design and chemical space exploration, significantly improving molecular generation efficiency. However, existing generation models focus only on the molecular structure itself, lacking the ability to perceive the three-dimensional space of target protein binding sites, which remains a significant limitation in targeted drug design. Furthermore, optimizing multiple parameters such as molecular activity, solubility, metabolic stability, and syntheticability is a complex, high-dimensional problem. Current algorithms are prone to getting trapped in local optima, struggling to find an equilibrium point, resulting in poor actual drug-like properties of the generated molecules.
[0003] The shortcomings and deficiencies of existing models are mainly reflected in: (1) lack of target structure recognition ability, resulting in low binding accuracy. (2) inability to achieve multi-objective collaborative optimization, resulting in performance imbalance. (3) poor syntheticability of generated molecules and insufficient drug diversity. (4) lack of physical feedback and reinforcement learning mechanisms, resulting in uncontrollable generation results. Summary of the Invention
[0004] To address the aforementioned problems, this invention provides a method and apparatus for generating structure-sensing small molecules based on pocket conditional diffusion.
[0005] This invention provides the following technical solution: a method for generating structure-sensing small molecules based on pocket conditional diffusion, comprising the following steps:
[0006] S1. Extract the target protein structure data, which includes the three-dimensional structure data of the target protein pocket and the target protein pocket embedding vector;
[0007] S2. Input the target protein pocket embedding vector into the pocket-ligand generator to obtain several candidate molecules, including ligand three-dimensional structure data and ligand feature vectors.
[0008] S3. Input the three-dimensional structure data of the target protein pocket and the three-dimensional structure data of the ligand into the affinity scoring model to screen out candidate molecules that meet the affinity requirements.
[0009] Furthermore, the three-dimensional structural data of the target protein pocket includes the spatial coordinates of all atoms contained in the target protein and the corresponding atom types; the target protein pocket embedding vector includes: total number of atoms, number of heavy atoms (non-hydrogen atoms), number of carbon atoms, number of nitrogen atoms, number of oxygen atoms, number of sulfur atoms, percentage of other elements, mean of X-axis coordinate, mean of Y-axis coordinate, mean of Z-axis coordinate, standard deviation of coordinates, and pocket radius.
[0010] Furthermore, the pocket-ligand generator is constructed through the following steps:
[0011] (1) Construction of the conditional coding layer
[0012] The ligand feature vectors and protein pocket embedding vectors of the protein structure in the training dataset are respectively passed through a linear projection layer and mapped to a unified model dimension d to obtain the ligand embedding representation X1 and the protein pocket embedding representation X2.
[0013] The ligand feature vector is obtained by encoding based on the ligand structure;
[0014] The protein pocket embedding vector includes: total number of atoms, number of heavy atoms (non-hydrogen atoms), number of carbon atoms, number of nitrogen atoms, number of oxygen atoms, number of sulfur atoms, percentage of other elements, mean of X-axis coordinate, mean of Y-axis coordinate, mean of Z-axis coordinate, standard deviation of coordinates, and pocket radius;
[0015] (2) It adopts an autoregressive architecture based on Transformer, which includes multiple layers of cross-attention modules;
[0016] In each layer of the cross-attention module, the ligand embedding representation X1 generates the query vector Q by multiplying by the transpose of the weight matrix Wq; the protein pocket embedding representation X2 generates the bond vector K by multiplying by the transpose of the weight matrix Wk and the value vector V by multiplying by the transpose of the weight matrix Wv.
[0017] The query vector Q is multiplied by the transpose of the key vector K, and the cross-attention weights are calculated by performing Softmax normalization. Then, the cross-attention weights are summed with the value vector V and passed through a residual connection and a normalization layer to output the updated ligand embedding vector.
[0018] The updated ligand embedding vectors output by the cross-attention modules of each layer are aggregated to obtain a ligand representation that fully integrates pocket geometry and chemical constraints; and a trained pocket-ligand generator is obtained.
[0019] Furthermore, the affinity scorer model is constructed through the following steps:
[0020] (1) Construct a joint graph based on the training dataset, which includes three-dimensional structure data of protein pockets and corresponding three-dimensional structure data of ligands, as well as the affinity scores of protein pockets and ligands.
[0021] The three-dimensional structural data of the protein pocket includes the spatial coordinates of all atoms contained in the protein and their corresponding atom types;
[0022] The ligand three-dimensional structure data includes the spatial coordinates of all atoms contained in the ligand, their corresponding atom types, and chemical bond connection information;
[0023] The three-dimensional structure data of the protein pocket and the corresponding three-dimensional structure data of the ligand are merged into a unified node set. The node set includes a matrix of unified node coordinates generated by the spatial coordinates of the atoms of the spliced protein pocket and the spatial coordinates of the atoms of the corresponding ligand, and a sequence of unified node types generated by the atomic types of the spliced protein pocket three-dimensional structure data and the atomic types of the corresponding ligand three-dimensional structure data.
[0024] Based on a preset distance threshold, construct ligand internal edge connections, protein pocket internal edge connections, and protein pocket-ligand cross-edge connections, and aggregate all the connection relationships of the above three types of edge connections into a unified edge index matrix;
[0025] Radial basis function distance encoding is performed on the inter-atomic distance of each edge to generate an edge feature matrix;
[0026] Create a protein pocket-ligand union graph, which includes a node set, an edge index matrix, and an edge feature matrix;
[0027] The protein pocket-ligand union graph is input into a graph neural network based on the SE(3) equivariant mechanism, which contains N cascaded equivariant message passing layers;
[0028] The following operations are performed on all levels from level 1 to level N:
[0029] Combining edge features: Based on the node coordinates and the edge index matrix, calculate the Euclidean norm of the atomic distance of each edge to obtain the spatial distance of the edge, and combine it with the edge feature matrix to form the edge input vector;
[0030] Calculate attention weights: Pass the edge input vector through an attention multilayer perceptron, output a scalar attention score, and normalize the score to the [0, 1] interval using Softmax to obtain the attention weights;
[0031] Update node features: while maintaining SE(3) equivariance: for each node, multiply the edge input vector with the corresponding attention weight to obtain a weighted edge input vector, and average the weighted edge input vector with the edge input vector to obtain new node coordinates;
[0032] Update edge features to pass to the next layer; recalculate new edge input vectors based on the new node coordinates;
[0033] After N iterations, the affinity scorer model is obtained.
[0034] Furthermore, the internal edge connections of the ligand are based on the chemical bond connection information of the ligand molecule, with each pair of bonding atoms as an edge;
[0035] The edges inside the protein pocket are connected to calculate the Euclidean distance between all atomic pairs inside the protein pocket, construct a distance matrix, and filter atomic pairs whose distance is less than a threshold and greater than zero.
[0036] The protein pocket-ligand cross-edge connection calculates the Euclidean distance between the protein pocket atom and the ligand atom and constructs a distance matrix, then filters out protein pocket-ligand atom pairs whose distance is less than a threshold.
[0037] A structure-sensing small molecule generation device based on pocket conditional diffusion, comprising at least:
[0038] The extraction module is used to extract target protein structure data, which includes the three-dimensional structure data of the target protein pocket and the target protein pocket embedding vector.
[0039] The generation module is used to input the target protein pocket embedding vector into the pocket-ligand generator to obtain several candidate molecules, including ligand three-dimensional structure data and ligand feature vectors.
[0040] The scoring module inputs the three-dimensional structure data of the target protein pocket and the three-dimensional structure data of the ligand into the affinity scoring model to screen out candidate molecules that meet the affinity requirements.
[0041] An electronic device, comprising:
[0042] One or more processors;
[0043] Memory, used to store one or more programs;
[0044] When the one or more programs are executed by the one or more processors, the one or more processors perform the methods described above.
[0045] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described above.
[0046] The beneficial effects of this invention are as follows:
[0047] By combining deep learning, molecular dynamics, reinforcement learning, and multi-task discriminator technology, a closed-loop process of target-guided molecular generation and evaluation is achieved, thereby automatically generating small molecule compounds with high binding activity, good drug-likeness, and syntheticability. Attached Figure Description
[0048] Figure 1 This is a flowchart of the method of the present invention;
[0049] Figure 2 This is a flowchart of the pocket-ligand generator of the present invention;
[0050] Figure 3 This is a schematic diagram of the affinity scoring model of the present invention;
[0051] Figure 4 This is a schematic diagram of the modules of the device of the present invention;
[0052] Figure 5 A schematic diagram of the candidate molecule structure is obtained for the generated part. Detailed Implementation
[0053] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0054] This invention proposes a structure-aware small molecule generation method based on pocket conditional diffusion. On one hand, this invention introduces a pocket conditional constraint mechanism into the diffusion generation model. Specifically, during molecule generation, the geometric features of the target protein binding pocket are used as conditional input, and the spatial orientation and chemical structure of the generated molecule are controlled through conditional diffusion inversion. Compared with traditional non-structure-aware models, this invention can automatically learn the distribution, polarity, and volume characteristics of amino acid residues in the protein pocket, ensuring a high degree of matching between the generated molecule and the target binding cavity in three-dimensional space, significantly improving the accuracy of binding energy prediction and docking stability. On the other hand, this invention simultaneously optimizes the molecule's binding energy, ADMET performance, synthetic accessibility (SA), and drug-likeness (QED) by introducing a multi-task discriminator and Pareto front optimization strategy. The dynamic weighting of each sub-network of the discriminator can automatically adjust the target proportion according to the training phase, effectively avoiding the performance bias caused by traditional single scalar optimization methods. This significantly improves the practicality and druggability of molecule design, making the generated molecule not only theoretically strong but also pharmacokineticly feasible and production-operable. This invention combines AutoDockVina docking energy feedback with EGNN geometric representation learning during the generation process, fusing energy constraints at the molecular mechanics level with deep learning representations to ensure that the generated molecules are structurally chemically plausible and energy-stable. The generated molecules significantly outperform traditional diffusion models in terms of physical feasibility and structural plausibility. This combination establishes a closed-loop correction mechanism between theoretical predictions and physical realizability, thereby generating small molecule compounds with realistic structures and reasonable energy levels.
[0055] In model training, this invention employs a reinforcement learning (RL) dynamic feedback optimization module, where the output of the multi-task discriminator and the molecular docking results jointly constitute the reward function, updating the generated model parameters in real time during training. Unlike existing diffusion or Transformer models with fixed objective functions, this method forms a "generation-evaluation-optimization-regeneration" loop mechanism, dynamically adjusting the generation direction based on molecular physical properties to avoid mode collapse. Model training stability is improved by approximately 35%; optimal performance is achieved within 50% of the training epochs. This mechanism significantly reduces computational resource consumption and model tuning costs, demonstrating high efficiency and practicality.
[0056] like Figure 2 The pocket-ligand generator of the present invention is constructed through the following steps:
[0057] (1) Construction of the conditional coding layer
[0058] The ligand feature vectors and protein pocket embedding vectors of the protein structure in the training dataset are respectively passed through a linear projection layer and mapped to a unified model dimension d to obtain the ligand embedding representation X1 and the protein pocket embedding representation X2.
[0059] The training dataset consists of 19,037 experimentally determined protein-ligand complexes obtained from the PDBbind-Plus database. Each protein-ligand complex includes: protein structure, ligand structure, and affinity tag.
[0060] The ligand feature vector is obtained by encoding based on the ligand structure;
[0061] The ligand structure was encoded into a 2048-bit ligand feature vector using Morgan fingerprint (ECFP4);
[0062] The protein pocket embedding vector includes: total number of atoms, number of heavy atoms (non-hydrogen atoms), number of carbon atoms, number of nitrogen atoms, number of oxygen atoms, number of sulfur atoms, percentage of other elements, mean of X-axis coordinate, mean of Y-axis coordinate, mean of Z-axis coordinate, standard deviation of coordinates, and pocket radius;
[0063] The feature vectors of protein pockets extracted from the protein structure include: total number of atoms, number of heavy atoms (non-hydrogen atoms), number of carbon atoms, number of nitrogen atoms, number of oxygen atoms, number of sulfur atoms, percentage of other elements, mean of X-axis coordinate, mean of Y-axis coordinate, mean of Z-axis coordinate, standard deviation of coordinates, and pocket radius, totaling 12. All continuous variables are z-score standardized to generate a 2060-dimensional feature vector usable for training.
[0064] (2) An autoregressive architecture based on Transformer is adopted, which includes multiple cross-attention modules (CrossAttention 1 to CrossAttention N);
[0065] In each layer's cross-attention module, the ligand embedding representation X1 (dimension m×d, where m is the ligand sequence length) generates the query vector Q (dimension n×dq) by multiplying it by the transpose of the weight matrix Wq; the protein pocket embedding representation X2 (dimension n×d, where n is the pocket embedding dimension) generates the bond vector K (dimension m×dk) by multiplying it by the transpose of the weight matrix Wk, and generates the value vector V (dimension m×dv) by multiplying it by the transpose of the weight matrix Wv.
[0066] Multiply the query vector Q by the transpose matrix K of the key vector K. T The cross-attention weights are calculated using Softmax normalization. Then, the cross-attention weights are summed with the value vector V and passed through a residual connection and a normalization layer to output the updated ligand embedding vector.
[0067] The updated ligand embedding vectors output by the cross-attention modules of each layer are aggregated to obtain a ligand representation that fully integrates pocket geometry and chemical constraints; and a trained pocket-ligand generator is obtained.
[0068] As Figure 3 , the affinity scoring model of the present invention is constructed through the following steps:
[0069] (1) Construct a joint graph based on the training data set, where the training data set includes the three-dimensional structure data of the protein pocket, the corresponding three-dimensional structure data of the ligand, and the affinity scores of the corresponding protein pocket and ligand;
[0070] The three-dimensional structure data of the protein pocket includes the spatial coordinates of all atoms contained in the protein and the corresponding atom types; <000015=4>The three-dimensional structure data of the ligand includes the spatial coordinates of all atoms contained in the ligand, the corresponding atom types, and the chemical bond connection information;
[0072] Merge the three-dimensional structure data of the protein pocket and the corresponding three-dimensional structure data of the ligand into a unified node set. The node set includes a matrix of unified node coordinates generated by splicing the spatial coordinates of the atoms of the protein pocket and the spatial coordinates of the corresponding atoms of the ligand, and, the atom types of the spliced three-dimensional structure data of the protein pocket and the atom types of the corresponding three-dimensional structure data of the ligand, generating a sequence of unified node types;
[0073] According to the preset distance threshold Construct the internal edge connection of the ligand, the internal edge connection of the protein pocket, and the cross-edge connection between the protein pocket and the ligand, and aggregate all the connection relationships of the above three types of edge connections into a unified edge index matrix;
[0074] The internal edge connection of the ligand is based on the chemical bond connection information of the ligand molecule, and each pair of bonded atoms is used as an edge; for example, aromatic ring C(sp2)-C(sp2) (adjacent carbons in the benzene ring); aliphatic chain: C(sp3)-C(sp3)
[0075] The internal edge connection of the protein pocket calculates the Euclidean distance between all atom pairs inside the protein pocket, constructs a distance matrix, and filters out atom pairs with a distance less than the threshold and greater than zero;
[0076] For all atom pairs (u,v) in the protein, calculate the Euclidean distance:
[0077] d(u,v) = ||Xp[u] - Xp[v]||2 = sqrt(Δx^2 + Δy^2 + Δz^2);
[0078] Form a distance matrix Dp, where Dp[u,v] = d(u,v), and the diagonal is 0. Select a threshold r.
[0079] For all (u,v), if 0 < Dp[u,v] ≤ r, then establish a two-way edge (u→v) and (v→u)
[0080] Protein pocket-ligand cross-edge connections are used to calculate the Euclidean distance between protein pocket atoms and ligand atoms and construct a distance matrix. Protein pocket-ligand atom pairs with a distance less than a threshold are then selected.
[0081] Calculate the distance from all ligand atoms to protein atoms:
[0082] d(i,j)=||Xl[i]-Xp[j]||2;
[0083] Form a distance matrix Dc, where Dc[i,j]=d(i,j).
[0084] Radial basis function (RBF) distance encoding is performed on the interatomic distances of each edge to generate an edge feature matrix;
[0085] Let the number of basis functions k = 50, the cutoff radius be the threshold size, and the center be ck = k·(cutoff / (k-1));
[0086] Let the width σ = cutoff / k. For any distance r, the k-th dimension is:
[0087] φk(r)=exp(-((r-c_k) / σ)^2);
[0088] Use rbf(r) = [φ0(r), ..., φ(k-1)(r)] as the distance encoding for this edge;
[0089] The interatomic distance is calculated using the spatial coordinates of the atoms in the protein pocket and the corresponding spatial coordinates of the atoms in the ligand. The spatial coordinates of the two atoms are a = (x1, y1, z1) and b = (x2, y2, z2), respectively.
[0090] d(a,b)=sqrt((x1-x2)^2+(y1-y2)^2+(z1-z2)^2);
[0091] Create a protein pocket-ligand union graph, which includes a node set, an edge index matrix, and an edge feature matrix;
[0092] The protein pocket-ligand joint graph is input into a graph neural network (EGNN) based on the SE(3) equivariant mechanism, which contains N cascaded equivariant message passing layers;
[0093] The following operations are performed on all levels from level 1 to level N:
[0094] Combining edge features: Based on the node coordinates and the edge index matrix, calculate the Euclidean norm of the atomic distance of each edge to obtain the spatial distance of the edge, and combine it with the edge feature matrix to form the edge input vector;
[0095] Calculate attention weights: Pass the edge input vector through an attention multilayer perceptron, output a scalar attention score, and normalize the score to the [0, 1] interval using Softmax to obtain the attention weights;
[0096] Update node features: while maintaining SE(3) equivariance: for each node, multiply the edge input vector with the corresponding attention weight to obtain a weighted edge input vector, and average the weighted edge input vector with the edge input vector to obtain new node coordinates;
[0097] Update the edge feature to pass it to the next layer; recalculate the new edge input vector based on the new node coordinates;
[0098] After N iterations, the affinity scorer model is obtained.
[0099] After N iterations, the node features and coordinates are gradually refined to capture the geometric (interatomic distance, angle, spatial arrangement of residues and ligands, three-dimensional morphology of the protein pocket, and position of the ligand in the pocket) and chemical information (hydrogen bond donor and acceptor, hydrophobicity, electronegativity, and aromatic interactions) of the pocket-ligand interaction.
[0100] The embodiments of the present invention will be further described below with reference to examples.
[0101] Example 1: This example establishes a feature extraction and pocket condition model targeting the PRRSV nucleocapsid protein (N protein), which can generate small molecule inhibitors that highly match the viral RNA binding site.
[0102] like Figure 1 As shown, this embodiment includes the following steps:
[0103] S1. Extract the structural data of the target protein (PRRSV nucleocapsid protein), which includes the three-dimensional structural data of the target protein pocket and the target protein pocket embedding vector. The three-dimensional structural data of the target protein pocket includes the spatial coordinates of all atoms contained in the target protein and the corresponding atom types. The target protein pocket embedding vector includes: total number of atoms, number of heavy atoms (non-hydrogen atoms), number of carbon atoms, number of nitrogen atoms, number of oxygen atoms, number of sulfur atoms, percentage of other elements, mean of X-axis coordinate, mean of Y-axis coordinate, mean of Z-axis coordinate, standard deviation of coordinates, and pocket radius.
[0104] z-score normalization is applied to all continuous variables of the target protein pocket embedding vector to generate a 3D feature vector;
[0105] S2. Input the target protein pocket embedding vector into the pocket-ligand generator to obtain candidate molecules (some of which are shown below). Figure 5 (As shown); the analysis of candidate molecules is as follows:
[0106]
[0107]
[0108] Note: [A] is the set of candidate molecules generated by the pocket-ligand generator of this invention; [B] is the set of candidate molecules generated by the existing transformer model.
[0109] S3. Input the target protein pocket three-dimensional structure data and ligand three-dimensional structure data (three-dimensional structure data of candidate molecules and ligand feature vectors; in this embodiment, candidate molecules are encoded as 2048-bit ligand feature vectors using Morgan fingerprint (ECFP4)) into the affinity scoring model to screen out the candidate molecule with the highest affinity (-9.2 kcal / mol), whose structural formula is as follows:
[0110]
[0111] The affinity of the optimal ligand molecule generated and screened using the existing transformer model is only -7.3 kcal / mol. The affinity of the ligand molecule screened by this invention is increased by 27%, which is a significant improvement.
[0112] In summary, this invention establishes a feature extraction and pocket condition model using PRRSV nucleocapsid protein (N protein) as the target, which can generate small molecule inhibitors that highly match the viral RNA binding site. It can be seen that this method has good scalability and can be transferred to drug design tasks targeting other viruses or human proteins.
[0113] Therefore, this invention is not only applicable to the research and development of drugs for animal diseases, but can also be extended to the development of human antiviral drugs, tumor target and protein complex inhibitors.
[0114] Example 2
[0115] like Figure 4 A structure-sensing small molecule generation device based on pocket conditional diffusion, comprising at least:
[0116] Extraction module 1 is used to extract target protein structure data, which includes the three-dimensional structure data of the target protein pocket and the target protein pocket embedding vector;
[0117] Generation module 2 is used to input the target protein pocket embedding vector into the pocket-ligand generator to obtain several candidate molecules, including ligand three-dimensional structure data and ligand feature vectors;
[0118] The scoring module 3 inputs the three-dimensional structure data of the target protein pocket and the three-dimensional structure data of the ligand into the affinity scoring model to screen out candidate molecules that meet the affinity requirements.
[0119] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0120] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 application according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0121] Accordingly, this application also provides an electronic device, including:
[0122] One or more processors;
[0123] Memory, used to store one or more programs;
[0124] When the one or more programs are executed by the one or more processors, the one or more processors perform the methods described above.
[0125] Accordingly, this application also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of any of the above methods.
[0126] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can also be implemented in other ways. The method and system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0127] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0128] On the other hand, a computer-readable storage medium stores computer instructions thereon, which, when executed by a processor, implement the steps of the above-described method. When the computer program is executed by the processor, it implements the method as described in any of the first aspects above. If the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, 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, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0129] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A structure-aware small molecule generation method based on pocket condition diffusion, characterized in that, The method comprises the following steps: S1, extracting target protein structure data, including target protein pocket three-dimensional structure data and target protein pocket embedding vector; S2, inputting the target protein pocket embedding vector into a pocket-ligand generator to obtain a plurality of candidate molecules, including ligand three-dimensional structure data and ligand feature vector; S3, inputting the target protein pocket three-dimensional structure data and the ligand three-dimensional structure data into an affinity scorer model to screen out candidate molecules meeting the affinity requirement.
2. The structure-aware small molecule generation method based on pocket condition diffusion according to claim 1, wherein, The target protein pocket three-dimensional structure data comprises spatial coordinates and corresponding atom types of all atoms contained in the target protein; The target protein pocket embedding vector comprises total atom number, heavy atom number (non-hydrogen atom number), carbon atom number, nitrogen atom number, oxygen atom number, sulfur atom number, other element proportion, X-axis coordinate mean value, Y-axis coordinate mean value, Z-axis coordinate mean value, coordinate standard deviation, and pocket radius.
3. The method of claim 1, wherein, The pocket-ligand generator is constructed by the following steps: (1) Condition coding layer construction The ligand feature vector and the protein pocket embedding vector of the protein structure in the training data set are respectively mapped to a unified model dimension through a linear projection layer to obtain ligand embedding representation and protein pocket embedding representation; The ligand feature vector is obtained by coding according to the ligand structure; The protein pocket embedding vector comprises total atom number, heavy atom number, carbon atom number, nitrogen atom number, oxygen atom number, sulfur atom number, other element proportion, X-axis coordinate mean value, Y-axis coordinate mean value, Z-axis coordinate mean value, coordinate standard deviation, and pocket radius; (2) A Transformer-based autoregressive architecture is adopted, which contains a plurality of cross-attention modules; In each layer of the cross-attention module, the ligand embedding representation generates a Query vector Q by multiplying the transpose matrix of the weight matrix Wq; the protein pocket embedding representation generates a key vector K by multiplying the transpose matrix of the weight matrix Wk and generates a value vector V by multiplying the transpose matrix of the weight matrix Wv; The Query vector Q is multiplied by the transpose matrix of the key vector K, and the cross-attention weight is calculated by Softmax normalization, then the cross-attention weight is summed with the value vector V through a residual connection and a normalization layer, and an updated ligand embedding vector is outputted; The updated ligand embedding vectors outputted by each layer of the cross-attention module are aggregated to obtain a ligand representation fully fused with pocket geometry and chemical constraints; and a trained pocket-ligand generator is obtained.
4. The method of claim 1, wherein the method is based on pocket condition diffusion for structure-aware small molecule generation. The affinity scorer model is constructed by the following steps: (1) A joint graph is constructed according to a training data set, the training data set comprising protein pocket three-dimensional structure data and corresponding ligand three-dimensional structure data and affinity scores of the corresponding protein pocket and ligand; The protein pocket three-dimensional structure data comprises spatial coordinates and corresponding atom types of all atoms contained in the protein; The ligand three-dimensional structure data comprises spatial coordinates and corresponding atom types of all atoms contained in the ligand and chemical bond connection information; combining the protein pocket three-dimensional structure data and the corresponding ligand three-dimensional structure data into a unified node set, the node set including spatial coordinates of atoms of the spliced protein pocket and spatial coordinates of atoms of the corresponding ligand, generating a matrix of unified node coordinates, and atom types of the spliced protein pocket three-dimensional structure data and atom types of the corresponding ligand three-dimensional structure data, generating a sequence of unified node types; constructing ligand internal edge connections, protein pocket internal edge connections and protein pocket-ligand cross edge connections according to a preset distance threshold, and aggregating all connection relationships of the three types of edge connections into a unified edge index matrix; radial basis function distance encoding is performed on the interatomic distance of each edge to generate an edge feature matrix; creating a protein pocket-ligand joint graph, which includes the node set, the edge index matrix and the edge feature matrix; inputting the protein pocket-ligand joint graph into a graph neural network based on SE(3) equivariant mechanism, the network including N layers of concatenated equivariant message passing layers; the first layer to the Nth layer all perform the following operations: combining edge features: calculating the Euclidean norm of the interatomic distance of each edge according to the node coordinates and the edge index matrix to obtain the spatial distance of the edge, and combining the edge feature matrix to form an edge input vector; calculating attention weights: passing the edge input vector through an attention multilayer perceptron to output a scalar attention score and normalizing the score to the [0, 1] interval through Softmax to obtain attention weights; updating node features: while maintaining SE(3) equivariance: for each node, multiply the edge input vector by the corresponding attention weight to obtain a weighted edge input vector, and aggregate the weighted edge input vector and the edge input vector to obtain new node coordinates; updating edge features to pass to the next layer: based on the new node coordinates, recalculate the new edge input vector; after N layers of iteration, an affinity scorer model is obtained.
5. The method of claim 4, characterized in that, The ligand internal edge connection is based on the bond connection information of the ligand molecule, and each pair of bonding atoms is regarded as an edge. The protein pocket internal edge connection calculates the Euclidean distance between all pairs of atoms in the protein pocket to construct a distance matrix, and screens the atom pairs with a distance less than a threshold and greater than zero; The protein pocket-ligand cross edge connection calculates the Euclidean distance between the protein pocket atoms and the ligand atoms and constructs a distance matrix, and screens the protein pocket-ligand atom pairs with a distance less than a threshold.
6. A structure-aware small molecule generation device based on pocket conditional diffusion, at least comprising: an extraction module for extracting target protein structure data, including target protein pocket three-dimensional structure data and target protein pocket embedding vectors; a generation module for inputting the target protein pocket embedding vectors into a pocket-ligand generator to obtain a plurality of candidate molecules, including ligand three-dimensional structure data and ligand feature vectors; a scoring module for inputting the target protein pocket three-dimensional structure data and the ligand three-dimensional structure data into an affinity scorer model to screen candidate molecules meeting the affinity requirements.
7. An electronic device, comprising: comprising: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement a method according to any of claims 1-5.
8. A computer readable storage medium having stored thereon computer instructions, wherein, The instructions, when executed by a processor, implement steps of a method according to any of claims 1-5.