Protein-ligand binding pose prediction method based on bilinear attention network

Abstract

The present invention provides a protein-ligand binding posture prediction method based on a bilinear attention network, constructs a mathematical model that can capture the interaction between protein and ligand and maintain the physical and chemical constraints in the molecule, uses a bilinear attention network to learn the deep interaction between protein and ligand, and models the physical constraints into the model through a triangular constraint model to assist in the feature learning of protein and ligand. In addition, since the interaction information contains information about protein and ligand, the present invention uses a gated self-attention mechanism to learn the information of individual proteins and ligands from the interaction features, achieving good performance in the protein-ligand docking posture prediction task and improving the interpretability of protein and ligand feature learning. Compared with existing technologies, the present invention can improve the performance and robustness of protein-ligand docking posture prediction, thereby providing suggestions and references for drug development.

Description

Technical Field

[0001] The present invention relates to the fields of computational biology and deep learning technology, and more specifically, to a method for predicting protein-ligand binding posture based on a bilinear attention network. Background Art

[0002] Over the past few decades, great efforts have been made to study the structure and function of proteins. The structure of biological entities has a significant impact on their function, however, predicting the structure of biological entities from both experimental and computational aspects faces huge challenges. The emergence of the AlphaFold deep learning algorithm has greatly promoted the development of the field of structural biology, which has had a profound impact on drug discovery. In the drug discovery process, screening ligands through wet experiments is costly and time-consuming, and computational prediction of the binding affinity between proteins and ligands can facilitate drug screening. However, the simple binding affinity value cannot fully explain the interaction between proteins and ligands. In order to more comprehensively study the mechanism of protein-ligand interaction, it is necessary to predict the docking posture of proteins and ligands, but this is also much more difficult than predicting their binding affinity.

[0003] Protein-ligand docking is a widely used computational method for predicting and analyzing interactions between proteins and small molecule ligands, playing a key role in drug discovery and virtual drug screening. Traditional docking methods employ physics-based scoring functions and search algorithms to explore chemical space and complete the docking process. For example, AutoDock Vina utilizes efficient search algorithms and scoring functions to explore the conformational space of ligands. SMINA combines optimized search algorithms and scoring functions to accurately predict binding modes. GNINA enhances SMINA by incorporating a learned three-dimensional convolutional neural network (3D CNN) for scoring. QVina-W is a blind docking software that builds on the speed-optimized QuickVina 2 with advanced algorithms to efficiently explore ligand binding modes. GLIDE is a docking method that combines rough positioning, torsional flexibility energy optimization, and Monte Carlo sampling to achieve accurate ligand docking. However, these traditional docking methods rely heavily on exhaustive conformational sampling of ligands and proteins within the vast chemical space, which is computationally time-consuming and resource-intensive.

[0004] Traditional methods for predicting docking pose structures mostly rely on physics-inspired scoring functions and extensive conformational sampling. Some deep learning methods focus on learning more accurate scoring functions, but due to their use of a sampling scoring framework, their inference speed is often slower. Tankbind reduces the conformational sampling burden by predicting a protein-ligand distance matrix and then using an optimization algorithm to convert the distance matrix into docking pose structures. However, the predicted protein-ligand distance matrix may not be a valid Euclidean distance matrix, which can lead to problems in the second-stage conversion process. Deepdock uses the distance between protein and ligand to construct an energy function based on distance likelihood. It finds the lowest energy based on distance and then uses rotatable bonds, dihedral angles, etc. to generate docking structures based on distance. The physical constraints of the molecules can make the generated structure more consistent with the real situation, but the equivariance of the protein and ligand is still not maintained, resulting in the inability to fully learn the 3D spatial information of the protein and ligand. Equibind is an equivariant model that directly predicts the structural coordinates of the binding pose. It uses a graph neural network to refine the ligand conformation and then uses a key point alignment mechanism to rotate and translate the refined ligand into the pocket. Its speed is improved, but the prediction accuracy is not accurate enough. Tankbind and Equibind both make predictions in Cartesian coordinate space. Since each atomic coordinate is predicted separately, this may cause the molecules to fall apart or not conform to the physical and chemical constraints in the final coordinate generation results. These models do not fully consider the physical and chemical constraints in the molecules and the deep interactions between proteins and ligands.

[0005] In addition, the existing patent documents also disclose a three-dimensional protein-ligand activity prediction method based on the attention mechanism. By constructing a molecular three-dimensional information extraction network, molecular characterization is performed, and the three-dimensional information of the protein-ligand binding posture is extracted by combining the spherical information transfer method and physically meaningful features; by constructing a three-dimensional graph neural network model based on the attention mechanism, protein-ligand interactions are predicted, including conformation prediction and binding affinity prediction, and finally the prediction results are output; since proteins and molecules are a very complex dynamic system, maintaining the physical and chemical constraints in the system when modeling protein and molecular systems is a huge challenge. The model in the existing technology still cannot accurately capture the interaction between proteins and ligands and maintain the physical and chemical constraints in the molecules. Summary of the Invention

[0006] In order to overcome the defects of the above-mentioned prior art in inaccurate prediction due to insufficient consideration of the physicochemical constraints in the molecules and the deep interactions between proteins and ligands, the present invention provides a protein-ligand binding posture prediction method based on a bilinear attention network, which can learn the deep interactions between proteins and ligands, as well as the physicochemical constraints, and further predict the docking posture of proteins and ligands, thereby providing suggestions and references for drug development.

[0007] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0008] A protein-ligand binding pose prediction method based on a bilinear attention network, comprising the following steps:

[0009] S1: Obtain initial information of protein and ligand;

[0010] The protein initial information includes: a molecular map of the protein and atomic chemical characteristics of the protein;

[0011] The initial information of the ligand includes: a molecular diagram of the ligand and atomic chemical characteristics of the ligand;

[0012] S2: performing feature extraction on the initial information of the protein and the initial information of the ligand respectively, and obtaining the atomic embedding of the protein and the atomic embedding of the ligand respectively;

[0013] S3: The atomic embedding of the protein and the atomic embedding of the ligand are jointly input into a pre-trained bilinear attention network to obtain the deep interaction features between the protein and the ligand;

[0014] S4: adding triangular physical constraints to the deep interaction features between the protein and the ligand to obtain the interaction features after adding the triangular physical constraints;

[0015] S5: Inputting the interaction features after adding the triangular physical constraints into a pre-trained gated self-attention network for information extraction to obtain protein extraction information, ligand extraction information, and protein-ligand interaction extraction information;

[0016] S6: Repeat steps S4 to S5 several times, and perform bilinear information fusion on the protein extraction information, ligand extraction information, protein-ligand interaction extraction information, protein atomic embedding, and ligand atomic embedding obtained last time to obtain fusion features;

[0017] S7: calculating a distance matrix between the protein and the ligand and a confidence level of binding between the protein and the ligand based on the fusion features;

[0018] S8: Obtain several binding pockets on the protein surface for binding to the ligand, obtain the optimal binding pocket based on the confidence of the binding between the protein and the ligand, and obtain the docking posture of the protein and the ligand based on the distance matrix between the protein and the ligand to complete the prediction of the binding posture of the protein and the ligand.

[0019] Preferably, in step S2, a pre-trained graph vector network (GVP) is used to extract features from protein information to obtain the atomic embedding M of the protein. p ,satisfy: in, is the atomic embedding of the jth protein, d is the embedding dimension, and q is the number of atoms in the protein;

[0020] The pre-trained graph isomorphism network GIN is used to extract the features of the initial ligand information and obtain the atomic embedding M of the ligand. l ,satisfy: in, is the atomic embedding of the i-th ligand, and s is the number of atoms in the ligand.

[0021] Preferably, in step S3, the atoms of the protein are embedded into M p and the atoms of the ligand are embedded in M l The two are jointly input into the pre-trained bilinear attention network to obtain the deep interaction features z between protein and ligand ij , specifically:

[0022]

[0023] Among them, z ij is the deep interaction feature between the i-th ligand and the j-th protein; is the total weight parameter, and are the weight parameters of protein and ligand, respectively, satisfying and is the Hadamard product; σ(·) is the activation function.

[0024] Preferably, step S4 includes:

[0025] Get the distance between atoms inside the protein and the distances between atoms within the ligand The pre-trained triangle constraint model is expressed as:

[0026]

[0027]

[0028] gate(zij )=σ(MLP(z ij ))

[0029] in, is a triangular physical constraint; and are z ij The first and second gate linear transformations; MLP is a multi-layer perceptron; gate is a gate function;

[0030] Deep interaction characteristics of the protein and ligand ij Add corresponding triangle physical constraints Get the interaction feature z′ after adding triangular physical constraints ij :

[0031]

[0032] Among them, gate is the gate function.

[0033] Preferably, step S5 includes:

[0034] The interaction feature z′ after adding the triangular physical constraint ij Input into a pre-trained gated self-attention network for information extraction; the pre-trained gated self-attention network includes a self-attention mechanism and a gate mechanism;

[0035] According to the self-attention mechanism, the interaction feature z′ after adding the triangular physical constraint is obtained. ij The linear mapping Q, K and V are used to obtain protein extraction information according to the following formula Ligand extraction information

[0036]

[0037] g ij =gate(z′ ij )

[0038] in, V is the information extracted from the interaction between protein and ligand; split(·) is a preset function that divides the data into three equal parts; is the interaction feature z′ after adding triangular physical constraints ij Linear mapping V;

[0039] The interaction extraction information z″ between protein and ligand is obtained according to the following formula ij :

[0040]

[0041]

[0042] in, is the interaction feature z′ after adding triangular physical constraints ij The linear mapping Q of is the interaction feature z′ after adding triangular physical constraints ij The linear map K.

[0043] Preferably, step S6 includes:

[0044] The protein extraction information obtained last time Ligand extraction information Extracting information about protein-ligand interactions ij , and the protein's atomic embedding M p and the atoms of the ligand are embedded in M l According to the following formula, bilinear information fusion is performed to obtain the fusion feature h:

[0045]

[0046] Where V is the preset protein parameter, which satisfies: U is the preset ligand parameter, which satisfies:

[0047] Preferably, in step S7, the distance matrix between the protein and the ligand is calculated according to the fusion feature h using the following formula:

[0048]

[0049] Among them, h ij is the fusion feature of the i-th ligand and the j-th protein;

[0050] According to the fusion feature h, the confidence of protein-ligand binding is calculated using the following formula:

[0051]

[0052] Among them, h ij is the fusion feature of the i-th ligand and the j-th protein.

[0053] Preferably, step S8 includes:

[0054] The P2Rank algorithm is used to obtain several binding pockets on the protein surface for binding to the ligand, and the first total loss function is set according to the confidence of the binding between the protein and the ligand. Obtain the optimal binding pocket;

[0055] Set the second total loss function according to the distance matrix between the protein and the ligand Iterative optimization is used to obtain the docking posture of the protein and the ligand, and to complete the prediction of the binding posture of the protein and the ligand.

[0056] Preferably, the first total loss function Includes: distance loss function and confidence loss function

[0057] The distance loss function for:

[0058]

[0059] Among them, δ(r) is a preset function. The smaller the distance between the ligand and the real binding pocket, the larger the value of δ(r), and 0<δ(r)≤1; D ij is the actual distance between protein and ligand when bound;

[0060] The confidence loss function for:

[0061]

[0062] Among them, ε is the preset marginal value;

[0063] The first total loss function for:

[0064]

[0065] Among them, α is the first hyperparameter.

[0066] Preferably, the second total loss function Including: distance loss function between molecules and local structure loss function

[0067] The intermolecular distance loss function for:

[0068]

[0069]

[0070] in, is the coordinate of the ith predicted ligand; is the j-th predicted protein coordinate;

[0071] The local structure loss function for:

[0072]

[0073]

[0074] in, is the distance matrix of ligand atoms; is the j-th predicted ligand coordinate; is the k-th predicted ligand coordinate;

[0075] The second total loss function include:

[0076]

[0077] Among them, β is the second hyperparameter.

[0078] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0079] The present invention provides a protein-ligand binding posture prediction method based on a bilinear attention network, which first obtains protein initial information and ligand initial information; performs feature extraction on the protein initial information and ligand initial information respectively, and obtains the atomic embedding of the protein and the atomic embedding of the ligand accordingly; inputs the atomic embedding of the protein and the atomic embedding of the ligand into a pre-trained bilinear attention network to obtain the deep interaction features between the protein and the ligand; adds triangular physical constraints to the deep interaction features between the protein and the ligand to obtain the interaction features after adding the triangular physical constraints; inputs the interaction features after adding the triangular physical constraints into a pre-trained gated self-attention network for Perform information extraction to obtain protein extraction information, ligand extraction information, and protein-ligand interaction extraction information; repeat the "triangular physical constraint and gated self-attention" steps several times, and perform bilinear information fusion on the protein extraction information, ligand extraction information, protein-ligand interaction extraction information, and protein atomic embedding and ligand atomic embedding obtained last time to obtain fusion features; calculate the distance matrix between the protein and the ligand and the confidence level of the protein-ligand binding based on the fusion features, obtain the docking posture of the protein and the ligand based on the distance matrix between the protein and the ligand and the confidence level of the protein-ligand binding, and complete the prediction of the protein-ligand binding posture;

[0080] The present invention constructs a mathematical model that can capture the interaction between proteins and ligands and maintain the physical and chemical constraints in the molecules. It uses a bilinear attention network to learn the deep interaction between proteins and ligands, and models the physical constraints into the model through a triangular constraint model to assist in the feature learning of proteins and ligands. In addition, since the interaction information contains information about proteins and ligands, the present invention uses a gated self-attention mechanism to learn the information of individual proteins and ligands from the interaction features, achieving good performance in the task of protein and ligand docking pose prediction and improving the interpretability of protein and ligand feature learning. Compared with existing technologies, the present invention can improve the performance and robustness of protein and ligand docking pose prediction, thereby providing suggestions and references for drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 Flowchart of a method for predicting protein-ligand binding posture based on a bilinear attention network provided in Example 1.

[0082] Figure 2 This is an architecture diagram of a protein-ligand binding posture prediction method based on a bilinear attention network provided in Example 2.

[0083] Figure 3 This is a schematic diagram of the triangular constraint model provided in Example 2.

[0084] Figure 4 This is a diagram of the gated self-attention network structure provided in Example 2.

[0085] Figure 5 This is a comparison chart of the RMSD quantile ablation experiment results provided in Example 2.

[0086] Figure 6 This is a comparison chart of the percentage ablation experimental results with RMSD less than 2° and 5° provided in Example 2. DETAILED DESCRIPTION

[0087] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0088] In order to better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size;

[0089] It is understandable to those skilled in the art that some well-known structures and descriptions thereof may be omitted in the drawings.

[0090] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0091] Example 1

[0092] like Figure 1 As shown, this embodiment provides a method for predicting protein-ligand binding posture based on a bilinear attention network, comprising the following steps:

[0093] S1: Obtain initial information of protein and ligand;

[0094] The protein initial information includes: a molecular map of the protein and atomic chemical characteristics of the protein;

[0095] The initial information of the ligand includes: a molecular diagram of the ligand and atomic chemical characteristics of the ligand;

[0096] S2: performing feature extraction on the initial information of the protein and the initial information of the ligand respectively, and obtaining the atomic embedding of the protein and the atomic embedding of the ligand respectively;

[0097] S3: The atomic embedding of the protein and the atomic embedding of the ligand are jointly input into a pre-trained bilinear attention network to obtain the deep interaction features between the protein and the ligand;

[0098] S4: adding triangular physical constraints to the deep interaction features between the protein and the ligand to obtain the interaction features after adding the triangular physical constraints;

[0099] S5: Inputting the interaction features after adding the triangular physical constraints into a pre-trained gated self-attention network for information extraction to obtain protein extraction information, ligand extraction information, and protein-ligand interaction extraction information;

[0100] S6: Repeat steps S4 to S5 several times, and perform bilinear information fusion on the protein extraction information, ligand extraction information, protein-ligand interaction extraction information, protein atomic embedding, and ligand atomic embedding obtained last time to obtain fusion features;

[0101] S7: calculating a distance matrix between the protein and the ligand and a confidence level of binding between the protein and the ligand based on the fusion features;

[0102] S8: Obtain several binding pockets on the protein surface for binding to the ligand, obtain the optimal binding pocket based on the confidence of the binding between the protein and the ligand, and obtain the docking posture of the protein and the ligand based on the distance matrix between the protein and the ligand to complete the prediction of the binding posture of the protein and the ligand.

[0103] In the specific implementation process, the first step is to obtain a dataset. The commonly used dataset is the PDBbind dataset, which includes structural data collected from the Protein Data Bank (PDB) and related experimental measurements. The PDBbind dataset also provides structural information about the protein-ligand complex, including the atomic coordinates of the protein and the structural and chemical information of the ligand. In addition, the PDBbind dataset contains the experimentally determined binding affinity of the protein-ligand complex.

[0104] Then, feature extraction is performed on the initial protein information and ligand information in the PDBbind dataset, respectively, to obtain the atomic embedding of the protein and the atomic embedding of the ligand. In this embodiment, this step extracts the two pieces of information of each protein and ligand, each of which has its own molecular graph. The molecular topology graph includes multiple nodes (atoms) and contains chemical features. Then, a graph neural network is used to learn the chemical features and the topological information of the molecular graph based on the molecular graph to obtain a preliminary representation, namely the atomic embedding.

[0105] The atomic embeddings of the protein and the ligand are jointly input into a pre-trained bilinear attention network to obtain deep interaction features between the protein and the ligand. This embodiment uses a bilinear attention network to capture the interaction between the ligand and the protein. The protein and the ligand are first mapped to the same feature space, and then the Hadamard product is used to obtain the interaction features between the protein and the ligand.

[0106] Add triangular physical constraints to the obtained deep interaction features between the protein and the ligand to obtain the interaction features after adding triangular physical constraints. When the protein binds to the ligand, the distances between the atoms in the protein and the distances between the atoms in the ligand are fixed. Therefore, it is important to consider this constraint in the model. To achieve this goal, this example uses a triangular constraint layer for modeling.

[0107] The interaction features after adding triangular physical constraints are then input into a pre-trained gated self-attention network for information extraction to obtain protein extraction information, ligand extraction information, and protein-ligand interaction extraction information.

[0108] This example uses gated self-attention to capture deep interaction features between proteins and ligands. Since protein-ligand interaction features are obtained through bilinear attention between the protein and ligand, they also contain information about the protein and ligand themselves. However, traditional self-attention cannot simultaneously learn the individual information of the protein and ligand from the protein-ligand interaction features. Therefore, traditional self-attention is modified by adding a gate to help the model learn the individual information of the protein and ligand from the protein-ligand interaction features. In gated self-attention, adding a gate mechanism to self-attention can learn the individual features of the protein and ligand from the interaction features and output a vector containing their respective information.

[0109] Repeat the process of adding triangular physical constraints and gated self-attention network extraction several times, and perform bilinear information fusion on the last obtained protein extraction information, ligand extraction information, protein-ligand interaction extraction information, and protein atomic embedding and ligand atomic embedding to obtain fused features. After deep interaction features and adding physical and chemical constraints, the interaction features are used as weight information, and the information of individual proteins and ligands extracted from the interaction features is extracted to further fuse the protein and ligand features.

[0110] Calculate the distance matrix between protein and ligand and the confidence of protein-ligand binding based on the fusion features;

[0111] Obtain several binding pockets on the protein surface for ligand binding, and obtain the optimal binding pocket based on the confidence level of the binding between the protein and the ligand. In protein-ligand docking, there may be multiple potential pockets on the protein surface. To solve this problem, this example uses the P2Rank algorithm to generate several possible binding pockets on the protein surface. However, after the ligand binds to one pocket, the possibility of the ligand binding to other pockets is greatly reduced. To address this issue, this method uses confidence level to consider the probability of the ligand binding to each binding pocket.

[0112] Finally, the docking posture of the protein and ligand is obtained according to the distance matrix between the protein and the ligand, completing the prediction of the binding posture of the protein and the ligand;

[0113] The process of this method is to input the characteristics of the protein and molecule, learn the final representation through the model, and then predict the distance matrix between the protein and molecule based on the representation. However, the distance matrix is ​​not the final ligand pose, but only represents the distance between the two. Therefore, it is necessary to optimize the obtained docking pose to obtain the most suitable binding pose, and together with the optimal binding pocket, jointly characterize the binding process between the protein and the ligand.

[0114] This method constructs a mathematical model that can capture the interaction between proteins and ligands and maintain the physical and chemical constraints in the molecules. It uses a bilinear attention network to learn the deep interaction between proteins and ligands, and models the physical constraints into the model through a triangular constraint model to assist in the feature learning of proteins and ligands. In addition, since the interaction information contains information about proteins and ligands, this method uses a gated self-attention mechanism to learn the information of individual proteins and ligands from the interaction features, achieving good performance in the task of protein-ligand docking pose prediction and improving the interpretability of protein-ligand feature learning. Compared with existing technologies, this method can improve the performance and robustness of protein-ligand docking pose prediction, thereby providing suggestions and references for drug development.

[0115] Example 2

[0116] This embodiment provides a method for predicting protein-ligand binding posture based on a bilinear attention network, comprising the following steps:

[0117] S1: Obtain initial information of protein and ligand;

[0118] The protein initial information includes: a molecular map of the protein and atomic chemical characteristics of the protein;

[0119] The initial information of the ligand includes: a molecular diagram of the ligand and atomic chemical characteristics of the ligand;

[0120] S2: performing feature extraction on the initial information of the protein and the initial information of the ligand respectively, and obtaining the atomic embedding of the protein and the atomic embedding of the ligand respectively;

[0121] S3: The atomic embedding of the protein and the atomic embedding of the ligand are jointly input into a pre-trained bilinear attention network to obtain the deep interaction features between the protein and the ligand;

[0122] S4: adding triangular physical constraints to the deep interaction features between the protein and the ligand to obtain the interaction features after adding the triangular physical constraints;

[0123] S5: Inputting the interaction features after adding the triangular physical constraints into a pre-trained gated self-attention network for information extraction to obtain protein extraction information, ligand extraction information, and protein-ligand interaction extraction information;

[0124] S6: Repeat steps S4 to S5 several times, and perform bilinear information fusion on the protein extraction information, ligand extraction information, protein-ligand interaction extraction information, protein atomic embedding, and ligand atomic embedding obtained last time to obtain fusion features;

[0125] S7: calculating a distance matrix between the protein and the ligand and a confidence level of binding between the protein and the ligand based on the fusion features;

[0126] S8: Obtain several binding pockets on the protein surface for binding to the ligand, obtain the optimal binding pocket based on the confidence of the binding between the protein and the ligand, and obtain the docking posture of the protein and the ligand based on the distance matrix between the protein and the ligand, thereby completing the prediction of the binding posture of the protein and the ligand;

[0127] In step S2, the pre-trained graph vector network (GVP) is used to extract protein information and obtain the protein atomic embedding M. p ,satisfy: in, is the atomic embedding of the jth protein, d is the embedding dimension, and q is the number of atoms in the protein;

[0128] The pre-trained graph isomorphism network GIN is used to extract the features of the initial ligand information and obtain the atomic embedding M of the ligand. l ,satisfy: in, is the atomic embedding of the i-th ligand, and s is the number of atoms of the ligand;

[0129] In step S3, the atoms of the protein are embedded into M p and the atoms of the ligand are embedded in M l The two are jointly input into the pre-trained bilinear attention network to obtain the deep interaction features z between protein and ligand ij , specifically:

[0130]

[0131] Among them, z ij is the deep interaction feature between the i-th ligand and the j-th protein; is the total weight parameter, and are the weight parameters of protein and ligand, respectively, satisfying and is the Hadamard product; σ(·) is the activation function;

[0132] The step S4 comprises:

[0133] Get the distance between atoms inside the protein and the distances between atoms within the ligand The pre-trained triangle constraint model is expressed as:

[0134]

[0135]

[0136] gate(z ij )=σ(MLP(z ij ))

[0137] in, is a triangular physical constraint; and are z ij The first and second gate linear transformations; MLP is a multi-layer perceptron; gate is a gate function;

[0138] Deep interaction characteristics of the protein and ligand ij Add corresponding triangle physical constraints Get the interaction feature z′ after adding triangular physical constraints ij :

[0139]

[0140] Among them, gate is the gate function;

[0141] The step S5 comprises:

[0142] The interaction feature z′ after adding the triangular physical constraint ij Input into a pre-trained gated self-attention network for information extraction; the pre-trained gated self-attention network includes a self-attention mechanism and a gate mechanism;

[0143] According to the self-attention mechanism, the interaction feature z′ after adding the triangular physical constraint is obtained. ij The linear mapping Q, K and V are used to obtain protein extraction information according to the following formula Ligand extraction information

[0144]

[0145] g ij =gate(z′ ij )

[0146] in, V is the information extracted from the interaction between protein and ligand; split(·) is a preset function that divides the data into three equal parts; is the interaction feature z′ after adding triangular physical constraints ij Linear mapping V;

[0147] The interaction extraction information z″ between protein and ligand is obtained according to the following formula ij :

[0148]

[0149]

[0150] in, is the interaction feature z′ after adding triangular physical constraints ij The linear mapping Q of is the interaction feature z′ after adding triangular physical constraints ij Linear mapping K;

[0151] The step S6 comprises:

[0152] The protein extraction information obtained last time Ligand extraction information Extracting information about protein-ligand interactions ij , and the protein's atomic embedding M p and the atoms of the ligand are embedded in M l According to the following formula, bilinear information fusion is performed to obtain the fusion feature h:

[0153]

[0154] Where V is the preset protein parameter, which satisfies: U is the preset ligand parameter, which satisfies:

[0155] In step S7, the distance matrix between the protein and the ligand is calculated according to the fusion feature h using the following formula:

[0156]

[0157] Among them, h ij is the fusion feature of the i-th ligand and the j-th protein;

[0158] According to the fusion feature h, the confidence of protein-ligand binding is calculated using the following formula:

[0159]

[0160] Among them, h ij is the fusion feature of the i-th ligand and the j-th protein;

[0161] The step S8 comprises:

[0162] The P2Rank algorithm is used to obtain several binding pockets on the protein surface for binding to the ligand, and the first total loss function is set according to the confidence of the binding between the protein and the ligand. Obtain the optimal binding pocket;

[0163] Set the second total loss function according to the distance matrix between the protein and the ligand Iterative optimization obtains the docking posture of protein and ligand, and completes the prediction of protein and ligand binding posture;

[0164] The first total loss function Includes: distance loss function and confidence loss function

[0165] The distance loss function for:

[0166]

[0167] Among them, δ(r) is a preset function. The smaller the distance between the ligand and the real binding pocket, the larger the value of δ(r), and 0<δ(r)≤1; D ij is the actual distance between protein and ligand when bound;

[0168] The confidence loss function for:

[0169]

[0170] Among them, ∈ is the preset marginal value;

[0171] The first total loss function for:

[0172]

[0173] Wherein, α is the first hyperparameter. In this embodiment, α=1. The first loss function is used to predict the distance matrix between proteins and molecules.

[0174] The second total loss function Including: distance loss function between molecules and local structure loss function

[0175] The intermolecular distance loss function for:

[0176]

[0177]

[0178] in, is the coordinate of the ith predicted ligand; is the j-th predicted protein coordinate; The loss of distance between molecules is taken into account, that is, the distance between the atoms of the optimized small molecule and the protein atoms should be as close as possible to the distance between the two in the predicted distance matrix;

[0179] The local structure loss function for:

[0180]

[0181]

[0182] in, is the distance matrix of ligand atoms, obtained using RDKit; is the j-th predicted ligand coordinate; is the k-th predicted ligand coordinate; The distance between atoms in the generated small molecule should be considered to be as similar as possible to the internal distance of the original small molecule;

[0183] The second total loss function include:

[0184]

[0185] Among them, β is the second hyperparameter; the second loss function is used to obtain the final coordinates using an optimization method.

[0186] In the specific implementation process, Figure 2 As shown, first obtain the initial information of protein and ligand, that is, obtain the molecular graph of protein and atomic chemical characteristics X p , and the molecular diagram of the ligand and atomic chemical characteristics X l ; Molecular diagram of ligand Indicates that the molecular graph of protein is constructed using k-neighbor graph on protein residues. express;

[0187] In order to consider the equivariance of proteins, this embodiment uses the graph vector network (GVP) to learn protein information and obtain the atomic embedding M of the protein. p ,satisfy: in, is the atomic embedding of the jth protein, d is the embedding dimension, and q is the number of atoms in the protein;

[0188] In order to better learn the embedding of each node through the molecular graph of the ligand, this embodiment uses the graph isomorphism network GIN to perform message passing and aggregation on the molecular graph, thereby obtaining the atomic embedding M of the ligand l ,satisfy: in, is the atomic embedding of the i-th ligand, and s is the number of atoms of the ligand;

[0189] Use bilinear attention network to capture the deep interaction between protein and ligand to obtain interaction feature z ij , z ij ∈R q×s The calculation formula is as follows:

[0190]

[0191] Among them, z ij is the deep interaction feature between the i-th ligand and the j-th protein; is the total weight parameter, and are the weight parameters of protein and ligand, respectively, satisfying and is the Hadamard product; σ(·) is the activation function; through the above formula, the bilinear attention interaction can be regarded as first mapping the representation to a shared feature space, and then learning a vector about the Hadamard product and the vector Weighted interactions; this method is used to learn the deep interactions between proteins and ligands;

[0192] When a protein binds to a ligand, the distances between atoms within the protein The distance between the atoms inside the ligand is fixed, so it is important to consider this constraint in the model; in order to be able to model physical constraints into the model, a triangular constraint model is used to add physical constraints to the interaction feature z ij First, we obtain the distances between atoms inside the protein and the distances between atoms within the ligand like Figure 3 As shown, the triangular constraint model is expressed as:

[0193]

[0194]

[0195] gate(z ij )=σ(MLP(z ij ))

[0196] in, is a triangular physical constraint; and are z ij The first and second gate linear transformations; MLP is a multi-layer perceptron; gate is a gate function;

[0197] The distance between the internal atoms of the protein and the ligand is interacted with the embedding of another pair of atoms to obtain the triangular constraint information, and the deep interaction characteristics of the protein and the ligand are analyzed. ij Add corresponding triangle physical constraints Get the interaction feature z′ after adding triangular physical constraints ij , the calculation formula is as follows:

[0198]

[0199] Among them, gate is the gate function, which is a learnable weight parameter used to help integrate the triangular constraint information with the interaction information. The triangular constraint layer uses a set of fixed distance constraints to learn the interaction between the ligand and the protein. These constraints can be determined based on whether the atom is a protein or a ligand and the distance to ensure that the interaction between the ligand and the protein is reasonable, thereby better capturing the interaction information between the protein and the ligand.

[0200] In the interaction features, the information of individual proteins and ligands will be learned at the same time, but the traditional self-attention network cannot learn the individual information. To solve this problem, this method adds a gate on the basis of the traditional self-attention network to help learn the interaction features, protein and ligand information; Figure 4 As shown, the interaction feature z′ after adding triangular physical constraints is obtained according to the self-attention mechanism ij The linear mapping Q, K and V are used to obtain protein extraction information according to the following formula Ligand extraction information

[0201]

[0202] g ij =gate(z′ ij )

[0203] Among them, g ij Implementation of a gating mechanism to capture individual protein and ligand information from interaction features; V is the information extracted from the interaction between protein and ligand; split(·) is a preset function that divides the data into three equal parts; is the interaction feature z′ after adding triangular physical constraints ij Linear mapping V;

[0204] The interaction extraction information z″ between protein and ligand is obtained according to the following formula ij :

[0205]

[0206]

[0207] in, is the self-attention weight of the interaction feature; is the interaction feature z′ after adding triangular physical constraints ij The linear mapping Q of is the interaction feature z′ after adding triangular physical constraints ij Linear mapping K;

[0208] exist Figure 2 In the figure, N indicates that there are N modules, each of which is connected in sequence by "triangle constraint and gated self-attention". In the middle module, only z is used for input and output. Only in the last module, the two w of gated self-attention and the final output z are input into the bilinear information fusion.

[0209] After learning the interaction features of proteins and ligands many times, it is necessary to fuse the weighted protein and ligand information to obtain the final embedding; the protein extraction information obtained last time is Ligand extraction information Extracting information about protein-ligand interactions ij , and the protein's atomic embedding M p and the atoms of the ligand are embedded in M l According to the following formula, bilinear information fusion is performed to obtain the fusion feature h:

[0210]

[0211] Where V is the preset protein parameter, which satisfies: U is the preset ligand parameter, which satisfies: Both V and U are learnable parameters;

[0212] According to the fusion feature h, the distance matrix between protein and ligand is calculated using the following formula

[0213]

[0214] Among them, h ij is the fusion feature of the i-th ligand and the j-th protein;

[0215] According to the fusion feature h, the confidence of protein-ligand binding is calculated using the following formula

[0216]

[0217] Among them, h ijis the fusion feature of the i-th ligand and the j-th protein;

[0218] After obtaining the above distance matrix and confidence, the docking pose of the protein and ligand can be generated. This method obtains the docking pose through two stages. The first stage is to generate the corresponding distance matrix between the protein and the molecule, which is the above step. The second stage is to finally generate the molecular pose, that is, the spatial coordinates of each atom of the molecule, through the distance matrix. This method generates the final spatial coordinates of the small molecule, that is, its pose, from the distance matrix predicted in the first stage through an iterative optimization process. The specific steps are as follows:

[0219] The P2Rank algorithm is used to obtain 10 binding pockets on the protein surface for binding to the ligand, and the first total loss function is set according to the confidence of the binding between the protein and the ligand. Obtain the optimal binding pocket; this method uses confidence loss to help the model accurately identify the pocket that needs to be docked, thereby improving the model's robustness to the binding pocket;

[0220] Finally, the second total loss function is set according to the distance matrix between the protein and the ligand Iterative optimization is performed to obtain the docking pose of the protein and ligand, and the prediction of the binding pose of the protein and ligand is completed. In this embodiment, the RMSD is calculated by comparing the predicted ligand coordinates (i.e., the optimal docking pose) with the actual ligand coordinates to evaluate the prediction performance. The smaller the RMSD, the more accurate the docking.

[0221] In addition, this embodiment also verifies the effectiveness of this method through an ablation experiment. The specific experimental process is as follows:

[0222] 1) Data collection: Experiments were performed on the PDBbind v2020 dataset, which contains 19,443 protein-ligand complexes. The time segmentation strategy described in EquiBind was used, that is, the dataset was divided according to the year of accumulation of protein-ligand complex structures; the training and validation sets included structures accumulated before 2019, while the test set included structures accumulated after 2019; by eliminating a portion of structures that could not be processed using RDKit, the training set included 17,787 structures, the validation set included 968 structures, and the test set included 363 structures; Torchdrug was used to obtain the atomic chemical features X of proteins and ligands respectively. p and X l ;

[0223] 2) Model building: Follow the steps described above to build the model and input the molecular graphs and atomic chemical features of the protein and ligand into the network model;

[0224] 3) Selection of validation methods and evaluation metrics: This experiment uses a five-fold cross-validation approach and uses the root mean square deviation (RMSD) to evaluate model performance. In addition, quantiles and mean values ​​are used to evaluate the model's predictive performance across different ranges of results.

[0225] 4) Parameter setting and effect evaluation: The dimension of protein and ligand node embeddings was set to 128. In the distance matrix-based ligand pose generation, the Adam optimizer was used with a learning rate of 0.2 and 8000 iterations. Since an RMSD below 2° is acceptable, this experiment calculated the percentage of predicted ligand poses with an RMSD below 2° and below 5° as additional evidence of the model's ability to predict ligand poses.

[0226] 5) Remove the bilinear attention, triangle constraint and gate self-attention in this method and conduct ablation experiments respectively. The experimental results are as follows: Figure 5 and 6 As shown, it can be seen that the three modules designed by this method can greatly improve the performance of the protein-ligand docking pose prediction model;

[0227] This method constructs a mathematical model that can capture the interaction between proteins and ligands and maintain the physical and chemical constraints in the molecules. It uses a bilinear attention network to learn the deep interaction between proteins and ligands, and models the physical constraints into the model through a triangular constraint model to assist in the feature learning of proteins and ligands. In addition, since the interaction information contains information about proteins and ligands, this method uses a gated self-attention mechanism to learn the information of individual proteins and ligands from the interaction features, achieving good performance in the task of protein-ligand docking pose prediction and improving the interpretability of protein-ligand feature learning. Compared with existing technologies, this method can improve the performance and robustness of protein-ligand docking pose prediction, thereby providing suggestions and references for drug development.

[0228] The same or similar reference numerals correspond to the same or similar components;

[0229] The terms used in the drawings to describe positional relationships are for illustrative purposes only and should not be construed as limiting this patent;

[0230] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A protein-ligand binding posture prediction method based on a bilinear attention network, characterized in that: The following steps are involved: S1: Obtain initial information of protein and ligand; The protein initial information includes: a molecular map of the protein and atomic chemical characteristics of the protein; The initial information of the ligand includes: a molecular diagram of the ligand and atomic chemical characteristics of the ligand; S2: Feature extraction is performed on the initial protein information and the initial ligand information respectively, and the atomic embedding of the protein and the atomic embedding of the ligand are obtained accordingly, including: Use the pre-trained graph vector network (GVP) to extract protein information and obtain the atomic embedding of the protein ,satisfy: ,in, is the atomic embedding of the jth protein, d is the embedding dimension, and q is the number of atoms in the protein; Use the pre-trained graph isomorphism network GIN to extract the features of the initial ligand information and obtain the atomic embedding of the ligand ,satisfy: ,in, is the atomic embedding of the i-th ligand, and s is the number of atoms of the ligand; S3: The atomic embedding of the protein and the atomic embedding of the ligand are jointly input into a pre-trained bilinear attention network to obtain the deep interaction features between the protein and the ligand; Embedding the atoms of the protein and atomic embedding of ligands The two proteins are jointly input into a pre-trained bilinear attention network to obtain the deep interaction features between proteins and ligands. , specifically: in, is the deep interaction feature between the i-th ligand and the j-th protein; is the total weight parameter, and are the weight parameters of protein and ligand, respectively, satisfying and ; For Hadamard; is the activation function; S4: Adding triangular physical constraints to the deep interaction features between the protein and the ligand to obtain the interaction features after adding the triangular physical constraints, including: Get the distance between atoms inside the protein , and the distances between atoms within the ligand , the pre-trained triangle constraint model is expressed as: in, is a triangular physical constraint; and They are The first and second gate linear transformations; MLP is a multi-layer perceptron; gate is a gate function; Deep characterization of the protein-ligand interaction Add corresponding triangle physical constraints , get the interactive features after adding triangular physical constraints : Among them, gate is the gate function; S5: Inputting the interaction features after adding the triangular physical constraints into a pre-trained gated self-attention network for information extraction to obtain protein extraction information, ligand extraction information, and protein-ligand interaction extraction information; S6: Repeat steps S4 to S5 several times, and perform bilinear information fusion on the protein extraction information, ligand extraction information, protein-ligand interaction extraction information, protein atomic embedding, and ligand atomic embedding obtained last time to obtain fusion features; S7: calculating a distance matrix between the protein and the ligand and a confidence level of binding between the protein and the ligand based on the fusion features; S8: Obtain several binding pockets on the protein surface for binding to the ligand, obtain the optimal binding pocket based on the confidence of the binding between the protein and the ligand, and obtain the docking posture of the protein and the ligand based on the distance matrix between the protein and the ligand to complete the prediction of the binding posture of the protein and the ligand.

2. A protein-ligand binding posture prediction method based on a bilinear attention network according to claim 1, characterized in that: The step S5 comprises: The interactive features after adding triangular physical constraints Input into a pre-trained gated self-attention network for information extraction; the pre-trained gated self-attention network includes a self-attention mechanism and a gate mechanism; Obtaining interactive features after adding triangular physical constraints based on the self-attention mechanism The linear mapping Q, K and V are used to obtain protein extraction information according to the following formula , ligand extraction information : in, V for extracting information about protein-ligand interactions; It is a preset function that divides the data into three equal parts; Interaction features after adding triangular physical constraints Linear mapping V; The interaction extraction information between protein and ligand is obtained according to the following formula : in, Interaction features after adding triangular physical constraints The linear mapping Q of Interaction features after adding triangular physical constraints The linear map K.

3. The method for predicting protein-ligand binding posture based on a bilinear attention network according to claim 2, characterized in that: The step S6 comprises: The protein extraction information obtained last time , ligand extraction information , protein-ligand interaction extraction information , and the atomic embedding of proteins and atomic embedding of ligands According to the following formula, bilinear information fusion is performed together to obtain fusion features : in, is the preset protein parameter, satisfying: ; is the preset ligand parameter, satisfying: .

4. A protein-ligand binding posture prediction method based on a bilinear attention network according to claim 3, characterized in that: In step S7, according to the fusion feature , the distance matrix between protein and ligand is calculated using the following formula : in, is the fusion feature of the i-th ligand and the j-th protein; According to the fusion features , the confidence of protein-ligand binding is calculated using the following formula : in, is the fusion feature of the i-th ligand and the j-th protein.

5. The method for predicting protein-ligand binding posture based on a bilinear attention network according to claim 4, characterized in that: The step S8 comprises: The P2Rank algorithm is used to obtain several binding pockets on the protein surface for binding to the ligand, and the first total loss function is set according to the confidence of the binding between the protein and the ligand. Obtain the optimal binding pocket; Set the second total loss function according to the distance matrix between the protein and the ligand Iterative optimization is used to obtain the docking posture of the protein and the ligand, and to complete the prediction of the binding posture of the protein and the ligand.

6. A protein-ligand binding posture prediction method based on a bilinear attention network according to claim 5, characterized in that: The first total loss function Includes: distance loss function and confidence loss function ; The distance loss function for: in, It is a preset function. The smaller the distance between the ligand and the real binding pocket, The larger the value of ; is the actual distance between protein and ligand when bound; The confidence loss function for: in, is the preset marginal value; The first total loss function for: in, is the first hyperparameter.

7. The method for predicting protein-ligand binding posture based on a bilinear attention network according to claim 5, characterized in that: The second total loss function Including: distance loss function between molecules and local structure loss function ; The intermolecular distance loss function for: in, is the coordinate of the ith predicted ligand; is the j-th predicted protein coordinate; The local structure loss function for: in, is the distance matrix of ligand atoms; is the j-th predicted ligand coordinate; is the k-th predicted ligand coordinate; The second total loss function include: in, is the second hyperparameter.

Images