Deep learning based protein working conformation determination method

Abstract

The application discloses a protein working conformation determination method based on deep learning, and is used for solving the problems of insufficient screening precision of protein alternative conformation under the condition of limited docking times and docking score deviation caused by inherent properties of ligands; the application obtains a non-dominant alternative conformation set and a calibration ligand set of the same protein, constructs a limited docking score matrix and a ligand knowledge graph, generates a ligand inherent scoring baseline by using a knowledge graph relationship path reasoning network, deducts the ligand inherent scoring baseline in a VAE-based matrix completion model and completes a residual docking score matrix, determines a working conformation according to a conformation applicability score and outputs a three-dimensional structure file; and the application can improve the working conformation determination accuracy under the condition of limited docking budget, and provides an input structure for virtual screening.

Description

Technical Field

[0001] This invention relates to the field of protein working conformation determination technology, and more particularly to a method for protein working conformation determination based on deep learning. Background Technology

[0002] Virtual screening based on protein structure typically requires determining the working conformation of the receptor first. Due to the prevalence of binding-induced adaptation and flexible protein motion, the same protein often corresponds to multiple spatial conformations that may participate in ligand recognition. Different conformations differ in binding pocket geometry, polarity distribution, accessible surfaces, and the exposure status of key residues, leading to variations in molecular docking results. In the drug discovery process, candidate conformations can originate from crystal structures, molecular dynamics sampling, or multi-objective optimization results. Therefore, determining the suitable working conformation from multiple candidate conformations before entering large-scale virtual screening is a crucial step in the receptor preparation phase.

[0003] In existing technologies, common approaches first collect multiple candidate conformations of the same protein, then select a set of known active or representative ligands, perform molecular docking around a pre-defined binding pocket, and choose the conformation to be used for subsequent virtual screening based on the average docking score, ranking results, enrichment performance, or consensus evaluation results. To reduce computational overhead, some approaches only perform docking on a subset of positions in the conformation-ligand combination, and then use molecular descriptors, similarity relationships, matrix factorization, autoencoder models, or graph learning models to infer the unobserved positions. Ligand-side information is typically represented by fragment composition, molecular core, and physicochemical properties, while conformation-side information is typically represented by pocket shape, residue contact, and spatial distribution characteristics. These are then combined with docking observations to complete conformation evaluation and ranking.

[0004] Existing methods often directly use the original docking scores during the conformation comparison stage. The scoring bias caused by differences in ligand structure and the binding differences caused by conformation are often mixed in the same scoring matrix, making it difficult to form a consistent conformation evaluation result. In scenarios where the number of docking attempts is limited, the observed scores are usually sparse and unevenly distributed, with insufficient observations for some conformations and some ligands. Direct averaging, direct sorting, or performing routine completion based on the original scores can easily introduce the inherent tendency of ligands to receive high or low scores into the working conformation selection process, leading to a deviation between the virtual screening input conformations and the actual pocket fit.

[0005] Therefore, a method for determining the working conformation of proteins that can overcome the shortcomings of the prior art is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for determining the working conformation of a protein based on deep learning. The core technical problem to be solved by this application is: how to determine the working conformation that can stably represent the binding pocket adaptability from multiple candidate conformations of the same protein in a virtual screening scenario under the condition of limited molecular docking times, and effectively remove the interference of the inherent scoring bias of the ligand on the conformation selection result during the conformation evaluation process.

[0007] The protein working conformation determination method based on deep learning according to embodiments of the present invention includes:

[0008] S1. Obtain the non-dominant candidate conformations of the same protein, generate a set of candidate conformations, obtain calibration ligands, and generate a set of calibration ligands;

[0009] S2. Under the constraint of the upper limit of docking times, molecular docking is performed on the candidate conformation set and the calibration ligand set to generate a finite docking score matrix;

[0010] S3. Construct a ligand knowledge graph based on the calibration ligand set. The ligand knowledge graph includes ligand nodes, fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes. It also includes relation edges describing the correspondence between ligands and fragments, ligands and molecular nuclei, and ligands and physicochemical property binning. Train the relation path reasoning network of the knowledge graph based on the finite docking scoring matrix to generate the inherent scoring baseline of each calibration ligand. Summarize the inherent scoring baselines of the ligands to generate a ligand baseline table and output the ligand baseline table as a docking scoring decomposition constraint.

[0011] S4. Construct a VAE-based matrix completion model based on the candidate conformation set and the calibration ligand set. In the decoding output and reconstruction error calculation stage of the VAE-based matrix completion model, introduce a ligand baseline table. Perform ligand intrinsic scoring baseline subtraction on the docking scores in the finite docking score matrix to obtain the conformation residual docking score. The VAE-based matrix completion model uses the conformation residual docking score as the learning target to perform matrix completion prediction and generate the residual docking score completion matrix.

[0012] S5. Calculate the conformation suitability score for each candidate conformation based on the residual docking score completion matrix, and generate the conformation suitability score.

[0013] S6. Determine the candidate conformations that meet the maximum score condition based on the conformation suitability score, and generate the working conformation.

[0014] S7. Output the 3D structure file corresponding to the working conformation for virtual screening.

[0015] Optionally, S1 includes:

[0016] Obtain three-dimensional structure files of multiple conformations of the same protein generated by multi-objective optimization, as well as the multi-objective values ​​corresponding to each conformation, to form a conformation candidate library;

[0017] Based on the conformation candidate library, non-dominant determination is performed on multi-objective values ​​and dominated conformations are eliminated to form a non-dominant conformation record.

[0018] The binding pocket region definition is determined based on the non-dominant conformation record. The binding pocket region definition includes reference ligand coordinates or preset key residue coordinates. Based on the binding pocket region definition, the binding pocket neighborhood atoms are extracted from the conformations in the non-dominant conformation record and the file format is unified to generate a set of candidate conformations.

[0019] Obtain the ligand structure library, calculate the ligand structure similarity based on the ligand structure library, and select representative ligands by clustering under the constraint of a preset upper limit on the number of ligands to generate a calibration ligand set.

[0020] Optionally, S2 includes:

[0021] Based on the upper limit constraint of docking times and the set of candidate conformations, the docking search space parameters corresponding to each candidate conformation are determined, forming a search space parameter set;

[0022] Based on the upper limit constraint of docking times, the search space parameter set, the set of candidate conformations and the set of calibration ligands, the required pairings of conformations and ligands for molecular docking are calculated, and a list of docking tasks is generated.

[0023] Based on the docking task list and the search space parameter set, molecular docking is performed on each conformation and ligand pair combination to obtain docking scores and summarize them to form a docking score record set.

[0024] Based on the docking score record set, the docking scores are written into a matrix index with the candidate conformation set as rows and the calibration ligand set as columns. Missing positions not included in the docking task list are marked with missing flags, thus generating a finite docking score matrix.

[0025] Optionally, S3 includes:

[0026] Based on the calibration ligand set, structural analysis is performed on each calibration ligand to determine the corresponding ligand node, fragment node, molecular core node, and physicochemical property binning node, and the node correspondence is generated.

[0027] Based on the node correspondence, a ligand knowledge graph is constructed, with ligand nodes, fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes as graph nodes. A set of relational edges is generated based on the node correspondence, and the set of relational edges is written into the ligand knowledge graph.

[0028] Non-missing docking scores are extracted from the finite docking score matrix, and the non-missing docking scores are aggregated according to the ligand nodes corresponding to the calibration ligands to generate a ligand observation score set.

[0029] Based on the ligand observation score set, a summary statistic of the ligand observation score is calculated for each ligand node to generate a ligand training label set;

[0030] Using the ligand knowledge graph and the ligand training label set as training inputs, the knowledge graph relational path reasoning network is trained so that the knowledge graph relational path reasoning network outputs the inherent scoring baseline of the ligand for each ligand node in the ligand knowledge graph, generating a set of inherent scoring baselines of the ligands.

[0031] Based on the intrinsic scoring baseline set of ligands, the intrinsic scoring baseline of each calibrated ligand is written into the ligand baseline table according to the order of the calibrated ligand set, and the correspondence between the ligand baseline table and the calibrated ligand set forms the docking score decomposition constraint.

[0032] Optionally, S4 includes:

[0033] The row and column indices of the finite docking score matrix are determined based on the candidate conformation set and the calibration ligand set, and a VAE-based matrix completion model is constructed accordingly. This allows the VAE-based matrix completion model to generate docking score prediction matrices for the row and column indices during the decoding output and reconstruction error calculation stages.

[0034] Based on the ligand baseline table and according to the column index, the intrinsic scoring baseline of each calibrated ligand is extracted to form the intrinsic scoring baseline vector of the ligand. The intrinsic scoring baseline vector of the ligand is then used as the docking score decomposition constraint input to the VAE-based matrix completion model.

[0035] Perform ligand intrinsic scoring baseline subtraction on the non-missing docking scores in the finite docking scoring matrix to obtain the conformational residual docking scores corresponding to the non-missing positions, thus forming the conformational residual docking score training data.

[0036] In the decoding output and reconstruction error calculation stage, the ligand intrinsic scoring baseline subtraction is performed on the docking score prediction values ​​corresponding to non-missing positions in the docking score prediction matrix to obtain the predicted conformational residual docking score, and the reconstruction error is calculated based on the conformational residual docking score training data and the predicted conformational residual docking score.

[0037] The VAE-based matrix completion model is updated based on the reconstruction error, and matrix completion prediction is performed with the conformational residual docking score as the learning objective to obtain the docking score prediction matrix at all positions.

[0038] The docking score prediction values ​​at all positions in the docking score prediction matrix are subtracted from the ligand-inherent scoring baseline according to the column index to generate a residual docking score completion matrix.

[0039] Optionally, S5 includes:

[0040] The correspondence between candidate conformations and matrix rows is established based on the row indices of the candidate conformation set and the residual docking score completion matrix, and the correspondence between calibration ligands and matrix columns is established based on the column indices of the calibration ligand set and the residual docking score completion matrix.

[0041] For each candidate conformation in the candidate conformation set, all residual docking scores corresponding to the row of the candidate conformation are read from the residual docking score completion matrix, and a residual docking score sequence is formed according to the column order corresponding to the calibration ligand set.

[0042] An aggregation calculation is performed on the residual docking score sequence, which includes summing the residual docking score sequence and dividing it by the number of ligands in the calibration ligand set to obtain the conformation suitability score corresponding to the candidate conformation.

[0043] The conformation suitability scores corresponding to each conformation in the candidate conformation set are summarized in the order of the candidate conformation set to generate a conformation suitability score.

[0044] Optionally, S6 includes:

[0045] Based on the conformation suitability score, perform a traversal comparison of all score values ​​to determine the maximum score value among the conformation suitability scores;

[0046] The maximum score index set is determined based on the position of the maximum score value in the conformation suitability score. Then, the corresponding candidate conformations are extracted from the candidate conformation set based on the maximum score index set to form the maximum score candidate set.

[0047] When the maximum score candidate set contains multiple candidate conformations, the maximum score candidate set is filtered according to the order of the candidate conformation set, and the candidate conformation with the highest order is determined as the candidate conformation that satisfies the maximum score condition.

[0048] The candidate conformation that satisfies the maximum score condition is determined as the working conformation.

[0049] Optionally, the S7 includes:

[0050] Receive the three-dimensional coordinate information and atom type information corresponding to the working conformation, and organize the three-dimensional coordinate information and atom type information according to the preset field order to form a structure for writing data;

[0051] Based on the structure writing data, a three-dimensional structure file corresponding to the working conformation is generated. The three-dimensional structure file adopts the protein data bank format or the molecular structure data format.

[0052] Output a 3D structure file and use it as the input structure file for virtual filtering.

[0053] Optionally, a ligand knowledge graph is used as the structural input, and the embedding vectors of ligand nodes in the calibration ligand set are used as the initial input. The embedding vectors of ligand nodes are 64-dimensional. The knowledge graph relational path reasoning network includes a three-layer graph reasoning layer and a readout network. Each graph reasoning layer gathers information from adjacent fragment nodes, adjacent molecular nucleus nodes, and adjacent physicochemical property binning nodes for ligand nodes. Each graph reasoning layer contains 64 neuron channels. The ligand node embedding vectors output by the three-layer graph reasoning layer are input to the readout network. The readout network is a two-layer fully connected structure with 32 neurons and 1 neuron, respectively, used to output the ligand intrinsic scoring baseline corresponding to each calibration ligand.

[0054] Optionally, the VAE-based matrix completion model includes a conformation encoder, a ligand encoder, and a decoder. The conformation encoder receives a 256-dimensional conformational structure feature vector, which is passed through a fully connected layer of 128 neurons and a fully connected layer of 64 neurons to output a 32-dimensional conformational latent representation. The ligand encoder receives a 256-dimensional ligand structure feature vector, which is passed through a fully connected layer of 128 neurons and a fully connected layer of 64 neurons to output a 32-dimensional ligand latent representation. The decoder receives a 64-dimensional vector formed by concatenating the conformational latent representation and the ligand latent representation, which is passed through a fully connected layer of 64 neurons and a fully connected layer of 1 neuron to output a predicted docking score. The reconstruction error is calculated only for the predicted docking score corresponding to the non-missing position in the finite docking score matrix and the conformational residual docking score obtained after deducting the intrinsic scoring baseline of the ligand from the observed docking score.

[0055] The beneficial effects of this invention are:

[0056] (1) This invention proposes an improved docking score decomposition method. Instead of directly performing unified learning on the original docking scores within a finite docking score matrix, it first constructs a ligand knowledge graph based on a set of calibrated ligands, including ligand nodes, fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes. Then, it uses a knowledge graph relational path reasoning network to generate the inherent scoring baseline for each calibrated ligand, and introduces this baseline as a constraint for docking score decomposition into the subsequent matrix completion process. This invention extracts the column offset caused by the inherent properties of the ligands in the original scores through the correspondence between ligand structural composition information and observed docking scores, so that conformation evaluation is no longer directly affected by the inherent tendency of different ligands to obtain high or low scores. Compared with existing methods that directly use the original docking scores for conformation ranking or rely solely on simple mean correction, this invention can establish scoring baselines at multiple levels—ligand fragments, molecular nucleus, and physicochemical properties—enhancing the consistency of scoring calibers for different ligands and making subsequent conformation applicability judgments more focused on reflecting differences in the conformational binding environment.

[0057] (2) This invention proposes a novel residual docking score matrix completion method. In the VAE-based matrix completion model, the structural features of candidate conformations and ligands are encoded as latent representations, and then the decoder predicts the full matrix docking score. Simultaneously, the inherent scoring baseline of the ligands is uniformly deducted in the decoding output and reconstruction error calculation stages, so that the model learning objective is transformed from the original docking score to the conformational residual docking score. This invention thus separates the systematic bias on the ligand side from the adaptation differences on the conformation side, preserving the observed information in the finite docking score matrix while avoiding the direct assignment of pseudo-supervision values ​​to missing positions. Compared with existing matrix factorization methods or conventional VAE completion methods that directly face the original sparse score matrix, this invention is more conducive to the stable learning of the real interaction differences between conformations and ligands under limited observation conditions, and improves the model's adaptability to sparse docking data by calculating the reconstruction error only for non-missing positions.

[0058] (3) This invention proposes a holistic method for determining working conformations, which connects non-dominant conformation screening, representative ligand selection, docking task allocation with constraints, knowledge graph relational path reasoning, residual matrix completion, and conformation applicability aggregation evaluation into a unified processing flow. First, this invention uses non-dominant criteria to eliminate conformations that have been proven to be inferior by multi-objective values. Then, under the constraint of the upper limit of docking times, it performs balanced task allocation on the conformation and ligand dimensions, and introduces search space constraints to suppress the excessive occupation of docking resources by conformations with significantly large search volumes. Subsequently, it performs unified aggregation on each candidate conformation based on the residual docking score completion matrix to determine the working conformation and output a three-dimensional structure file. Compared with existing schemes that rely on full docking, empirical conformation selection, or direct decision-making based on only a small number of observation scores, this invention can improve the balance of the effective observation distribution within a limited computational budget, and makes the final determination result take into account the premise of multi-objective optimization, the coverage of ligand structure, and the universal applicability at the residual score level. Therefore, it is more suitable as a method for determining the virtual screening input structure. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of a deep learning-based method for determining the working conformation of proteins proposed in this invention. Detailed Implementation

[0061] In Example 1, reference Figure 1 A deep learning-based method for determining the working conformation of proteins, comprising:

[0062] S1. Obtain the non-dominant candidate conformations of the same protein, generate a set of candidate conformations, obtain calibration ligands, and generate a set of calibration ligands;

[0063] S2. Under the constraint of the upper limit of docking times, molecular docking is performed on the candidate conformation set and the calibration ligand set to generate a finite docking score matrix;

[0064] S3. Construct a ligand knowledge graph based on the calibration ligand set. The ligand knowledge graph includes ligand nodes, fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes. It also includes relation edges describing the correspondence between ligands and fragments, ligands and molecular nuclei, and ligands and physicochemical property binning. Train the relation path reasoning network of the knowledge graph based on the finite docking scoring matrix to generate the inherent scoring baseline of each calibration ligand. Summarize the inherent scoring baselines of the ligands to generate a ligand baseline table and output the ligand baseline table as a docking scoring decomposition constraint.

[0065] S4. Construct a VAE-based matrix completion model based on the candidate conformation set and the calibration ligand set. In the decoding output and reconstruction error calculation stage of the VAE-based matrix completion model, introduce a ligand baseline table. Perform ligand intrinsic scoring baseline subtraction on the docking scores in the finite docking score matrix to obtain the conformation residual docking score. The VAE-based matrix completion model uses the conformation residual docking score as the learning target to perform matrix completion prediction and generate the residual docking score completion matrix.

[0066] S5. Calculate the conformation suitability score for each candidate conformation based on the residual docking score completion matrix, and generate the conformation suitability score.

[0067] S6. Determine the candidate conformations that meet the maximum score condition based on the conformation suitability score, and generate the working conformation.

[0068] S7. Output the 3D structure file corresponding to the working conformation for virtual screening.

[0069] In this embodiment, step S1 includes:

[0070] This step involves organizing the original conformational results from multi-objective optimization and the ligand structure library into conformational and ligand objects under a unified index. The original conformational results are denoted as... ,in A three-dimensional structural file representing a conformation of the same protein. Representation of Conformation The corresponding multi-objective value vector, Indicates the number of multi-objective values. The number of original conformations is represented by , and the ligand structure library is denoted as . ,in This represents a candidate ligand structure record. Indicates the number of candidate ligands, for Each First, unify the target direction. Preserve the original direction for the multi-objective values ​​that need to be minimized, and convert the multi-objective values ​​that need to be maximized into a unified comparison direction. Then, perform a non-dominant determination on all conformations. When a conformation... No worse than the conformation across all multi-objective values Furthermore, at least one multi-objective value is superior to the conformation. At that time, the configuration Determined as a dominated conformation and deleted, the records retained after filtering are written as follows: , This indicates the number of non-dominant conformations. This processing rule can remove conformations that have been proven to be inferior by multi-objective values ​​before the allocation of finite docking attempts, thus avoiding the row indices of the finite docking score matrix being occupied by invalid conformations.

[0071] right Each conformation in the model is defined with a uniform binding pocket region definition, denoted as […]. , Given the coordinates of a reference ligand or preset key residues, when a reference ligand is present, its geometric center is used as the pocket center, and the pocket bounding box is determined by the atomic distribution of the reference ligand. When a reference ligand is absent, the geometric center of the preset key residue coordinates is used as the pocket center, and the pocket bounding box is formed using preset side lengths. Then, spatial filtering is performed on the atomic coordinates in each conformation, comparing the distance from the atom to the pocket center with a distance threshold. The comparison is performed, and atoms are retained if the distance condition is met, and deleted if the distance condition is not met, resulting in a truncated pocket neighborhood structure file. Distance threshold It can be taken as the fixed boundary after the pocket surround box is expanded, for The atoms retained in the original document undergo further document standardization, which includes adding hydrogen atoms, standardizing charge marking, standardizing residue numbering, standardizing chain identification, and standardizing coordinate fields. The document format is standardized to the structural document format used for molecular docking. This process yields a sequence of pocket neighborhood structural documents arranged in conformational order. ;

[0072] In the pocket neighborhood structure file Based on this, a 256-dimensional conformational structure feature vector is generated for each conformation. 256-dimensional conformational structure feature vector Composed of eight 32-dimensional sub-blocks, dimensions 1 to 32 represent the voxel occupancy distribution within the pocket, with the number of atoms counted according to a fixed spatial grid around the pocket center. Dimensions 33 to 64 represent the accessible surface distribution within the pocket, with the accessible surface area counted according to the same spatial grid. Dimensions 65 to 96 represent the hydrophobic atom distribution, counting the spatial frequency of carbon atoms and hydrophobic labeled atoms. Dimensions 97 to 128 represent the hydrogen bond donor atom distribution, dimensions 129 to 160 represent the hydrogen bond acceptor atom distribution, dimensions 161 to 192 represent the positively charged atom distribution, dimensions 193 to 224 represent the negatively charged atom distribution, and dimensions 225 to 256 represent the key residue contact distribution, counting the contact frequency between different residue types within the pocket and the neighborhood of the pocket center. Calculated under the same spatial grid rules for the eight 32-dimensional sub-blocks, a conformational structural feature vector sequence consistent with the pocket boundary is obtained. , constructive structural feature vector It is directly used as the conformational encoding input for the VAE-based matrix completion model;

[0073] right Each candidate ligand first undergoes structural normalization, which includes unifying aromaticity labeling, bond order, charge state, and stereochemical labeling. Each normalized ligand generates a 2048-bit circular fingerprint to represent its local connectivity environment. The structural similarity between any two ligands is calculated using the intersection-union ratio (IU) of their circular fingerprints. A similarity matrix is ​​constructed based on the structural similarities of all candidate ligands. Then, clustering is performed under a preset upper limit on the number of ligands. The clustering method uses a center representative selection approach based on the similarity matrix. When the number of clusters reaches the preset upper limit, the ligand with the smallest distance to the cluster center is selected as the representative ligand from each cluster, resulting in a calibration ligand sequence arranged in representative order. , Indicating the number of calibration ligands, in one feasible example, By using a representative ligand selection method instead of a random sampling method, the calibrated ligand sequences can cover more fragment types, molecular core types, and physicochemical property ranges, thus providing a complete source for the construction of relation edges in the ligand knowledge graph.

[0074] Determining the calibration ligand sequence Then, for each ligand Continue generating structural descriptions for use in the ligand knowledge graph; fragment records are denoted as... ,in Indicates ligand The first one obtained after BRICS splitting A segment, Indicates ligand The number of fragments, recorded in the molecular nucleus, is denoted as , Using the Bemis-Murcko rule from ligands Extracted from, the physicochemical properties are recorded in bins as follows: ,in The binning indicates the molecular weight. This indicates the compartment containing the octanol-water partition coefficient. The bin containing the number of hydrogen bond donors. The binning indicates the number of hydrogen bond acceptors. Discrete binning labels are obtained by comparing the four types of physicochemical properties with the preset binning boundary table, and fragment records are generated. Molecular nucleus recording Record of physicochemical properties in separate bins It is directly used to construct ligand nodes, fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes in the ligand knowledge graph, and to establish the correspondence edges between ligands and fragments, ligands and molecular nucleus, and ligands and physicochemical property binning.

[0075] In addition to the map construction description, a 256-dimensional ligand structure feature vector is generated for each calibration ligand. 256-dimensional ligand structure feature vector It consists of six fixed sub-blocks: Dimensions 1 to 64 are fragment occurrence codes, recording the occurrence of fragments according to a preset 64-frequency fragment vocabulary; Dimensions 65 to 128 are molecular nucleus codes, recording the molecular nucleus category according to a preset 64-frequency molecular nucleus vocabulary; Dimensions 129 to 160 are functional group counting codes, counting the occurrence of acidic groups, basic groups, amide groups, ether groups, halogen substituents, and aromatic ring substituents; Dimensions 161 to 192 are atom type counting codes, counting the distribution of carbon, nitrogen, oxygen, sulfur, phosphorus, halogens, and charged atoms; Dimensions 193 to 224 are topological path length distributions, counting the number of atom pairs whose bond path lengths fall into different intervals; Dimensions 225 to 256 are physicochemical property binning codes, expanding the binning results of molecular weight, octanol water partition coefficient, number of hydrogen bond donors, and number of hydrogen bond acceptors into fixed-length discrete codes, and the ligand structure feature vector sequence is written as follows. ligand structure feature vector It can be directly used as the ligand encoding input for VAE-based matrix completion models;

[0076] This step generates a sequence of pocket neighborhood structure files arranged according to a uniform conformation index. and the sequence of eigenvectors of the structure And form a calibration ligand sequence arranged according to a uniform ligand index. Fragment Record Sequence Molecular nucleus recording sequence Physicochemical properties binning record sequence and ligand structure feature vector sequence In an implementable example, , each and each All dimensions are 256, and the knowledge graph relation path reasoning network uses... , , and Constructing a ligand knowledge graph and using a VAE-based matrix completion model. , and By establishing coded inputs for conformations and ligands, the resulting input objects have a unified pocket boundary on the conformation side and a unified structural coverage on the ligand side, which can correspond to the row and column indices of a finite docking scoring matrix.

[0077] In this embodiment, step S2 includes:

[0078] The goal of this step is to organize the candidate conformation set and calibration ligand set into a finite docking scoring matrix under the constraint of an upper limit on the number of docking attempts. This finite docking scoring matrix is ​​then fixed as the matrix input skeleton shared by the knowledge graph relational path reasoning network and the VAE-based matrix completion model. The pocket neighborhood structure file sequence is denoted as... ,in Indicates the first The pocket neighborhood structure file corresponding to each candidate conformation. The number of candidate conformations is indicated by the calibration ligand sequence denoted as . ,in Indicates the first One calibration ligand, The number of calibration ligands is indicated, with the maximum number of docking attempts denoted as [value]. , This represents the total number of allowed conformation-ligand pair combinations. To ensure that the finite docking score matrix contains observed elements in both the conformation and ligand dimensions, a preset value is used. In an implementable example, , The total number of possible conformations and ligand pair combinations is 2048. ;

[0079] according to When determining the docking search space parameters for each candidate conformation, start with any pocket neighborhood structure file. Read the three-dimensional coordinates of all retained atoms to form a pocket set of atom coordinates. ,in Indicates the first The three-dimensional coordinates of each retained atom This indicates the number of atoms to retain. Then, the maximum and minimum coordinate values ​​are read in each of the three coordinate directions. The midpoint between the maximum and minimum values ​​in any direction is determined as the center coordinate of the pocket. The difference between the maximum and minimum values ​​in any direction is increased by twice the fixed boundary value. The search edge length is determined, and the boundary is fixed. This represents the safe boundary for expanding beyond the pocket coordinate range, and then the three search side lengths are compared with the preset minimum side length. and preset maximum side length Comparison, in less than Updated in time , in greater than Updated in time Thus forming the first The 6-dimensional docking search space parameter vector corresponding to each candidate conformation ,in , , These represent the three coordinate components of the pocket center. , , These represent the search edge lengths in the three coordinate directions, and the corresponding lengths of all candidate conformations. Arranged in conformational order to form a sequence of search space parameters. ;

[0080] Based on the upper limit constraint on the number of docking attempts, , and When calculating the required conformation-ligand pairwise combinations, first construct the set of all possible pairwise combinations. And then according to Determine the actual number of tasks , Pick ,Will The target number of allocations for each candidate conformation is obtained by distributing them equally among the number of candidate conformations. ,Will The target number of allocations for each calibration ligand is obtained by distributing the ligands evenly according to their quantity. When there is a remainder after equal distribution, the remainder is distributed according to... The order of the arrangement is filled in sequentially to fill in the first few. Up, and according to The order of the arrangement is filled in sequentially to fill in the first few. Above, the first Before generating the task record, the selected conformations and ligands are paired and written as follows: , No. The wheelset has not yet entered Any pair combination Calculate the selection value and determine the first pair based on the pair with the largest selection value. Task records:

[0081] ;

[0082] in, Indicates the first In the rotation selection, the conformation and ligand pair up. Indicates the index of candidate conformations. Indicates the candidate calibration ligand index. Indicates the task generation round. , This represents the set of all possible conformations paired with ligands. Indicates the first The conformations and ligands selected before the wheel are paired together. Indicates the first The rounds have already been assigned to the first Number of tasks for each alternative conformation Indicates the first The rounds have already been assigned to the first Number of tasks per calibration ligand Indicates the first The number of times each candidate conformation is assigned to a target. Indicates the first The number of target allocations for each calibrated ligand. Represents the search space constraint coefficient. To preset dimensionless parameters, , , They represent the first Each candidate conformation corresponds to a parameter vector in the docking search space. The lengths of the three search sides, The index represents the candidate conformation, used to determine the maximum search volume among all candidate conformations. The first and second terms constrain the balanced distribution of conformation and ligand dimensions, respectively. The third term uses the normalized search volume to suppress excessive usage of candidate conformations with significantly large search space in the task list. If multiple pairs of combinations have the same selection value, the selection is first determined by... The order of arrangement determines the candidate conformation index, then according to... The order of arrangement determines the calibration ligand index, and the ligands are arranged according to the round. Write them sequentially to obtain the docking task list. ;

[0083] in accordance with When performing molecular docking, first perform calibration on any calibrated ligand. Generate docking ligand records , Includes 3D coordinates, atom type, charge, and rotatable bond markers; any task record Corresponding to one molecular docking call, the molecular docking call is based on As a receptor input, to As a ligand input, with As input to the search space, Multiple candidate combination poses are sampled within a limited space, and the original scoring sequence corresponding to the candidate combination poses is written as follows: ,in Indicates the first The original scores of each candidate combined with the posture. This indicates the number of candidate docking attitudes. If the original score uses the binding energy direction, the original score with the smallest value is selected and a unified direction transformation is performed to obtain the docking score. If the original score already satisfies the condition that a larger value indicates a more favorable combination, then the original score with the largest value is directly selected as the matching score. Arrange all tasks in order. To form a docking scoring record sequence ;

[0084] When writing the docking score into the matrix index, first press The arrangement order is fixed by the matrix row index, then according to... Using a fixed matrix column index for arrangement, a finite docking scoring matrix is ​​constructed. During matrix initialization, all positions are marked as missing. , This indicates that molecular docking was not performed, and the process continues. , put the first Matching score in the record Write matrix position Matrix elements Indicates the first The alternative conformations and the first Docking scores between calibration ligands, and pairwise combinations not appearing in the docking task list. Maintain missing markers In an implementable example, The dimension is Thus formed The finite observation boundary is defined by the observed docking scores and missing markers. The knowledge graph relational path reasoning network then... Using the observed docking scores as the training signal source, the VAE-based matrix completion model... As the input skeleton for matrix completion, the same row and column indices are used to complete the conformational residual docking scoring learning.

[0085] In this embodiment, step S3 includes:

[0086] The structural information in the calibration ligand set is organized into a ligand knowledge graph, and the non-missing docking scores in the finite docking score matrix are organized into training signals for the knowledge graph relational path reasoning network. The calibration ligand sequence is written as... , Indicates the first One calibration ligand, The finite docking score matrix represents the number of calibration ligands. , Indicates the first The alternative conformations and the first Docking score between calibration ligands Indicates the number of alternative conformations. The combined processing objective of the ligand knowledge graph and the knowledge graph relationship path reasoning network is to extract the column offsets affected by the inherent properties of the ligands in the finite docking scoring matrix into ligand inherent scoring baselines that can be referenced by column, and then write the ligand inherent scoring baselines into a ligand baseline table for direct reading by the VAE-based matrix completion model in the decoding output and reconstruction error calculation stages.

[0087] For any calibrated ligand First, structural standardization is performed. Structural standardization includes unifying aromaticity labeling, bond order, charge state, and stereochemical labeling. After structural standardization is completed, calibration ligands are then... Structural parsing is performed, and fragment parsing uses the BRICS splitting rule to segment breakable connection positions, resulting in a sequence of fragment nodes. , Indicates the first The first calibration ligand corresponding to the first Each fragment node Indicating the number of fragment nodes, the molecular nucleus analysis uses the Bemis-Murcko rule to extract the ring system and connection backbone, obtaining the number of nodes. Molecular nucleus nodes corresponding to each calibration ligand The physicochemical property binning analysis first calculates the molecular weight, octanol-water partition coefficient, number of hydrogen bond donors, and number of hydrogen bond acceptors, and then compares them with the preset binning boundary table to obtain four physicochemical property binning nodes. , , , , and then for the first Establish a ligand node for each calibrated ligand. and ligand nodes With fragment node sequence Molecular nucleus node Four physicochemical property-specific binning nodes , , , The correspondence is written into the node correspondence record. After all calibrated ligands have completed structural analysis, the standard expressions of the same fragments are bound to the same fragment node, the standard expressions of the same molecular core are bound to the same molecular core node, and the binning tags of the same physicochemical properties are bound to the same physicochemical property binning node.

[0088] A ligand knowledge graph is constructed based on the node correspondence records. The ligand knowledge graph contains ligand node sequences. Fragment node sequences, molecular nucleus node sequences, and physicochemical property binning node sequences are used. Relationship edges are categorized into three types: ligand-fragment relationship edges, ligand-molecular nucleus relationship edges, and ligand-physicochemical property binning relationship edges. For any calibrated ligand... ligand nodes With all fragment nodes Write the ligand-fragment relationship edge, and add the ligand node. With molecular nucleus node Write the ligand-molecule nucleus relationship edge, and add the ligand node. With four physicochemical property separation nodes , , , Write the binning relationship edges between ligands and physicochemical properties. The ligand knowledge graph is stored using a combination of node table and adjacency table, so that any ligand node can access the corresponding fragment node, molecular nucleus node and physicochemical property binning node along the relationship edge.

[0089] Based on the limited docking scoring matrix When extracting the training signal, traverse all matrix elements column-wise, and perform a column-wise traversal for any calibration ligand index. Extract the first The score for all non-missing docking points in the column is obtained as the first result. ligand observation score set corresponding to each ligand node Since the task allocation rule under the constraint of the upper limit on the number of docking attempts has ensured that each calibration ligand corresponds to at least one observed position, any All are not empty. Based on the node correspondence in the ligand knowledge graph, the direct observation scores and the associated observation scores corresponding to shared fragment nodes, shared molecular nucleus nodes, and shared physicochemical property binning nodes are collectively organized into ligand training tags. Ligand training labels for each ligand node Calculate using the following formula:

[0090] ;

[0091] In the formula, Indicates the first The ligand training labels corresponding to each ligand node. Indicates the calibration ligand index. Indicates the first The set of ligand observation scores corresponding to each ligand node. This represents a docking score value in the ligand observation score set. Indicates the constraint coefficient of the fragment node. To preset dimensionless parameters, Represents the constraint coefficient of the molecular core node. To preset dimensionless parameters, Indicates the constraint coefficient of the binning nodes representing the physicochemical properties. To preset dimensionless parameters, Indicates the first The fragment node sequence corresponding to each calibration ligand. Represents a sequence of fragment nodes One of the fragment nodes, Represents fragment nodes The linked set of calibration ligand indices Indicates the first Each calibration ligand corresponds to a molecular core node. Indicates the node with the molecular parent nucleus The linked set of calibration ligand indices Indicates the first The first calibration ligand corresponding to the first Each physicochemical property is divided into bin nodes. Binning index for physicochemical properties , Represents the binning nodes related to physical and chemical properties The linked set of calibration ligand indices This indicates the calibration ligand index connected to the specified node. , , , , This indicates the number of elements in the corresponding set. According to this calculation method, the ligand training labels... It is composed of the direct observation docking score, the average observation score of the associated ligands corresponding to the shared fragment nodes, the average observation score of the associated ligands corresponding to the shared molecular nucleus nodes, and the average observation score of the associated ligands corresponding to the shared physicochemical property bin nodes. All quantities involved in the calculation are docking score values, the number of set elements, or preset dimensionless parameters. , , Includes at least the current calibrated ligand index any It includes at least one docking score, so the denominator will not have a zero value;

[0092] Ligand training tags After its formation, the knowledge graph relational path reasoning network begins training. This knowledge graph relational path reasoning network is denoted as... , The parameters of the knowledge graph relational path reasoning network are represented. The input of the knowledge graph relational path reasoning network consists of three parts: the first part is the node table and adjacency table of the ligand knowledge graph; the second part is the initial embedding vector of all nodes; and the third part is the ligand training label sequence. For any ligand node First read the first Ligand structure feature vectors corresponding to each calibration ligand ligand structure feature vector It consists of six fixed sub-blocks: dimensions 1 to 64 encode fragment occurrence, dimensions 65 to 128 encode the molecular nucleus, dimensions 129 to 160 encode functional group counts, dimensions 161 to 192 encode atomic type counts, dimensions 193 to 224 encode the topological path length distribution, and dimensions 225 to 256 encode the physicochemical properties. Then, the 256-dimensional ligand structure feature vector is further... The initial embedding vectors for ligand nodes are projected through a linear mapping layer. Discrete indices are then established for fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes based on the standard expressions for fragments, molecular nucleus, and binning labels, respectively. The initial 64-dimensional node embedding vectors are then read from the trainable embedding table. The resulting network input includes: The initial embedding vector sequences of 64-dimensional ligand nodes, fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes, as well as three types of relational edge adjacency lists;

[0093] The knowledge graph relation path reasoning network consists of three graph reasoning layers. Each layer contains 64 neuron channels, a corresponding relation weight matrix, and a bias, and employs rectified linear units. The data flow order of the graph reasoning layer is fixed in two segments. The first segment performs message reception for fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes. The message source is the 64-dimensional embedding vector of the ligand node connected to the specified node in the adjacency list. The relation weight matrix is ​​linearly transformed for different relation types, and then all transformation results received by the same node are summed and the bias is added to obtain the updated non-ligand node embedding vector. The second segment performs message reception for any ligand node. Perform adjacency aggregation, with the aggregation source being ligand nodes. The connected fragment node embedding vectors, molecular nucleus node embedding vectors, and physicochemical property binning node embedding vectors, and the 64-dimensional embedding vectors from different sources are summed after linear transformation by the relation weight matrix, plus the bias, and then passed through a rectified linear unit to obtain the 64-dimensional ligand node embedding vector of the current layer. The three-layer graph inference layer is executed in series in the same order, so that each ligand node obtains a 64-dimensional final ligand node embedding vector at the end of the third layer, which integrates fragment composition, molecular nucleus similarity, and physicochemical property binning similarity.

[0094] Read the network record as , This indicates the readout network parameters. The readout network employs a two-layer fully connected structure. The first layer receives the 64-dimensional final ligand node embedding vector and outputs a 32-dimensional hidden vector. The second layer receives the 32-dimensional hidden vector and outputs a scalar for any ligand node. The 64-dimensional final ligand nodes are embedded into the vector input readout network. , obtained the The intrinsic scoring baseline of each calibrated ligand. The training phase will With ligand training tags Alignment is performed using the average of the squared differences among all ligand nodes as the mean squared error, and the parameters of the knowledge graph relation path inference network are updated according to the backpropagation order. Reading network parameters The inference phase involves recursively performing the same forward propagation process on all ligand nodes, including fragment node embedding tables, molecular nucleus node embedding tables, and physicochemical property binning node embedding tables, to obtain the intrinsic scoring baseline sequence of ligands arranged in the order of calibration ligands. ;

[0095] Will According to the calibration ligand sequence The order of arrangement is written as a ligand baseline table. ligand baseline table for 32-dimensional vector, in an implementable example, ligand baseline table The Middle Inherent scoring baseline for ligands at each position Limited docking scoring matrix The Strict column alignment is used to read the matrix completion model based on VAE columns. Furthermore, in the decoding output and reconstruction error calculation stage, the ligand-inherent scoring baseline is deducted from the docking score.

[0096] In this embodiment, step S4 includes:

[0097] A VAE-based matrix completion model is used to perform matrix completion on a finite docking score matrix under limited observation conditions. It separates the intrinsic ligand scoring baselines from the original docking scores within the finite docking score matrix, and assigns a set of candidate conformations to a sequence of conformational structure feature vectors in a fixed order. ,in Indicates the first Each candidate conformation corresponds to a 256-dimensional conformational structure feature vector. The 256-dimensional field consists of the distribution of pocket voxel occupancy, pocket accessible surface, hydrophobic atoms, hydrogen bond donor atoms, hydrogen bond acceptor atoms, positively charged atoms, negatively charged atoms, and key residue contacts. The calibration ligand set corresponds to a ligand structure feature vector sequence in a fixed order. ,in Indicates the first Each calibration ligand corresponds to a 256-dimensional ligand structure feature vector. The 256-dimensional field consists of fragment occurrence encoding, molecular nucleus encoding, functional group count encoding, atom type count encoding, topological path length distribution encoding, and physicochemical property binning encoding. The finite docking scoring matrix is ​​denoted as... ,in Indicates the first The alternative conformations and the first Docking score between calibration ligands Indicates the position where molecular docking was not performed; the ligand baseline is denoted as... ,in Indicates the first The intrinsic scoring baseline of each calibrated ligand;

[0098] First, determine the row index sequence of the finite docking scoring matrix based on the order of the candidate conformation set and the calibration ligand set. With column index sequence Then, based on the limited docking scoring matrix Construct the observation mask matrix ,when season ,when season , by all satisfied The matrix positions form a set of non-missing position indices. The resulting input skeleton includes: arranged by row index Constructive structural features, arranged by column index ligand structural features and dimensions are Limited docking scoring matrix Dimensions Observation mask matrix and dimension are The intrinsic scoring baseline vector of the ligand;

[0099] The VAE-based matrix completion model consists of a conformational encoder. Ligand encoder and decoder Composition, in which Indicates the parameters of the conformation encoder. Indicates ligand encoder parameters, This represents the decoder parameters; the conformational encoder receives any 256-dimensional conformational structure feature vector. The data passes through a fully connected layer with 128 neurons and then a fully connected layer with 64 neurons, each containing weights, biases, and rectified linear units, before outputting a 32-dimensional conformational mean vector. and 32-dimensional conformational log-variance vector Subsequently, sampling vectors are drawn from a 32-dimensional standard Gaussian distribution. ,Will After conversion to standard deviation vector and Multiply by elements, then multiply by... Adding elements together, we get the first... The 32-dimensional latent representation of each candidate conformation The ligand encoder receives any 256-dimensional ligand structure feature vector. It employs the same two-layer fully connected structure as the conformational encoder, and outputs a 32-dimensional ligand mean vector. and 32-dimensional ligand log-variance vector Combined with a 32-dimensional standard Gaussian sampling vector Generate the first 32-dimensional ligand latent representation for each calibrated ligand ;

[0100] Decoder for any matrix position During location prediction, the 32-dimensional latent representation of the conformation is used. Potential representation with 32-dimensional ligands The vectors are concatenated sequentially into a 64-dimensional interaction vector, which is then input into a decoder. The decoder contains a fully connected layer with 64 neurons and an output layer with 1 neuron. The output position is... Predicted docking score For all and all Repeat the same process to obtain a finite docking score matrix. Dimensionally consistent docking score prediction matrix The forward path of the VAE-based matrix completion model consists of four segments: conformation feature encoding, ligand feature encoding, latent representation sampling, and positional interactive decoding. The predicted value at each position in the matrix is ​​generated according to the same row and column indices.

[0101] Ligand baseline table Instead of forming a separate new prediction branch, it is directly embedded into the decoding output and reconstruction error calculation stage according to the column index, first according to the column index sequence. Reading the ligand baseline table All This forms the intrinsic scoring baseline vector of the ligand. For any non-missing position From a limited docking scoring matrix Read the observation docking score Then, from the intrinsic scoring baseline vector of the ligand... Column index Read ,Will deduct The observational conformational residual docking score was then obtained. ,all The conformational residual docking scoring training data is constructed by indexing non-missing locations, and then the same non-missing locations are used to construct training data for the non-missing location docking score. Predicted docking score Perform ligand intrinsic scoring baseline subtraction under the same column index to obtain the predicted conformational residual docking score. The reconstruction error is not calculated directly from the original docking score, but from the docking score of the observed conformation residual and the docking score of the predicted conformation residual.

[0102] Reconstruction error Determine by the following formula:

[0103] ;

[0104] In the formula, Indicates reconstruction error. Indicates the number of alternative conformations. Indicates the number of calibration ligands. Represents the observation mask matrix In position The value of , Represents a finite docking scoring matrix In position The observation and docking score, Ligand baseline table The Middle The inherent scoring baseline of each calibrated ligand. Represents the docking score prediction matrix In position The predicted docking score, Indicates the alternative conformation index. This indicates the calibration ligand index; the molecular part is only for... The cumulative error of non-missing positions is normalized by the number of non-missing positions in the denominator. Therefore, missing positions will not be written into the pseudo-supervision value. All quantities involved in the calculation are docking score values, mask values ​​and ligand intrinsic scoring baseline values, and do not include abstract quantities that cannot be directly obtained.

[0105] Based on reconstruction error For the parameters of the conformation encoder ligand encoder parameters and decoder parameters Perform backpropagation updates and repeat the training loop of "forward prediction, ligand intrinsic scoring baseline subtraction, reconstruction error calculation, and parameter update" until the preset number of training rounds is reached. After training is complete, perform a full set of matrix positions. Perform forward prediction again to obtain the docking score prediction matrix covering the entire matrix range. Then by column index sequence right Perform ligand-inherent scoring baseline subtraction at all locations: for any location Read the predicted docking score. and the intrinsic scoring baseline of the ligands in the corresponding column ,Will deduct The residual docking score completion value was then obtained. ,all Matrix row index sequence and matrix column index sequence After arrangement, a residual docking score completion matrix is ​​formed. In a feasible example, the residual docking score completion matrix The dimension is Residual docking scoring completion matrix Each matrix element has been subtracted according to the ligand-specific scoring baseline of its corresponding column, therefore the residual docking score completion matrix is ​​complete. The differences between different columns in the same row are those related to conformation under a unified standard.

[0106] In this embodiment, step S5 includes:

[0107] The residual docking score completion matrix inherits the full matrix prediction results from the VAE-based matrix completion model. The processing objective is to uniformly aggregate the matrix elements that have already undergone ligand intrinsic scoring baseline subtraction according to the candidate conformation dimension, generating directly comparable conformation suitability scores. The candidate conformation set follows the pocket neighborhood structure file sequence. , Indicates the first The pocket neighborhood structure file corresponding to each candidate conformation. This indicates the number of candidate conformations; the calibration ligand set follows the calibration ligand sequence. , Indicates the first One calibration ligand, The residual docking score completion matrix is ​​used to indicate the number of calibration ligands. , Indicates the first The alternative conformations and the first The residual docking score completion value of each calibrated ligand after subtracting the inherent scoring baseline, due to All positions in the matrix have been predicted using a VAE-based matrix completion model. Missing markers are no longer included; every row in the matrix contains them. A residual docking score completion value that can directly participate in aggregation;

[0108] To ensure strict consistency between matrix rows and columns and object indices, a correspondence record between candidate constructs and matrix rows is first established. Correspondence record The establishment rule is: according to the set of candidate conformations The order of arrangement binds the first candidate conformation to the residual docking score completion matrix. The first line binds the second alternative conformation to the second line, until the first line is bound to the second line. The candidate construct is bound to the first Rows, based on corresponding records Establish row index mapping Row index mapping The generation rule is as follows: for any candidate conformation In the corresponding relationship record Find a unique record And index the rows in the record. As an alternative conformation The matrix row index is returned, from which we obtain The same method was used to establish the correspondence between calibration ligands and matrix columns. Correspondence record The establishment rule is as follows: according to the calibration ligand set The order of arrangement binds the first calibration ligand to the residual docking score completion matrix. In column 1, bind the second calibration ligand to column 2, and so on, until the first ligand is bound to column 2. The calibration ligand is bound to the first Columns, then records based on the corresponding relationships. Establish column index mapping Column index mapping The generation rule is: for any calibration ligand In the corresponding relationship record Find a unique record And index the columns in the record. As a calibration ligand The matrix column index is returned as the value, from which we obtain After constructing using this search expression, any matrix position All can be uniquely located to the first The alternative conformations and the first The combination relationship of each calibrated ligand will not result in the rearrangement of the matrix row index or the matrix column index during the aggregation phase.

[0109] In row index mapping and column index mapping After the setup is complete, read the corresponding matrix rows for each candidate conformation in the candidate conformation set one by one, and for any candidate conformation... First, use row index mapping Positioning residual docking scoring completion matrix The corresponding row, then according to the calibration ligand set The order in which the column index mappings are called sequentially. From column 1 to column 2 The column reads all residual docking score completion values ​​to form the first... The residual docking scoring sequence corresponding to each alternative conformation Residual docking scoring sequence for dimensional vector, Indicates the first The alternative configurations are in the first The residual docking score completion values ​​on the corresponding columns of each calibration ligand are due to the residual docking score completion matrix. All columns have undergone ligand-inherent scoring baseline subtraction, and the residual docking scoring sequence has been obtained. All components are under a unified scoring caliber, and the residual docking scoring sequence This reflects the distribution of conformational relevance scores for the same candidate conformation across all calibrated ligand dimensions;

[0110] The generation of conformation suitability scores is fixed in the residual docking score sequence. The aggregation calculation position is determined by summing the results sequentially according to the calibrated ligand set columns and then dividing by the number of calibrated ligands. Conformation suitability score for each alternative conformation Determine by the following formula:

[0111] ;

[0112] In the formula, Indicates the first The conformation suitability score for each alternative conformation. Indicates the alternative conformation index. Indicates the number of calibration ligands. Indicates the calibration ligand index. Represents the residual docking score sequence The Middle The residual docking score completion value at each position Represents the residual docking score completion matrix The middle is composed of alternative conformations Corresponding matrix rows and calibration ligands The residual docking score completion value at the position jointly determined by the corresponding matrix columns. Indicate alternative conformations Matrix row index, Indicates calibrated ligands Matrix column index, Indicates the first The pocket neighborhood structure file corresponding to each candidate conformation. Indicates the first For each calibration ligand, a conformational suitability score is calculated according to this rule. Completely completed matrix by residual docking score The aggregation is composed of elements from the same row, without introducing new compensation terms, column weighting, or additional normalization terms. The aggregation source is always limited to all rows corresponding to the specified candidate conformations. Individual residual docking score completion value;

[0113] Repeatedly perform matrix row reading, residual docking score sequence formation, and aggregation calculation on all candidate conformations to obtain conformation suitability score sequences that correspond one-to-one with the candidate conformation set. Then according to the set of alternative conformations The original arrangement order will be all Write them sequentially as conformational suitability score vectors Conformation applicability scoring vector for A dimensional vector, in an implementable example, Conformation applicability scoring vector The dimension is 64, and the conformational applicability score vector The Middle Record of the value and corresponding relationship of each position The Middle The candidate conformations are strictly aligned, therefore the conformation suitability score vector It retains both the order of the candidate conformation set and the residual docking score completion matrix. The row index semantics ensure that, throughout the entire processing, the underlying numerical values ​​involved in the conformation suitability score calculation are always limited to the residual docking score completion matrix. Matrix elements With the number of calibration ligands The original docking score in the finite docking score matrix is ​​no longer called, nor is the intrinsic ligand scoring baseline in the ligand baseline table, thus making the conformation suitability score vector... The calculation method is strictly maintained at the level of conformational residual docking scoring.

[0114] In this embodiment, step S6 includes:

[0115] The goal of S6 processing is to determine a unique working conformation from the candidate conformation set, provided that the conformation suitability scoring has been standardized to the conformation residual docking scoring caliber. The candidate conformation set follows the pocket neighborhood structure file sequence. ,in Indicates the first The pocket neighborhood structure file corresponding to each candidate conformation. This indicates the number of alternative conformations; the conformation suitability score follows the conformation suitability score vector. ,in Indicates the first Each candidate conformation has a conformation suitability score, and the conformation suitability score vector is... With alternative conformation set Maintaining a one-to-one correspondence, the vector position index is completely consistent with the order of the candidate conformations, therefore the first... Each rating value Unique Corresponding Number One alternative configuration ;

[0116] Before performing the maximum score determination, first fix the traversal index sequence. At the start of the traversal, the applicability score of the first conformation is determined. Write as the current maximum rating value And set index 1 as the current maximum rating index, then proceed according to index from 2 to... Read one by one in sequence When read in Greater than the current maximum score At that time, update the current maximum score to And update the current maximum rating index to the index. When read in Equal to the current maximum score At that time, retain the current maximum rating value unchanged, and set the index. Append records to the maximum rating index set when reading in. Less than the current maximum score At that time, no update is performed; all comparisons are completed according to the traversal rules. After reading each score, the maximum score among all conformational suitability scores is obtained. ;

[0117] At the maximum score Once determined, it is used according to the conformational suitability scoring vector. Recheck all location indexes Generate the maximum score index set Maximum rating index set Each index in the table represents a value associated with the maximum rating. Positions that are completely equal, based on conformational applicability scoring vector With alternative conformation set A one-to-one correspondence will be established for the maximum set of rating indices. Each index in Mapping to the corresponding alternative conformations This forms the largest candidate set of scores. Maximum score candidate set All candidate conformations satisfy the same maximum score condition, which is the maximum score candidate set. The number of elements is equal to the maximum score index set. The number of elements;

[0118] When the maximum score candidate set When only one candidate conformation is available, the unique candidate conformation is directly determined as the working conformation. To ensure consistency in the judgment criteria across various scenarios, this implementation still checks the maximum score candidate set according to the original arrangement order of the candidate conformation set. The member positions are checked specifically according to the set of candidate conformations. The arrangement order is scanned from index 1 forward to determine the candidate conformations corresponding to the current position. Does it belong to the maximum score candidate set? When the first candidate set belonging to the maximum score is scanned When an alternative conformation is found, the scan is stopped immediately, and the current position index is recorded as the final working index. Because the scanning order is consistent with the arrangement order of the candidate conformation set, the final working index... It must be the largest set of rating indices. The index with the smallest value, which is also the largest rating index that appears first in the sorted order, and then based on... From the set of alternative conformations Extract the corresponding candidate conformations and will Determined as the working designation;

[0119] When the maximum score candidate set When multiple candidate conformations are included, and the conformation suitability scores corresponding to these multiple candidate conformations are exactly the same, it is impossible to further distinguish them based solely on the score values. Therefore, this implementation does not introduce new scoring items, nor does it call additional data from the finite docking score matrix, residual docking score completion matrix, or ligand baseline table. Instead, it directly uses the set of candidate conformations. The sorting order is used as the sole filtering criterion. During filtering, the order from index 1 to index 2 is still used. The linear scan rule applies to any index. ,like If so, skip the corresponding candidate conformation. ,like Then the index Immediately designated as the final working index And stop filtering, according to this rule, the largest set of candidate scores. The candidate conformation with the highest order in the sequence is uniquely retained, resulting in the largest set of candidate scores. All other candidate conformations transition from the candidate state to the unselected state, thus obtaining and Each is unique, thus avoiding multiple working constructs when the maximum scores are tied;

[0120] In a feasible example, the number of alternative conformations Conformation applicability scoring vector After the traversal and comparison are completed, if the maximum score value is... If the indexes appear simultaneously in indices 12, 18, and 41, then the maximum set of score indices is written as: The maximum candidate set for ratings is written as Then according to the set of alternative conformations When performing a linear scan based on the order of the sequences, the first index hit with the highest score is 12, therefore the final working index is... The working construct is If the maximum score If it only appears at index 27, then the maximum score index set is written as The maximum candidate set for ratings is written as The final working index is The working construct is ;

[0121] After adopting the above judgment process, the data actually used for calculation and comparison in S6 only includes the conformational suitability score vector. Alternative conformation set Maximum score Maximum score index set Maximum score candidate set and final working index The entire screening process revolves around the conformational suitability scoring vector. Expanding upon this process, without further transforming the conformation suitability score, introducing new model predictions, or altering the original order of the candidate conformation set, the working conformation is determined through a processing chain of maximum score determination—parallel index extraction—screening according to the original order. It is uniquely identified under the same scoring criteria and the same indexing system.

[0122] In this embodiment, step S7 includes:

[0123] The goal of S7 processing is to transform the already defined working construct... It is directly converted into a 3D structure file that can be read by a virtual screening program. The data actually used by S7 only includes the working conformation. Internally stored 3D coordinate information, atom type information, and preset output format markers , This indicates the output format of the 3D structure file. This represents the protein data bank format. Molecular structure data format, working conformation The pocket neighborhood atom records are stored using an atomic sequential storage method, allowing the number of atoms that can be read in atomic order. 3D coordinate vector sequence and atomic type sequence , Representing working constructs The number of atoms, Indicates the first The three-dimensional coordinate vector of an atom. , , They represent the first The three-dimensional coordinates of each atom. Indicates the first The atom type label for each atom, which uses element symbol encoding;

[0124] Receive working configuration Then, according to the working concept Given the existing atomic arrangement, read the atomic records one by one, starting from the first atom, until the... End of atom, index any atom Synchronously read three-dimensional coordinate vectors With atom type marker Then, generate the write sequence number according to the reading order. , Indicates the first The file is written with a sequence number of atoms, following an incrementing natural number rule. The first atom is written with a sequence number of 1, and the last atom is written with a sequence number of 0. Based on this, a structure is formed and data is written. , Indicates the first The structure of each atom is written into a record, and the structure is written into data. Each record consists of underlying writable fields, containing no scoring variables or model intermediate variables. Coordinate fields are uniformly retained to three decimal places, atom type fields uniformly use element symbols, and write sequence number fields uniformly use decimal integers. After this process, the structured data is written. Already with the working concept The atomic order remains strictly consistent, allowing direct entry into the text writing stage;

[0125] when At that time, data was written to the structure according to the fixed field order of the protein data bank format. Perform a line-by-line transformation and write a record to any structure. Write the record name, sequence number, atom name, residue name, chain identifier, and residue sequence number in sequence. coordinate, coordinate, Coordinates, occupancy rate, temperature factor, and element symbol are all recorded under the fixed name "ATOM". The atom name field is directly used. The characters are aligned to a four-character width. The residue name is fixed as "PKT", the chain identifier is fixed as "A", the residue number is fixed as 1, the occupancy rate is fixed as 1.00, the temperature factor is fixed as 0.00, and the element symbol field is directly written in... For all arrive The same process is repeated to write the structure records, generating atomic record text in protein data bank format arranged in atomic order. After all atomic records are written, an end marker "END" is appended to the end of the file, thus generating the working conformation. The corresponding protein data bank format 3D structure file;

[0126] when At that time, the structure data was written according to the organization rules of the molecular structure data format. The transformation is performed by first writing to the header information area. The first line of the header information area uses the working configuration identifier text, which is obtained by concatenating the working configuration index with a fixed prefix. Then, the count area is written to the count area, which contains the number of atoms. The number of bonds is set to 0 because S7 actually receives only three-dimensional coordinate information and atom type information, without additional chemical bond information, and then writes it into the atomic block, recording any structure. Write in the order of atomic block fields according to the molecular structure data format , , and atom type markers Furthermore, the mass difference field, formal charge field, stereo mark field, valence state field, and mapping field are all uniformly set to 0, for all... arrive The same process is repeated for writing the structure records to form complete atomic block text. Since the number of bonds in the counting region is 0, no bond block content is written to the file. After the atomic block writing is completed, a structure end marker "MEND" is appended to the end of the file, thereby generating the working conformation. The corresponding molecular structure data format is a three-dimensional structure file;

[0127] After the 3D structure file is generated, it is written to the storage medium. If so, the file extension should be written as ".pdb". The file extension is then written as ".sdf". The filename is generated by combining the working construct index and the format extension, so that the 3D structure file and the working construct are linked. To maintain a unique correspondence, after writing is complete, an integrity check is performed on the 3D structure file. The integrity check includes three rules: First, the number of atomic records and the number of atoms. First, consistency is required. Second, each atomic record must contain both an atomic type field and a 3D coordinate field. Third, the end of the file must contain a string containing the atomic type field and a 3D coordinate field. When all three rules are met, the corresponding end marker registers the 3D structure file as the input structure file for virtual filtering. After this processing, the working conformation is obtained. The spatial distribution and type distribution of atoms are fixed and saved in the form of standardized files. The virtual screening program can directly read all atomic records according to the file format specifications and perform structure matching and screening calculations.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for determining the working conformation of proteins based on deep learning, characterized in that, include: S1. Obtain the non-dominant candidate conformations of the same protein, generate a set of candidate conformations, obtain calibration ligands, and generate a set of calibration ligands; S2. Under the constraint of the upper limit of docking times, molecular docking is performed on the candidate conformation set and the calibration ligand set to generate a finite docking score matrix; S3. Construct a ligand knowledge graph based on the calibration ligand set. The ligand knowledge graph includes ligand nodes, fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes. It also includes relation edges describing the correspondence between ligands and fragments, ligands and molecular nuclei, and ligands and physicochemical property binning. Train the relation path reasoning network of the knowledge graph based on the finite docking scoring matrix to generate the inherent scoring baseline of each calibration ligand. Summarize the inherent scoring baselines of the ligands to generate a ligand baseline table and output the ligand baseline table as a docking scoring decomposition constraint. S4. Construct a VAE-based matrix completion model based on the candidate conformation set and the calibration ligand set. In the decoding output and reconstruction error calculation stage of the VAE-based matrix completion model, introduce a ligand baseline table. Perform ligand intrinsic scoring baseline subtraction on the docking scores in the finite docking score matrix to obtain the conformation residual docking score. The VAE-based matrix completion model uses the conformation residual docking score as the learning target to perform matrix completion prediction and generate the residual docking score completion matrix. S5. Calculate the conformation suitability score for each candidate conformation based on the residual docking score completion matrix, and generate the conformation suitability score. S6. Determine the candidate conformations that meet the maximum score condition based on the conformation suitability score, and generate the working conformation. S7. Output the 3D structure file corresponding to the working conformation for virtual screening.

2. The method for determining the working conformation of a protein based on deep learning according to claim 1, characterized in that, S1 includes: Obtain three-dimensional structure files of multiple conformations of the same protein generated by multi-objective optimization, as well as the multi-objective values ​​corresponding to each conformation, to form a conformation candidate library; Based on the conformation candidate library, non-dominant determination is performed on multi-objective values ​​and dominated conformations are eliminated to form a non-dominant conformation record. The binding pocket region definition is determined based on the non-dominant conformation record. The binding pocket region definition includes reference ligand coordinates or preset key residue coordinates. Based on the binding pocket region definition, the binding pocket neighborhood atoms are extracted from the conformations in the non-dominant conformation record and the file format is unified to generate a set of candidate conformations. Obtain the ligand structure library, calculate the ligand structure similarity based on the ligand structure library, and select representative ligands by clustering under the constraint of a preset upper limit on the number of ligands to generate a calibration ligand set.

3. The method for determining the working conformation of a protein based on deep learning according to claim 1, characterized in that, S2 include: Based on the upper limit constraint of docking times and the set of candidate conformations, the docking search space parameters corresponding to each candidate conformation are determined, forming a search space parameter set; Based on the upper limit constraint of docking times, the search space parameter set, the set of candidate conformations and the set of calibration ligands, the required pairings of conformations and ligands for molecular docking are calculated, and a list of docking tasks is generated. Based on the docking task list and the search space parameter set, molecular docking is performed on each conformation and ligand pair combination to obtain docking scores and summarize them to form a docking score record set. Based on the docking score record set, the docking scores are written into a matrix index with the candidate conformation set as rows and the calibration ligand set as columns. Missing positions not included in the docking task list are marked with missing flags, thus generating a finite docking score matrix.

4. The method for determining the working conformation of a protein based on deep learning according to claim 1, characterized in that, S3 includes: Based on the calibration ligand set, structural analysis is performed on each calibration ligand to determine the corresponding ligand node, fragment node, molecular core node, and physicochemical property binning node, and the node correspondence is generated. Based on the node correspondence, a ligand knowledge graph is constructed, with ligand nodes, fragment nodes, molecular nucleus nodes, and physicochemical property binning nodes as graph nodes. A set of relational edges is generated based on the node correspondence, and the set of relational edges is written into the ligand knowledge graph. Non-missing docking scores are extracted from the finite docking score matrix, and the non-missing docking scores are aggregated according to the ligand nodes corresponding to the calibration ligands to generate a ligand observation score set. Based on the ligand observation score set, a summary statistic of the ligand observation score is calculated for each ligand node to generate a ligand training label set; Using the ligand knowledge graph and the ligand training label set as training inputs, the knowledge graph relational path reasoning network is trained so that the knowledge graph relational path reasoning network outputs the inherent scoring baseline of the ligand for each ligand node in the ligand knowledge graph, generating a set of inherent scoring baselines of the ligands. Based on the intrinsic scoring baseline set of ligands, the intrinsic scoring baseline of each calibrated ligand is written into the ligand baseline table according to the order of the calibrated ligand set, and the correspondence between the ligand baseline table and the calibrated ligand set forms the docking score decomposition constraint.

5. The method for determining the working conformation of a protein based on deep learning according to claim 1, characterized in that, S4 include: The row and column indices of the finite docking score matrix are determined based on the candidate conformation set and the calibration ligand set, and a VAE-based matrix completion model is constructed accordingly. This allows the VAE-based matrix completion model to generate docking score prediction matrices for the row and column indices during the decoding output and reconstruction error calculation stages. Based on the ligand baseline table and according to the column index, the intrinsic scoring baseline of each calibrated ligand is extracted to form the intrinsic scoring baseline vector of the ligand. The intrinsic scoring baseline vector of the ligand is then used as the docking score decomposition constraint input to the VAE-based matrix completion model. Perform ligand intrinsic scoring baseline subtraction on the non-missing docking scores in the finite docking scoring matrix to obtain the conformational residual docking scores corresponding to the non-missing positions, thus forming the conformational residual docking score training data. In the decoding output and reconstruction error calculation stage, the ligand intrinsic scoring baseline subtraction is performed on the docking score prediction values ​​corresponding to non-missing positions in the docking score prediction matrix to obtain the predicted conformational residual docking score, and the reconstruction error is calculated based on the conformational residual docking score training data and the predicted conformational residual docking score. The VAE-based matrix completion model is updated based on the reconstruction error, and matrix completion prediction is performed with the conformational residual docking score as the learning objective to obtain the docking score prediction matrix at all positions. The docking score prediction values ​​at all positions in the docking score prediction matrix are subtracted from the ligand-inherent scoring baseline according to the column index to generate a residual docking score completion matrix.

6. The method for determining the working conformation of a protein based on deep learning according to claim 1, characterized in that, S5 include: The correspondence between candidate conformations and matrix rows is established based on the row indices of the candidate conformation set and the residual docking score completion matrix, and the correspondence between calibration ligands and matrix columns is established based on the column indices of the calibration ligand set and the residual docking score completion matrix. For each candidate conformation in the candidate conformation set, all residual docking scores corresponding to the row of the candidate conformation are read from the residual docking score completion matrix, and a residual docking score sequence is formed according to the column order corresponding to the calibration ligand set. An aggregation calculation is performed on the residual docking score sequence, which includes summing the residual docking score sequence and dividing it by the number of ligands in the calibration ligand set to obtain the conformation suitability score corresponding to the candidate conformation. The conformation suitability scores corresponding to each conformation in the candidate conformation set are summarized in the order of the candidate conformation set to generate a conformation suitability score.

7. The method for determining the working conformation of a protein based on deep learning according to claim 1, characterized in that, S6 include: Based on the conformation suitability score, perform a traversal comparison of all score values ​​to determine the maximum score value among the conformation suitability scores; The maximum score index set is determined based on the position of the maximum score value in the conformation suitability score. Then, the corresponding candidate conformations are extracted from the candidate conformation set based on the maximum score index set to form the maximum score candidate set. When the maximum score candidate set contains multiple candidate conformations, the maximum score candidate set is filtered according to the order of the candidate conformation set, and the candidate conformation with the highest order is determined as the candidate conformation that satisfies the maximum score condition. The candidate conformation that satisfies the maximum score condition is determined as the working conformation.

8. The method for determining the working conformation of a protein based on deep learning according to claim 1, characterized in that, S7 includes: Receive the three-dimensional coordinate information and atom type information corresponding to the working conformation, and organize the three-dimensional coordinate information and atom type information according to the preset field order to form a structure for writing data; Based on the structure writing data, a three-dimensional structure file corresponding to the working conformation is generated. The three-dimensional structure file adopts the protein data bank format or the molecular structure data format. Output a 3D structure file and use it as the input structure file for virtual filtering.

9. The method for determining the working conformation of a protein based on deep learning according to claim 4, characterized in that, Using a ligand knowledge graph as the structural input and the embedding vectors of ligand nodes in the calibrated ligand set as the initial input, the embedding vectors of ligand nodes are 64-dimensional. The knowledge graph relational path reasoning network includes a three-layer graph reasoning layer and a readout network. Each graph reasoning layer gathers information from adjacent fragment nodes, adjacent molecular nucleus nodes, and adjacent physicochemical property binning nodes for ligand nodes. Each graph reasoning layer contains 64 neuron channels. The ligand node embedding vectors output by the three-layer graph reasoning layer are input to the readout network, which is a two-layer fully connected structure with 32 and 1 neurons respectively, used to output the ligand intrinsic scoring baseline corresponding to each calibrated ligand.

10. The method for determining the working conformation of a protein based on deep learning according to claim 5, characterized in that, The VAE-based matrix completion model includes a conformation encoder, a ligand encoder, and a decoder. The conformation encoder receives a 256-dimensional conformational structure feature vector, which is passed through a fully connected layer of 128 neurons and a fully connected layer of 64 neurons to output a 32-dimensional conformational latent representation. The ligand encoder receives a 256-dimensional ligand structure feature vector, which is passed through a fully connected layer of 128 neurons and a fully connected layer of 64 neurons to output a 32-dimensional ligand latent representation. The decoder receives a 64-dimensional vector formed by concatenating the conformational latent representation and the ligand latent representation, which is passed through a fully connected layer of 64 neurons and a fully connected layer of 1 neuron to output a predicted docking score. The reconstruction error is calculated only for the predicted docking score corresponding to the non-missing position in the finite docking score matrix and the conformational residual docking score obtained after deducting the intrinsic scoring baseline of the ligand from the observed docking score.

Images