A compound virtual screening method and system for a specific target

Abstract

The present application belongs to the technical field of compound virtual screening, and particularly relates to a compound virtual screening method and system for specific target points. The present application is executed in parallel through three rapid filtering means, eliminates compounds that do not meet the three rapid filtering means, and reduces the calculation cost of subsequent screening. By setting an upper limit of computing power for the to-be-scheduled stage, dynamically adjusting the retention ratio based on the real computing power consumption of the completed batches, and distributing the unspent computing power difference to the subsequent screening stage, the computing power consumption of each screening stage is always controlled. By parallelly sending the compound-target pair into multiple different docking scoring methods, dynamically assigning weights according to historical verification performance, and weighted summing to obtain a multi-method consensus docking score, the single scoring method bias is reduced, and the accuracy and reliability of the rearrangement result are improved. By calculating a multi-dimensional comprehensive score, the hit compounds are comprehensively evaluated in multiple dimensions, and a more reliable candidate compound list is provided.

Description

Technical Field

[0001] This invention belongs to the field of virtual screening technology for compounds, specifically relating to a virtual screening method and system for compounds targeting specific targets. Background Technology

[0002] Computer-aided drug design is an important technology in the field of new drug development. Structure-based virtual screening simulates the interaction between compounds and target proteins on a computer, and can quickly screen out candidate compounds with potential activity from a large compound library, thereby significantly reducing the cost and time of experimental screening.

[0003] In existing virtual screening methods, compounds are initially filtered through drug-likeness rules or ADMET property predictions. Molecular docking is then performed on the filtered compounds, and the matched compounds are ranked and output based on the docking scores. However, the screening criteria of a single filtering method may be too strict, resulting in the premature elimination of a large number of potentially active compounds at an early stage, leading to an excessively high false negative rate. During the screening process, the retention ratios for each screening stage are usually determined once before the screening begins, and cannot be dynamically adjusted according to changes in computing power consumption during the actual calculation process. This results in the computing power budget not being fully utilized, or key screening steps being forced to be interrupted when computing power is insufficient. Existing virtual screening methods typically rely on a single scoring function in the molecular docking scoring stage. A single scoring function may perform well for a specific type of compound-target binding mode, but poorly for other types, which can easily lead to unreliable rearrangement results and increase false positive and false negative rates. In the post-processing stage of the hit compounds, the top few compounds with the highest docking scores are usually directly output. A large number of structurally similar compounds are output at the same time, resulting in insufficient structural diversity of the hit compounds. Summary of the Invention

[0004] The purpose of this invention is to provide a virtual screening method and system for compounds targeting specific targets, in order to solve the problems of the prior art.

[0005] The specific technical solution adopted by this invention is as follows: A virtual screening method for compounds targeting specific targets includes the following steps: Receive the three-dimensional structural data of the target site, combined with the pocket definition and the library of compounds to be screened; The compounds in the compound library were subjected to a cascaded coarse screening using three rapid filtration methods: drug-likeness rule filtering, lethal ADMET early screening, and pharmacophore matching based on the target binding pocket, to obtain coarsely screened surviving compounds. Standard molecular docking was performed on each of the coarse-screened surviving compounds, and the compounds with the highest ranking according to the standard docking were retained to obtain the middle-screened surviving compounds; The multi-method consensus docking scoring algorithm is called one by one for the surviving compounds in the intermediate screening to obtain their multi-method consensus docking scores; the surviving compounds in the intermediate screening are rearranged according to their scores, and the compounds with the highest scores are retained to obtain the compounds hit in the fine screening. Online dynamic computing power scheduling is performed on the three-level screening consisting of cascaded coarse screening, standard molecular docking and multi-method consensus docking scoring, so that the computing power consumption of the three-level screening is controlled within the available computing power budget, and the number of surviving compounds in the three-level screening does not increase at each level along the screening direction, and the unit computing cost of a single compound increases at each level along the screening direction. The skeleton clustering of the finely screened hit compounds is performed, and a representative compound is selected from each cluster to remove structural redundancy and obtain diverse hit compounds; For diverse hit compounds, binding strength index, drugability index and structural novelty index are calculated, and the three-dimensional indexes are weighted according to the set combination weights to obtain the multidimensional comprehensive score of each diverse hit compound. The diverse hit compounds are then ranked according to the multidimensional comprehensive score. The output contains a list of diversity hits for sorted diversity hit compounds.

[0006] Preferably, the online dynamic computing power scheduling for the three-level screening consisting of cascaded coarse screening, standard molecular docking, and multi-method consensus docking scoring includes: Set a computing power limit for at least one scheduled level and divide the compounds to be evaluated in the scheduled level into several batches for batch-by-batch evaluation. During the batch-by-batch evaluation process, based on the actual computing power consumption of the completed batches, before the scheduled level has finished evaluating all its compounds, dynamically adjust the retention ratio of the scheduled level so that the actual computing power consumption of the scheduled level is constrained within the preset computing power limit. When the actual computing power consumption of a certain scheduling level is lower than its preset computing power limit, the unconsumed computing power difference of the scheduling level is carried over to the screening level after the scheduling level, so as to increase the computing power limit of the screening level after the scheduling level, and relax the retention ratio of the screening level after the scheduling level according to the increased computing power limit.

[0007] Preferably, the step of sequentially calling the multi-method consensus docking scoring algorithm to obtain the multi-method consensus docking score for each surviving compound screened includes: The compound and the target compound-target pair were fed in parallel into a variety of different docking scoring methods, including force field-based scoring functions, experience-based scoring functions, and knowledge-based scoring functions. After normalizing the scores obtained from different docking scoring methods, the weights of each docking scoring method are dynamically allocated based on their validation performance on historical target systems. Then, each normalized score is assigned a different weight and the scores are weighted and summed to obtain the multi-method consensus docking score for the surviving compounds in the mid-screening.

[0008] Preferably, the skeleton clustering of the finely screened hit compounds includes: The skeleton clustering uses the Murcko skeleton as the clustering basis, and compounds with similarity to the Murcko skeleton higher than the dynamic similarity threshold are grouped into the same cluster. The dynamic similarity threshold is dynamically adjusted according to the required level of diversity in the selected targets. When higher structural diversity is required, the dynamic similarity threshold is lowered, and when higher hit enrichment density is required, the dynamic similarity threshold is raised.

[0009] Preferably, the step of selecting a representative compound from each cluster, removing structural redundancy, and obtaining diverse hit compounds includes: The compound with the best multi-method consensus docking score within each cluster is selected as the representative compound of that cluster. When multiple fine-screened compounds with optimal multi-method consensus docking scores exist within the same cluster, secondary selection is performed based on the druggability indicators of the compounds within that cluster.

[0010] Preferably, the calculation of binding strength index, druggability index, and structural novelty index for diverse hit compounds includes: The multi-method consensus docking score of diverse hit compounds is used as an indicator of binding strength. By predicting the absorption, distribution, metabolism, excretion, and toxicity dimensions of diverse hit compounds, a comprehensive druggability prediction result for diverse hit compounds is obtained; the comprehensive druggability prediction result of diverse hit compounds is used as a druggability indicator. The structural similarity between the diverse hit compounds and known active compounds is calculated, and the structural novelty evaluation result is calculated based on the structural similarity. The lower the structural similarity, the higher the structural novelty evaluation result. The structural novelty evaluation result is used as the structural novelty index.

[0011] Preferably, the drug-likeness rule filtering is used to screen compounds based on their drug-likeness dimension; the drug-likeness rules include at least one of the Lipinski five rules, the Veber rule, and the Ghose filtering rule, and when a compound does not meet any of the drug-likeness rules, the compound is directly eliminated; The lethal ADMET early screening is used to predict the lethal druggability dimension of a compound; the lethal druggability dimension includes at least one of strong hERG cardiotoxicity risk, strong mutagenicity risk and significant water solubility defect. When any lethal druggability defect is predicted, the compound is directly eliminated. Pharmacophore matching based on target binding pockets is used to pre-screen the chemical feature matching degree between compounds and target binding pockets; the pharmacophore matching score between compounds and binding pockets is calculated, and when the pharmacophore matching score between a compound and its binding pocket is lower than a preset matching threshold, the compound is directly eliminated.

[0012] Preferably, based on the output sorted diversity hit list, the confirmed hit compounds in the diversity hit list are subjected to similarity expansion search in chemical space, and compounds with similar structures to the confirmed hits are recalled to form a new set of compounds to be screened. Then, a new screening process is performed on the new set of compounds to be screened. Output funnel statistics for each level of filtering, and also output the reliability of the diversity hit list.

[0013] A virtual screening system for compounds targeting specific targets includes the following modules: The input module is used to receive the three-dimensional structural data of the target point, combined with the pocket definition and the compound library to be screened; The cascaded coarse screening module is used to perform cascaded coarse screening of compounds in the compound library using three rapid filtering methods: drug-likeness rule filtering, lethal ADMET early screening, and pharmacophore matching based on target binding pocket, to obtain coarsely screened surviving compounds. The intermediate sieve module is used to perform standard molecular docking on each of the coarse sieve surviving compounds, and retain the compounds that are ranked first according to the standard docking, thus obtaining the intermediate sieve surviving compounds; The fine screening module is used to call the multi-method consensus docking scoring algorithm one by one on the surviving compounds of the intermediate screening to obtain their multi-method consensus docking scores; the surviving compounds of the intermediate screening are rearranged according to their scores, and the compounds with the highest scores are retained according to their consensus docking scores to obtain the fine screening hit compounds; The online dynamic computing power scheduling module is used to perform online dynamic computing power scheduling on the three-level screening consisting of cascaded coarse screening, standard molecular docking and multi-method consensus docking scoring, so that the computing power consumption of the three-level screening is controlled within the available computing power budget, and the number of surviving compounds in the three-level screening does not increase at each level along the screening direction, and the unit computing cost of a single compound increases at each level along the screening direction. The hit enrichment and redundancy removal module is used to perform skeleton clustering on the finely screened hit compounds, and select a representative compound from each cluster to remove structural redundancy and obtain diverse hit compounds. The multidimensional hit scoring module is used to calculate the binding strength index, drugability index, and structural novelty index for diverse hit compounds, and to calculate the multidimensional comprehensive score of each diverse hit compound by weighting the three dimensions of the index according to the set combination weights, and to rank the diverse hit compounds according to the multidimensional comprehensive score. The output module is used to output a list of diversity hits containing sorted diversity hit compounds.

[0014] Preferably, the online dynamic computing power scheduling for the three-level screening consisting of cascaded coarse screening, standard molecular docking, and multi-method consensus docking scoring includes: Set a computing power limit for at least one scheduled level and divide the compounds to be evaluated in the scheduled level into several batches for batch-by-batch evaluation. During the batch-by-batch evaluation process, based on the actual computing power consumption of the completed batches, before the scheduled level has finished evaluating all its compounds, dynamically adjust the retention ratio of the scheduled level so that the actual computing power consumption of the scheduled level is constrained within the preset computing power limit. When the actual computing power consumption of a certain scheduling level is lower than its preset computing power limit, the unconsumed computing power difference of the scheduling level is carried over to the screening level after the scheduling level, so as to increase the computing power limit of the screening level after the scheduling level, and relax the retention ratio of the screening level after the scheduling level according to the increased computing power limit.

[0015] Beneficial effects This invention filters out compounds that do not meet the drug-likeness criteria in advance through drug-likeness rules, filters out compounds with lethal defects through lethal ADMET screening, and filters out compounds with weak binding ability to the target through pharmacophore matching based on the target binding pocket. The three filtering methods are executed in parallel, and compounds are retained only when they simultaneously meet the retention conditions of all three filtering methods. This effectively removes compounds that do not meet the drug-likeness rules, have lethal drug-likeness defects, or do not match the target binding pocket, reducing the computational cost of subsequent screening.

[0016] This invention employs a three-stage cascaded funnel-shaped screening framework—cascaded coarse screening, standard molecular docking, and multi-method consensus docking score rearrangement—to set a computing power ceiling for at least one scheduled stage. The compounds to be evaluated in this stage are divided into several batches for batch-by-batch evaluation. During the evaluation process, based on the actual computing power consumption of completed batches, the retention ratio of the scheduled stage is dynamically adjusted before all its compounds have been evaluated, ensuring that the actual computing power consumption of the stage is constrained within the preset computing power ceiling. Furthermore, when the actual computing power consumption of a scheduled stage is lower than its preset computing power ceiling, the unused computing power difference is carried over to the screening stages following it, thereby increasing the computing power ceiling of those stages. The retention ratio of these stages is then relaxed based on the increased computing power ceiling. This ensures that the computing power consumption of each screening stage is always controlled, achieving a balance between screening throughput and hit quality within a limited computing power budget.

[0017] This invention feeds compound and target compound-target pair into multiple different docking scoring methods in parallel. After normalizing the scores obtained from different docking scoring methods, weights are dynamically assigned based on the validation performance of each different docking scoring method on historical target systems. The normalized scores are then weighted and summed to obtain a multi-method consensus docking score for the viable compounds in the mid-screening stage. This reduces the risk of misjudgment caused by the bias of a single scoring method, avoids false positives or false negatives caused by overestimating or underestimating binding strength due to a single method, and improves the accuracy and reliability of the rearrangement results.

[0018] This invention performs skeleton clustering on the finely screened compounds, grouping compounds with similarity to the Murcko skeleton higher than the dynamic similarity threshold into the same cluster. From each cluster, the compound with the best multi-method consensus docking score is selected as a representative compound. While retaining the best-performing compounds in each chemical structure category, structural redundancy is removed, ensuring rich structural diversity in the output compounds. In the multi-dimensional comprehensive scoring, a multi-dimensional comprehensive score is calculated by combining intensity, druggability, and structural novelty indices with predefined weighted combinations. This score ranks the diverse compounds, achieving a comprehensive evaluation of the compounds and providing a more reliable and valuable list of candidate compounds for subsequent experimental verification. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the online dynamic computing power scheduling of the present invention; Figure 3 This is a schematic diagram of the three-stage cascaded funnel-type screening framework of the present invention; Figure 4 This is a system module diagram of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. The following embodiments are used to illustrate the present invention, but should not be used to limit the scope of protection of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions, and simple improvements to the method of the present invention under the premise of the concept of the present invention are all within the scope of protection claimed by the present invention.

[0021] Example 1 Please refer to Figures 1-3 This embodiment provides a virtual screening method for compounds targeting specific targets, including the following steps: S1. Receive the three-dimensional structural data of the target site, combined with the pocket definition and the compound library to be screened; Among them, the three-dimensional structural data of the target point can be obtained through experimental methods such as X-ray crystal diffraction, nuclear magnetic resonance or cryo-electron microscopy, or through computational methods such as homology modeling; The binding pocket is defined as the active site region on the three-dimensional structure of the target for binding with the compound, including the boundary coordinates of the binding pocket, key residue information, and a description of the physicochemical properties of the binding pocket. The compound library to be screened can be derived from public databases, commercial compound libraries, or custom-synthesized sets of compounds, and the three-dimensional structural data and binding pocket definitions of the aforementioned target sites can be obtained from public databases.

[0022] S2. Simultaneously perform coarse screening on compounds in the compound library using three rapid filtration methods to obtain coarsely screened surviving compounds. Three rapid filtering methods include drug-like rule filtering, lethal ADMET early screening, and pharmacophore matching based on target binding pocket; S2.1, Drug-likeness rule filtering; Drug-likeness rule filtering is used to screen compounds based on their drug-likeness. Drug-likeness rules include at least one of the Lipinski five rules, the Veber rule, and the Ghose filtering rule. When a compound does not meet any of the drug-likeness rules, it is directly eliminated and will not proceed to subsequent steps such as standard molecular docking. The Lipinski five rules are as follows: the molecular weight of the compound is less than 500 Daltons, the lipid-water partition coefficient logP of the compound is less than 5, the number of hydrogen bond donors of the compound is less than 5, the number of hydrogen bond acceptors of the compound is less than 10, and the number of rotatable bonds of the compound is less than 10. When a compound meets all five conditions, the compound satisfies the Lipinski five rules; when a compound does not meet any of the five conditions, the compound does not satisfy the Lipinski five rules. The Veber rule includes: the polar surface area of ​​the compound is less than 140 square angstroms and the number of rotatable bonds in the compound is less than 10; when a compound meets both of the above conditions, the compound satisfies the Veber rule; when a compound does not meet either of the above two conditions, the compound does not satisfy the Veber rule. The Ghose filtering rules include: the molecular weight of the compound is between 160 and 480 Daltons, the lipid-water partition coefficient logP of the compound is between -0.4 and 5.6, the molar refractive index of the compound is between 40 and 130, and the number of atoms in the compound is between 20 and 70. When a compound meets all four of the above conditions, the compound meets the Ghose filtering rules; when a compound does not meet any of the above four conditions, the compound does not meet the Ghose filtering rules. S2.2, lethal ADMET early screening; The lethal ADMET early screening is used to predict the lethal druggability dimension of a compound. The lethal druggability dimension includes at least one of strong hERG cardiotoxicity risk, strong mutagenicity risk, and significant water solubility defects. When any lethal druggability defect is predicted, the compound is directly eliminated and will not proceed to subsequent steps such as standard molecular docking. A strong hERG cardiotoxicity risk is defined as a compound exhibiting inhibitory activity against hERG potassium ion channels, with a predicted IC50 value below 1 μmol / L. The IC50 value can be predicted using common methods in existing technologies. Specifically, methods based on quantitative structure-activity relationship (QSAR) models can be employed, using known hERG inhibitory activity data as a training set, extracting molecular descriptors of the compound as input features, and using machine learning algorithms to construct a regression model to output a predicted IC50 value. Alternatively, methods based on structural warnings can be used, matching known hERG toxic structural fragment libraries to output a risk score and estimate the IC50 value. Finally, methods based on molecular docking can be used, calculating the binding free energy between the compound and the hERG potassium ion channel to estimate the IC50 value. A strong mutagenic risk is defined as a compound being identified as mutagenic in an Ames test or a structure warning-based prediction model. A positive prediction result indicates a strong mutagenic risk. The structure warning-based prediction model employs commonly used existing technologies and can be based on a known library of mutagenic structural fragments. This library includes structure warning fragments from aromatic amines, nitroaromatics, epoxides, alkyl halides, α,β-unsaturated carbonyl compounds, quinones, nitrosamines, azides, and diazo compounds. The commonly used structure warning-based prediction model matches the compound's molecular structure with the known library of mutagenic structural fragments. If a match is found for any structure warning fragment, the compound is considered to have a mutagenic risk. A significant water solubility defect is defined as a compound whose predicted water solubility is below 10 micromoles per liter. The predicted water solubility of the compound is determined using a commonly used prediction model in existing technologies. Specifically, it employs a prediction method based on the quantitative structure-activity relationship model, using known experimental data on the compound's water solubility as a training set, extracting the compound's molecular descriptors as input features, and using a machine learning algorithm to construct a regression model to output the predicted water solubility value of the compound in micromoles per liter. Alternatively, empirical formulas such as the general solubility equation based on molecular structure can be used for estimation.

[0023] S2.3, Pharmacophore matching based on target binding pocket; Pharmacophore matching based on target binding pockets is used to pre-screen the chemical feature matching degree between compounds and target binding pockets. The calculation of the pharmacophore matching score between the compound and the binding pocket includes: extracting pharmacophore features of key residues in the binding pocket, including features such as hydrogen bond donors, hydrogen bond acceptors, hydrophobic centers, positive charge centers, negative charge centers, and aromatic ring centers, and recording the spatial coordinates of each feature; extracting pharmacophore features of the compound molecule and comparing the spatial positions of the pharmacophore features in the compound with those in the key residues of the binding pocket; if the distance between the two is less than a preset distance threshold of 2.0 Å and the feature types match, it is counted as a matching point; the total number of pharmacophore features of key residues in the binding pocket is taken as the total number of possible matching points; the pharmacophore matching score is equal to the number of matching points divided by the total number of possible matching points, and the pharmacophore matching score ranges from 0 to 1; The preset matching threshold is set by those skilled in the art according to the stringency of the screening, and the preset matching threshold ranges from 0.5 to 0.8. When the matching score between the compound and the pharmacophore of the binding pocket is lower than the preset matching threshold, the compound is directly eliminated and will not proceed to subsequent steps such as standard molecular docking. In practice, shape matching based on the target binding pocket can also be used to pre-screen the spatial matching degree between the compound and the target binding pocket. The calculation of the shape matching score between the compound and the binding pocket includes: extracting the three-dimensional shape features of the binding pocket, including geometric parameters such as the boundary coordinates, pocket depth, pocket volume, and pocket opening size of the binding pocket, and constructing a meshed surface model or point cloud model of the binding pocket; extracting the three-dimensional conformation of the compound molecule, and comparing the geometric complementarity between the three-dimensional conformation of the compound and the three-dimensional shape of the binding pocket, and using meshed integral to calculate the overlap volume between the surface of the compound molecule and the surface of the binding pocket; The shape matching score is equal to the overlap volume between the surface of the compound molecule and the surface of the binding pocket divided by the total volume of the compound molecule. The shape matching score ranges from 0 to 1. When the shape matching score between the compound and the binding pocket is lower than the preset matching threshold, the compound is directly eliminated and will not proceed to subsequent steps such as standard molecular docking.

[0024] Three rapid filtering methods—drug-likeness rule filtering, lethal ADMET early screening, and pharmacophore matching based on target binding pocket—were used in a cascade coarse screening of compounds in the compound library. Compounds in the library were retained only if they met the retention conditions of all three filtering methods simultaneously; compounds that did not meet the retention conditions of any one of the filtering methods were eliminated. This cascade coarse screening effectively removed compounds that did not conform to drug-likeness rules, had lethal drug-likeness defects, or did not match the target binding pocket, reducing the computational cost of subsequent screening.

[0025] S3. Perform standard molecular docking on each of the coarse-screened surviving compounds, and retain the compounds that are ranked first according to the standard docking to obtain the middle-screened surviving compounds. Standard molecular docking employs a force field-based molecular docking algorithm to place the three-dimensional conformations of coarsely screened surviving compounds one by one into the binding pocket of the target site, and calculate the binding free energy or binding affinity score between the compound and the target site as the standard docking score. The three-dimensional structures of the viable compounds from the coarse screening are extracted, and atomic types, charges, bond sequences, and rotatable bond information are assigned to them to generate ligand parameter files for the compounds. The three-dimensional structures of the target sites are extracted, and atomic types, charges, and residue protonation states are assigned to them to generate receptor parameter files for the target sites. The force field parameters are calculated using the Amber force field to determine the interaction energies between atoms. Based on the active site of the target or the known ligand binding location, a search space is defined, including the center coordinates of the binding pocket and the search range. The search range is a cubic region with a side length of 15 to 30 angstroms, with the center of the binding pocket as the origin. The three-dimensional conformations of the coarsely screened surviving compounds are placed into binding pockets, and a genetic algorithm is used to search for conformations. The position, orientation, and dihedral angle of the rotatable bonds of the compounds in the binding pockets are optimized to generate multiple docking postures; 10 to 50 docking postures are generated for each compound. For each generated docking attitude, the binding free energy between the compound and the target site was calculated. The binding free energy includes van der Waals interaction energy, electrostatic interaction energy, hydrogen bond interaction energy, desolvation effect energy, and conformational entropy change energy. All energies were calculated using the standard potential energy function and parameters in the Amber force field. From all docking attitudes, the docking attitude with the lowest binding free energy is selected as the optimal docking attitude, and the optimal binding free energy value is used as the standard docking score of the compound. The standard docking score is expressed in kilocalories per mole, and the lower the value, the stronger the binding between the compound and the target. The compounds that survived the coarse screening were sorted according to the standard docking criteria, and the compounds that ranked higher were retained as the compounds that survived the medium screening.

[0026] S4. Call the multi-method consensus docking scoring algorithm on each of the surviving compounds in the intermediate screening to obtain their multi-method consensus docking scores; rearrange the surviving compounds in the intermediate screening according to their scores, and retain the compounds with the highest scores according to their consensus docking scores to obtain the compounds that were hit in the fine screening. The compound and the target compound-target pair were fed in parallel into a variety of different docking scoring methods, including force field-based scoring functions, experience-based scoring functions, and knowledge-based scoring functions. The force field-based scoring function uses the potential energy parameters in the Amber force field to calculate the van der Waals interaction energy, electrostatic interaction energy, hydrogen bond interaction energy, and desolvation effect energy between the compound and the target. The energy is then weighted and summed according to the preset weights of the force field. The preset weights can be set to 1.0, 1.0, 0.5, 0.8, and 0.3 respectively to obtain the binding free energy score. Based on an experience-based scoring function, features such as the number of hydrogen bonds, hydrophobic contact area, number of rotatable bonds, aromatic ring stacking, and metal coordination between the compound and the target are extracted. Each feature is assigned a preset weight using a pre-fitted linear regression model, and the weighted sum is used to obtain the binding affinity score. The preset weights obtained by the pre-fitted linear regression model are 0.2, 0.01, -0.1, 0.3, and 0.5, respectively. A knowledge-based scoring function is used to construct a database of distance distributions between protein-ligand complex atoms. The frequency of atomic pairs between C, N, O, and S protein atoms and C, N, O, S, P, and halogen ligands within different distance intervals is statistically analyzed to construct an atomic pair preference potential energy matrix. Different distance intervals can be set to a range of 0 to 10 Å with a step size of 0.1 Å. The atomic pair distances between the compound and the target are input into the atomic pair preference potential energy matrix. The distance-dependent potential energy of each atomic pair is summed by looking up a table to obtain the binding preference score. The weight coefficients of the atomic pair preference potential energy matrix are obtained by Boltzmann transformation from the statistical frequencies: protein C-ligand C has a weight coefficient of 0.05, protein C-ligand O has a weight coefficient of 0.02, protein N-ligand O has a weight coefficient of 0.15, protein O-ligand N has a weight coefficient of 0.15, protein S-ligand S has a weight coefficient of 0.08, and protein C-ligand halogen has a weight coefficient of 0.03. The scores obtained from different docking scoring methods are normalized using the min-max normalization method. After normalization, the scores of each scoring method are aligned to the highest possible value, eliminating the differences in dimensions and numerical ranges between different scoring methods and mapping the scores of each scoring method to the same dimension and direction. Based on the validation performance of different docking scoring methods on historical target systems, weights are dynamically allocated, and each normalized score is assigned a different dynamically allocated weight before being weighted and summed to obtain the multi-method consensus docking score of the surviving compounds in the mid-screening. In this embodiment, the initial score weights of the force field-based scoring function, the experience-based scoring function, and the knowledge-based scoring function can be set to 0.4, 0.35, and 0.25, respectively. Dynamic weight allocation is performed based on the validation performance of different docking scoring methods on historical target systems. This includes extracting the binding characteristics between each surviving compound in the middle screening and the target binding pocket to form a feature vector. The feature vector includes the number of hydrogen bond matching points between the compound and the binding pocket, the hydrophobic contact area between the compound and the binding pocket, the number of rotatable bonds, and the molecular weight. For each surviving compound in the intermediate screening, the prediction confidence of each docking scoring method for that compound is calculated based on its feature vector. Its functional relationship is:

[0027] In the formula, The value represents the prediction confidence level of this docking scoring method for the compound, ranging from 0 to 1; The first eigenvector One eigenvalue; For the first The weight coefficients corresponding to each feature are determined by a pre-trained neural network model; This is the offset, determined by a pre-trained neural network model; The Sigmoid function is used to map the output to a range between 0 and 1. The confidence calculation model for each docking scoring method is pre-trained based on a training set and a training target. The training set consists of compound-target complexes containing known binding data, with each complex having a known binding free energy or binding affinity. The training target is the relationship between the prediction error and the feature vector of each docking scoring method; the smaller the error, the higher the confidence. The confidence calculation model adopts a fully connected neural network structure. Its input layer contains 4 neurons, corresponding to the 4 feature values ​​in the feature vector, namely the number of hydrogen bond matching points, hydrophobic contact area, number of rotatable bonds, and molecular weight. The fully connected neural network contains one hidden layer with 8 neurons, which uses the ReLU activation function. The output layer contains 1 neuron, which uses the Sigmoid activation function and outputs a confidence value ranging from 0 to 1. During training, the training set is randomly divided into a training subset and a validation subset in an 8:2 ratio. The training subset is used to update model parameters, while the validation subset is used to monitor whether the model is overfitting. Mean squared error is used as the loss function, and gradient descent is performed using the Adam optimizer. The initial learning rate is set to 0.001, the batch size is 64, and the maximum number of training epochs is 200. Training is terminated early when the loss function value on the validation subset no longer decreases in 10 consecutive training epochs to prevent overfitting. After training, the model with the smallest loss function value on the validation subset is saved as the final confidence calculation model. Through the above training process, each docking scoring method obtains an independent confidence calculation model. The independent confidence calculation model can output the prediction confidence of the scoring method for the current surviving compound based on the feature vector of the current surviving compound in the screening. Based on the confidence level of each docking scoring method, weights are dynamically assigned. The weight assignment formula is as follows:

[0028] The dynamic weight of this docking scoring method ranges from 0 to 1, and the sum of the dynamic weights of all methods is 1. The confidence level of this docking scoring method in predicting this compound is given. The confidence level of the force field-based scoring function. The confidence level of the experience-based scoring function. The confidence level of the knowledge-based scoring function. To avoid the denominator being zero, it is a very small constant; in this embodiment, it is 10.-8 ; By dynamically allocating weights, the scoring method trained on the historical target system, which has a higher confidence in predicting current compounds, can make a greater contribution in the current rearrangement, thus reducing the risk of misjudgment caused by the bias of a single scoring method. The surviving compounds from the intermediate screening are rearranged according to their scores, and the compounds with the highest scores are retained according to the consensus docking scores to obtain the compounds that are hit in the fine screening. Specifically, the top 10% of the compounds can be retained from the ranking results, and the settings can be dynamically adjusted according to the rigor of the screening and the size of the compound library.

[0029] S5. Online dynamic computing power scheduling is performed on the three-level screening consisting of cascade coarse screening, standard molecular docking, and multi-method consensus docking scoring. A computing power limit is set for at least one scheduling level, and the compounds to be evaluated in the scheduling level are divided into several batches for batch-by-batch evaluation; wherein, the scheduling level is a set of compounds that have been retained after a certain screening step but have not yet entered the next computationally intensive step, and the compounds in the scheduling level are in a state of waiting to be scheduled. During the batch-by-batch evaluation process, based on the actual computing power consumption of the completed batches, it is determined whether the actual computing power consumption of the completed batches is lower than the preset computing power limit. When the actual computing power consumption of a certain scheduling level is not lower than its preset computing power limit, the retention ratio of the scheduling level is dynamically adjusted before the scheduling level has finished evaluating all its compounds, so that the actual computing power consumption of the scheduling level is constrained within the preset computing power limit. When the actual computing power consumption of a certain scheduling level is lower than its preset computing power limit, the unconsumed computing power difference of the scheduling level is carried over to the screening level after the scheduling level, so as to increase the computing power limit of the screening level after the scheduling level, and relax the retention ratio of the screening level after the scheduling level according to the increased computing power limit. Specifically, when the computing power limit for the standard molecular docking stage is 1000 units, with each unit being one CPU core hour, and there are 10,000 viable compounds to be evaluated in the coarse screening, these 10,000 compounds are divided into 10 batches of 1000 compounds each. After evaluating the first two batches of 2000 compounds, if the actual computing power consumption is found to be 150 units, which is lower than the proportionally allocated budget of 200 units, the remaining 50 units of computing power difference can be carried over to subsequent batches, thereby allowing subsequent batches to retain more compounds and increasing the sensitivity of the intermediate screening stage. If the actual computing power consumption is higher than the budget, the retention ratio is dynamically reduced to ensure that computing power is not over-consumed.

[0030] By using online dynamic computing power scheduling, it is possible to maximize screening throughput and hit quality within a limited computing power budget, while ensuring that the computing power consumption of each screening stage is always under control.

[0031] S6. Perform skeleton clustering on the finely screened hit compounds, and select a representative compound from each cluster to remove structural redundancy and obtain diverse hit compounds; The skeleton clustering uses the Murcko skeleton as the clustering basis, and compounds with similarity to the Murcko skeleton higher than the dynamic similarity threshold are grouped into the same cluster. The Murcko skeleton refers to the core ring structure and the atoms connecting the ring systems that remain after removing all side chain atoms from a compound molecule. It is the smallest skeleton structure composed of the ring system and the bridging atoms connecting the ring systems. For a compound with two benzene rings connected by a carbon atom, its Murcko skeleton consists of the two benzene rings and the carbon atom connecting them. For a compound containing only one benzene ring, its Murcko skeleton is the benzene ring itself. The dynamic similarity threshold is dynamically adjusted according to the required level of diversity in the screening targets. When higher structural diversity is required, the dynamic similarity threshold is lowered, and when higher hit enrichment density is required, the dynamic similarity threshold is raised. The hit enrichment density is the proportion of compounds with high scores in the same cluster. If the screening objective is to obtain the most diverse range of hit compounds, the dynamic similarity threshold is set to a lower value such as 0.6, so that only compounds with highly similar structures are grouped into the same cluster; if the screening objective is to obtain as many hit compounds as possible, the dynamic similarity threshold is set to a higher value such as 0.8, so that more compounds with similar structures are grouped into the same cluster.

[0032] From each cluster, a representative compound is selected, and structural redundancy is removed to obtain diverse hit compounds. This includes selecting the fine-screened hit compound with the best multi-method consensus docking score within each cluster as the representative compound of that cluster; when there are multiple fine-screened hit compounds with the best multi-method consensus docking scores within the same cluster, secondary selection is performed based on the druggability index of the compounds within that cluster; ensuring that the compound with the best score and the best druggability in each cluster is retained, thereby improving the overall quality of hit compounds while ensuring hit diversity.

[0033] S7. For the diverse hit compounds, calculate the binding strength index, drugability index, and structural novelty index; and calculate the multidimensional comprehensive score of each diverse hit compound by weighting the three dimensions of the index according to the set combination weights, and rank the diverse hit compounds according to the multidimensional comprehensive score. The multi-method consensus docking score of diverse hit compounds is directly used as an indicator of binding strength. The higher the multi-method consensus docking score, the stronger the binding between the compound and the target. By predicting the absorption, distribution, metabolism, excretion, and toxicity dimensions of diverse hit compounds, a comprehensive druggability prediction result for diverse hit compounds is obtained; the comprehensive druggability prediction result of diverse hit compounds is used as a druggability indicator. Among them, by using a prediction model based on quantitative structure-activity relationship model for diverse hit compounds, the molecular descriptors such as molecular weight, lipid-water partition coefficient logP, polar surface area, number of hydrogen bond donors and acceptors, and number of rotatable bonds of the compounds are extracted from the training sets constructed for each dimension as input features. The random forest algorithm is used to construct a regression model and output the corresponding predicted values. Specifically, the training set for predicting the absorption dimension consists of known experimental data such as intestinal absorption rate, bioavailability, and Caco-2 cell permeability of compounds; the training set for predicting the distribution dimension consists of known experimental data such as plasma protein binding rate, apparent volume of distribution, and blood-brain barrier permeability of compounds; the training set for predicting the metabolism dimension consists of known experimental data such as metabolic half-life, CYP450 enzyme inhibition IC50 value, and metabolic pathway of compounds; the training set for predicting the excretion dimension consists of known experimental data such as plasma clearance rate, half-life, and renal clearance of compounds; and the training set for predicting the toxicity dimension consists of known experimental data such as hepatotoxicity, nephrotoxicity, genotoxicity, hERG inhibition, and skin sensitization of compounds. The predicted values ​​for the five dimensions were mapped to a range of 0 to 1 using a min-max normalization method. After normalization, a weighted sum was performed to obtain a quantified drugability score. The weights for the predictions of absorption, distribution, metabolism, excretion, and toxicity were 0.25, 0.2, 0.2, 0.15, and 0.2, respectively. A higher drugability score indicates that the compound is more suitable as a candidate drug. The structural similarity between the diverse hit compounds and known active compounds is calculated, and the structural novelty evaluation result is calculated based on the structural similarity. The lower the structural similarity, the higher the structural novelty evaluation result. The structural novelty evaluation result is used as the structural novelty index. Among them, the known active compounds can be derived from public databases such as ChEMBL and PubChem or active compounds reported in existing literature. Structural similarity can be calculated using the Tanimoto coefficient based on molecular fingerprints. The molecular structure of the compound is extracted to generate an ECFP4 molecular fingerprint with a length of 1024 bits. For each pair of compounds, the Tanimoto coefficient of their molecular fingerprint is calculated. Structural similarity based on Tanimoto coefficient The calculation formula is:

[0034] In the formula, For structural similarity, This represents the total number of sites with a value of 1 in the fingerprint of compound A. This represents the total number of sites with a value of 1 in the fingerprint of compound B. This represents the total number of sites that are simultaneously 1 in the fingerprints of compounds A and B. The structural novelty evaluation result is calculated by structural similarity, and the result is the sum of the novelty evaluation result and the structural similarity. =1; then the higher the novelty evaluation result, which is used as an indicator of structural novelty, the greater the structural difference between the diversity hit compound and the known active compound, and the more likely it is to discover a completely new chemical structure category. The multidimensional comprehensive score is calculated as follows: before calculating the multidimensional comprehensive score, the binding strength index, drugability index, and structural novelty index are normalized respectively. A minimum-maximum normalization method can be used to map the values ​​of each index to the range of 0 to 1, thereby eliminating dimensional differences. The binding strength index, drugability index, and structural novelty index are multiplied by their respective combination weights and then summed. The combination weights can be set according to the screening objective. If the screening objective prioritizes binding strength, the weight of the binding strength index can be set to 0.5, and the weights of the drugability index and structural novelty index can each be set to 0.25. If the screening objective prioritizes drugability, the weight allocation can be adjusted accordingly. Therefore, based on the multidimensional comprehensive score of each diverse hit compound, the diverse hit compounds are ranked; through the multidimensional comprehensive score, the comprehensive value of each diverse hit compound can be comprehensively evaluated, avoiding the one-sidedness caused by relying on only a single dimension indicator.

[0035] S8. Output a list of diversity hits containing sorted diversity hit compounds; Based on the output sorted diversity hit list, compounds that have been experimentally verified to be active or compounds ranked high in the multidimensional comprehensive score are selected as confirmed hit compounds. Similarity expansion search is performed on the confirmed hit compounds in chemical space to recall compounds with similar structures to the confirmed hits to form a new subset of compounds to be screened. A new screening process is then performed on this new subset of compounds to be screened. Specifically, the complete screening process from S2 to S7 is performed, but the compound library in S1 is replaced with the new subset of compounds to be screened. Similarity-based extended retrieval employs Tanimoto similarity based on molecular fingerprints. As a retrieval basis, a recall similarity threshold is set to control the number and structural diversity of recalled compounds. The recall similarity threshold can be set according to the screening requirements, and the typical value range is 0.6 to 0.9. When it is necessary to recall highly similar compounds, a higher threshold such as 0.85 is set; when it is necessary to expand the recall scope and increase structural diversity, a lower threshold such as 0.6 is set. Through iterative expansion, more candidate compounds with similar structures but potentially higher activity can be discovered around the confirmed hit compounds, further enriching the hit compound library; Output funnel statistics for each level of filtering, and also output the reliability of the diversity hit list.

[0036] The funnel statistics include the number of compounds screened at each stage, the rejection rate, and the retention rate. This allows for further evaluation of the efficiency of the screening process and the rationality of the retention strategies at each stage. The reliability of the screening results is quantified based on the consistency of the consensus scoring of multiple methods. When the ranking results of the compounds by each scoring method are similar, the given scores are close in the normalized range, and the evaluation direction of the binding mode of the compounds is consistent, all of which consider the binding to be good or poor, the scoring results of each scoring method are considered to be highly consistent, indicating that the binding mode of the hit compound is relatively reliable and has high reliability. Conversely, when the scoring results of each scoring method have a large dispersion, it indicates that the binding mode of the hit compound is uncertain and the reliability is low.

[0037] Example 2 Please refer to Figure 4 This embodiment provides a virtual screening system for compounds targeting specific targets, used to execute a virtual screening method for compounds targeting specific targets as described in Embodiment 1, including the following modules: The input module is used to receive the three-dimensional structural data of the target point, combined with the pocket definition and the compound library to be screened; The cascaded coarse screening module is used to perform cascaded coarse screening of compounds in the compound library using three rapid filtering methods: drug-likeness rule filtering, lethal ADMET early screening, and pharmacophore matching based on target binding pocket, to obtain coarsely screened surviving compounds. The intermediate sieve module is used to perform standard molecular docking on each of the coarse sieve surviving compounds, and retain the compounds that are ranked first according to the standard docking, thus obtaining the intermediate sieve surviving compounds; The fine screening module is used to call the multi-method consensus docking scoring algorithm one by one on the surviving compounds of the intermediate screening to obtain their multi-method consensus docking scores; the surviving compounds of the intermediate screening are rearranged according to their scores, and the compounds with the highest scores are retained according to their consensus docking scores to obtain the fine screening hit compounds; The online dynamic computing power scheduling module is used to perform online dynamic computing power scheduling on the three-level screening consisting of cascaded coarse screening, standard molecular docking and multi-method consensus docking scoring, so that the computing power consumption of the three-level screening is controlled within the available computing power budget, and the number of surviving compounds in the three-level screening does not increase at each level along the screening direction, and the unit computing cost of a single compound increases at each level along the screening direction. The hit enrichment and redundancy removal module is used to perform skeleton clustering on the finely screened hit compounds, and select a representative compound from each cluster to remove structural redundancy and obtain diverse hit compounds. The multidimensional hit scoring module is used to calculate the binding strength index, drugability index, and structural novelty index for diverse hit compounds, and to calculate the multidimensional comprehensive score of each diverse hit compound by weighting the three dimensions of the index according to the set combination weights, and to rank the diverse hit compounds according to the multidimensional comprehensive score. The output module is used to output a list of diversity hits containing sorted diversity hit compounds.

[0038] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A virtual screening method for compounds targeting specific targets, characterized in that, Includes the following steps: Receive the three-dimensional structural data of the target site, combined with the pocket definition and the library of compounds to be screened; The compounds in the compound library were subjected to a cascaded coarse screening using three rapid filtration methods: drug-likeness rule filtering, lethal ADMET early screening, and pharmacophore matching based on the target binding pocket, to obtain coarsely screened surviving compounds. Standard molecular docking was performed on each of the coarse-screened surviving compounds, and the compounds with the highest ranking according to the standard docking were retained to obtain the middle-screened surviving compounds; The multi-method consensus docking scoring algorithm was called on each of the surviving compounds in the intermediate screening to obtain their multi-method consensus docking scores; The surviving compounds from the intermediate screening are rearranged according to their scores, and the compounds with the highest scores are retained according to the consensus docking scores to obtain the compounds that are hit in the fine screening. Online dynamic computing power scheduling is performed on the three-level screening consisting of cascaded coarse screening, standard molecular docking and multi-method consensus docking scoring, so that the computing power consumption of the three-level screening is controlled within the available computing power budget, and the number of surviving compounds in the three-level screening does not increase at each level along the screening direction, and the unit computing cost of a single compound increases at each level along the screening direction. The skeleton clustering of the finely screened hit compounds is performed, and a representative compound is selected from each cluster to remove structural redundancy and obtain diverse hit compounds; For diverse hit compounds, binding strength index, drugability index and structural novelty index are calculated, and the three-dimensional indexes are weighted according to the set combination weights to obtain the multidimensional comprehensive score of each diverse hit compound. The diverse hit compounds are then ranked according to the multidimensional comprehensive score. The output contains a list of diversity hits, sorted by the number of diversity hits. The online dynamic computing power scheduling for the three-level screening consisting of cascade coarse screening, standard molecular docking, and multi-method consensus docking scoring includes: setting a computing power upper limit for at least one scheduling level, and dividing the compounds to be evaluated in the scheduling level into several batches for batch-by-batch evaluation; During the batch-by-batch evaluation process, based on the actual computing power consumption of completed batches, the retention ratio of the scheduled level is dynamically adjusted before all compounds to be evaluated are evaluated, so that the actual computing power consumption of the scheduled level is constrained within the preset computing power limit. When the actual computing power consumption of a scheduled level is lower than its preset computing power limit, the unconsumed computing power difference of the scheduled level is carried over to the screening level following the scheduled level to increase the computing power limit of the screening level following the scheduled level, and the retention ratio of the screening level following the scheduled level is relaxed according to the increased computing power limit.

2. The virtual screening method for compounds targeting specific targets according to claim 1, characterized in that, The process of sequentially calling the multi-method consensus docking scoring algorithm to obtain the multi-method consensus docking score for each surviving compound screened in the intermediate screening includes: The compound and the target compound-target pair were fed in parallel into a variety of different docking scoring methods, including force field-based scoring functions, experience-based scoring functions, and knowledge-based scoring functions. After normalizing the scores obtained from different docking scoring methods, the weights of each docking scoring method are dynamically allocated based on their validation performance on historical target systems. Then, each normalized score is assigned a different weight and the scores are weighted and summed to obtain the multi-method consensus docking score for the surviving compounds in the mid-screening.

3. The virtual screening method for compounds targeting specific targets according to claim 1, characterized in that, The skeleton clustering of the selected compounds includes: The skeleton clustering uses the Murcko skeleton as the clustering basis, and compounds with similarity to the Murcko skeleton higher than the dynamic similarity threshold are grouped into the same cluster. The dynamic similarity threshold is dynamically adjusted according to the required level of diversity in the selected targets. When higher structural diversity is required, the dynamic similarity threshold is lowered, and when higher hit enrichment density is required, the dynamic similarity threshold is raised.

4. The virtual screening method for compounds targeting specific targets according to claim 1, characterized in that, The process of selecting a representative compound from each cluster, removing structural redundancy, and obtaining diverse hit compounds includes: The compound with the best multi-method consensus docking score within each cluster is selected as the representative compound of that cluster. When multiple fine-screened compounds with optimal multi-method consensus docking scores exist within the same cluster, secondary selection is performed based on the druggability indicators of the compounds within that cluster.

5. The virtual screening method for compounds targeting specific targets according to claim 1, characterized in that, For the diverse hit compounds, the calculation of binding strength, druggability, and structural novelty indices includes: The multi-method consensus docking score of diverse hit compounds is used as an indicator of binding strength. By predicting the absorption, distribution, metabolism, excretion, and toxicity dimensions of diverse hit compounds, a comprehensive druggability prediction result for diverse hit compounds is obtained; the comprehensive druggability prediction result of diverse hit compounds is used as a druggability indicator. The structural similarity between the diverse hit compounds and known active compounds is calculated, and the structural novelty evaluation result is calculated based on the structural similarity. The lower the structural similarity, the higher the structural novelty evaluation result. The structural novelty evaluation result is used as the structural novelty index.

6. The virtual screening method for compounds targeting specific targets according to claim 1, characterized in that: Drug classability rule filtering is used to screen compounds based on their drug classability. Drug classability rules include at least one of the Lipinski five rules, the Veber rule, and the Ghose filtering rule. When a compound does not meet any of the drug classability rules, the compound is directly eliminated. The lethal ADMET early screening is used to predict the lethal druggability dimension of a compound; the lethal druggability dimension includes at least one of strong hERG cardiotoxicity risk, strong mutagenicity risk and significant water solubility defect. When any lethal druggability defect is predicted, the compound is directly eliminated. Pharmacophore matching based on target binding pockets is used to pre-screen the chemical signature matching degree between compounds and target binding pockets. The pharmacophore matching score between the compound and the binding pocket is calculated. When the pharmacophore matching score between the compound and the binding pocket is lower than the preset matching threshold, the compound is directly eliminated.

7. The virtual screening method for compounds targeting specific targets according to claim 1, characterized in that: Based on the output sorted diversity hit list, similarity expansion search is performed on the confirmed hit compounds in the chemical space to recall compounds with similar structures to the confirmed hits to form a new set of compounds to be screened. Then, a new screening process is performed on this new set of compounds to be screened. Output funnel statistics for each level of filtering, and also output the reliability of the diversity hit list.

8. A virtual screening system for compounds targeting specific targets, characterized in that, Includes the following modules: The input module is used to receive the three-dimensional structural data of the target point, combined with the pocket definition and the compound library to be screened; The cascaded coarse screening module is used to perform cascaded coarse screening of compounds in the compound library using three rapid filtering methods: drug-likeness rule filtering, lethal ADMET early screening, and pharmacophore matching based on target binding pocket, to obtain coarsely screened surviving compounds. The intermediate sieve module is used to perform standard molecular docking on each of the coarse sieve surviving compounds, and retain the compounds that are ranked first according to the standard docking, thus obtaining the intermediate sieve surviving compounds; The fine screening module is used to call the multi-method consensus docking scoring algorithm on each of the surviving compounds in the intermediate screening to obtain their multi-method consensus docking scores. The surviving compounds from the intermediate screening are rearranged according to their scores, and the compounds with the highest scores are retained according to the consensus docking scores to obtain the compounds that are hit in the fine screening. The online dynamic computing power scheduling module is used to perform online dynamic computing power scheduling on the three-level screening consisting of cascaded coarse screening, standard molecular docking and multi-method consensus docking scoring, so that the computing power consumption of the three-level screening is controlled within the available computing power budget, and the number of surviving compounds in the three-level screening does not increase at each level along the screening direction, and the unit computing cost of a single compound increases at each level along the screening direction. The hit enrichment and redundancy removal module is used to perform skeleton clustering on the finely screened hit compounds, and select a representative compound from each cluster to remove structural redundancy and obtain diverse hit compounds. The multidimensional hit scoring module is used to calculate the binding strength index, drugability index, and structural novelty index for diverse hit compounds, and to calculate the multidimensional comprehensive score of each diverse hit compound by weighting the three dimensions of the index according to the set combination weights, and to rank the diverse hit compounds according to the multidimensional comprehensive score. The output module is used to output a list of diversity hits containing sorted diversity hit compounds.

9. The virtual screening system for compounds targeting specific targets according to claim 8, characterized in that, The online dynamic computing power scheduling for the three-level screening consisting of cascaded coarse screening, standard molecular docking, and multi-method consensus docking scoring includes: Set a computing power limit for at least one scheduled level and divide the compounds to be evaluated in the scheduled level into several batches for batch-by-batch evaluation. During the batch-by-batch evaluation process, based on the actual computing power consumption of the completed batches, before the scheduled level has finished evaluating all its compounds, dynamically adjust the retention ratio of the scheduled level so that the actual computing power consumption of the scheduled level is constrained within the preset computing power limit. When the actual computing power consumption of a certain scheduling level is lower than its preset computing power limit, the unconsumed computing power difference of the scheduling level is carried over to the screening level after the scheduling level, so as to increase the computing power limit of the screening level after the scheduling level, and relax the retention ratio of the screening level after the scheduling level according to the increased computing power limit.

Images