Molecular virtual screening methods and related equipment based on artificial intelligence

Abstract

The present disclosure relates to a molecular virtual screening method based on artificial intelligence and related equipment, the method comprising: obtaining SMILES of multiple molecules to be screened and a series of ligand molecules, calculating the molecular descriptors, parent core structures and substituent groups of the series of ligand molecules; generating a 3D conformation of the ligand molecule and performing molecular docking with the 3D conformation of the target protein, and calculating its protein-ligand interaction; obtaining molecular activity experimental data of each ligand molecule; training an XGBoost model by constituting the molecular descriptors, parent core structures, substituent groups, protein-ligand interactions and corresponding molecular activity experimental data of the ligand molecule as training data; incorporating the trained XGBoost model into a SHAP algorithm for computational analysis to obtain the SHAP values ​​of various input data; formulating molecular screening rules based on the SHAP values; and screening molecules that meet the requirements from multiple molecules to be screened based on the molecular screening rules.

Description

Technical Field

[0001] The present disclosure relates to the technical field of drug-assisted design, and in particular, to a molecular virtual screening method based on artificial intelligence and related equipment. Background Art

[0002] Drug development is a large and competitive market, with a growing demand for rapid discovery and development of new drugs. The traditional drug development process requires a lot of time and resources, and there are high costs and uncertainties. Therefore, the market demand for efficient, intelligent and precise drug development methods is gradually increasing. In recent years, with the continuous improvement of computer performance, the application of artificial intelligence in drug design has developed rapidly.

[0003] Molecular virtual screening is a key step in drug design, which aims to use computer technology to screen large compound libraries to identify potential drug candidate molecules. This method can avoid the high cost and time consumption of actual synthesis and testing of compounds, and improve the efficiency and success rate of drug research and development. In related technologies, molecular virtual screening uses artificial intelligence technology to improve screening efficiency and accuracy, but it is limited by the input data of the artificial intelligence model and the interpretability of the model itself, and the accuracy of molecular virtual screening needs to be further improved. Summary of the invention

[0004] The purpose of the present disclosure is to provide an artificial intelligence-based molecular virtual screening method and related equipment, aiming to expand the input data dimension and enhance the interpretability of the model to improve the accuracy of molecular virtual screening.

[0005] In order to achieve the above-mentioned object, the present disclosure provides a first aspect of a molecular virtual screening method based on artificial intelligence, the method comprising:

[0006] Obtain SMILES of a series of ligand molecules and multiple molecules to be screened;

[0007] Calculating the molecular descriptors, the parent core structure and the substituent groups relative to the parent core structure of the series of ligand molecules based on the SMILES of the series of ligand molecules;

[0008] Selecting the 3D conformation of the target protein of each ligand molecule, generating the 3D conformation of the ligand molecule and performing molecular docking with the 3D conformation of the target protein, and calculating the protein-ligand interaction between the ligand molecule and the target protein after molecular docking, wherein the protein-ligand interaction includes amino acid residues and protein-ligand interaction type;

[0009] Obtaining molecular activity experimental data of each of the ligand molecules;

[0010] The molecular descriptor of the ligand molecule, the core structure, the substituent group relative to the core structure, the protein-ligand interaction, and the corresponding molecular activity experimental data constitute training data, and the XGBoost model is trained based on the training data and the parameters are adjusted to obtain the trained XGBoost model;

[0011] Incorporating the trained XGBoost model into the SHAP algorithm for computational analysis to obtain SHAP values ​​of the molecular descriptor, the parent core structure, the substituent group, and the protein-ligand interaction, wherein the SHAP value is used to indicate the contribution of the corresponding data to the molecular activity;

[0012] Select at least one of a molecular descriptor, a core structure, a substitution group, and a protein-ligand interaction as a target screening feature, and formulate a molecular screening rule for the target screening feature based on the SHAP value;

[0013] Based on the SMILES of the molecule to be screened, molecules that meet the molecular screening rules are screened from the multiple molecules to be screened.

[0014] Optionally, the calculation of the molecular descriptors, the core structure and the substituent groups relative to the core structure of the series of ligand molecules based on the SMILES of the series of ligand molecules comprises:

[0015] Calculating a molecular descriptor of each of the ligand molecules in the series of ligand molecules based on the SMILES of the series of ligand molecules, wherein the molecular descriptor includes molecular low-dimensional structure information and molecular 3D structure information;

[0016] Based on the FMCS algorithm of the SMILES, cluster analysis is performed on the series of ligand molecules to obtain one or more types of ligand molecule sets, and the maximum common substructure of each type of ligand molecule set is calculated as the mother core structure of the type of ligand molecule set;

[0017] For each ligand molecule in each type of ligand molecule set, the chemical bond between the ligand molecule and the parent core structure of the ligand molecule set is broken to obtain a substitution group of the ligand molecule relative to the parent core structure.

[0018] Optionally, the molecular low-dimensional structure information includes molecular 0D structure information, molecular 1D structure information, molecular 2D structure information and molecular property descriptors.

[0019] Optionally, the molecular 0D structural information includes molecular weight and chemical bond type, the molecular 1D structural information includes the number of hydrogen bond donors, the number of hydrogen bond acceptors, and polar surface area, the molecular 2D structural information includes molecular shape and hybridization type, and the molecular property descriptors include lipid solubility, water solubility, and drug-like properties.

[0020] Optionally, the molecular 3D structure information includes molecular shape, three-dimensional atomic space coordinates, shape complexity, and best fit plane.

[0021] Optionally, the SHAP value is used to indicate the contribution of the corresponding data to the molecular activity, including:

[0022] When the SHAP value is positive, the SHAP value indicates that the contribution of the corresponding data to the molecular activity is positive, and when the SHAP value is negative, the SHAP value indicates that the contribution of the corresponding data to the molecular activity is negative; the absolute value of the SHAP value indicates the degree of contribution of the corresponding data to the molecular activity.

[0023] Optionally, the selecting at least one of a molecular descriptor, a core structure, a substitution group, and a protein-ligand interaction as a target screening feature, and formulating a molecular screening rule for the target screening feature based on the SHAP value, comprises:

[0024] A molecular descriptor with a positive SHAP value is selected as a target screening feature, and a target value range of the SHAP value of the selected molecular descriptor is determined as at least a part of the molecular screening rule.

[0025] Optionally, the selecting at least one of a molecular descriptor, a core structure, a substitution group, and a protein-ligand interaction as a target screening feature, and formulating a molecular screening rule for the target screening feature based on the SHAP value, comprises:

[0026] At least one of the core structures, substituent groups, and protein-ligand interactions with positive SHAP values ​​is selected as a target screening feature, and a rule that the molecules to be screened must have the target screening feature is formulated as at least a part of the molecular screening rule.

[0027] A second aspect of the present disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect.

[0028] A third aspect of the present disclosure provides an electronic device, including:

[0029] a memory having a computer program stored thereon;

[0030] A processor is used to execute the computer program in the memory to implement the steps of any one of the methods described in the first aspect above.

[0031] Through the above technical solution, the SMILES of the molecules to be screened and the series of ligand molecules are obtained, and then the molecular descriptors, parent core structures, substitution groups and multi-dimensional input data of the series of ligand molecules are calculated, and the XGBoost model is trained and adjusted together with the obtained molecular activity test data. The trained XGBoost model is incorporated into the SHAP algorithm for structure-activity relationship analysis to explain the contribution of each input data of the XGBoost model to the molecular activity, and molecular virtual screening rules are formulated according to the explained contribution for molecular virtual screening. A multi-dimensional feature representation system is used to provide more comprehensive and in-depth information support for structure-activity relationship analysis. Combined with the high-precision XGBoost model, the accuracy and robustness of structure-activity relationship analysis in molecular virtual screening are greatly improved. Based on the interpretation framework of the SHAP algorithm, the model is interpreted in a fine-grained manner, which facilitates a clear understanding of the contribution of each feature to the model prediction, and thus better understands the model decision-making process, formulates more accurate molecular screening rules, and improves the accuracy of molecular screening.

[0032] Other features and advantages of the present disclosure will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings are used to provide a further understanding of the present disclosure and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present disclosure but do not constitute a limitation of the present disclosure. In the accompanying drawings:

[0034] Figure 1 The present invention is a flowchart of a molecular virtual screening method based on artificial intelligence according to an exemplary embodiment.

[0035] Figure 2 The figure is a flowchart of multi-dimensional feature data calculation according to an exemplary embodiment.

[0036] Figure 3 is another flow chart of a molecular virtual screening method based on artificial intelligence according to an exemplary embodiment.

[0037] Figure 4 It is a block diagram of a molecular virtual screening device based on artificial intelligence according to an exemplary embodiment.

[0038] Figure 5 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0039] The specific implementation of the present disclosure is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described herein is only used to illustrate and explain the present disclosure, and is not used to limit the present disclosure.

[0040] Structure-Activity Relationship (SAR) analysis is a key step in drug design. It guides the design and optimization of new drugs by studying the relationship between the chemical structure of a compound and its biological activity. By performing structure-activity relationship analysis on compound molecules, the design of drug molecules can be guided, such as virtual screening of drug molecules. Traditional SAR analysis methods usually require a large amount of experimental data and complex calculation fitting, and the process is cumbersome and time-consuming. With the development of machine learning and deep learning algorithms, the application of AI in SAR analysis has gradually become popular. AI technology can quickly process and analyze a large amount of compound data to obtain the structure-activity relationship of a series of compounds, and use it to predict the activity and toxicity of potential candidate drugs and optimize the molecular structure. Compared with traditional methods, AI-driven SAR analysis can significantly shorten the drug development cycle, reduce costs, and increase the success rate.

[0041] However, despite the remarkable progress made by artificial intelligence in SAR analysis, the relevant technology still has some important defects and deficiencies, such as biased and limited data quality, limited generalization performance, and single dimension of feature description. In the field of medicine, accuracy is not the only indicator to measure the computational model. The degree of explanation of the model decision process and prediction output results is also very important. Existing artificial intelligence algorithms are often considered to be "black box models", that is, the internal operation of the algorithm is opaque. Because of their complex structure, it is difficult for humans to accurately grasp the decision-making process in model construction, which leads to a series of serious problems such as trust, responsibility, adjustability and ethics.

[0042] Some of the latest interpretable machine learning algorithms can solve the "black box" problem. Interpretable algorithms are divided into intrinsic interpretability and ex post interpretability based on different stages of artificial intelligence algorithms. There are usually two methods for intrinsic interpretability: choosing a simple model and reducing the number of input features. Interpretability is achieved by limiting the complexity of the model and building a simple model, such as linear regression and decision trees, but this method often sacrifices model accuracy; ex post interpretability refers to the use of some explanatory methods to analyze and infer the decision process / prediction results after the AI ​​model is trained and built. This method is very flexible and can adapt to different problems. Among them, Shapley additive exPlanation (SHAP) is one of the most popular ex post interpretation methods in recent years.

[0043] Based on this, this invention relies on the SHAP method and creatively constructs a molecular screening method based on an interpretable drug structure-activity analysis system that integrates the core of artificial intelligence, uses SHAP as the explanation logic, and uses multi-dimensional feature description as the supporting framework. It aims to expand the input dimension of SAR analysis in the process of drug molecule screening and enhance the interpretability of the model to overcome the shortcomings of the existing technology.

[0044] Figure 1 is a flow chart of a molecular virtual screening method based on artificial intelligence according to an exemplary embodiment, see Figure 1 , the method comprising:

[0045] S101, obtaining SMILES of a series of ligand molecules and multiple molecules to be screened.

[0046] S102, calculating the molecular descriptors, the parent core structure and the substituent groups relative to the parent core structure of the series of ligand molecules based on the SMILES of the series of ligand molecules.

[0047] S103, selecting the 3D conformation of the target protein of each ligand molecule, generating the 3D conformation of the ligand molecule and performing molecular docking with the 3D conformation of the target protein, and calculating the protein-ligand interaction between the ligand molecule and the target protein after molecular docking, wherein the protein-ligand interaction includes amino acid residues and protein-ligand interaction type.

[0048] S104, obtaining molecular activity experimental data of each of the ligand molecules.

[0049] S105, the molecular descriptor of the ligand molecule, the core structure, the substituent groups relative to the core structure, the protein-ligand interaction and the corresponding molecular activity experimental data are combined into training data, and an XGBoost model is trained based on the training data and parameters are adjusted to obtain a trained XGBoost model.

[0050] S106, incorporating the trained XGBoost model into the SHAP algorithm for computational analysis to obtain SHAP values ​​of the molecular descriptor, the parent core structure, the substitution group, and the protein-ligand interaction, wherein the SHAP value is used to indicate the contribution of the corresponding data to the molecular activity.

[0051] S107, selecting at least one of a molecular descriptor, a core structure, a substitution group, and a protein-ligand interaction as a target screening feature, and formulating a molecular screening rule for the target screening feature based on the SHAP value.

[0052] S108, screening molecules that meet the molecular screening rules from the multiple molecules to be screened based on the SMILES of the molecules to be screened.

[0053] In step S101, a collection of SMILES (Simplified molecular input line entry system) containing multiple compounds can be obtained from an internal compound library, a commercial compound library, etc. These compounds will be used as molecules to be screened in subsequent screening. One or more groups of molecules with similar structures can also be selected as a series of ligand molecules. SMILES is a specification that uses ASCII strings to describe molecular structures, mainly used to represent the structures of organic molecules and other chemical substances. SMILES uses a series of characters to represent atoms, bonds, and molecular structures.

[0054] After step S101 is completed, step S102 may be executed. In step S102, see Figure 2 , the SMILES of a series of ligand molecules can be used as input to calculate the molecular descriptors, the core structure, and the substituent groups of the ligand molecules relative to the core structure.

[0055] Specifically, see, optionally, Figure 2 In step S102, the molecular descriptors, the core structure and the substituent groups relative to the core structure of the series of ligand molecules are calculated based on the SMILES of the series of ligand molecules, including:

[0056] S1021, calculating the molecular descriptor of each ligand molecule in the series of ligand molecules based on the SMILES of the series of ligand molecules, wherein the molecular descriptor includes molecular low-dimensional structure information and molecular 3D structure information.

[0057] S1022, based on the FMCS algorithm of the SMILES, cluster analysis is performed on the series of ligand molecules to obtain one or more classes of ligand molecule sets, and the maximum common substructure of each class of ligand molecule sets is calculated as the mother core structure of the class of ligand molecule sets.

[0058] S1023, for each ligand molecule in each type of ligand molecule set, breaking the chemical bond between the ligand molecule and the parent core structure of the ligand molecule set of the same type, to obtain a substitution group of the ligand molecule relative to the parent core structure.

[0059] In step S1021, the molecular descriptor includes molecular low-dimensional structure information and molecular 3D structure information, wherein the molecular low-dimensional structure information includes molecular 0D structure information, molecular 1D structure information, molecular 2D structure information and molecular property descriptors, and the molecular 0D structure information can be described by 0D descriptors, which are mainly composed of structural descriptors and quantitative descriptors, such as molecular weight, number of atoms, number of various rings and number of double bonds. The molecular 1D structure information can be described by 1D descriptors, which are composed of structural fragments or molecular fingerprints, such as some specific groups such as keto, azido, amino, etc. The molecular 2D structure information can be described by 2D descriptors, such as topological descriptors. Molecular property descriptors are used to describe the basic properties of molecules such as some physical and chemical properties.

[0060] In view of the structure-activity relationship analysis requirements of the present invention, in order to further expand the dimension of input data, optionally, the molecular 0D structural information includes molecular weight and chemical bond type, the molecular 1D structural information includes the number of hydrogen bond donors, the number of hydrogen bond acceptors, and the polar surface area, the molecular 2D structural information includes molecular shape and hybridization type, and the molecular property descriptors include fat solubility, water solubility, and drug-like properties. In a specific embodiment, the molecular low-dimensional structural information includes a total of 210 types of feature data, which can be input into the calculation module for calculation.

[0061] The molecular 3D structural information is used to describe the three-dimensional structure of the molecule. Optionally, the molecular 3D structural information includes molecular shape, three-dimensional atomic space coordinates, shape complexity, and best fit plane.

[0062] By calling the relevant algorithm in the prior art, the SMILES of each ligand molecule in the series of ligand molecules can be directly used to calculate its corresponding molecular descriptor mentioned above.

[0063] After step S1021 is completed, step S1022 can be executed. In step S1022, the frequent maximum common substructure FMCS (Frequent Maximum Common Substructure) algorithm is a method used in chemical informatics to identify the maximum common substructure between a group of molecules. The maximum common substructure (MCS) problem is modeled as a graph isomorphism problem, and subgraph enumeration and subgraph isomorphism testing are used to handle multiple MCS problems. The maximum common substructure is the maximum structure that each ligand molecule in each type of ligand molecule set has. Using the SMILES of a series of ligand molecules as the input of the FMCS algorithm, the series of ligand molecules can be clustered and analyzed first to obtain one or more types of ligand molecule combinations, and each type of ligand molecule set has a similar structure. Then, the maximum common substructure of each type of ligand molecule set is calculated as its parent core structure. Of course, in possible implementations, other calculation methods can also be used to calculate the parent core structure.

[0064] After step S1022 is completed, step S1023 is executed. In step S1023, for each ligand molecule in each type of ligand molecule set, the chemical bonds between these ligand molecules and the parent core structure of the series of ligand molecules to which they belong are broken to obtain the substitution group of the ligand molecule relative to the parent core structure.

[0065] After step S102 is completed, step S103 may be executed. Molecular docking is a process in which two or more molecules recognize each other through geometric matching and energy matching. Figure 2 In step S103, the target protein of each ligand molecule can be selected by using relevant calculation software, and its 3D conformation can be obtained. The 3D conformation of the target protein can be pre-built into the protein database. Generate the 3D conformation of the ligand molecule, and perform molecular docking between the 3D conformation of the ligand molecule and the 3D conformation of the target protein. After the molecular docking calculation is completed, the protein-ligand interaction between the ligand molecule and the target protein can be calculated. Among them, the protein-ligand interaction can include amino acid residues and protein-ligand interaction types. For example, (94|CYS-HygB) represents the hydrogen bond formed between the ligand and the 94th cysteine, cysteine ​​is an amino acid residue, and the hydrogen bond is a protein-ligand interaction type.

[0066] In step S104, the molecular activity experimental data is the molecular activity data obtained through experiments, which is used to characterize the molecular activity. This data is difficult to obtain through computer simulation calculations, and it is necessary to obtain its real data through experiments. Step S104 can be performed after step S103 or before step S101, as long as step S104 is performed before step S105, and the present invention does not make specific restrictions on this.

[0067] Traditional SAR analysis mainly focuses on the two-dimensional structural information of compounds, while ignoring other important features. The present invention adopts a multi-dimensional feature representation system, integrating the multi-level features of protein-ligand, including the two-dimensional / three-dimensional structural properties of molecules, the physicochemical properties of proteins and molecules, and the interaction between proteins and ligands, expanding the feature input of existing SAR analysis, greatly enhancing the comprehensiveness and depth of system information, and providing more comprehensive and in-depth information support for structure-activity relationship analysis in molecular virtual screening.

[0068] After step S104 is completed, step S105 may be executed. Figure 3In step S105, the molecular descriptor of the ligand molecule calculated in the above step S102, the core structure, the substituent group relative to the core structure, the protein-ligand interaction calculated in step S103, and the molecular activity experimental data obtained in step S104 can be used as training data to form a training set. The training set includes multiple groups of training data, each group of training data includes molecular descriptors, core structures, substituent groups, protein-ligand interactions, and molecular activity experimental data corresponding to the same ligand molecule, wherein the molecular descriptors, core structures, substituent groups, and protein-ligand interactions can be used as training samples, and the molecular activity experimental data can be used as training labels, and then the XGBoost (eXtreme Gradient Boosting) model is trained by machine learning, and the trained XGBoost model is obtained after the training termination condition is reached. The trained XGBoost model has the ability to predict the corresponding molecular activity using molecular descriptors, core structures, substituent groups, and protein-ligand interactions as input data.

[0069] As an ensemble learning algorithm based on gradient boosting decision trees, the XGBoost model shows excellent performance when processing tabular data. Its model effect is comparable to the top level of machine learning, and sometimes even exceeds the performance of deep learning models. Although large language models (LLMs) have made significant progress in the field of text data processing in recent years, XGBoost's accuracy is still better than many deep learning models and LLMs when processing small-scale but high-quality tabular data. At the same time, it is also significantly ahead of LLMs in terms of interpretability. It has the advantages of scalability, high efficiency, strong generalization ability, and avoidance of overfitting. The model algorithm uses a strong learner built with classification and regression trees as basic components in an iterative manner, which can greatly improve the accuracy and robustness of SAR analysis. In view of this, XGBoost is undoubtedly a more ideal choice when conducting molecular virtual screening of drug molecules based on structure-activity relationship analysis.

[0070] After step S105 is completed, step S106 may be executed. Figure 3 In step S106, the result of XGBoost model training is used as a new round of input and incorporated into the SHAP algorithm to deeply analyze the model. The intrinsic information of the model is comprehensively analyzed and mined from multiple angles, and activity information is mined based on SAR analysis. Specifically, it can include: identifying positive factors that promote activity improvement, analyzing negative factors that lead to reduced activity, determining the key factors that have the most significant impact on activity, and identifying secondary factors that have relatively small impact on activity.

[0071] The SHAP (SHapley Additive exPlanations) algorithm is based on game theory and is a model-independent machine learning interpretation method that provides a solution for calculating the feature contribution of a single prediction of any machine learning model. Based on the SHAP algorithm, the contribution of each input data of the machine learning model to the output data can be calculated and represented by the SHAP value (Shapley value). In step S106, after the XGBoost model is incorporated into the SHAP algorithm for calculation and analysis, the SHAP values ​​of the molecular descriptors, parent core structure, substitution groups, and protein-ligand interactions can be calculated. Structure-activity relationship analysis can be performed using these SHAP values. The SHAP value can be used to represent the contribution of the input data of XGBoost to the output data, i.e., the molecular activity. When the SHAP value is positive, it can be indicated that the corresponding data contributes positively to the molecular activity. When the SHAP value is negative, it can be indicated that the corresponding data contributes negatively to the molecular activity; the absolute value of the SHAP value represents the contribution of the corresponding data to the molecular activity.

[0072] Based on the prediction model, by introducing the SHAP interpretation framework, the prediction results of the model can be interpreted and visualized in a fine-grained manner. Researchers can clearly understand the contribution of each feature to the model prediction, thereby better understanding the intrinsic relationship between the structural information and activity of drug molecules and protein targets.

[0073] After step S106 is completed, step S107 may be executed. Figure 3 In step S107, one or more of the molecular descriptors, the core structure, the substituent groups and the protein-ligand interaction can be selected as the target screening features according to the actual molecular screening requirements. For example, all of these four types of feature information can be used as the target screening features, and then the molecular screening rules are formulated according to their SHAP values. For example, the value range of the molecular descriptor or the presence or absence of the core structure, the substituent groups and the protein-ligand interaction can be set. Specifically, when the target screening feature includes the molecular descriptor, in step S107, at least one of the molecular descriptors, the core structure, the substituent groups and the protein-ligand interaction is selected as the target screening feature. The molecular screening rules for the target screening feature are formulated based on the SHAP value, which may include:

[0074] A molecular descriptor with a positive SHAP value is selected as a target screening feature, and a target value range of the SHAP value of the selected molecular descriptor is determined as at least a part of the molecular screening rule.

[0075] When the target screening feature includes a core structure, a substituent group or a protein-ligand interaction, in step S107, at least one of the molecular descriptor, the core structure, the substituent group, and the protein-ligand interaction is selected as the target screening feature, and a molecular screening rule for the target screening feature is formulated based on the SHAP value, which may include:

[0076] At least one of the core structures, substituent groups, and protein-ligand interactions with positive SHAP values ​​is selected as a target screening feature, and a rule that the molecules to be screened must have the target screening feature is formulated as at least a part of the molecular screening rule.

[0077] Specifically, a series of key molecular activity information was collected through the above steps. For example, descriptors such as PEOE_VSA6, SPS, 3D_OverLap and 3D_USRCAT were identified as positive factors, which are helpful to improve the activity of the molecule; while descriptors such as VSA_EState4, SMR_VSA10 and FpDensityMorgan showed negative factors, which may reduce the activity of the molecule. Based on the indications of these descriptors, we should consider ligand molecules with molecular structures corresponding to the aforementioned positive factors when conducting molecular screening. For the analysis of the skeleton and substituents, if the skeleton structure of skeleton_a is more helpful to improve the activity of the molecule than that of skeleton_b, skeleton_a or its bioisosteric analogs are preferred as the skeleton. In the protein-ligand interaction analysis, if the hydrogen bond formed by the ligand with cysteine ​​at position 94 (94|CYS-HygB) and the salt bridge formed with aspartic acid at position 101 (101|ASP-SaB) both show significant activity enhancement effects, consideration is given to retaining or enhancing the interaction between 94|CYS-HygB and 101|ASP-SaB.

[0078] For key activity information, we can adopt the strategy of leveraging strengths and avoiding weaknesses, modify and organize them into an "activity rule list" to serve as molecular screening rules for downstream tasks. For example, we can select PEOE_VSA6, SPS, 3D_OverLap, 3D_USRCAT, VSA_EState4, FpDensityMorgan, 94|CYS-HygB and 101|ASP-SaB as target screening features, and organize the "activity rule list" as follows:

[0079] (1), "PEOE_VSA6": 24±5;

[0080] (2), "SPS": 10±1;

[0081] (3), "3D_OverLap": 0.8±0.2;

[0082] (4), "3D_USRCAT": 0.4±0.2;

[0083] (5) "VSA_EState4": >3 or <2;

[0084] (6), "FpDensityMorgan": < 0.8;

[0085] (7), "94|CYS-HygB": True;

[0086] (8), "101|ASP-SaB": True.

[0087] Among them, PEOE_VSA6: (Partial Equalization of Orbital Electronegativity -Volume Surface Area 6) is a descriptor based on charge distribution and volume surface area. It uses partial charge (PEOE) to calculate the volume surface area (VSA) of a specific region in a molecule. PEOE_VSA6 specifically refers to the 6th volume surface area segment, which is usually used to describe the surface area within a certain charge range in a molecule.

[0088] SPS: (Surface Polar Surface Area) SPS describes the polar surface area of ​​a molecule. It calculates the sum of the surface areas of polar atoms (such as oxygen, nitrogen, etc.) on the molecule surface.

[0089] 3D_OverLap: describes the degree of overlap between two molecules in three-dimensional space. It is usually used to compare the three-dimensional structural similarity of two molecules.

[0090] 3D_USRCAT: (Ultrafast Shape Recognition with Categorical descriptors) 3D_USRCAT is a fast 3D shape recognition method that combines shape and chemical features. It describes molecules by calculating their shape features (such as volume, surface, etc.) and chemical features (such as hydrogen bond donors / acceptors, hydrophobicity, etc.).

[0091] VSA_EState4: (Volume Surface Area - Estate 4) VSA_EState4 is a descriptor based on the EState metric and the volume surface area (VSA). The EState metric (Electrotopological State) is a parameter that describes the electronic topological state of atoms in a molecule, taking into account the charge distribution of the atoms and the surrounding atomic environment. VSA_EState4 specifically refers to the 4th volume surface area segment, which is used to describe the surface area within a specific EState range in a molecule.

[0092] FpDensityMorgan: is a descriptor based on Morgan fingerprints (also known as ECFC fingerprints, Extended Connectivity Fingerprints). Morgan fingerprints are bit vectors that describe the structure of a molecule and are generated by recursively extending the neighborhood of an atom. FpDensityMorgan describes the density of Morgan fingerprints, i.e. the number of fingerprint bits per unit length.

[0093] For the above parameters, please refer to the above “Activity Rule List”. Items (1) to (6) give the value ranges. “True” in items (7) and (8) indicates that there must be a corresponding protein-ligand interaction.

[0094] After step S107 is completed, step S108 can be executed. In step S108, if the target screening feature includes a molecular descriptor, the SMILES of the molecule to be screened can be used to calculate the SHAP value of its molecular descriptor, and the SHAP value can be compared with the SHAP value range of the corresponding target screening feature in the molecular screening rule. If the SHAP value is within the range, the SHAP value of the molecular descriptor meets the requirements of the molecular screening rule. If the target screening features include parent core structure, substitution groups, protein-ligand interaction, and the molecular screening rules require corresponding structural features, then the SMILES calculation of the molecule to be screened can be used to determine whether the molecule to be screened has these structural features. If it contains all the target screening features required by the molecular screening rules, it meets the requirements.

[0095] Continuing with the above example, referring to the above “Activity Rule List”, it is required that the value ranges of “PEOE_VSA6”, “SPS”, “3D_OverLap”, “3D_USRCAT”, “VSA_EState4”, and “FpDensityMorgan” of the screened molecules are within the range specified in the “Activity Rule List”, and have protein-ligand interactions corresponding to “94|CYS-HygB” and “101|ASP-SaB”. For the molecules to be screened, their related molecular descriptors, parent core structures, substitution groups, and protein-ligand interaction information can be calculated first, and then compared with the various screening rules in the molecular screening rules to screen and filter the molecules. In this way, molecules with potential high activity can be identified from the compound library, thereby improving the efficiency of drug discovery.

[0096] Through the above technical solution, after obtaining the molecules to be screened and the series of ligand molecules, the molecular descriptors, parent core structures, substitution groups and multi-dimensional input data of protein-ligand interactions of the series of ligand molecules are calculated, and the XGBoost model is trained and adjusted together with the obtained molecular activity test data. The trained XGBoost model is incorporated into the SHAP algorithm for structure-activity relationship analysis to explain the contribution of each input data of the XGBoost model to the molecular activity, and molecular virtual screening rules are formulated according to the explained contribution for molecular virtual screening. A multi-dimensional feature representation system is used to provide more comprehensive and in-depth information support for structure-activity relationship analysis, combined with a high-precision XGBoost model, greatly improving the accuracy and robustness of structure-activity relationship analysis in molecular virtual screening, and based on the interpretation framework of the SHAP algorithm, the model is interpreted in a fine-grained manner, which facilitates a clear understanding of the contribution of each feature to the model prediction, enhances the transparency and interpretability of the model, and thus better understands the model decision-making process, improves the trust in the prediction results, and thus provides strong support for molecular screening, formulates more accurate molecular screening rules, and improves the accuracy of molecular screening.

[0097] Figure 4 is a block diagram of a molecular virtual screening device based on artificial intelligence according to an exemplary embodiment, such as Figure 4 As shown, the device comprises:

[0098] SMILES acquisition module, used to obtain SMILES of a series of ligand molecules and multiple molecules to be screened;

[0099] A calculation module, used for calculating the molecular descriptors, the parent core structure and the substituent groups relative to the parent core structure of the series of ligand molecules based on the SMILES of the series of ligand molecules;

[0100] A molecular docking module, used for selecting the 3D conformation of the target protein of each ligand molecule, generating the 3D conformation of the ligand molecule and performing molecular docking with the 3D conformation of the target protein, and calculating the protein-ligand interaction between the ligand molecule and the target protein after molecular docking, wherein the protein-ligand interaction includes amino acid residues and protein-ligand interaction type;

[0101] A molecular activity acquisition module, used to acquire molecular activity experimental data of each of the ligand molecules;

[0102] A training module, for forming training data from the molecular descriptor of the ligand molecule, the parent core structure, the substituent groups relative to the parent core structure, the protein-ligand interaction, and the corresponding molecular activity experimental data, training an XGBoost model based on the training data and adjusting parameters to obtain a trained XGBoost model;

[0103] An analysis module, for incorporating the trained XGBoost model into the SHAP algorithm for computational analysis, and obtaining SHAP values ​​of the molecular descriptor, the parent core structure, the substituent group, and the protein-ligand interaction, wherein the SHAP value is used to indicate the contribution of the corresponding data to the molecular activity;

[0104] A screening rule module, for selecting at least one of a molecular descriptor, a core structure, a substitution group, and a protein-ligand interaction as a target screening feature, and formulating a molecular screening rule for the target screening feature based on the SHAP value;

[0105] The molecular screening module is used to screen molecules that meet the molecular screening rules from the multiple molecules to be screened based on the SMILES of the molecules to be screened.

[0106] Optionally, the calculation module is further used for:

[0107] Calculating a molecular descriptor of each of the ligand molecules in the series of ligand molecules based on the SMILES of the series of ligand molecules, wherein the molecular descriptor includes molecular low-dimensional structure information and molecular 3D structure information;

[0108] Based on the FMCS algorithm of the SMILES, cluster analysis is performed on the series of ligand molecules to obtain one or more types of ligand molecule sets, and the maximum common substructure of each type of ligand molecule set is calculated as the mother core structure of the type of ligand molecule set;

[0109] For each ligand molecule in each type of ligand molecule set, the chemical bond between the ligand molecule and the parent core structure of the ligand molecule set is broken to obtain a substitution group of the ligand molecule relative to the parent core structure.

[0110] Optionally, the molecular low-dimensional structure information includes molecular 0D structure information, molecular 1D structure information, molecular 2D structure information and molecular property descriptors.

[0111] Optionally, the molecular 0D structural information includes molecular weight and chemical bond type, the molecular 1D structural information includes the number of hydrogen bond donors, the number of hydrogen bond acceptors, and polar surface area, the molecular 2D structural information includes molecular shape and hybridization type, and the molecular property descriptors include lipid solubility, water solubility, and drug-like properties.

[0112] Optionally, the molecular 3D structure information includes molecular shape, three-dimensional atomic space coordinates, shape complexity, and best fit plane.

[0113] Optionally, the SHAP value is used to indicate the contribution of the corresponding data to the molecular activity, including:

[0114] When the SHAP value is positive, the SHAP value indicates that the contribution of the corresponding data to the molecular activity is positive, and when the SHAP value is negative, the SHAP value indicates that the contribution of the corresponding data to the molecular activity is negative; the absolute value of the SHAP value indicates the degree of contribution of the corresponding data to the molecular activity.

[0115] Optionally, the screening rule module is further used to:

[0116] A molecular descriptor with a positive SHAP value is selected as a target screening feature, and a target value range of the SHAP value of the selected molecular descriptor is determined as at least a part of the molecular screening rule.

[0117] Optionally, the screening rule module is further used to:

[0118] At least one of the core structures, substituent groups, and protein-ligand interactions with positive SHAP values ​​is selected as a target screening feature, and a rule that the molecules to be screened must have the target screening feature is formulated as at least a part of the molecular screening rule.

[0119] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0120] Figure 5 FIG. 1 is a block diagram of an electronic device according to an exemplary embodiment. Figure 5 As shown, the electronic device 700 may include: a processor 701 , a memory 702 . The electronic device 700 may also include one or more of a multimedia component 703 , an input / output (I / O) interface 704 , and a communication component 705 .

[0121] The processor 701 is used to control the overall operation of the electronic device 700 to complete all or part of the steps in the above-mentioned molecular virtual screening method based on artificial intelligence. The memory 702 is used to store various types of data to support the operation of the electronic device 700, which may include instructions for any application or method used to operate on the electronic device 700, and application-related data, such as contact data, messages sent and received, pictures, audio, video, etc. The memory 702 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Read-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, disk or optical disk. The multimedia component 703 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 702 or sent through the communication component 705. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 704 provides an interface between the processor 701 and other interface modules, and the other interface modules may be keyboards, mice, buttons, etc. These buttons may be virtual buttons or physical buttons. The communication component 705 is used for wired or wireless communication between the electronic device 700 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IOT, eMTC, or other 5G, etc., or a combination of one or more of them, is not limited here. Therefore, the corresponding communication component 705 may include: Wi-Fi module, Bluetooth module, NFC module, etc.

[0122] In an exemplary embodiment, the electronic device 700 can be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to perform the above-mentioned artificial intelligence-based molecular virtual screening method.

[0123] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, and when the program instructions are executed by a processor, the steps of the above-mentioned molecular virtual screening method based on artificial intelligence are implemented. For example, the computer-readable storage medium can be the above-mentioned memory 702 including program instructions, and the above-mentioned program instructions can be executed by the processor 701 of the electronic device 700 to complete the above-mentioned molecular virtual screening method based on artificial intelligence.

[0124] The preferred embodiments of the present disclosure are described in detail above in conjunction with the accompanying drawings; however, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept of the present disclosure, a variety of simple modifications can be made to the technical solution of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

[0125] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present disclosure will not further describe various possible combinations.

[0126] In addition, various embodiments of the present disclosure may be arbitrarily combined, and as long as they do not violate the concept of the present disclosure, they should also be regarded as the contents disclosed by the present disclosure.

Claims

1. A molecular virtual screening method based on artificial intelligence, characterized in that: The method comprises: Obtain SMILES of a series of ligand molecules and multiple molecules to be screened; Calculating the molecular descriptors, the parent core structure and the substituent groups relative to the parent core structure of the series of ligand molecules based on the SMILES of the series of ligand molecules; Selecting the 3D conformation of the target protein of each ligand molecule, generating the 3D conformation of the ligand molecule and performing molecular docking with the 3D conformation of the target protein, and calculating the protein-ligand interaction between the ligand molecule and the target protein after molecular docking, wherein the protein-ligand interaction includes amino acid residues and protein-ligand interaction type; Obtaining molecular activity experimental data of each of the ligand molecules; The molecular descriptor of the ligand molecule, the core structure, the substituent group relative to the core structure, the protein-ligand interaction, and the corresponding molecular activity experimental data constitute training data, and the XGBoost model is trained based on the training data and the parameters are adjusted to obtain the trained XGBoost model; Incorporating the trained XGBoost model into the SHAP algorithm for computational analysis to obtain SHAP values ​​of the molecular descriptor, the parent core structure, the substituent group, and the protein-ligand interaction, wherein the SHAP value is used to indicate the contribution of the corresponding data to the molecular activity; Select at least one of a molecular descriptor, a core structure, a substitution group, and a protein-ligand interaction as a target screening feature, and formulate a molecular screening rule for the target screening feature based on the SHAP value; Based on the SMILES of the molecule to be screened, a molecule that meets the molecular screening rule is screened from the multiple molecules to be screened; Wherein, the selecting at least one of the molecular descriptor, the core structure, the substitution group, and the protein-ligand interaction as the target screening feature, and formulating the molecular screening rule about the target screening feature based on the SHAP value, comprises: Selecting molecular descriptors with positive SHAP values ​​as target screening features, and determining a target value range of the SHAP values ​​of the selected molecular descriptors as at least a part of the molecular screening rule; At least one of the core structures, substituent groups, and protein-ligand interactions with positive SHAP values ​​is selected as a target screening feature, and a rule that the molecules to be screened must have the target screening feature is formulated as at least a part of the molecular screening rule.

2. The method according to claim 1, characterized in that The molecular descriptors, the parent core structure and the substituent groups relative to the parent core structure of the series of ligand molecules are calculated based on the SMILES of the series of ligand molecules, including: Calculating a molecular descriptor of each of the ligand molecules in the series of ligand molecules based on the SMILES of the series of ligand molecules, wherein the molecular descriptor includes molecular low-dimensional structure information and molecular 3D structure information; Based on the FMCS algorithm of the SMILES, cluster analysis is performed on the series of ligand molecules to obtain one or more types of ligand molecule sets, and the maximum common substructure of each type of ligand molecule set is calculated as the mother core structure of the type of ligand molecule set; For each ligand molecule in each type of ligand molecule set, the chemical bond between the ligand molecule and the parent core structure of the ligand molecule set is broken to obtain a substitution group of the ligand molecule relative to the parent core structure.

3. The method according to claim 2, characterized in that The molecular low-dimensional structure information includes molecular 0D structure information, molecular 1D structure information, molecular 2D structure information and molecular property descriptors.

4. The method according to claim 3, characterized in that The molecular 0D structural information includes molecular weight and chemical bond type, the molecular 1D structural information includes the number of hydrogen bond donors, the number of hydrogen bond acceptors, and polar surface area, the molecular 2D structural information includes molecular shape and hybridization type, and the molecular property descriptors include lipid solubility, water solubility, and drug-like properties.

5. The method according to claim 2, characterized in that: The molecular 3D structural information includes molecular shape, three-dimensional atomic spatial coordinates, shape complexity, and best fitting plane.

6. The method according to claim 1, characterized in that The SHAP value is used to indicate the contribution of the corresponding data to the molecular activity, including: When the SHAP value is positive, the SHAP value indicates that the contribution of the corresponding data to the molecular activity is positive, and when the SHAP value is negative, the SHAP value indicates that the contribution of the corresponding data to the molecular activity is negative; the absolute value of the SHAP value indicates the degree of contribution of the corresponding data to the molecular activity.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 6 are implemented.

8. An electronic device, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the method according to any one of claims 1 to 6.

Images