Drug molecule screening and optimizing method based on artificial intelligence prediction

Abstract

The invention relates to the technical field of computer-aided drug design, in particular to a drug molecule screening and optimizing method based on artificial intelligence prediction, which comprises the following steps: S1, obtaining a dynamic protein conformation set and molecular multi-dimensional characterization: obtaining a dynamic conformation set of a target protein and a physicochemical property spatial distribution diagram of a binding pocket of the dynamic conformation set, a two-dimensional molecular map topological structure and three-dimensional conformation coordinates of the drug molecules are obtained; s2, multi-modal fusion prediction is carried out; s3, generating interpretable optimization guidance; and S4, automatic iterative optimization: performing batch prediction and screening on the new candidate molecular structure, taking the screened optimal molecule as a new starting point, repeatedly executing the interpretability optimization guidance generation step and the step until an iteration termination condition is met, and outputting a final optimized molecule list. Through the multi-modal fusion deep learning model, the interaction strength of the drug molecules and the target protein can be quickly and accurately predicted, and the screening efficiency of the drug molecules is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided drug design, and in particular to a drug molecule screening and optimization method based on artificial intelligence prediction. BACKGROUND

[0002] In the process of drug development, compound activity screening and optimization is a crucial but extremely resource and time-consuming link. The existing technology usually adopts high-throughput screening or virtual screening method based on molecular docking, but these methods have the following problems: Traditional virtual screening relies on fixed protein crystal structure, which cannot consider the dynamic changes of protein conformation and the induced fit effect in the process of drug molecule and protein binding, resulting in low prediction accuracy.

[0003] Most of the existing AI-based prediction models are black box models, which only give the predicted value of binding affinity, lack specific and interpretable guidance for molecular optimization direction, and chemists cannot understand why the model gives a certain prediction result, so it is difficult to effectively optimize the molecular structure accordingly.

[0004] Most AI models only use a single type of data (such as two-dimensional molecular graph or three-dimensional structure) for training, and fail to effectively integrate multi-modal data (such as molecular dynamics simulation trajectory, physicochemical property distribution of binding pocket, etc.), resulting in limited generalization ability of the model and decreased prediction performance on new target points or new skeleton molecules.

[0005] The process of molecular optimization is often one-way, that is, a batch of molecules are screened and handed over to experts for review, lacking a closed-loop system based on real-time feedback and automatic iterative optimization of prediction results.

[0006] Therefore, there is an urgent need in the art for an intelligent drug molecule screening and optimization method that can integrate multi-dimensional data, provide interpretable optimization guidance, and realize a closed-loop iteration. SUMMARY

[0007] Based on the above purpose, the present application provides a drug molecule screening and optimization method based on artificial intelligence prediction, comprising the following steps: S1: Dynamic protein conformation set and molecular multi-dimensional representation acquisition: acquiring a dynamic conformation set of a target protein and a physicochemical property space distribution map of a binding pocket thereof, and acquiring a two-dimensional molecular graph topology structure and a three-dimensional conformation coordinate of a drug molecule; S2: Multi-modal fusion prediction: inputting the two-dimensional molecular graph topology structure of the drug molecule, the three-dimensional conformation coordinate of the drug molecule, and the physicochemical property space distribution map of one dynamic conformation of the target protein into a pre-trained multi-channel deep learning model, and outputting a continuous numerical value representing the predicted interaction strength between the drug molecule and the target protein; S3: Explanability-optimized guidance generation: for the lead compound molecule to be optimized, calculate the gradient of the predicted interaction strength values on the atoms and chemical bonds in the two-dimensional molecular graph of the lead compound molecule, identify the key chemical structure region according to the gradient, and generate a structure optimization guide containing specific modification suggestions and expected weight improvement based on a chemical rule base; S4: Automated iterative optimization: automatically generate a new candidate molecular structure based on the structure optimization guide, perform batch prediction and screening on the new candidate molecular structure, and repeat the steps of the explainability-optimized guidance generation and this step with the screened preferred molecule as the new starting point until the iteration termination condition is met, and output the final optimized molecule list.

[0008] Preferably, the specific process of obtaining the dynamic conformation set of the target protein and the physicochemical property space distribution map of the binding pocket thereof in the dynamic protein conformation set and molecular multi-dimensional characterization acquisition includes: S1.1: Obtain the static three-dimensional structure coordinate file of the target protein from the protein structure database; S1.2: In the environment of molecular simulation software, place the static three-dimensional structure coordinates of the target protein in a solvent environment simulating physiological conditions, apply Newton's law of motion for molecular dynamics simulation calculation, and thus obtain a dynamic conformation trajectory containing the conformation of the target protein at different time points; S1.3: Perform conformation analysis on the dynamic conformation trajectory, calculate the root mean square deviation value of the backbone atoms of each frame of conformation relative to the initial conformation, and simultaneously calculate the hydrophobic cavity volume value of the binding pocket region in each frame of conformation; use unsupervised machine learning clustering algorithm to cluster the feature vector set composed of the root mean square deviation value and the hydrophobic cavity volume value; select the conformation with high frequency of occurrence from each cluster group as the representative target protein conformation, and collectively constitute the dynamic conformation set; S1.4: For each representative target protein conformation in the dynamic conformation set, establish a three-dimensional grid coordinate system in the binding pocket region thereof; use computational chemistry method to numerically solve electrostatic equation at each three-dimensional grid point to obtain the electrostatic potential energy value of the point, obtain the hydrophobic parameter value of the point by querying the atomic type parameter library, and judge the hydrogen bond donor or acceptor ability of the point by analyzing the geometric arrangement and properties of the protein atoms near the grid point; map the calculation results of each point in the entire three-dimensional grid space into a discretized three-dimensional data matrix respectively, and thus generate the electrostatic potential energy distribution map, the hydrophobic property distribution map and the hydrogen bond donor / acceptor site distribution map of the binding pocket.

[0009] Preferably, in the multi-modal fusion prediction, the construction and training process of the multi-channel deep learning model includes: S2.1: constructing the multi-channel deep learning model, which comprises three parallel data processing channels: a molecular graph channel for processing the two-dimensional molecular graph topology, which adopts a multi-layer graph convolutional neural network to extract the molecular topological feature vector by iteratively aggregating the information of atomic nodes and their adjacent chemical bonds; a molecular conformation channel for processing the three-dimensional conformation coordinates, which adopts a three-dimensional convolutional neural network to extract the molecular spatial feature vector by performing convolution operations on the three-dimensional voxelized molecular coordinates; a protein pocket channel for processing the physicochemical property spatial distribution map, which adopts a three-dimensional convolutional neural network to extract convolution features for each type of physicochemical property distribution map, and then fuses the extracted features to obtain the pocket physicochemical feature vector; S2.2: designing a feature fusion layer to perform a concatenation operation on the molecular topological feature vector and the molecular spatial feature vector to form a unified molecular global feature vector; then performing an element-wise multiplication operation on the molecular global feature vector and the pocket physicochemical feature vector to simulate the interaction strength between the molecule and the pocket at different spatial positions, and outputting a fused feature vector; S2.3: inputting the fused feature vector into a regression prediction module composed of fully connected layers, which finally outputs a continuous numerical value of the predicted interaction strength; S2.4: during model training, using known binding affinity experimental data as true label values to construct a training sample set; using a backpropagation algorithm to minimize the mean square error between the model prediction value and the true label value as the optimization objective, while adding L2 norm regularization constraint on the model weight parameters to prevent overfitting, and iteratively updating all trainable parameters in the multi-channel deep learning model until the model loss function converges.

[0010] Preferably, in the interpretable optimization guide generation, the specific process of generating a structure optimization guide includes: S3.1: inputting the structure information of the lead compound molecule into the trained multi-channel deep learning model and performing a forward propagation calculation to obtain the predicted interaction strength value; S3.2: performing a backpropagation calculation to obtain the gradient of the predicted interaction strength value with respect to each atomic node feature vector in the two-dimensional molecular graph of the lead compound molecule; calculating the norm of each atomic node gradient vector to obtain the significance score of each atomic node; S3.3: sorting according to the significance score, and selecting atomic nodes with a ranking higher than a set threshold and chemical bonds directly connected thereto as the key chemical structure region; S3.4: Query a pre-defined chemical functional group replacement rule base which records the electronic isosteric relationship, steric complementarity and synthetic accessibility information between functional groups; match feasible functional group replacement, addition or deletion schemes from the chemical functional group replacement rule base according to the chemical environment of the key chemical structure region, forming multiple specific modification suggestions; S3.5: For each of the specific modification suggestions, the calculation process of the expected improvement weight is as follows: first, estimate the change amount of the molecular descriptor after the execution of the modification suggestion; then, input the change amount into a pre-trained auxiliary prediction model, which can predict the potential improvement amplitude of binding affinity according to the descriptor change; finally, weight and fuse the predicted improvement amplitude and the significance score of the key chemical structure region to generate the final expected improvement weight.

[0011] Preferably, in the automatic iterative optimization, the process of automatically generating a new candidate molecular structure comprises: S4.1: Analyze each of the specific modification suggestions in the structure optimization guide and convert it into a specific chemical reaction operation instruction, which includes the identification of the reaction center atom, the chemical bond to be broken, and the identification of the new functional group to be connected; S4.2: In a database containing known synthetic reaction templates, retrieve the synthetic reaction rules matching the chemical reaction operation instruction; S4.3: According to the matched synthetic reaction rules and the structure of the lead compound molecule, call a molecular graph processing algorithm to automatically perform bond breaking and formation, generating a preliminary new molecular structure; S4.4: Perform three-dimensional conformation optimization calculation on the preliminary new molecular structure, and minimize its steric hindrance energy using molecular mechanics force field to obtain lower energy and stable three-dimensional conformation coordinates, thereby completing the generation of a new candidate molecular structure.

[0012] Preferably, in the automatic iterative optimization, the calculation process of the pre-defined synthetic feasibility evaluation function for screening the new candidate molecular structure comprises: analyzing the molecular complexity of the new candidate molecular structure, counting the number of chiral centers and the fusion degree of ring systems; retrieving the synthetic reaction database, simulating its possible synthetic path, and estimating the reaction yield of the key step; inputting the molecular complexity, the number of chiral centers, the fusion degree of ring systems and the estimated reaction yield as input features into a classification model trained by historical synthesis data, which outputs a synthetic feasibility score representing the synthetic difficulty level.

[0013] Preferably, in the automatic iterative optimization, the judgment method of the iteration termination condition is: in each iteration cycle, record the highest predicted interaction strength value in all the new candidate molecular structures generated by this iteration; calculate the relative percentage of improvement of this value and the highest predicted interaction strength value in the last iteration; set an initial threshold and a decay coefficient, if the relative percentage of improvement of continuous multiple iterations is less than the current threshold, it is judged that the iteration termination condition is met; at the same time, after each iteration, the current threshold is decayed using the decay coefficient to realize adaptive tightening termination standard.

[0014] Preferably, after the S4.3 generates the preliminary new molecular structure, before the three-dimensional conformation optimization, a diversity control strategy is introduced: for all the preliminary new molecular structures generated by the same specific modification suggestion, calculate the similarity matrix based on the molecular fingerprint; adopt a clustering algorithm to divide the preliminary new molecular structures into multiple clusters; randomly select a representative molecular structure from each cluster to enter the subsequent three-dimensional conformation optimization calculation and prediction screening process to ensure the diversity of exploring the chemical space.

[0015] Preferably, in the S1.3, the unsupervised machine learning clustering algorithm is a density clustering algorithm, and the neighborhood radius parameter is determined by analyzing the Euclidean distance distribution between all vectors in the feature vector set, specifically by selecting a specific quantile value in the distance distribution as the value of the neighborhood radius parameter.

[0016] Preferably, in the S3.5, the pre-trained auxiliary prediction model is a gradient boosting decision tree model, which is trained by: collecting a large number of known functional group replacement cases and the binding affinity change values caused by them; calculating the change amount of a series of physical and chemical descriptors of the molecules before and after the functional group replacement in each case; using the change amount of the descriptors as input features, and using the change value of the binding affinity as the prediction target, using the gradient boosting algorithm to train the gradient boosting decision tree model.

[0017] The beneficial effects of the present application are: 1、The drug molecule intelligent screening and optimization method proposed in the present application can model the dynamic conformation change of proteins by combining molecular dynamics simulation data, so as to more accurately capture the induced fitting effect in the binding process of drug molecules and proteins. This not only solves the limitation of traditional methods which only rely on fixed protein crystal structures, but also significantly improves the prediction accuracy of molecular screening.

[0018] 2、The drug molecule intelligent screening method in the application is based on an interpretable artificial intelligence model, which not only gives a prediction value of the binding affinity of the drug molecule and the protein, but also provides specific optimization guidance, including which structural features need to be adjusted and how to adjust, so as to help chemists understand the prediction process of the model and provide a clear direction for molecular structure optimization. Through this interpretable feedback, chemists can more effectively optimize and improve the molecular structure.

[0019] 3、The application fuses multiple types of data, including molecular dynamics simulation trajectories, molecular three-dimensional structures, physicochemical property distributions of binding pockets, etc., to construct a multi-modal fusion intelligent screening and optimization model. This model not only improves the prediction performance on new target points or new skeleton molecules, but also significantly enhances the generalization ability of the model, which can be more widely applied to different drug research and development projects.

[0020] 4、The application introduces a closed-loop feedback system based on the prediction results, which can feed back the results in the molecular optimization process to the model in real time and automatically perform iterative optimization. In this way, the model can continuously adjust the optimization direction according to real-time feedback information, forming an automated and efficient molecular optimization process. Through this closed-loop system, the optimization process is better controlled and improved, effectively shortening the drug research and development cycle and reducing resource and time consumption. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0022] Fig. 1 The step flow chart of the method of the application; Fig. 2 The step flow chart of the construction and training process of the multi-channel deep learning model of the method of the application; Fig. 3 The step flow chart of the determination method of the iteration termination condition of the method of the application. DETAILED DESCRIPTION

[0023] The application will be described in detail below in combination with the drawings and specific embodiments. It should be noted here that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the application.

[0024] Please refer to Figs. 1-3The embodiment of the present application provides a drug molecule screening and optimization method based on artificial intelligence prediction. Firstly, in S1, a dynamic conformation set of a target protein is obtained from experimental data or computational simulation. The dynamic conformations are obtained by molecular dynamics simulation, reflecting the dynamic behavior of the protein under different conditions. The physicochemical property space distribution map of the binding pocket of the protein can provide detailed functional region information of the target protein. The map describes the electron density, charge distribution, and polarity and hydrophobicity of different regions, which are crucial for predicting the binding of drug molecules to target proteins.

[0025] In addition, the two-dimensional molecular graph topology and three-dimensional conformation coordinates of the drug molecule are also obtained. The two-dimensional molecular graph topology can show the connection relationship between atoms in the molecule, while the three-dimensional conformation coordinates show the specific configuration of the drug molecule in space, providing possible binding modes of the drug molecule.

[0026] By obtaining dynamic protein conformations and multi-dimensional representations of molecules, the present application can capture the variability of proteins and the dynamic binding behavior of drug molecules in three-dimensional space, and can more accurately simulate and predict drug-target interactions compared to traditional methods.

[0027] In S2, the two-dimensional molecular graph, three-dimensional conformation coordinates of the drug molecule, and physicochemical property space distribution map of the dynamic conformation of the target protein are input into a pre-trained multi-channel deep learning model. The model uses multi-modal data to learn the complex relationship between drug molecules and proteins, and outputs a continuous numerical value representing the interaction strength between the drug molecule and the target protein. This value represents the binding affinity of the drug molecule to the protein and can be used for subsequent screening and optimization.

[0028] Multi-modal fusion prediction can more comprehensively and accurately evaluate the binding affinity of drug molecules by combining different types of data (two-dimensional graph, three-dimensional conformation, and protein physicochemical properties), overcoming the limitations of single data source in traditional methods and improving the accuracy and efficiency of drug screening.

[0029] In S3, for the lead compound molecule to be optimized, the gradient of the predicted interaction strength to each atom and chemical bond in the two-dimensional graph of the molecule is calculated. The gradient calculation process helps to identify the key chemical structure regions that have the greatest impact on the interaction strength. These regions are usually the focus of drug molecule optimization. Through modification suggestions provided by the chemical rule base, structure optimization guidance is generated, including how to modify the molecular structure, change which chemical bonds, add which functional groups, etc. optimization direction, while giving the expected improvement weight.

[0030] The explainability optimization guide provides clear optimization suggestions for chemists, avoiding the "black box" model. By targeting the optimization of molecules, researchers can adjust the molecular structure and obtain predictable optimization results, thereby improving the efficiency of drug development.

[0031] S4 generates new candidate molecular structures based on the structure optimization guide generated in the previous step, and performs batch prediction and screening on them. The preferred molecules screened will be used as the starting point for the next iteration, and the explainability optimization guide generation step and the automated optimization step will be repeated until the preset iteration termination condition is met, such as the optimization effect reaching the expected target or the number of iterations reaching the upper limit.

[0032] The automated iterative optimization can achieve closed-loop feedback and self-optimization by continuously generating and screening new candidate molecules. This process avoids the cumbersome steps of manual intervention, speeds up the optimization process of drug molecules, and can fine-tune the optimization results in each iteration to ultimately obtain the best candidate molecule. Through automation, the efficiency and accuracy of drug molecule screening and optimization are greatly improved.

[0033] Based on artificial intelligence technology, the present application combines multi-modal data fusion, explainability guidance and automated iterative optimization, breaking through the limitations of traditional drug molecule screening and optimization methods, and improving the prediction accuracy, optimization efficiency and operability. By implementing a closed-loop feedback system, the continuity and automation of the optimization process are ensured, significantly shortening the drug development cycle and reducing resource consumption, greatly promoting the intelligent and precise process of drug development.

[0034] In one possible implementation, first, download the static three-dimensional structure coordinate file of the target protein from the public or private protein structure database (such as the PDB database). The structure file contains the spatial coordinates of all atoms of the protein and their type information, providing the basic conformation for subsequent molecular simulation.

[0035] The static three-dimensional structure is imported into the molecular simulation software environment and placed in a solvent environment under simulated physiological conditions, such as water solution, with appropriate temperature and ion concentration settings. Using molecular dynamics simulation methods, the motion trajectory of each atom in time evolution is calculated according to Newton's law of motion, thereby obtaining the dynamic conformation trajectory of the protein at different time points. This step can capture the flexibility and variability of the protein under physiological conditions.

[0036] The root mean square deviation (RMSD) value of each frame of protein backbone atoms relative to the initial conformation is calculated to reflect the overall structural change amplitude of the protein. At the same time, the hydrophobic cavity volume of the binding pocket region is calculated to measure the binding capacity of the pocket. The RMSD value of each frame of conformation and the hydrophobic cavity volume form a feature vector set, and an unsupervised machine learning clustering algorithm is used for grouping. In each cluster, the conformation with higher frequency is selected as the representative target protein conformation, thereby constructing a dynamic conformation set that can reflect the main dynamic characteristics of the protein in a limited number of representative conformations.

[0037] For each representative target protein conformation, a three-dimensional grid coordinate system is established in the binding pocket region. Through computational chemistry methods, the electrostatic potential energy distribution of each grid point is solved by solving the electrostatic problem. At the same time, by querying the atomic type parameter library, the hydrophobic parameter value of each grid point is obtained; according to the geometric arrangement and properties of the protein atoms around the grid point, the hydrogen bond donor or acceptor ability of the position is judged. The calculation results of all grid points are discretized into a three-dimensional matrix to generate electrostatic potential energy distribution map, hydrophobic property distribution map and hydrogen bond donor / acceptor site distribution map, providing detailed spatial information for subsequent drug molecule and protein binding prediction.

[0038] Through the above steps, the dynamic conformation information of the target protein and the spatial distribution of the physicochemical properties of the binding pocket can be systematically obtained. Compared with the traditional method of using only static protein structure, this method fully considers the flexibility change of the protein and the heterogeneity of the binding pocket, thereby significantly improving the accuracy of drug molecule and protein interaction prediction. The dynamic conformation set and physicochemical property distribution map provide high-quality input for the multi-modal deep learning model, ensuring the reliability of the subsequent screening and optimization steps, and providing scientific basis for structure optimization, which helps to improve drug research and development efficiency, shorten development cycle and reduce experimental resource consumption.

[0039] In one possible implementation, S2.1 involves constructing a multi-channel deep learning model containing three parallel data processing channels. Each channel processes different types of data features: Molecular graph channel: This channel is specifically designed to process the two-dimensional topological structure of molecules. Graph convolutional neural network (GCN) is used to aggregate the information of atomic nodes and their adjacent chemical bonds in the molecule. Each iteration, GCN encodes the relationship between atoms and chemical bonds to extract feature vectors of the molecular graph.

[0040] Molecular conformation channel: This channel processes the three-dimensional conformation coordinates of molecules. Three-dimensional convolutional neural network (3DCNN) is used for convolution operation on three-dimensional voxelized molecular coordinates to extract the structural features of molecules in space. In this way, the model can capture the stereo information of the molecular space.

[0041] Protein pocket channel: This channel is specifically designed to handle the spatial distribution of physicochemical properties of protein binding pockets. For each physicochemical property (such as electrostatic potential energy, hydrophobicity, etc.), a three-dimensional convolutional neural network extracts convolutional features from the distribution map, and finally fuses the extracted features to obtain the physicochemical feature vector of the pocket.

[0042] In S2.2, the molecular feature vectors obtained from the molecular graph channel and the molecular conformation channel are spliced to form a unified "molecular global feature vector". Next, this molecular global feature vector is element-wise multiplied with the pocket physicochemical feature vector obtained from the protein pocket channel. This operation can simulate the interaction strength between the molecule and the protein pocket at different spatial positions. The final output fusion feature vector will contain multi-dimensional information about the interaction between the molecule and the protein binding pocket.

[0043] In S2.3, the fused feature vector is input into a regression prediction module composed of fully connected layers. The role of this module is to output the predicted interaction strength based on the input fusion feature vector. This output value is a continuous numerical value representing the affinity or binding strength of the molecule and the protein.

[0044] In S2.4, during the training process, known experimental data (such as binding affinity experimental data) are used as true label values, and a training sample set is constructed. Using the backpropagation algorithm, the model is optimized by minimizing the mean square error (MSE) between the model's predicted value and the true label value. At the same time, to prevent overfitting, L2 norm regularization constraints are used to penalize the model's weight parameters, ensuring the model's generalization ability. Through repeated iterative optimization, the model's loss function converges, completing the training.

[0045] By using a multi-channel deep learning model, this method can simultaneously process data features at three different levels of molecular graph, molecular three-dimensional conformation, and protein pocket physicochemical properties. Compared with traditional single data source methods, the use of multi-modal fusion can comprehensively capture multi-dimensional information of molecular and protein interactions, greatly improving the accuracy of prediction. The design of the feature fusion layer simulates the interaction strength between the molecule and the protein pocket, enabling the model to accurately predict the binding affinity of drug molecules, providing stronger support for drug screening and optimization. In addition, by regularizing and precisely training the model, the overfitting problem can be effectively avoided, ensuring the reliability and generalization ability of the model on new data. Overall, this technology can significantly improve the efficiency of drug development, shorten the development cycle, and provide important data support for precise design of new drugs.

[0046] In one possible implementation, first, the structural information of the lead compound molecule is input into the trained multi-channel deep learning model for forward propagation calculation. In this way, the model will output a predicted interaction strength value based on the input molecular information, which reflects the binding affinity of the molecule to the target protein. Forward propagation calculation is the most basic step in the model prediction process, providing preliminary prediction results for subsequent optimization.

[0047] Next, a backpropagation calculation is performed to obtain the gradient of the predicted interaction strength value with respect to each atomic node feature vector in the two-dimensional molecular graph of the lead compound molecule. This gradient represents the contribution of each atomic node to the prediction result. By calculating the norm of the gradient vector of each atomic node, the significance score of the node can be obtained. The higher the significance score, the greater the influence of the atomic node on the prediction result, and therefore it is more important in the optimization process.

[0048] According to the calculated significance score, all atomic nodes are sorted, and nodes ranked higher than a set threshold and their directly connected chemical bonds are selected. These nodes and their chemical bonds are marked as "key chemical structure regions", i.e. parts that need special attention and modification in the optimization process of the drug molecule. The core purpose of this step is to focus on the regions that have the greatest impact on the activity of the molecule, ensuring that the optimization work can concentrate resources to improve the effect.

[0049] Query the pre-defined chemical functional group replacement rule library. The rule library records information such as electronic effect, spatial stereoeffect, and synthetic feasibility between functional groups. According to the chemical environment of the key chemical structure region, select feasible functional group replacement, addition or deletion schemes from the rule library. These schemes will form multiple specific modification suggestions to help improve the structure of the drug molecule in order to achieve higher binding affinity.

[0050] For each specific modification suggestion, first, the change in molecular descriptors after the modification is estimated. Molecular descriptors such as molecular weight, polarity, hydrophobicity, etc. can quantitatively reflect the changes in the properties of the molecule. Next, these changes are input into the pre-trained auxiliary prediction model, which can predict how the binding affinity of the drug molecule will be improved according to the changes in the descriptors. Finally, the predicted improvement and the significance score of the corresponding key chemical structure region are weighted and fused to obtain the expected improvement weight of each modification suggestion. This process helps to select the most likely structural modification scheme that can bring significant optimization effect.

[0051] Through the interpretable optimization guidance, not only the efficiency of drug molecule optimization is improved, but also the transparency and operability of the optimization process are enhanced, providing strong support for drug design and screening.

[0052] In one possible implementation, each specific modification suggestion generated in the structure optimization guide needs to be parsed first. Each suggestion includes a modification plan for the existing drug molecule structure, such as the functional group to be replaced or added, the chemical bond to be broken, and the identification of the new functional group. The goal of parsing is to convert each modification suggestion into specific chemical reaction operation instructions. This instruction can be understood as a specific change rule for the molecular structure, providing clear guidance for subsequent synthesis reactions and molecular structure optimization.

[0053] The parsed chemical reaction operation instructions are input into a database containing known synthesis reaction templates for retrieval. These reaction templates contain a variety of verified chemical reaction rules that can be applied in actual synthesis. In this way, it can be ensured that the selected modification plan is not only theoretically feasible, but also has an operable synthesis path. The goal of database retrieval is to find a matching synthesis reaction rule to guide how to break or form chemical bonds on the existing molecular framework.

[0054] After determining the matching synthesis reaction rule, the breaking and forming of chemical bonds are automatically executed by calling the molecular graph processing algorithm. This process automatically calculates the changes in chemical bonds in the molecule by simulating the path of chemical reactions. The molecular graph processing algorithm can accurately break and form new chemical bonds in the molecular graph according to the reaction templates in the rule library, thereby generating a preliminary new molecular structure. The automation of this step makes the design of drug molecules more efficient, avoiding complex calculations and repeated verification in manual operations.

[0055] After the preliminary generation of the new molecular structure, three-dimensional conformation optimization is also needed. This process uses molecular mechanics force field minimization techniques to adjust the three-dimensional structure of the molecule by optimizing intermolecular interactions and steric hindrance, so that it is in the lowest energy state. This optimization calculation adjusts the positions of all atoms in the molecule to ensure its stable spatial structure and avoid high-energy unstable conformations caused by unreasonable steric repulsion. Finally, the optimized three-dimensional conformation is the new candidate molecular structure, which meets the requirements of drug design and has lower energy and better stability.

[0056] The automated iterative optimization method based on artificial intelligence not only improves the efficiency of drug molecule screening and optimization, but also improves the accuracy of molecular design, providing strong technical support for drug research and development.

[0057] In one possible implementation, the new candidate molecular structure needs to be analyzed first to calculate its molecular complexity. Molecular complexity is an important indicator to measure the molecular structure, usually including the size of the molecule, the types of functional groups, and the connection between atoms in the molecule. Molecules with high molecular complexity often contain multiple ring systems, multiple chiral centers, or more complex three-dimensional structures, which may increase the challenge of synthesis. Therefore, by quantifying the molecular complexity, the synthesis difficulty of the molecule can be preliminarily evaluated.

[0058] Chiral center is a key structural feature in a molecule, affecting whether the molecule has optical activity. This step counts the number of chiral centers in the new candidate molecule. The number of chiral centers directly affects the synthesis difficulty of the molecule, because with each additional chiral center, the synthesis steps and difficulty will significantly increase, and more complex reaction conditions and reagents may be required.

[0059] The ring system structure in a molecule is closely related to the feasibility of its synthesis path. The degree of ring system fusion refers to how the ring structures in the molecule are connected and nested with each other. Highly fused ring system structures usually increase the difficulty of synthesis, because such structures can cause large steric hindrance, increasing the complexity of reaction steps. In this step, the system will analyze and evaluate the degree of ring system fusion, providing an important basis for synthesis feasibility evaluation.

[0060] After analyzing the molecular complexity, chiral center, and ring system fusion degree, the system will retrieve the synthesis reaction database to simulate possible synthesis paths of the molecule. This step identifies the most likely synthesis path by comparing with historical synthesis data, and estimates the reaction yield of each key step based on known reaction data and experience. The reaction yield directly affects the success rate of synthesis, and a lower reaction yield means a higher risk of failure in the synthesis process.

[0061] Finally, after the above analysis and estimation, all data (molecular complexity, number of chiral centers, ring system fusion degree, and reaction yield) are input as input features into a classification model trained on historical synthesis data. The classification model uses machine learning algorithms to train based on the synthesis results in historical data, and can output a synthesis feasibility score representing the synthesis difficulty level of the molecule. The score can be a continuous value or a discrete level value, reflecting whether the synthesis of the molecule is feasible and how difficult it is.

[0062] This AI-based synthesis feasibility evaluation method can provide researchers with scientific synthesis difficulty prediction during drug molecule screening, not only improving the accuracy and efficiency of prediction, but also saving a lot of time and resources for drug research and development.

[0063] In one possible implementation, in each iteration, the system predicts the interaction strength of all newly generated candidate molecular structures with the target receptor or other molecules. During this process, the system records the highest predicted interaction strength value from all new candidates, which represents the most likely binding strength of the molecule to the target in that iteration.

[0064] The highest predicted interaction strength value of each iteration is compared with the previous iteration's highest value. The system calculates the relative improvement percentage of the current iteration compared to the previous one, which is the difference between the current value and the previous value divided by the previous value. This percentage reflects the progress of the current iteration in optimization, and a higher improvement percentage usually means that the optimization of the candidate molecules has made significant progress.

[0065] The system sets an initial threshold value to determine whether the relative improvement percentage has reached the expected level of optimization. At the same time, to ensure that the system does not rely too much on the initial settings after a long time, the system also sets a decay coefficient for the threshold value. The decay coefficient controls the gradual reduction of the threshold value, making the termination criteria of the system more stringent as the number of iterations increases.

[0066] If the relative improvement percentage of consecutive iterations is less than the current threshold value, the system determines that the iteration termination condition is met, i.e., the optimization process enters a state of stagnation and no longer continues iteration. This judgment method can effectively avoid over-iteration, saving computational resources and time, and avoiding falling into local optimization.

[0067] After each iteration, the system decays the current threshold value according to the decay coefficient, making it gradually more stringent. This means that even if the relative improvement percentage is small, a higher standard must be met to continue optimization as the iteration progresses. The decay mechanism allows the system to have more refined requirements for optimization progress, ensuring that the final optimization effect is more significant.

[0068] This AI-based automated iterative optimization method realizes the efficiency and accuracy of the drug molecule screening and optimization process through intelligent termination condition judgment and adaptive decay mechanism. This technology not only improves the efficiency of drug optimization, but also significantly reduces computational cost and resource consumption.

[0069] In one possible implementation, in S4.3, based on specific modification suggestions, the system generates a set of preliminary new molecular structures through algorithms. The generation of these molecular structures is usually based on the modification strategy of the target molecule, such as adding or replacing specific groups, functional groups, etc., so that the newly generated molecules may have optimization effects in terms of chemical properties and biological activity.

[0070] Before performing the three-dimensional conformation optimization, to ensure the diversity of chemical space, the system first calculates the molecular fingerprints of these preliminary new molecular structures. The molecular fingerprint is a binary or numerical feature vector used to characterize the molecular structure, which can reflect the basic structural units in the molecule. The system generates a similarity matrix by calculating the similarity between these molecular fingerprints, which quantifies the similarity between each molecular structure.

[0071] According to the similarity matrix, the system uses clustering algorithms to divide all the preliminary new molecular structures into multiple clusters. The goal of the clustering algorithm is to group similar molecules into one category, and the molecular structures in different clusters are significantly different, so as to cover a wider chemical space. Common clustering methods include hierarchical clustering, K-means clustering, etc. Selecting the appropriate algorithm can ensure the rationality and effectiveness of the division.

[0072] In each cluster, the system randomly selects a representative molecular structure. This representative molecule is the most representative of the cluster's molecular characteristics and can effectively represent the chemical properties of the entire cluster. The selected representative molecular structure will enter the subsequent three-dimensional conformation optimization calculation and prediction screening process, which aims to further optimize the molecular structure and enhance its interaction ability with the target.

[0073] Through the above process, the system can ensure that the molecules selected from each cluster have different chemical characteristics, thereby increasing the diversity of candidate molecules. The diversity control strategy avoids the problem of high similarity among all candidate molecules, making the optimization search range more extensive and enabling the exploration of more potential effective molecular structures.

[0074] By introducing the diversity control strategy, the similarity matrix based on molecular fingerprints and clustering algorithms, the system can effectively ensure the diversity of candidate molecules in chemical space, thereby improving the efficiency, accuracy and success rate of drug molecule screening and optimization, and ultimately providing more potential effective candidate molecules for drug development.

[0075] In one possible implementation, in S1.3, the unsupervised machine learning clustering algorithm used is a density clustering algorithm. Density clustering is an algorithm that groups data points based on their density distribution in space. Common algorithms include DBSCAN (Density-Based Spatial Clustering of Applications with Noise), etc. Unlike traditional clustering algorithms, density clustering does not require the number of clusters to be specified in advance, but identifies clusters by judging the local density of data points, thereby being able to handle irregularly shaped clusters and being suitable for handling complex molecular feature data in drug molecule screening.

[0076] Before applying the density clustering algorithm, the system first needs to construct a set of feature vectors, which consists of multiple molecular feature vectors, each representing different properties of a molecule, such as size, polarity, lipophilicity, and other chemical properties. Next, the system analyzes the Euclidean distance distribution between these feature vectors to understand the similarity between each vector. Euclidean distance is a measure of the straight-line distance between two vectors, with a smaller Euclidean distance indicating that the two vectors are closer in feature space, and vice versa.

[0077] One of the key parameters of the density clustering algorithm is the neighborhood radius, which defines the range of "neighbors" for a point. The neighborhood radius determines how many points around a point will be considered to belong to the same cluster during clustering. To accurately cluster, the value of the neighborhood radius needs to be reasonably selected. In this method, the value of the neighborhood radius is determined by analyzing the Euclidean distance distribution between all vectors in the feature vector set. Specifically, first calculate the Euclidean distance between all vectors, get a distance distribution graph, then select a specific quantile value from the distribution as the value of the neighborhood radius. For example, the 90th percentile distance in the distribution can be selected as the neighborhood radius, which can ensure that only high-density areas are considered during clustering, avoiding misjudgment of too sparse points as the same cluster.

[0078] After determining the neighborhood radius parameter, the density clustering algorithm will start clustering the feature vector set. During clustering, the system will determine whether a point belongs to a cluster based on the number of points within its neighborhood radius. If there are enough points around a point, it will be assigned to a cluster; otherwise, the point may be marked as noise or an isolated point. In this way, the density clustering algorithm can automatically discover different clusters based on the distribution of data, ensuring that similar molecules are classified into the same category.

[0079] The density clustering algorithm selects the appropriate neighborhood radius by analyzing the Euclidean distance between feature vectors, thereby achieving accurate clustering of drug molecules. This technology can ensure diversity coverage, improve accuracy, and adaptively handle complex molecular data in drug molecule screening, ultimately improving the efficiency and success rate of drug screening.

[0080] In one possible implementation, in step S3.5, a large number of known functional group replacement cases need to be collected first. These cases usually come from experimental data or known drug molecules in literature. These functional group replacement cases involve the replacement of different functional groups (such as amino, hydroxyl, alkyl, etc.) in the molecular structure, and the impact of this replacement on the binding affinity of the molecule to the target protein. The binding affinity change value is determined by experiment or calculation, usually expressed as the difference in binding energy before and after the functional group replacement.

[0081] For each functional group replacement case, the change in a series of physical and chemical descriptors of the molecule before and after replacement needs to be calculated. The physical and chemical descriptors include the molecular weight, polarity, lipophilicity, number of hydrogen bond acceptors and donors, hydrophobicity, electronic structure, etc. of the molecule. These descriptors can reflect the chemical properties and biological activities of the molecule under different conditions. By comparing the changes in the descriptors before and after replacement, the influence of functional group replacement on the characteristics of the molecule can be obtained.

[0082] The calculated descriptor change values are used as input features, and the binding affinity change values are used as prediction targets to construct a training data set. Each training sample includes the change in the descriptors and the corresponding binding affinity change value. These data can be accumulated from multiple functional group replacement cases to form a training data set containing a large number of samples.

[0083] Using the constructed training data set, a gradient boosting decision tree model is trained using a gradient boosting algorithm. Gradient boosting is an ensemble learning method that builds multiple weak learners (usually decision trees) and combines these learners through weighting to enhance the predictive ability of the model. During each training, the model focuses on the errors in the previous round of prediction and gradually improves the prediction accuracy by adjusting the weights of the decision trees.

[0084] During the model training process, the parameters of the gradient boosting decision tree model need to be optimized, such as the depth of the tree, the learning rate, and the number of iterations, to improve the prediction accuracy of the model. In addition, cross-validation and other methods are used to verify the generalization ability of the model to ensure the stability of the model on new data.

[0085] By using the gradient boosting decision tree model to learn and predict the relationship between functional group replacement and binding affinity change, strong support can be provided in drug molecule screening and optimization. This method not only improves the accuracy and efficiency of drug optimization, but also reduces the waste of experimental resources, which helps to accelerate the development process of new drugs.

[0086] The present application encompasses any substitutions, modifications, equivalent methods and solutions made to the essence and scope of the present application. In order for the public to have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without these details by those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0087] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can also be made, which should be considered as the protection scope of the present application.

Claims

1. A method for drug molecule screening and optimization based on artificial intelligence prediction, characterized in that, Includes the following steps: S1: Acquisition of dynamic protein conformation sets and multidimensional molecular characterization: Acquire the dynamic conformation sets of target proteins and the spatial distribution map of the physicochemical properties of their binding pockets, and obtain the two-dimensional molecular topology and three-dimensional conformation coordinates of drug molecules. S2: Multimodal fusion prediction: The two-dimensional molecular graph topology of the drug molecule, the three-dimensional conformation coordinates of the drug molecule, and the spatial distribution map of the physicochemical properties of the target protein in a dynamic conformation are input into a pre-trained multi-channel deep learning model, and the output is a continuous value representing the predicted interaction strength between the drug molecule and the target protein. S3: Interpretable Optimization Guidance Generation: For the lead compound molecule to be optimized, calculate the gradient of the predicted interaction strength value with respect to the atoms and chemical bonds in the two-dimensional molecular diagram of the lead compound molecule, identify key chemical structure regions based on the gradient, and generate structural optimization guidance containing specific modification suggestions and expected weight enhancement based on the chemical rule base; S4: Automated Iterative Optimization: Based on the structure optimization guidance, new candidate molecular structures are automatically generated. The new candidate molecular structures are batch predicted and screened. The selected preferred molecules are used as new starting points to repeat the interpretability optimization guidance generation step and this step until the iteration termination condition is met, and the final optimized molecular list is output.

2. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 1, characterized in that, In the acquisition of the dynamic protein conformation set and molecular multidimensional characterization, the specific process of obtaining the dynamic conformation set of the target protein and the spatial distribution map of the physicochemical properties of its binding pocket includes: S1.1: Obtain the static three-dimensional structural coordinate file of the target protein from the protein structure database; S1.2: In the molecular simulation software environment, the static three-dimensional structural coordinates of the target protein are placed in a solvent environment simulating physiological conditions, and Newton's laws of motion are applied to perform molecular dynamics simulation calculations, thereby obtaining a dynamic conformational trajectory containing the conformations of the target protein at different time points. S1.3: Perform conformational analysis on the dynamic conformation trajectory, calculate the root mean square deviation of the skeleton atoms of each frame conformation relative to the initial conformation, and simultaneously calculate the hydrophobic cavity volume of the binding pocket region in each frame conformation; use an unsupervised machine learning clustering algorithm to cluster the feature vector set composed of the root mean square deviation and the hydrophobic cavity volume; select the conformations with high frequency of occurrence from each cluster group as representative target protein conformations, which together constitute the dynamic conformation set; S1.4: For each representative target protein conformation in the dynamic conformation set, a three-dimensional grid coordinate system is established in its binding pocket region; using computational chemistry methods, the electrostatic equation is numerically solved at each three-dimensional grid point to obtain the electrostatic potential energy value at that point; the hydrophobic parameter value at that point is obtained by querying the atom type parameter library; and the hydrogen bond donor or acceptor ability is determined by analyzing the geometric arrangement and properties of protein atoms near the grid point; the calculation results of each point in the entire three-dimensional grid space are mapped into a discretized three-dimensional data matrix, thereby generating the electrostatic potential energy distribution map, hydrophobic property distribution map, and hydrogen bond donor / acceptor site distribution map of the binding pocket.

3. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 1, characterized in that, In the multimodal fusion prediction, the construction and training process of the multi-channel deep learning model includes: S2.1: Construct the multi-channel deep learning model, which includes three parallel data processing channels: a molecular graph channel for processing the two-dimensional molecular graph topology, which uses a multi-layer graph convolutional neural network to extract the molecular topological feature vector by iteratively aggregating information of atomic nodes and their adjacent chemical bonds; a molecular conformation channel for processing the three-dimensional conformation coordinates, which uses a three-dimensional convolutional neural network to extract the molecular spatial feature vector by performing convolution operations on the three-dimensional voxelized molecular coordinates; and a protein pocket channel for processing the spatial distribution map of physicochemical properties, which uses a three-dimensional convolutional neural network to extract convolutional features for each physicochemical property distribution map, and then fuses the extracted features to obtain the pocket physicochemical feature vector. S2.2: Design a feature fusion layer to concatenate the molecular topological feature vector and the molecular spatial feature vector to form a unified molecular global feature vector; then, perform element-wise multiplication of the molecular global feature vector with the pocket physicochemical feature vector to simulate the interaction strength between the molecule and the pocket at different spatial locations, and output a fused feature vector. S2.3: The fused feature vector is input into a regression prediction module composed of fully connected layers, which ultimately outputs a continuous value of the predicted interaction strength; S2.4: During model training, known binding affinity experimental data are used as the true label values ​​to construct a training sample set; the backpropagation algorithm is used to minimize the mean square error between the model's predicted values ​​and the true label values ​​as the optimization objective, while adding L2 norm regularization constraints on the model weight parameters to prevent overfitting, and iteratively updating all trainable parameters in the multi-channel deep learning model until the model loss function converges.

4. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 1, characterized in that, In the generation of interpretable optimization guidance, the specific process of generating structural optimization guidance includes: S3.1: Input the structural information of the lead compound molecule into the trained multi-channel deep learning model and perform a forward propagation calculation to obtain the predicted interaction strength value; S3.2: Perform backpropagation calculation to obtain the gradient of the predicted interaction strength value with respect to the eigenvector of each atomic node in the two-dimensional molecular graph of the lead compound molecule; calculate the norm of the gradient vector of each atomic node to obtain the significance score of each atomic node; S3.3: Sort the data according to the significance scores, select the atomic nodes with a ranking higher than the set threshold and the chemical bonds directly connected to them, and mark them as the key chemical structure regions; S3.4: Query the predefined chemical functional group substitution rule library, which records the electronic isostatic relationship, stereocomplementarity, and synthetic accessibility information between functional groups; based on the chemical environment of the key chemical structure region, match feasible functional group substitution, addition, or deletion schemes from the chemical functional group substitution rule library to form multiple specific modification suggestions; S3.5: For each specific modification suggestion, the calculation process of the expected enhancement weight is as follows: First, estimate the change in the molecular descriptor after implementing the modification suggestion; then, input the change into a pre-trained auxiliary prediction model, which can predict the potential enhancement of binding affinity based on the descriptor change; finally, weight and fuse the predicted enhancement with the significance score of the key chemical structure region to generate the final expected enhancement weight.

5. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 1, characterized in that, In the automated iterative optimization, the process of automatically generating new candidate molecular structures includes: S4.1: Analyze each specific modification suggestion in the structure optimization guide and convert it into specific chemical reaction operation instructions, the instructions including the identifier of the reaction center atom, the chemical bond to be broken, and the identifier of the new functional group to be connected; S4.2: Retrieve a synthetic reaction rule that matches the chemical reaction operation instruction from a database containing known synthetic reaction templates; S4.3: Based on the matched synthesis reaction rules and the structure of the lead compound molecule, the molecular graphics processing algorithm is invoked to automatically execute the breaking and formation of chemical bonds, generating a preliminary new molecular structure; S4.4: Perform three-dimensional conformation optimization calculations on the preliminary new molecular structure, minimize its spatial steric hindrance energy using molecular mechanical force field, obtain low-energy and stable three-dimensional conformation coordinates, thereby completing the generation of a new candidate molecular structure.

6. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 1, characterized in that, In the automated iterative optimization, the calculation process of the predefined synthetic feasibility assessment function used to screen the new candidate molecular structures includes: analyzing the molecular complexity of the new candidate molecular structures, counting the number of chiral centers and the degree of fusion of ring systems; searching the synthetic reaction database, simulating possible synthetic paths, and estimating the reaction yield of key steps; and inputting the molecular complexity, the number of chiral centers, the degree of fusion of ring systems, and the estimated reaction yield as input features into a classification model trained on historical synthetic data, which outputs a synthetic feasibility score representing the level of synthetic difficulty.

7. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 1, characterized in that, In the automated iterative optimization, the method for determining the iteration termination condition is as follows: in each iteration cycle, record the highest predicted interaction strength value among all the new candidate molecular structures generated in this iteration; Calculate the percentage increase in this value relative to the highest predicted interaction strength value in the previous iteration; An initial threshold and a decay coefficient are set. If the relative improvement percentage of multiple consecutive iterations is less than the current threshold, the iteration termination condition is determined to be met. At the same time, after each iteration, the decay coefficient is used to decay the current threshold to achieve adaptive tightening of the termination criterion.

8. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 5, characterized in that, After generating the preliminary new molecular structure in S4.3 and before performing three-dimensional conformation optimization, a diversity control strategy is introduced: for all the preliminary new molecular structures generated by the same specific modification suggestion, their similarity matrix based on molecular fingerprint is calculated; a clustering algorithm is used to divide the preliminary new molecular structures into multiple clusters; a representative molecular structure is randomly selected from each cluster to enter the subsequent three-dimensional conformation optimization calculation and prediction screening process to ensure the diversity of the explored chemical space.

9. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 2, characterized in that, In S1.3, the unsupervised machine learning clustering algorithm is a density clustering algorithm, and its neighborhood radius parameter is determined by analyzing the Euclidean distance distribution between all pairs of vectors in the feature vector set. Specifically, a specific quantile value in the distance distribution is selected as the value of the neighborhood radius parameter.

10. The method for drug molecule screening and optimization based on artificial intelligence prediction according to claim 4, characterized in that, In step S3.5, the pre-trained auxiliary prediction model is a gradient boosting decision tree model, which is trained by: collecting a large number of known functional group substitution cases and the resulting changes in binding affinity; calculating the changes in a series of physicochemical descriptors of the molecule before and after the functional group substitution in each case; using the changes in the descriptors as input features and the changes in binding affinity as the prediction target, and using the gradient boosting algorithm to train the gradient boosting decision tree model.

Images