A method and system for generating a ligand based on pocket electron density

Abstract

This invention discloses a method and system for generating ligands based on pocket electron density. The method includes: obtaining the target pocket electron density; inputting the target pocket electron density into a preset optimal ligand density inference model to generate a predicted ligand electron density; inputting the predicted ligand electron density into a preset conditional generation model to generate multiple similar ligand electron densities; using a preset analytical model to analyze the multiple similar ligand electron densities to generate target atom information; and constructing a target ligand molecule based on the target atom information. The technical solution provided by this invention improves the binding affinity between the generated ligand molecule and the protein pocket.

Description

Technical Field

[0001] This invention relates to the field of pharmaceutical component analysis technology, and specifically to a method and system for generating ligands based on pocket electron density. Background Technology

[0002] In the field of drug design, generating ligand molecules that bind to specific target proteins has always been a hot research topic, with numerous studies in recent years focusing on SMILE structures and 2D graph structures. However, these molecular representations lack spatial structural information, especially when researching targeted drugs (targeted drugs are drugs or their formulations endowed with targeting capabilities. Their purpose is to enable the drug or its carrier to target specific lesion sites and accumulate or release the active ingredient at the target site). These molecular representations often result in molecules whose size and shape are difficult to match with the target. Therefore, in recent research, 3D molecules based on 3D logical cloud representations, combined with deep learning, have become a new hot topic. However, to distinguish different atoms and chemical properties, existing technologies use multiple channels to represent different element types or multiple channels to represent chemical properties (hydrogen bond acceptors, hydrogen bond donors, aromatic systems, etc.) in the input of deep learning models, without fully considering target characteristics. Therefore, how to accurately generate corresponding small molecule ligands while considering the characteristics of the target protein remains an unsolved problem. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method and system for generating ligands based on pocket electron density, thereby improving the binding affinity between the generated ligand small molecules and the pocket.

[0004] According to a first aspect, a method for generating ligands based on pocket electron density, the method comprising:

[0005] Obtain the electron density of the target pocket;

[0006] The target pocket electron density is input into a preset optimal ligand density inference model to generate a predicted ligand electron density. The preset optimal ligand density inference model is trained and generated based on the real pocket electron density and the real ligand electron density.

[0007] The predicted ligand electron density is input into a preset conditional generation model to generate multiple similar ligand electron densities. The preset conditional generation model is trained and generated based on small molecule samples.

[0008] The electron density of the multiple similar ligands is analyzed using a preset analytical model to generate target atom information. The preset analytical model is generated based on training with small molecule samples. The target atom information includes the atomic position and atom type of the target atom.

[0009] Construct the target ligand molecule based on the target atom information.

[0010] Optionally, the step of training and generating the preset optimal ligand density inference model based on the true pocket electron density and the true ligand electron density includes:

[0011] Obtain training samples, which include the true pocket electron density and the true ligand electron density, wherein the true ligand electron density is a label of the true pocket electron density;

[0012] The training samples are input into the generator in the generative adversarial network to obtain the second predicted ligand electron density.

[0013] Positive samples are obtained by connecting the actual ligand electron density with the actual pocket electron density, and negative samples are obtained by connecting the second predicted ligand electron density with the actual pocket electron density.

[0014] The positive samples are labeled as true, and the negative samples are labeled as false.

[0015] The positive and negative samples are respectively input into the discriminator in the generative adversarial network to obtain the discrimination result, and the result of the preset loss function is calculated based on the error between the discrimination result and the corresponding label;

[0016] Adjust the parameters of the generator and / or the parameters of the discriminator based on the result of the preset loss function;

[0017] Determine whether the number of training iterations and / or the result of the preset loss function meet the preset conditions; if the number of training iterations and / or the result of the preset loss function do not meet the preset conditions, repeat the steps from obtaining training samples to adjusting the parameters of the generator and / or the parameters of the discriminator according to the result of the preset loss function until the number of training iterations and / or the result of the preset loss function meet the preset conditions;

[0018] The generator that meets the preset conditions is used as the optimal ligand density inference model.

[0019] Optionally, obtaining training samples includes:

[0020] Discretize the actual pocket electron density and the corresponding actual ligand electron density to obtain discrete samples;

[0021] The training samples are obtained by normalizing and spatially transforming the discrete samples.

[0022] Optionally, the result of calculating a preset loss function based on the error between the discrimination result and the corresponding label includes:

[0023] The result of the first loss function is calculated based on the similarity between the gradient of the second predicted ligand electron density and the gradient of the true ligand electron density. The first loss function is used to adjust the parameters of the generator.

[0024] The result of a second loss function is calculated based on the similarity between the true ligand electron density at the reference position and the second predicted ligand electron density. The second loss function is used to adjust the parameters of the generator. The reference position is the location of non-covalent interactions in the true ligand electron density.

[0025] Based on the degree of balance between the discrimination result and the true result, the result of the third loss function is calculated. The third loss function is used to adjust the parameters of the discriminator and the generator simultaneously.

[0026] Based on the similarity between the second predicted ligand electron density and the true ligand electron density, the result of the fourth loss function is calculated, which is used to adjust the parameters of the generator.

[0027] The results of the first loss function, the second loss function, the third loss function, and the fourth loss function are added together according to the types of the generator and the discriminator to obtain the result of the preset loss function.

[0028] Optionally, the step of training and generating the preset conditional generation model based on small molecule samples includes:

[0029] Training samples were generated based on the conditions described for generating small molecule samples;

[0030] A variational autoencoder is trained based on the conditional training samples to learn the probability distribution of the conditional training samples;

[0031] The variational autoencoder at the end of the learning process is used as the conditional generative model.

[0032] Optionally, the step of generating conditional training samples based on the small molecule samples includes:

[0033] Obtain the structural model of the small molecule sample;

[0034] Statistically analyze the atomic coordinates in the structural model.

[0035] Transform the atomic coordinates from real space to frequency domain space to generate structure factors;

[0036] The structure factor is adjusted to return to the real space at a preset resolution to generate electron density grid information, and the electron density grid information is used as the conditional training sample. The preset resolution is a preset range for the summation of frequency domain spatial vector components.

[0037] Optionally, after generating target atom information by analyzing the electron densities of the multiple similar ligands using a preset analytical model, the method further includes:

[0038] The atomic positions and atomic types are corrected according to preset chemical rules.

[0039] According to a second aspect, a ligand generation system based on pocket electron density, the system comprising:

[0040] The information acquisition module is used to obtain the electron density of the target pocket;

[0041] The ligand electron density inference module is used to input the target pocket electron density into a preset optimal ligand electron density inference model to generate a predicted ligand electron density. The preset optimal ligand electron density inference model is trained and generated based on the real pocket electron density and the real ligand electron density.

[0042] The ligand electron density extension module is used to input the predicted ligand electron density into a preset conditional generation model to generate multiple similar ligand electron densities. The preset conditional generation model is trained and generated based on small molecule samples.

[0043] The atom analysis module is used to analyze the electron density of the multiple similar ligands using a preset analysis model to generate target atom information. The preset analysis model is generated based on small molecule samples, and the target atom information includes the atomic position and atom type of the target atom.

[0044] The ligand generation module is used to construct the target ligand molecule based on the target atom information.

[0045] According to a third aspect, an electronic device includes:

[0046] A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect, or any alternative embodiment of the first aspect.

[0047] According to a fourth aspect, embodiments of the present invention provide a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions for causing the computer to perform the method described in the first aspect, or any alternative embodiment of the first aspect.

[0048] The technical solution of this invention has the following advantages:

[0049] This invention provides a method and system for generating ligands based on pocket electron density. The method first trains an optimal ligand density inference model using real pocket electron densities and real ligand electron densities, thereby inferring a relatively good predicted ligand electron density from the pocket electron density. Then, a conditional generation model is used to generate multiple similar ligand densities to the predicted ligand density, thus expanding the analytical sample. Finally, these similar ligand densities are input into an analytical model to obtain the atomic coordinates and atom types of atoms in the target ligand molecule. The obtained atoms are then connected according to chemical rules to obtain the target ligand molecule. The target ligand molecule obtained based on electron density using this method contains complete pocket topological features, resulting in a more accurate target ligand molecule that binds more tightly to the protein pocket. In the field of drug development analysis, precisely binding pocket-targeted ligands significantly improve drug efficacy. Attached Figure Description

[0050] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram illustrating the steps of a method for generating ligands based on pocket electron density according to an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of a ligand generation system based on pocket electron density according to an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention; Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0056] like Figure 1As shown in the figure, the present invention provides a method for generating ligands based on pocket electron density, which specifically includes the following steps:

[0057] Step S101: Obtain the target pocket electron density. Specifically, proteins are composed of various amino acids, and the amino acids that can bind to ligands constitute the pocket, i.e., the target protein. Small molecule drugs are fabricated targeting these pockets, i.e., ligand molecules that can bind to the molecular pocket are generated. To ensure a higher fit between the generated ligand molecules and the pocket, it is crucial to fully consider the pocket's characteristics. Obtaining electron density requires considering factors such as atomic positions and element types. Therefore, in this embodiment of the invention, electron density that simultaneously includes element type and chemical property information is used to represent the target molecule. This representation method is superior to using 3D logic clouds for characterizing chemical properties and saves computational resources. Furthermore, molecules generated from the pocket electron density can well consider the fit between the small molecule and the pocket shape, and have a clear focus on the non-covalent interaction (NCI) region, which greatly affects activity, while also generating legitimate molecular conformations. Therefore, before generating the target ligand molecule, it is necessary to first obtain the electron density of the pocket in the target molecule, i.e., the target pocket electron density. The methods for obtaining the target pocket electron density include, but are not limited to, experimental methods and computational methods.

[0058] Step S102: Input the target pocket electron density into a preset optimal ligand density inference model to generate a predicted ligand electron density. The preset optimal ligand density inference model is trained based on the actual pocket electron density and the actual ligand electron density. Specifically, in the field of deep learning, unsupervised learning has always been a research hotspot. By acquiring the feature information of a known object and using an unsupervised machine learning model to generate a new object similar to the known object, it has important application value in various scientific fields. Regarding the technical field of this invention, using an unsupervised learning model to generate a ligand molecule that conforms to the features of a known molecular pocket (protein pocket) and can bind to the known molecular pocket, using the electron density of the known molecular pocket (i.e., the quantity used to represent the feature of the molecular pocket), has significant application value in the process of new drug development. This invention embodiment establishes an optimal ligand density inference model based on, but not limited to, a Generative Adversarial Networks (GAN) model, realizing the function of generating a predicted ligand density corresponding to the pocket electron density. Then, the predicted ligand density is analyzed to obtain the ligand molecule corresponding to the predicted ligand density.

[0059] Step S103: Input the predicted ligand electron density into a preset conditional generation model to generate multiple similar ligand electron densities. The preset conditional generation model is trained based on small molecule samples. Specifically, due to the difficulty in obtaining paired data such as pocket-ligand structures, there is a lack of electron density data for pocket-ligand structures. Consequently, the number of predicted ligand densities of the same structure but with different manifestations obtained through the optimal ligand density inference model is limited. On the other hand, since the predicted ligand density obtained based on GAN implicitly models the distribution of latent variables, we do not know the specific distribution of latent variables after the entire GAN training is completed. Therefore, we cannot arbitrarily generate the desired ligand electron density. We can only arbitrarily input noise, and the network's generated data looks like the data in the training dataset. Such data has limitations, so directly using the prediction results to construct ligand molecules inevitably leads to errors. Therefore, how to reasonably expand the obtained prediction results is a problem that needs to be solved. In the process of new drug development, the more ligand molecules that meet the target characteristics are generated, the more beneficial it is for subsequent experimental analysis, and the higher the reliability of the conclusions. This invention is based on a variational autoencoder (VAE) to establish a conditional generation model, but this invention is not limited thereto. This model is trained using a large number of small molecule samples, and a large number of single small molecule samples are readily available. Therefore, a complete variational autoencoder can be trained using a large number of training samples, thereby obtaining a variety of similar ligand electron densities based on a single ligand electron density. After obtaining the predicted ligand electron density in step S102, the conditional generation model is used to obtain multiple similar ligand electron densities that are very similar to the predicted ligand electron density and satisfy the target characteristics and electron density rules. Since VAE explicitly models the distribution of latent variables, we can obtain the specific distribution of latent variables after training the VAE, and therefore can specify the appearance of the generated data. This invention uses a GAN to first generate a relatively accurate predicted ligand electron density, and then uses a VAE to expand the predicted ligand electron density, thereby obtaining multiple legitimate ligand electron densities that satisfy the target characteristics. This expands the data representation form and improves the reliability and accuracy of subsequent ligand molecule generation by utilizing multi-channel data.

[0060] Step S104: Analyze the electron densities of multiple similar ligands using a preset analytical model to generate target atom information. The preset analytical model is trained based on small molecule samples. The target atom information includes the atomic position and atom type of the target atom. Specifically, in this embodiment of the invention, electron density is a graphical representation in three-dimensional space. Therefore, after obtaining multiple legitimate ligand electron density maps, by analyzing the intensity of the electron density values ​​in the map and their relationship with NCI, and by comprehensively comparing and eliminating error points from multiple analytical results, all target atoms contained in the target ligand molecule, along with their corresponding atomic positions and atom types, can be obtained. Then, according to the combination rules of chemical substances, the obtained atoms can be connected to obtain the final target ligand molecule. The analytical models used in this invention include, but are not limited to: Vnet and Transformer keypoint recognition models. These models are trained on a large number of small molecule samples. The input of this model is the electron density of the small molecule, and the output is the corresponding small molecule conformation (including atomic position and atom type). The specific process of identifying keypoints using the above recognition models is prior art and will not be elaborated upon in this invention.

[0061] Step S105: Construct the target ligand molecule based on the target atom information. Specifically, using the target atom information obtained in step S104, and utilizing the chemical linking rules, the relevant atoms can be linked together to construct the final target ligand molecule. In this embodiment of the invention, the atom linking is completed in the OpenBabel software, and the specific linking rules are existing technologies, which will not be elaborated upon in this invention.

[0062] Specifically, in one embodiment, a method for generating ligands based on pocket electron density further includes the following steps after step S104:

[0063] Step S106: Correct the atomic positions and atom types according to preset chemical rules. Specifically, although the key point identification technology in the analytical model is mature, and the accuracy of the atomic positions and atom types obtained by the analytical model is already high, errors in the analysis of atomic positions and atom types are unavoidable. Therefore, in one embodiment, the present invention further uses traditional chemical rules based on bond length and bond angle to correct the identified atomic positions and atom types, thereby improving the accuracy of the obtained target ligand molecules. In this embodiment of the present invention, the steps for correcting atomic positions and atom types according to preset chemical rules are as follows:

[0064] 1. Enter the atom position, element type, and confidence level (confidence level represents the probability that an atom exists at a certain position).

[0065] 2. Based on dbscan technology, the generated points are clustered, with points that are close in distance grouped together to reduce redundant points.

[0066] 3. For any cluster, select the point with the highest confidence level in the cluster as the starting point, and remove the other points in the cluster.

[0067] 4. Based on bond length and bond angle rules, select the neighboring clusters connected to the current cluster.

[0068] 5. Based on the bond length and bond angle rules, select and fine-tune points in adjacent clusters that meet the bond length and bond angle requirements, and remove other points in the cluster. The class with the most element categories in a cluster is the element type of that cluster.

[0069] 6. Repeat steps 4 and 5 until all clusters have been connected.

[0070] 7. Use OpenBabel software to connect the selected atomic points.

[0071] Specifically, in one embodiment, the training process for the optimal ligand density inference model in step S102 above includes the following steps:

[0072] Step S201: Obtain training samples, which include the true pocket electron density and the true ligand electron density, with the true ligand electron density serving as the label for the true pocket electron density. Specifically, before training, the training samples required for training the model are first obtained. In this embodiment of the invention, a ligand prediction model is established based on GAN. GAN includes a generator and a discriminator, with the true ligand electron density as the target. The generator needs to generate a predicted ligand electron density similar to the target based on the true pocket electron density. Therefore, a training sample contains both the true pocket electron density and the true ligand electron density, and the true ligand electron density is used as the label for the true pocket electron density, thereby enabling supervised training. Specifically, the methods for obtaining the true pocket electron density and the true ligand electron density include, but are not limited to, experimental methods and computational methods. Commonly used methods include: methods for obtaining electron density based on X-ray crystallography experiments, methods for obtaining electron density based on electron microscopy experiments, and methods for obtaining electron density based on quantum chemical calculations.

[0073] Step S202: Input the training samples into the generator in the generative adversarial network (GAN) to obtain the second predicted ligand electron density. Specifically, the GAN is trained during the adversarial process between the generator and the discriminator. The generator aims to generate increasingly realistic ligand electron densities, while the discriminator needs to determine whether the actual electron density is real and whether the predicted electron density generated by the generator is false. When the result of the second predicted ligand electron density is determined to be false, it indicates that the ligand electron density generated by the generator is not realistic enough, and the generator parameters are adjusted accordingly. When the discriminator determines that the result of the second predicted ligand electron density is real, it indicates that the discriminator's discrimination ability is not strong enough and it cannot distinguish the false density, so the discriminator parameters are adjusted accordingly. Furthermore, when the discriminator determines the actual ligand electron density, if the discrimination result is false, it indicates that the discriminator cannot distinguish the real electron density, and the discriminator parameters need to be adjusted. However, training stops when a preset condition is met, and the GAN can be considered to have completed training. Therefore, in this embodiment of the invention, the actual pocket electron density is input into the generator, and the obtained result, namely the second predicted ligand electron density, is used to train the discriminator, thereby realizing the adjustment of the parameters of the generator and the discriminator.

[0074] Step S203: Establish a connection between the true ligand electron density and the true pocket electron density to obtain positive samples, and establish a connection between the second predicted ligand electron density and the true pocket electron density to obtain negative samples; label the positive samples as true and label the negative samples as false. Specifically, when training the discriminator, the true ligand electron density and the second predicted ligand electron density are used. The discriminator fully considers the binding characteristics of the pocket and the ligand, and needs to determine that the second predicted ligand electron density generated by the generator is a false result, and the obtained true ligand electron density is a true result. Therefore, this step labels the two electron densities mentioned above.

[0075] Step S204: Input the positive and negative samples into the discriminator in the generative adversarial network to obtain the discrimination result, and calculate the result of the preset loss function based on the error between the discrimination result and the corresponding label. Specifically, based on whether the discriminator can make a correct judgment, and in conjunction with the discriminator principle described in step S202, calculate the result of the preset loss function using the error between the discrimination result and the corresponding label. When the preset loss function becomes smaller and smaller, it indicates that the training effect of the model is becoming more accurate, thereby ensuring that the preset loss function is smaller in the next result than in the previous result as much as possible. Adjust the parameters of the generator and / or discriminator according to this standard.

[0076] Step S205: Adjust the parameters of the generator and / or the discriminator based on the results of the preset loss function. Specifically, the principle behind parameter adjustment is explained in step S202, and will not be repeated here.

[0077] Step S206: Determine whether the results of the training iterations and / or the preset loss function meet the preset conditions; if the results of the training iterations and / or the preset loss function do not meet the preset conditions, repeat the steps from obtaining training samples to adjusting the parameters of the generator and / or the discriminator based on the results of the preset loss function. Specifically, after one parameter adjustment, if the preset conditions are not met, return to step S201 and repeat the operations of steps S201 to S205, using new positive samples and obtaining new negative samples, and using the new samples to adjust the parameters of the generator and discriminator again. The preset conditions include, but are not limited to, stopping training after completing a specified number of training iterations (e.g., executing 1000 training iterations to end training), stopping training when the loss function tends to stabilize as it decreases, and stopping training when the value of the loss function is less than a specified value. In this embodiment of the invention, training is stopped and the adjustment of the generator and discriminator parameters ends when the generator's loss function appears ten consecutive times and the result of each time changes by no more than 2% compared to the previous result.

[0078] Step S207: The generator that meets the preset conditions is used as the optimal ligand density inference model. Specifically, the generator that meets the preset conditions can realize the function of inputting the target pocket electron density and outputting the predicted ligand electron density. The discriminator only serves to train the generator and has no effect on the result of predicting the ligand electron density; therefore, the generator is used as the optimal ligand density inference model. In this embodiment of the invention, the pixel2pixel network from the GAN family is used for training in steps S201 to 206, but the invention is not limited thereto.

[0079] Specifically, in one embodiment, step S201 above includes the following steps:

[0080] Step S301: Discretize the true pocket electron density and the corresponding true ligand electron density to obtain discrete samples. Specifically, the electron density map is a continuous function in space. Training using electron density requires converting continuous data into discrete data. In this embodiment of the invention, the input and label (i.e., the true pocket electron density and the corresponding true ligand electron density) of the training samples are stored in a three-dimensional matrix at intervals of 0.5×0.5×0.5 (unit: angstroms) to obtain discrete training samples.

[0081] Step S302: Normalize and spatially transform the discrete samples to obtain training samples. Specifically, the electron density values ​​are typically between 1 and 100. To simplify the calculation of the neural network model and improve computational efficiency, the electron density values ​​in the discrete samples are normalized. Then, a random angle and direction are selected, and the center of the three-dimensional matrix is ​​used as the rotation center. The true pocket electron density and the true ligand electron density are rotated and translated in the same way, thereby increasing the representation of the training samples and improving the accuracy of model training.

[0082] Specifically, in one embodiment, the calculation of the preset loss function in step S204 above includes the following steps:

[0083] Step S303: Based on the similarity between the gradient of the second predicted ligand electron density and the gradient of the true ligand electron density, calculate the result of the first loss function. The first loss function is used to adjust the parameters of the generator. Specifically, the gradients of the second predicted ligand electron density and the true ligand electron density are calculated respectively. Then, a loss function is set according to the similarity of the gradients, and a preset similarity threshold is used. In this embodiment of the invention, the mean square error of the two gradients is used to characterize the similarity. Setting the mean square error to be less than the preset similarity threshold is one of the bases for stopping training. The loss function is set by the electron density gradient so that the generator can better learn the topological features of the density.

[0084] Step S304: Based on the similarity between the true ligand electron density and the second predicted ligand electron density at the reference position, calculate the result of the second loss function. The second loss function is used to adjust the parameters of the generator. The reference position is the location of non-covalent interactions in the true ligand electron density. Specifically, in this embodiment, NCILoss is set as the second loss function. First, a cube is used to represent the NCI region (i.e., the reference position). The center of the atom on the true ligand molecule that participates in NCI interaction is taken as the center of the cube, and the side length of the cube is x times the covalent radius (x is generally a value greater than 1). The resulting cube is the NCI region. Next, in the true ligand molecule, the similarity (e.g., MSE) between the true ligand electron density and the predicted ligand electron density within the NCI region is calculated. This is NCILoss. The error between the calculated similarity and the preset similarity is less than a specified value as one of the criteria for stopping training. By combining the NCI position with NCILoss, more intramolecular NCI topological features are considered, enabling the trained ligand inference model to infer a more complete predicted ligand electron density.

[0085] Step S305: Based on the balance between the discrimination result and the true result, calculate the result of the third loss function. The third loss function is used to simultaneously adjust the parameters of the discriminator and the generator. Specifically, the third loss function includes: a loss function based on the discrimination result error of positive samples and a loss function based on the discrimination result error of negative samples. Stabilizing the change in the loss function result is one of the criteria for stopping training. The principle is the same as in step S202. Common loss function models such as DiceIOU and PatchGAN Loss can be selected, but this invention is not limited to these. The specific calculation steps will not be elaborated here.

[0086] Step S306: Based on the similarity between the second predicted ligand electron density and the true ligand electron density, calculate the result of the fourth loss function. The fourth loss function is used to adjust the generator parameters. Specifically, the fourth loss function is set by the error between the electron density result generated by the generator and the true electron density label, further improving the criterion for stopping training and ensuring the accuracy of model training.

[0087] Step S307: Add the results of the first loss function, the second loss function, the third loss function and the fourth loss function according to the types of generator and discriminator to obtain the result of the preset loss function.

[0088] Specifically, this embodiment of the invention establishes model training standards using the four types of loss functions set in steps S303 to S306 above. Training stops when the generator and discriminator loss functions established based on the above four types of loss functions meet the preset conditions of the preset loss functions in step S206, or after a specified number of training iterations. The formulas for the loss functions of the generator and discriminator are as follows:

[0089] L G =L NCI (G(x p ))+L2(G(x p ))+L grad (G(x p ))+L CE (D(G(x p ),x p ),True)

[0090] The generator's loss function L G , where x p G(x) represents the pocket electron density input. p L represents the predicted ligand electron density generated based on the pocket electron density. NCI (G(x p )) represents the NCI loss, i.e., the second loss function; L2(G(x) p )) represents the density distribution loss, i.e., the fourth loss function; Lgrad (G(x p )) represents gradient loss, i.e., the first loss function; L CE (D(G(x p ),x p (),True) represents the cross-entropy loss compared to the result "true", i.e., the third loss function; D(G(x) p ),x p The result is obtained by inputting the predicted ligand electron density and pocket electron density into the discriminator.

[0091] L D =L CE (D(x l ,x p ),True)+L CE (D(G(x p ),x p ),False)

[0092] The loss function L of the discriminator D , where x l Represents the actual (positive) ligand electron density. L CE (D(x l ,x p The "True" function represents the cross-entropy loss between the true (positive) ligand electron density and the pocket electron density, and the result "True," i.e., the third loss function; L CE (D(G(x p ),x p ),False) is the cross-entropy loss between the predicted (false) ligand electron density and the pocket electron density and the discrimination result "false", i.e., the third loss function; D(x l ,x p The result is obtained by inputting the true ligand electron density and pocket electron density into the discriminator after establishing a connection.

[0093] Specifically, in one embodiment, the training process for the conditional generation model in step S103 includes the following steps:

[0094] Step S208: Generate training samples based on small molecule samples. Specifically, to improve the accuracy of the variational autoencoder-based generative model and facilitate its application in drug development, a large number of drug-like ligand small molecule samples are needed to train the variational autoencoder model. For each sample, a unique probability distribution is learned. Therefore, the more samples, the more probability distributions are obtained, allowing for the identification of more accurate probability distributions when new samples are received, and the generation of multiple legitimate ligand electron densities similar to the new samples, conforming to topological features and chemical rules. In one embodiment, training data is created using a drug-like dataset. First, software such as OpenBabel is used to sample the SMILE expressions representing ligand molecules into multiple reasonable molecular conformations. Then, the electron densities corresponding to the molecular conformations are calculated as the training set.

[0095] Step S209: Train a variational autoencoder (VAE) based on conditional training samples to learn the probability distribution of the conditional training samples. The VAE is trained using a large number of conditional training samples, taking the samples as input and themselves as the target. The output of the encoder and decoder is made to be infinitely close to itself, resulting in a set of encoders and decoders at the end of training. This set of encoders and decoders constitutes the variational autoencoder. Therefore, after learning, multiple similar ligand electron densities can be generated based on the predicted ligand electron density obtained from the GAN, simulating the probability distribution of the predicted ligand electron density. The specific process of using a VAE to learn multiple sample probability distributions is existing technology and will not be elaborated upon in this invention.

[0096] Step S210: Use the variational autoencoder at the end of the learning process as a conditional generative model.

[0097] Specifically, in one embodiment, step S208 above includes the following steps:

[0098] Step S211: Obtain the structural model of the small molecule sample. Specifically, using software such as OpenBabel, the small molecule sample (smiles expression) representing the ligand molecule is sampled into multiple reasonable molecular conformations (i.e., structural models). Then, the electron density corresponding to the molecular conformation is calculated. The results of a large number of calculated electron densities are used as the training set to obtain a set containing a large number of conditional training samples.

[0099] Step S212: Atom coordinates in the statistical structural model;

[0100] Step S213: Transform the atomic coordinates from real space to frequency domain space to generate the structure factor;

[0101] Step S214: Adjust the structure factor with a preset resolution to return to the real space to generate electron density grid information, and use the electron density grid information as a conditional training sample. The preset resolution is the summation range of the preset frequency domain spatial vector components.

[0102] Specifically, regarding steps S212-214 above, the electron densities of numerous small molecule structure models obtained through software cannot be obtained experimentally. Furthermore, electron densities calculated based on quantum chemistry focus too much on small molecule systems, often neglecting the connections between small molecules within macromolecular systems. Therefore, from a macroscopic perspective, the calculation results are often erroneous. Based on this, the annotation method provided by this invention offers a preferred solution for obtaining electron density: a method for calculating electron density based on crystallographic theory first transforms atomic coordinate information from real space (solid space) to frequency domain space. The diffraction of electromagnetic waves by the crystal lattice yields the pattern in the frequency domain space. The frequency domain space not only continues the symmetry of the crystal lattice and information such as molecular structure and physicochemical properties from real space, but also decomposes the "level of detail" of molecules in real space according to different "frequency" values. That is, high-frequency information in the frequency domain space reflects the details of molecules in real space, while low-frequency information reflects the coarse outline of molecules in real space. After filtering the information in the frequency domain space according to frequency, and then transforming it back to real space, molecular representations with "different levels of detail" are obtained. Specifically, the molecular coordinate information in real space is subjected to Fourier transform to obtain the result after frequency domain space transformation (referred to as structure factor in this embodiment of the invention). Then, the frequency domain space vector is subjected to inverse Fourier transform within a preset frequency range to return to real space. The value obtained by transforming the molecular coordinate information in real space to frequency domain space and then back to real space is the electron density grid information.

[0103] Specifically, a Fourier transform-based approach is used to convert atomic coordinate information from real space to frequency domain space to generate structure factors. First, the atomic coordinate information of the target molecular structure is obtained. Atomic coordinate information is a set of spatial coordinates of a large number of atoms in the target molecular structure. Then, the Fourier transform formula is used to convert the atomic coordinate information from real space to frequency domain space to generate structure factors. In this embodiment of the invention, the Fourier transform is used to calculate the structure factor based on the atomic coordinate information. The specific conversion formula is as follows:

[0104]

[0105] Where r represents atomic coordinate information, x, y, and z represent the three components of atomic coordinate information, and f nLet represent the diffraction factor of n different elements, s represent the frequency domain spatial vector, h, k, and l represent the three components of the frequency domain spatial vector, F(s) represent the structure factor of the s vector (including diffraction amplitude and phase information), and 2πi be an imaginary number. Furthermore, by introducing descriptions of the solvent and atomic vibrations over time, the above formula can be improved, resulting in a more accurate conversion. The improved formula is as follows:

[0106]

[0107] Where f(bulk solvent) is the solvent shell description function, bulksolvent is called the solvent shell, f(b) is the temperature factor description function, b is called the temperature factor, the temperature factor is used to measure the degree of atomic thermal motion in the crystal, and the solvent shell describes the contribution of the solvent to diffraction.

[0108] Then, the structure factor is adjusted with a preset resolution and returned to the real space to generate electron density grid information.

[0109] The specific calculation formula is as follows:

[0110]

[0111] In this system, r represents atomic coordinates, s represents the frequency domain spatial vector, h, k, and l represent the three components of the frequency domain spatial vector, F(s) is the structure factor of the s vector, ρ(r) represents the electron density lattice information, Vcell represents the unit cell volume, and 2πi is the imaginary number. The preset resolution is obtained by summing a preset range of h, k, and l. Specifically, the preset resolution is obtained by summing a preset range of the components of the frequency domain spatial vector. That is, by changing the summing range of h, k, and l, different resolutions of electron density can be obtained. The higher the resolution, the clearer the electron density map and the more accurate the representation of bonds between atoms; the lower the resolution, the blurrier the electron density map, but the more accurate the representation of the overall molecular framework. The method of calculating electron density based on crystallography not only considers the macroscopic connections between atoms in macromolecular systems but also obtains electron density maps with different levels of detail by adjusting different preset resolutions, thus enabling the acquisition of electron density maps under various resolution conditions. Subsequently, by dividing the target molecular structure into a predetermined number of lattice points and inputting the x, y, and z coordinates of these lattice points into the electron density lattice information function ρ(r), the electron density map is divided into discrete lattice points, thus obtaining the electron density lattice information. The final saddle point positions obtained from electron density lattice information at different resolutions also differ. Therefore, by analyzing and comparing the saddle point positions at high resolution and low resolution, noise can be eliminated from both macroscopic and microscopic perspectives, resulting in more accurate identification results.

[0112] By executing the above steps, this invention provides a method for generating ligands based on pocket electron density. First, the pocket electron density is input into a ligand inference model to obtain a predicted ligand electron density. Then, an extended inference model is used to obtain multiple legitimate ligand electron densities that match the predicted ligand electron density characteristics, increasing the sample size. Finally, an analytical model is used to analyze the key points in the legitimate ligand electron densities to obtain the atomic positions and types in the target ligand molecule to be generated. These atomic positions and types are then connected according to the linkage rules of chemical substances to obtain the target ligand molecule. The target ligand molecule obtained by this method, based on the characteristics of electron density, fully considers the topological features and spatial structure of the protein pocket, resulting in a higher and more accurate fit between the generated target ligand molecule and the target protein (pocket). Applying this method to the field of drug design will greatly reduce the design difficulty of ligand-based small molecule drugs, improve the accuracy of small molecule drugs anchoring to receptors, and thus improve therapeutic efficacy.

[0113] like Figure 2 As shown, this embodiment also provides a ligand generation system based on pocket electron density, the system comprising:

[0114] The information acquisition module 101 is used to acquire the electron density of the target pocket. For details, please refer to the relevant description of step S101 in the above method embodiment, which will not be repeated here.

[0115] The ligand electron density inference module 102 is used to input the target pocket electron density into a preset optimal ligand electron density inference model to generate a predicted ligand electron density. The preset optimal ligand electron density inference model is trained based on the actual pocket electron density and the actual ligand electron density. For details, please refer to the relevant description of step S102 in the above method embodiment, which will not be repeated here.

[0116] The ligand electron density expansion module 103 is used to input the predicted ligand electron density into a preset conditional generation model to generate multiple similar ligand electron densities. The preset conditional generation model is trained based on small molecule samples. For details, please refer to the relevant description of step S103 in the above method embodiments, which will not be repeated here.

[0117] The atom analysis module 104 is used to analyze the electron density of multiple similar ligands using a preset analysis model to generate target atom information. The preset analysis model is generated based on small molecule samples. The target atom information includes the atomic position and atom type of the target atom. For details, please refer to the relevant description of step S104 in the above method embodiment, which will not be repeated here.

[0118] The ligand generation module 105 is used to construct the target ligand molecule based on the target atom information. For details, please refer to the relevant description of step S105 in the above method embodiments, which will not be repeated here.

[0119] This invention provides a ligand generation system based on pocket electron density, used to execute the ligand generation method based on pocket electron density provided in the above embodiments. Its implementation and principle are the same, and detailed descriptions can be found in the relevant descriptions of the above method embodiments, which will not be repeated here.

[0120] Figure 3 An electronic device according to an embodiment of the present invention is shown. The device includes a processor 901 and a memory 902, which can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0121] Processor 901 can be a Central Processing Unit (CPU). Processor 901 can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0122] The memory 902, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the above method embodiments. The processor 901 executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions, and modules stored in the memory 902, thereby implementing the methods in the above method embodiments.

[0123] The memory 902 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor 901, etc. Furthermore, the memory 902 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 902 may optionally include memory remotely located relative to the processor 901, and these remote memories may be connected to the processor 901 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0124] One or more modules are stored in memory 902, and when executed by processor 901, they perform the methods described in the above method embodiments.

[0125] The specific details of the aforementioned electronic device can be understood by referring to the relevant descriptions and effects in the above method embodiments, and will not be repeated here.

[0126] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The implemented program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0127] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for generating ligands based on pocket electron density, characterized in that, The method includes: Obtain the electron density of the target pocket; The target pocket electron density is input into a preset optimal ligand density inference model to generate a predicted ligand electron density. The preset optimal ligand density inference model is trained and generated based on the real pocket electron density and the real ligand electron density. The predicted ligand electron density is input into a preset conditional generation model to generate multiple similar ligand electron densities. The preset conditional generation model is trained based on small molecule samples. The step of training the preset conditional generation model based on small molecule samples includes: obtaining the structural model of the small molecule sample; statistically analyzing the atomic coordinates in the structural model; converting the atomic coordinates from real space to frequency domain space to generate structure factors; adjusting the structure factors at a preset resolution to return to real space to generate electron density grid information, and using this electron density grid information as conditional training samples. The preset resolution is a preset range for summing the frequency domain space vector components. A variational autoencoder is trained based on the conditional training samples to learn the probability distribution of the conditional training samples. The variational autoencoder at the end of the learning process is used as the conditional generation model. The electron density of the multiple similar ligands is analyzed using a preset analytical model to generate target atom information. The preset analytical model is generated based on training with small molecule samples. The target atom information includes the atomic position and atom type of the target atom. Construct the target ligand molecule based on the target atom information.

2. The method according to claim 1, characterized in that, The steps for training and generating the preset optimal ligand density inference model based on the real pocket electron density and the real ligand electron density include: Obtain training samples, which include the true pocket electron density and the true ligand electron density, wherein the true ligand electron density is a label of the true pocket electron density; The training samples are input into the generator in the generative adversarial network to obtain the second predicted ligand electron density. Positive samples are obtained by connecting the actual ligand electron density with the actual pocket electron density, and negative samples are obtained by connecting the second predicted ligand electron density with the actual pocket electron density. The positive samples are labeled as true, and the negative samples are labeled as false. The positive and negative samples are respectively input into the discriminator in the generative adversarial network to obtain the discrimination result, and the result of the preset loss function is calculated based on the error between the discrimination result and the corresponding label; Adjust the parameters of the generator and / or the parameters of the discriminator based on the result of the preset loss function; Determine whether the number of training iterations and / or the result of the preset loss function meet the preset conditions; if the number of training iterations and / or the result of the preset loss function do not meet the preset conditions, repeat the steps from obtaining training samples to adjusting the parameters of the generator and / or the parameters of the discriminator according to the result of the preset loss function until the number of training iterations and / or the result of the preset loss function meet the preset conditions; The generator that meets the preset conditions is used as the optimal ligand density inference model.

3. The method according to claim 2, characterized in that, The acquisition of training samples includes: Discretize the actual pocket electron density and the corresponding actual ligand electron density to obtain discrete samples; The training samples are obtained by normalizing and spatially transforming the discrete samples.

4. The method according to claim 2, characterized in that, The result of calculating a preset loss function based on the error between the discrimination result and the corresponding label includes: The result of the first loss function is calculated based on the similarity between the gradient of the second predicted ligand electron density and the gradient of the true ligand electron density. The first loss function is used to adjust the parameters of the generator. The result of a second loss function is calculated based on the similarity between the true ligand electron density at the reference position and the second predicted ligand electron density. The second loss function is used to adjust the parameters of the generator. The reference position is the location of non-covalent interactions in the true ligand electron density. Based on the degree of balance between the discrimination result and the true result, the result of the third loss function is calculated. The third loss function is used to adjust the parameters of the discriminator and the generator simultaneously. Based on the similarity between the second predicted ligand electron density and the true ligand electron density, the result of the fourth loss function is calculated, which is used to adjust the parameters of the generator. The results of the first loss function, the second loss function, the third loss function, and the fourth loss function are added together according to the types of the generator and the discriminator to obtain the result of the preset loss function.

5. The method according to claim 1, characterized in that, After using a preset analytical model to analyze the electron densities of the multiple similar ligands to generate target atom information, the method further includes: The atomic positions and atomic types are corrected according to preset chemical rules.

6. A ligand generation system based on pocket electron density, characterized in that, The system includes: The information acquisition module is used to obtain the electron density of the target pocket; The ligand electron density inference module is used to input the target pocket electron density into a preset optimal ligand electron density inference model to generate a predicted ligand electron density. The preset optimal ligand electron density inference model is trained and generated based on the real pocket electron density and the real ligand electron density. A ligand electron density extension module is used to input the predicted ligand electron density into a preset conditional generation model to generate multiple similar ligand electron densities. The preset conditional generation model is trained based on small molecule samples. The step of training the preset conditional generation model based on small molecule samples includes: obtaining the structural model of the small molecule sample; statistically analyzing the atomic coordinates in the structural model; converting the atomic coordinates from real space to frequency domain space to generate structure factors; adjusting the structure factors at a preset resolution to return to real space to generate electron density grid information, and using the electron density grid information as conditional training samples, where the preset resolution is a preset range for summing frequency domain space vector components; training a variational autoencoder based on the conditional training samples to learn the probability distribution of the conditional training samples; and using the variational autoencoder at the end of the learning process as the conditional generation model. The atom analysis module is used to analyze the electron density of the multiple similar ligands using a preset analysis model to generate target atom information. The preset analysis model is generated based on small molecule samples, and the target atom information includes the atomic position and atom type of the target atom. The ligand generation module is used to construct the target ligand molecule based on the target atom information.

7. An electronic device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method as described in any one of claims 1-5.

Images