A multi-objective optimization molecule generation method and system combined with active learning

Abstract

The application belongs to the technical field of deep learning and active learning (AL) in artificial intelligence, and discloses a multi-objective optimization molecule generation method and system combined with active learning, constructs a multi-objective optimization molecule generation model and a property prediction model combined with active learning and trains the same; after the training is completed, the weights of the generation model encoder and the property prediction model are frozen; the output of the decoder is input to the decoder after being randomly sampled from the latent space z, and the output of the decoder is input to the prediction model to predict the property; the molecule generation model performs gradient descent optimization z relative to the latent space z through the output of the prediction model; the optimized z is input to the decoder again to obtain a molecule graph vector and perform validity correction to obtain a final optimized molecule. The application improves the molecule generation speed and the multi-objective optimization success rate, and is suitable for affinity optimization problems of various target proteins.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of deep learning and active learning (AL) in artificial intelligence, and more particularly relates to a multi-objective optimization molecule generation method and system combined with active learning. BACKGROUND

[0002] Currently, in drug discovery, finding drug-like molecules with high affinity to target proteins is a difficult and resource-intensive task. The number of drug-like compounds in the chemical space can be as high as 10 23 ~ 10 60 Therefore, searching for molecules with the desired target in such a vast space is a highly challenging task, often requiring a large amount of time and resources. And for a specific disease, finding candidate molecules for a specific protein is even more complex. Currently, this task is accomplished through large-scale compound screening experiments and repeated synthesis tests by drug chemistry experts.

[0003] Recent work proposes to use deep generative models to directly generate candidate drugs with ideal properties, thereby avoiding a large amount of experimental work. However, most of the work has focused on generating molecules related to a single property, such as calculating simple drug similarity (QED), logP of lipid-water partition coefficient. Since calculating binding affinity based on physics requires high computational cost, few works have generated molecules with high affinity to target proteins. In addition, in a real environment, drug discovery needs to consider multiple properties comprehensively, and the multi-objective molecule design problem is more challenging.

[0004] Molecule generation models are mainly divided into sequence-based generation models and graph-based generation models based on the way of molecular representation. Sequence-based models mainly use SMILES to describe molecules, such as using LSTM-based CVAE or reinforcement learning strategies to generate molecules with specific properties. Although SMILES is easy to process by computer, it cannot capture the similarity of molecular structure and often loses a lot of structural information. Moreover, in order to generate effective strings, the model also needs to learn rules unrelated to molecular structure, adding unnecessary burden to model training. Therefore, graph-based models are more popular.

[0005] Graph-based generation models usually have two ways: sequential generation of graphs and direct generation of entire graphs. Sequential generation incorporates chemical rules into the generation process to ensure the chemical validity of each step; direct generation can perform validity checking at the last step to filter valid molecules. Current generation models generate molecules quickly, but cannot effectively find molecules with target properties in their latent space, so pure molecule generation models cannot solve the molecule optimization problem.

[0006] Current methods for molecular property optimization mainly include reinforcement learning, Markov sampling or deep generative models guided by Gaussian processes. However, these methods need to call the property calculation function multiple times, and when the property to be optimized (such as binding affinity) requires a large amount of computing resources, it will cause the speed of training and generating molecules to be extremely slow. In addition, the high variance problem of reinforcement learning will also increase the difficulty of model training.

[0007] Prior art one, application number: CN201911122108.7 A drug molecule generation method based on adversarial imitation learning, comprising: constructing an effective drug molecule library; establishing an improved drug molecule generation model, including: designing and implementing a multi-task reinforcement learning module, designing and implementing an adversarial imitation learning module; model pre-training; executing a drug molecule generation process; generating candidate drug molecule results. It can effectively promote the optimization of drug molecule biochemical properties and improve the stability of model training, and obtain better drug molecules.

[0008] Prior art two, application number: CN202080073784.4 Workflow for generating compounds having biological activity against a particular biological target, a computer-implemented method can include: receiving an input of a biological target; receiving a generative model (such as a tensor reinforcement learning (GENTRL) model or other model) trained with reference compounds, wherein the reference compounds include: general compounds, compounds that modulate the biological target, and compounds that modulate biological molecules other than the biological target; generating a structure of a generated compound with the generative model; prioritizing the structure of the generated compound based on at least one criterion; processing the prioritized chemical structure of the generated compound through a Sammon mapping protocol to obtain a hit structure; and providing a chemical structure of the hit structure. One or more non-transitory computer-readable media storing instructions that, in response to being executed by one or more processors, cause a computer system to perform operations including performing the computer method described herein to provide a chemical structure of a generated hit structure through a generative model.

[0009] The prior art three, application number: CN202211513416.4 based on the specific target drug generation method, device and MaxFlow platform of graph neural network, method includes: obtaining target protein data and pre-drug molecule number, determining the box information of the target area, the box information includes the position and size of the target protein;The target protein data, pre-drug molecule number and box information are input into the pre-trained receptor molecule generation model to obtain the three-dimensional structure of the candidate drug;The three-dimensional structure of the candidate drug is combined with the target protein data to obtain the data structure of the target protein-drug consistent program database file;Wherein, the receptor molecule generation model is obtained by training the target protein sample and drug structure sample based on the graph neural network. The technical problems of the prior art that the specific target drug design is difficult and the accuracy of the drug model is poor are solved.

[0010] Based on the above analysis, it can be seen that when the optimization target requires high computational cost, most current methods will consume a lot of resources and greatly reduce the speed of generating molecules. Therefore, how to quickly explore the chemical space while reducing the computational cost and generate molecules with expected properties is still a problem to be solved.

[0011] Through the above analysis, the problems and defects of the prior art are:

[0012] (1) The method of directly generating candidate drugs with ideal properties using deep generative models mostly focuses on generating molecules related to a single property, and the physical-based method costs high in calculating binding affinity.

[0013] (2) The sequence-based model uses SMILES to describe molecules, which cannot capture the similarity of molecular structure and often loses a lot of structural information, adding unnecessary burden to model training.

[0014] (3) The current generation model cannot effectively find molecules with target properties in its hidden space, and the model for molecular optimization generally uses reinforcement learning algorithm. When the calculated properties require a lot of computing resources, it will cause the speed of training and generating molecules to be extremely slow, increasing the difficulty of model training. SUMMARY

[0015] In view of the above defects or improvement needs of the prior art, the present application provides a multi-objective optimization molecule generation method and system combined with active learning, which aims to solve the technical problems that the method of directly generating candidate drugs with ideal properties by the existing deep generation model is costly in terms of calculation of binding affinity based on physics, and the way of using SMILES to describe molecules by the sequence-based model cannot capture the similarity of molecular structure, resulting in the loss of a lot of structural information and the unnecessary burden on model training, and the existing generation model cannot effectively find molecules with target properties in its latent space, and the model for molecular optimization generally uses reinforcement learning algorithm, resulting in extremely slow speed of training and molecule generation when the calculated properties require a large amount of computing resources, and increasing the difficulty of model training.

[0016] To achieve the above-mentioned purpose, according to one aspect of the present application, a multi-objective optimization molecule generation method combined with active learning is provided, comprising: constructing and training a multi-objective optimization molecule generation model combined with active learning and a property prediction model; after training the molecule generation model and the property prediction model, freezing the weights of the generation model encoder and the property prediction model; using After random sampling from the latent space z and inputting into the decoder, the output of the decoder is input into the prediction model to predict the properties; the molecule generation model performs gradient descent optimization of z with respect to the latent space z through the output of the prediction model; the optimized z is input into the decoder again to obtain a molecular graph vector, and finally an effective correction is performed to obtain the final optimized molecule.

[0017] Preferably, the molecule generation model performs gradient descent optimization of z with respect to the latent space z through the output of the prediction model comprises: according to the set target, a set of property prediction models (g 1 ,g 2 ,…,g k ) and corresponding weights (w1, w2, …, w k ) are given, and the following loss is optimized using the Adam optimizer according to the obtained prediction results:

[0018]

[0019] In the formula, f dec represents the decoder, g represents the property prediction model, and w represents the given weight value.

[0020] Preferably, the property prediction model comprises two parts of a feature extraction network and a fully connected network connected in sequence.

[0021] The feature extraction network consists of three alternating linear layers and three ReLU activation layers. The first linear layer takes a molecular graph vector of length S as input and outputs a feature vector of length H1. The second linear layer takes a feature vector of length H1 as input and outputs a feature vector of length H2 as output. The third linear layer has input and output vectors of length H2. S is the length of the molecular graph vector, which is 38*(10+4*38)=6156. H1 and H2 are the lengths of the hidden layer vectors, with H1 having a size of 1024 and H2 having a size of 512.

[0022] The formula for the ReLU activation function is as follows:

[0023] RELU(x) = max(0,x);

[0024] The structure of the fully connected network consists of one Dropout layer and one linear layer. The random deactivation ratio of the Dropout layer is set to 0.1. The input of the linear layer is the feature vector of H2, and the output is the property prediction result.

[0025] Preferably, the property prediction model is trained through the following steps:

[0026] (1) Use the ZINC250k dataset as the unlabeled sample pool P U One thousand molecules are randomly selected from them as the initial training set D. L One thousand molecules were randomly selected as the independent test set D. Test From sample pool P U Delete the extracted molecules; calculate the molecular property label y using different property calculation methods based on the given target type. true Then use the training dataset D L Molecular vector diagrams are constructed from molecules;

[0027] Regarding the calculation of molecular properties, the QED value and the SA score of synthetic feasibility were calculated using built-in functions of the RDkit package, and the binding affinity value with the target protein was calculated using AutoDock-GPU.

[0028] (2) Input the n*S molecular graph vector matrix obtained in step (1) into the feature extraction network to obtain the n*H feature vector matrix T; input the feature vector matrix T into the fully connected network to obtain the final predicted property label vector y, with a length of n, where n represents the number of data.

[0029] (3) Based on the predicted property label vector y and the true property label vector y obtained in step (2) true The loss is calculated, and the model is iteratively trained using the loss. The Adam optimizer and the ReduceLROnPlateau function are used to automatically update the learning rate, and finally the trained property prediction model is obtained.

[0030] wherein the formula using the mean square error loss function is as follows:

[0031]

[0032] (4) From the sample pool P U Randomly extract pool_num molecules, and use the trained generative model to generate five thousand legal molecules passing through the validity corrector, and merge the molecules to construct an unlabeled data set D U , pool_num size is set to 50000 in the case of QED or SA property prediction model training, and is set to 5000 in the case of binding affinity property prediction model training; perform data processing on the unlabeled data set D U , obtain the molecular graph vector matrix M U ; use the trained property prediction model to predict the unlabeled data set, obtain the prediction result y and the parameters θ L of the penultimate layer neural network Dropout layer output, size n*H2;

[0033] (5) input the molecular graph vector matrix M U and the parameter matrix θ L obtained in step (4) into the active learning screening function; wherein the active learning screening function includes: oversampling based on the boundary of the Fisher information optimized maximum likelihood estimation error, selecting 2*batch samples from the data set D U ; using k-MEANS++ seed algorithm to cluster 2*batch samples, and screening out batch samples; batch size is set to 10000 in the case of QED or SA property prediction model training, and is set to 1000 in the case of binding affinity property prediction model training;

[0034] (6) calculate the property label of the screened batch samples, and add them to the training set D L ; if the screened sample belongs to the sample in the unlabeled sample pool P U , delete the unlabeled sample from the sample pool, and retrain the property prediction model;

[0035] (7) repeat steps (4) to (6) until the stopping condition is reached, and obtain the finally trained property prediction model, the stopping condition is set to repeat 10 times.

[0036] Preferably, the implementation process of the screening function includes:

[0037] 1) using the parameter matrix θ L , respectively calculate the unlabeled data set DU and training set D L Fisher information matrix I of D

[0038] Fisher matrix calculation:

[0039]

[0040]

[0041]

[0042] In the formula, F represents the selected Fisher matrix, and F0 only contains the training set at the beginning;

[0043] 2) Greedy oversampling is performed, 2*batch samples are added to F, and the calculation formula is as follows:

[0044]

[0045]

[0046] In the formula, D S represents the unlabeled sample set screened by oversampling;

[0047] 3) Using the k-MEANS++ seed algorithm, clustering operation is performed according to the molecular graph vector matrix M U in D S , and batch cluster centers are screened out as the final screened samples;

[0048] Using the k-MEANS++ algorithm, a sample is randomly selected from the sample set as the first cluster center, and k-1 new cluster centers are sequentially sampled from the sample set, wherein the probability of selecting each new cluster center is proportional to the square distance to the nearest cluster center.

[0049] According to another aspect of the application, a multi-objective optimization molecule generation method combined with active learning is provided, comprising the following steps:

[0050] Step 1, obtain SMILES data and perform preprocessing to generate an adjacency matrix and an annotation matrix, then combine them into a molecular graph vector matrix; input the molecular graph vector matrix into the GVAE model to train the generation model;

[0051] Step 2, after the molecular generation model is trained, sample molecules from the hidden space to generate molecules, and correct the effectiveness of the generated molecules to obtain legal molecules;

[0052] Step 3, construct an initial labeled data set according to the ZINC data set, and train a property prediction model;

[0053] Step four, according to the ZINC dataset and the generated legal molecule, after the data is processed into a molecular graph vector matrix, it is input into the trained prediction model to obtain the property prediction result;

[0054] Step five, combining the Fisher information and the query function of the K-MEANS++ algorithm, the last layer parameters of the prediction model and the molecular structure information are used for screening to obtain high-quality unlabeled data;

[0055] Step six, after the data is labeled, the data is added to the labeled data set, and the prediction model is retrained, and steps four to six are repeated until the iteration stopping condition is reached;

[0056] Step seven, according to the optimization condition, one or more trained predictors are used to optimize the latent space;

[0057] Step eight, input the optimized latent vector into the decoder to generate molecules, and perform validity check on the molecules to obtain the final optimized molecules.

[0058] Preferably, the VAE model structure is used to realize the molecular generation model, and the molecular generation model is trained by the following steps:

[0059] (1) Process SMILES data: normalize the SMILES in the ZINC250k dataset and convert it into a molecule; use the RDkit package to construct the graph representation matrix of the molecule, which represents the annotation matrix X of atomic features and the adjacency matrix A of bond features; use the annotation matrix of size N*A to represent each row as one-hot encoding of atoms, and use the adjacency matrix of size N*N*B to represent whether there is a bond between each row and each column of atoms and the one-hot encoding of the bond; perform dimension transformation and matrix splicing operation on the annotation matrix and the adjacency matrix to obtain the molecular graph vector M, which has a size of N*(A+(N*B)); wherein, N represents the number of atoms, and the maximum number of atoms in the ZINC250k dataset is 38; A represents the number of atomic types, which includes 9 atomic types plus a virtual atomic type 0; B represents the number of bond types, which is 4;

[0060] (2) input the molecular graph vector M obtained in step (1) into the encoder to obtain the mean vector μ of length m and the standard deviation vector σ of length m;

[0061] (3) obtain the mean vector μ and the standard deviation vector σ obtained in step (2), and sample ∈ from the standard normal distribution using the reparameterization technique according to the formula z=μ+σ⊙∈ to obtain the latent variable sequence z of length m;

[0062] (4) input the latent variable z obtained in step (3) into the decoder for decoding to obtain a generated molecular graph vector

[0063] (5) a lower bound loss function under evidence is calculated, and the loss function is used for iterative training of a generation model; an Adam optimizer is adopted, and finally a trained molecular generation model is obtained

[0064] The loss function Loss is:

[0065] Loss=E[logP(M|z)]-D KL [Q(z|M)||P(z)];

[0066] In the formula, the first half of the Loss formula represents a reconstruction error, and is realized by using a cross-entropy loss function; the second half represents a KL divergence; E represents an expected value; P(M|z) represents a prior distribution of the latent variable z, and represents an encoder; Q(z|M) represents a posterior distribution of the latent variable z derived according to a real molecular graph vector, and represents a decoder.

[0067] According to still another aspect of the present application, a multi-objective optimization molecular generation system combined with active learning is provided, comprising:

[0068] A first module is used for obtaining SMILES data and performing preprocessing to generate an adjacency matrix and an annotation matrix, and then combining the two into a molecular graph vector matrix; the molecular graph vector matrix is input into a GVAE model to train a generation model;

[0069] A second module is used for generating molecules by sampling from a latent space after the molecular generation model is trained, and correcting the generated molecules to obtain legal molecules;

[0070] A third module is used for constructing an initial labeled data set according to a ZINC data set, and training a property prediction model;

[0071] A fourth module is used for constructing an unlabeled data set according to the ZINC data set and the generated legal molecules, processing the data into a molecular graph vector matrix, and inputting the data into the trained prediction model to obtain a property prediction result;

[0072] A fifth module is used for screening by using the last layer of parameters of the prediction model and molecular structure information by combining a Fisher information and a K-MEANS++ algorithm query function, to obtain high-quality unlabeled data;

[0073] A sixth module is used for adding the data to a labeled data set after labeling the data, retraining the prediction model, and repeating the fourth module to the sixth module until an iteration stopping condition is reached;

[0074] A seventh module is configured to optimize the latent space using one or more trained predictors according to an optimization condition;

[0075] An eighth module is configured to input the optimized latent vector into a decoder to generate a molecule and perform an effectiveness check on the molecule to obtain a final optimized molecule.

[0076] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0077] First, in view of the technical problems existing in the prior art and the difficulty of solving the problems, the present application closely combines the technical solutions to be protected and the results and data obtained during the research and development process, and analyzes in detail and depth how the technical solutions of the present application solve the technical problems and some creative technical effects brought about after solving the problems. The specific description is as follows:

[0078] The present application improves the effect of generating molecules with expected properties from three aspects. On the one hand, the GVAE model extracts molecular features based on graph representation learning to better learn the molecular structure and not easily lose information, and an effectiveness corrector is used to ensure the chemical effectiveness of the generated molecules. On the other hand, an active learning screening strategy combining Fisher information matrix and molecular diversity is proposed to construct a high-quality labeled dataset, and the performance of the property prediction model is improved to improve the molecular optimization capability. Finally, the LIMO proposes a reverse optimization method by changing the input to affect the output in the desired way, and the optimization process uses the property prediction model instead of the property calculation function, greatly improving the speed of generating molecules with optimized properties.

[0079] The multi-objective optimization molecule generation model combined with active learning provided by the present application improves the problem that the existing generation model for molecular optimization consumes too many resources and generates molecules slowly when optimizing expensive properties, and at the same time improves the multi-objective optimization, i.e. the success rate of discovering drug-like molecules with high affinity to target proteins, and quickly explores the chemical space. The present application uses a graph-based VAE model for molecule generation, effectively utilizing the structural information of the molecule. The present application realizes an active learning screening strategy based on Fisher information and K-MEANS++ algorithm, and constructs a high-quality dataset with rich information and diversity, and improves the ability of the model to optimize molecules by improving the performance of the property predictor.

[0080] In the experiments of the application, the improvement in both single-target property optimization and multi-target property optimization is found by testing on the ZINC250k dataset. As a molecule generation model, GAVE can better extract molecular structure features and use an effectiveness modifier to ensure the generation of 100% effective molecules. Taking multi-target optimization as an example, the success rate of optimization is improved by about 5%, the affinity of the generated optimized molecules is optimal, and the generation and optimization of molecules are the fastest. The high-quality dataset constructed by the active learning method combining Fisher information and molecular diversity effectively improves the performance of the property prediction model, and finally the LIMO reverse optimization method is used to improve the molecular optimization speed, which improves the success rate of generating molecules with expected properties in three aspects.

[0081] Secondly, from the perspective of the product or as a whole, the technical effects and advantages of the technical solution to be protected by the application are described as follows:

[0082] The application uses the reverse optimization method proposed by LIMO and combines it with the graph generation model, changes the input to affect the output in the desired way, and the optimization process does not require property calculation functions, which improves the generation speed. The multi-target optimization molecule generation method of the application combined with active learning has universality, only needs to provide the structure and position information of the target protein, and does not require more professional knowledge, and is suitable for affinity optimization problems of various target proteins. The application proposes an active learning screening strategy combining Fisher information and diversity, uses Fisher information to select data with large information quantity, which can make the model learn key features faster, help the model converge faster, improve the model training efficiency, and use molecular structure information for K-MEANS++ clustering to select diversified data, which can reduce data redundancy and improve the generalization ability and robustness of the model.

[0083] Thirdly, the creativity of the claims of the application is also reflected in the following important aspects:

[0084] (1) The expected income and commercial value of the technical solution of the application after transformation are:

[0085] The application adopts the reverse optimization technology to generate optimized molecules in the current molecule generation model for target optimization, so the application of the application can accelerate the new drug research and development process, thereby reducing the new drug research and development time and cost.

[0086] The application proposes an active learning query function to construct a high-quality dataset, which improves the model performance and optimization success rate; a multi-target optimization method is realized, which can consider multiple targets at the same time, such as binding affinity, SA, QED, etc. Therefore, the application of the application is closer to the real situation, which can improve the quality of new drugs, reduce the failure rate of drug research and development, and reduce the research and development cost.

[0087] (2) The technical solution of the application fills the technical gap in the industry at home and abroad:

[0088] The application first combines Fisher information and diversity screening based on molecular structure, considers data information and diversity to construct a high-quality data set, and applies the advanced strategy of active learning to the field of molecule generation and optimization. The strategy of the application is universal and can be applied to other classification tasks or regression tasks in the drug field that need to reduce the cost of labeling.

[0089] The application first applies LIMO reverse optimization technology to the graph generation model, uses the structure information of the molecule to ensure the quality of the generated initial molecule, which is more conducive to subsequent molecule optimization, and improves the success rate of multi-objective optimization. BRIEF DESCRIPTION OF DRAWINGS

[0090] Figure 1 is a flow chart of the active learning combined multi-objective optimization molecule generation method provided by the embodiment of the application;

[0091] Figure 2 is a principle diagram of the active learning combined multi-objective optimization molecule generation method provided by the embodiment of the application;

[0092] Figure 3 is a line diagram of the active learning combined multi-objective optimization molecule generation method provided by the embodiment of the application;

[0093] Figure 4 is a network structure diagram of the molecule generation model provided by the embodiment of the application;

[0094] Figure 5 is a network structure diagram of the property prediction model provided by the embodiment of the application. DETAILED DESCRIPTION

[0095] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0096] In view of the problems existing in the prior art, the application provides a multi-objective optimization molecule generation method and system combined with active learning, which will be described in detail below in combination with the drawings.

[0097] I. Explanation of the embodiment. In order to enable those skilled in the art to fully understand how the application is specifically implemented, this part is an explanation of the embodiment of the technical scheme of the claim.

[0098] As Figure 1 shown, the multi-objective optimization molecule generation method combined with active learning provided by the embodiment of the application comprises the following steps:

[0099] S101, a multi-objective optimization molecule generation model combined with active learning and a property prediction model are constructed, and the multi-objective optimization molecule generation model and the property prediction model are trained;

[0100] S102, after the training of the multi-objective optimization molecule generation model and the property prediction model is completed, the weights of the generation model encoder and the property prediction model are frozen;

[0101] S103, using N(0 m , I m ) randomly sampled from the latent space z is input to the decoder, and the output of the decoder is used as input to predict the property by using the property prediction model, wherein m represents the sequence length of the latent vector;

[0102] S104, the gradient descent optimization z of the latent space z is performed through the output of the prediction model, and the optimized z is input to the decoder to obtain a molecule graph vector and perform validity correction to obtain a final optimized molecule.

[0103] The embodiment of the application gives a set of property prediction models (g 1 , g 2 ,..., g k ) and corresponding weights (w1, w2,..., w k ) according to the set target, and uses the Adam optimizer to optimize the following losses according to the obtained prediction results:

[0104]

[0105] Wherein, f dec represents the decoder, g represents the property prediction model, and w represents the given weight value.

[0106] The validity corrector provided by the embodiment of the application uses the validity correction module of the MoFlow molecule generation model. Compared with the commonly used validity checker realized based on rejection sampling, the validity correction module of the MoFlow molecule generation model is used to ensure the validity of the molecule by performing valence constraint checking and correction on each atom of the generated molecule, and the influence of formal charge N + , S + , O + is considered.

[0107] The principle diagram of the multi-objective optimization molecule generation method combined with active learning provided by the embodiment of the application is shown in Figure 2 .

[0108] As a preferred embodiment, as shown in Figure 3 The multi-objective optimization molecule generation method combined with active learning provided by the embodiment of the present application comprises the following steps:

[0109] Step 1, obtain SMILES data and perform preprocessing to generate an adjacency matrix and an annotation matrix, and then combine them into a molecular graph vector matrix; input the molecular graph vector matrix into a GVAE model to train the generation model;

[0110] Step 2, after the training of the molecular generation model is completed, sample molecules from the latent space, correct the effectiveness of the generated molecules, and thus obtain legal molecules;

[0111] Step 3, construct an initial labeled data set according to the ZINC data set, and train a property prediction model;

[0112] Step 4, construct an unlabeled data set according to the ZINC data set and the generated legal molecules, process the data into a molecular graph vector matrix, and then input it into the trained prediction model to obtain a property prediction result;

[0113] Step 5, use the last layer parameters of the prediction model and the molecular structure information to screen high-quality unlabeled data by combining the Fisher information and the query function of the K-MEANS++ algorithm;

[0114] Step 6, add the data to the labeled data set after labeling, retrain the prediction model, and repeat steps 4-6 until the iteration stopping condition is reached;

[0115] Step 7, optimize the latent space using one or more trained predictors according to the optimization condition;

[0116] Step 8, input the optimized latent vector into the decoder to generate molecules, and perform effectiveness checking on the molecules to obtain the final optimized molecules.

[0117] As shown in Figure 4 The embodiment of the present application uses a VAE model structure to implement the molecular generation model. Specifically, the molecular generation model is trained by the following steps:

[0118] (1.1) Process SMILES data. After normalizing the SMILES in the ZINC250k dataset, convert it to a molecule. Use the RDkit package to build a graph representation matrix of the molecule, that is, an annotation matrix X representing atomic features and an adjacency matrix A representing bond features. The annotation matrix (size N*A) represents each row as a one-hot encoding of the atom. The adjacency matrix (N*N*B) represents whether there is a bond between each row and each column of atoms and the one-hot encoding of the bond. Then, perform dimension transformation and matrix splicing operations on the annotation matrix and the adjacency matrix to obtain a molecular graph vector M, with a size of N*(A+(N*B)).

[0119] where N represents the number of atoms (the maximum number of atoms in the ZINC250k dataset is 38), A represents the number of atomic types (including 9 atomic types plus a virtual atomic type 0), and B represents the number of bond types (the number is 4).

[0120] The step (1.1) provided by the embodiment of the application has the advantage that the molecular graph table method is used to encode the molecule, and the structural information of the molecule is fully utilized.

[0121] (1.2) Input the molecular graph vector M obtained in step (1.1) into an encoder to obtain a mean vector μ with a length of m and a standard deviation vector σ with a length of m.

[0122] (1.3) Obtain the mean vector μ and the standard deviation vector σ obtained in step (1.2), and sample ∈ from the standard normal distribution using the reparameterization trick, and obtain a latent variable sequence z with a length of m according to the formula z=μ+σ⊙∈.

[0123] (1.4) Input the latent variable z obtained in step (1.3) into a decoder for decoding to obtain a generated molecular graph vector

[0124] (1.5) Calculate the evidence lower bound (ELBO) loss function, and use the loss function to iteratively train the generative model, use the Adam optimizer, and finally obtain the trained molecular generative model.

[0125] The loss function Loss used in the embodiment of the application is:

[0126] Loss=E[logP(M|z)]-D KL [Q(z|M)||P(z)]

[0127] where the first half of the formula represents the reconstruction error (implemented using a cross-entropy loss function), and the second half represents the KL divergence; E represents an expected value, P(M|z) represents a prior distribution of the latent variable z, represents an encoder, Q(z|M) represents a posterior distribution of the latent variable z derived from the real molecular graph vector, and represents a decoder.

[0128] As shown in Figure 5 , the property prediction model provided by the embodiment of the present application comprises two parts of a feature extraction network and a fully connected network connected in sequence.

[0129] The specific structure of the feature extraction network is three linear layers and three RELU activation layers connected alternately. The input of the first linear layer is a molecular graph vector with a length of S, and the output is a feature vector with a length of H1; the input of the second linear layer is a feature vector with a length of H1, and the input is a feature vector with a length of H2; the input and output vectors of the third linear layer all have a length of H2, S is the length of the molecular graph vector (its length is 38*(10+4*38) = 6156), H1 and H2 are the lengths of the hidden layer vectors (H1 is 1024, and H2 is 512).

[0130] The formula of the RELU activation function is as follows:

[0131] RELU(x) = max(0, x)

[0132] The specific structure of the fully connected network is one Dropout layer and one linear layer. The random inactivation ratio of the Dropout is set to 0.1. The input of the linear layer is a feature vector with a length of H2, and the output is a property prediction result.

[0133] Specifically, the property prediction model provided by the embodiment of the present application is obtained by the following steps:

[0134] (2.1) Taking the ZINC250k dataset as an unlabeled sample pool P U , randomly extracting one thousand molecules as an initial training set D L , and randomly extracting one thousand molecules as an independent test set D Test , deleting the extracted molecules from the sample pool P U , calculating the molecular property label y true according to a given target type using different property calculation methods. Then, the molecules in the training data set D L are used to construct a molecular vector graph. The way of constructing the molecular graph vector in step (2.1) provided by the embodiment of the present application is exactly the same as that in step (1.1), and will not be repeated here.

[0135] Among them, the QED value and the synthetic feasibility SA score are calculated by using the built-in function of the RDkit package, and the binding affinity value with the target protein is calculated by using AutoDock-GPU.

[0136] (2.2) The molecular graph vector matrix of n*S obtained in step (2.1) is input into the feature extraction network to obtain an n*H feature vector matrix T. The feature vector matrix T is input into the full connection network to obtain the final predicted property label vector y, and the length of y is n. Wherein n represents the number of data.

[0137] (2.3) According to the predicted property label vector y obtained in step (2.2) and the real property label vector y true The loss is calculated, and the loss is used to iteratively train the generated model, and the Adam optimizer is used, and the learning rate is automatically updated by using the ReduceLROnPlateau function. Finally, the trained property prediction model is obtained.

[0138] The mean square error loss function is used in the embodiment of the application, and the formula is as follows:

[0139]

[0140] (2.4) From the sample pool P U pool_num molecules are randomly extracted, and five thousand legal molecules passing through the validity corrector are generated by using the trained generated model, and the two are combined to construct the unlabeled data set D U , and pool_num is set to 50000 in the case of training the QED or SA property prediction model, and is set to 5000 in the case of training the binding affinity property prediction model. The unlabeled data set D U is processed in the same way as step (1.1), and the molecular graph vector matrix M U is obtained. The trained property prediction model is used to predict the unlabeled data set, and the prediction result y and the parameter matrix θ L output by the penultimate layer neural network (i.e. the Dropout layer) of the prediction model are obtained, and the size is n*H2.

[0141] (2.5) The molecular graph vector matrix M U and the parameter matrix θ L obtained in step (2.4) are input into the active learning screening function. The active learning screening function realized in the application mainly includes two steps. First, oversampling is performed based on the Fisher information optimization of the boundary of the maximum likelihood estimation (MLE) error, from the data set D USelect 2*batch samples from D

[0142] The screening function provided by the embodiment of the present application specifically implements the following steps:

[0143] First, the parameter matrix θ L is used to calculate the Fisher information matrix I of the unlabeled data set D U and the training set D L , and the calculation formula is as follows:

[0144] Fisher matrix calculation:

[0145]

[0146]

[0147] F represents the selected Fisher matrix, and at the beginning (F0), only the training set is included.

[0148] Then, greedy oversampling is performed, and 2*batch samples are added to F, and the calculation formula is as follows:

[0149]

[0150]

[0151] wherein D S represents the unlabeled sample set selected by oversampling.

[0152] Finally, the k-MEANS++ seed algorithm is used to perform clustering operation according to the molecular graph vector matrix M U corresponding to the samples in D S , and batch cluster centers are selected as the finally selected samples.

[0153] The basic idea of the k-MEANS++ algorithm is to first randomly select a sample from the sample set as the first cluster center, and then sequentially sample and select k-1 new cluster centers from the sample set, wherein the probability of selecting each new cluster center is proportional to the square distance from the nearest cluster center.

[0154] The active learning screening strategy implemented by step (2.5) has the advantages that, first, the Fisher information is used to represent the information amount of data, so that data with rich information amount is screened out; second, the k-MEANS++ clustering algorithm is used to ensure the diversity of the screened data.

[0155] (2.6) Calculate the property label of the screened batch samples, and add them to the training set D L In addition, if the screened samples belong to the samples in the unlabeled sample pool P U , the samples are deleted from the sample pool; then, the property prediction model is retrained in the same way as steps (2.2) and (2.3).

[0156] (2.7) Repeat steps (2.4) to (2.6) until the stopping condition is reached, and obtain the finally trained property prediction model, wherein the stopping condition is set to 10 iterations.

[0157] The multi-objective optimization molecule generation system combined with active learning provided by the embodiment of the application comprises:

[0158] A model training module is configured to construct a multi-objective optimization molecule generation model combined with active learning and a property prediction model, and train the multi-objective optimization molecule generation model and the property prediction model.

[0159] A model weight freezing module is configured to freeze the weights of the generation model encoder and the property prediction model after the training of the multi-objective optimization molecule generation model and the property prediction model is completed.

[0160] A property prediction module is configured to use N(0 m , I m ) to randomly sample from the latent space z and input it to the decoder, and use the property prediction model to predict the property by taking the output of the decoder as input.

[0161] A molecule optimization module is configured to perform gradient descent optimization of the latent space z by the output of the prediction model, input the optimized z into the decoder to obtain a molecule graph vector, and perform validity correction to obtain a final optimized molecule.

[0162] II. Application Embodiment. In order to prove the creativity and technical value of the technical scheme of the application, this part is an application embodiment of the technical scheme of the claims on a specific product or related technology.

[0163] The molecule generation model can be applied to new drug discovery and development in the field of drug research and development. Generally, drug research and development needs a large amount of molecular screening and binding affinity optimization, which is a long and tedious process. However, using the molecule generation model can generate molecules with good binding affinity more quickly and efficiently.

[0164] For example, consider the development of a new cancer treatment drug that needs to have a specific structure and biological activity to work in target cells. Using the present application, only a corresponding binding affinity prediction model needs to be trained according to the target cells, and then a plurality of property prediction models are combined and applied according to requirements for reverse optimization, so that a large number of drug-like molecules with good binding affinity can be generated in a short time. Once the best candidate drugs are determined on the computer, they can be synthesized and tested in the laboratory to verify their efficacy and safety. This method can greatly shorten the time of drug discovery and bring new treatments to patients more quickly. In addition, using the present application can also reduce the funds and resources required to develop new drugs, thereby bringing commercial value to pharmaceutical companies and medical institutions.

[0165] III. Evidence of the effects of the embodiments. The embodiments of the present application have achieved some positive effects in the process of research and development or use, and indeed have great advantages compared with the prior art. The following content is described in combination with data, graphs and the like in the test process.

[0166] The experimental environment of the present application: in the Ubuntu 18.04 operating system, the CPU is Intel(R) Core(TM) i9-7900X, the GPU is 2 NVIDIA 1080Ti 12GB, the memory is 64GB DDR4, and the algorithm of the present application is realized by using Pytorch programming. The specific settings are as follows: the batch size is 128, the learning rate of the generation model is 5e-5, the epoch is 300, the initial learning rate of the property prediction model is 0.1, the epoch is 100, the initial learning rate of the reverse optimization module is 0.1, and the epoch is 10.

[0167] The specific evaluation indexes are as follows:

[0168] Success rate: the percentage of generated molecules meeting the given target (single target is dissociation constant K D less than 10 3 nmol / l, and multiple targets are K D less than 10 3 nmol / l and QED>0.5 and SA<5.5).

[0169] Novelty rate: the percentage of molecules with a similarity of less than 0.4 compared with the nearest neighbor molecules in the training set that meet the requirements of the given target.

[0170] Diversity: the diversity of molecules that meet the requirements of the given target, calculated by the valley coefficient on the Morgan fingerprint.

[0171] PM: the product of the above three indexes, is a comprehensive evaluation index of the proposed method.

[0172] Time: the time required to generate molecules and perform target optimization. The value is in the unit of the fastest generation speed.

[0173] Best affinity: the dissociation constant K D of the molecule with the highest affinity to the target protein in the generated molecules, in nmol / l.

[0174] R 2 : the coefficient of determination, the value is between 0 and 1, the larger the value, the better the model fitting effect.

[0175] RMSE: root mean square error, the smaller the value, the better the model fitting effect.

[0176] In order to illustrate the effectiveness of the method of the present application and the improvement of molecular optimization, a plurality of optimization tests of different targets were carried out, and were compared with the current advanced molecular optimization model. In the single target experiment, one million molecules were generated by each method, and the best one thousand molecules were selected for comparison; in the multi-target experiment, forty thousand molecules were generated by each method, and the best five thousand molecules were selected for comparison. The single target optimization results of optimizing the binding affinity with human estrogen receptor (ESR1) are shown in Table 1, and the multi-target optimization results of optimizing the binding affinity with ESR1, drug similarity (QED) and syntheticity (SA) score are shown in Table 2.

[0177] Table 1 single target optimization results

[0178]

[0179] Table 2 multi-target optimization results

[0180]

[0181] Table 3 model performance test experiment results

[0182]

[0183] According to the single target optimization experiment results recorded in Table 1 and the multi-target experiment results recorded in Table 2, it can be seen that the multi-target optimization molecule generation model combining active learning proposed in the present application is superior to the existing method in PM value, required time and best affinity value.

[0184] In addition, Table 2 shows the results of multi-target optimization experiments of the property prediction model trained on the randomly sampled data set and the property prediction model trained on the data set constructed by the active learning method proposed in the present application, and Table 3 shows the results of the independent test set D TestThe results of the model performance test experiment. It can be seen that the active learning method is better than random sampling, and is more beneficial to the improvement of model performance and target optimization.

[0185] The multi-objective optimization molecular generation model combined with active learning provided by the embodiment of the application can better learn the molecular structure information by extracting the molecular features based on the graph representation learning of the GVAE model. The chemical effectiveness of the generated molecules is ensured by the effectiveness correction model. On the other hand, the active learning screening strategy combining the Fisher information and the molecular diversity is proposed, a diversified high-quality labeled data set is constructed, the performance of the property prediction model is improved, and the molecular optimization capability is improved. Finally, the LIMO is used to propose the reverse optimization method, the input is changed to affect the output in the expected way, the optimization process does not need the property calculation function, and the speed of generating molecules with optimized properties is greatly improved.

[0186] It should be noted that the embodiments of the application can be realized by hardware, software or a combination of software and hardware. The hardware part can be realized by special logic; the software part can be stored in the memory and executed by the appropriate instruction execution system, such as microprocessor or special designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control code, such as provided on carrier media such as magnetic disk, CD or DVD-ROM, programmable memory such as read-only memory (firmware) or data carrier such as optical or electronic signal carrier. The devices of the application and their modules can be realized by hardware circuit such as ultra-large scale integrated circuit or gate array, semiconductor such as logic chip, transistor, or programmable hardware device such as field programmable gate array, programmable logic device, etc., can also be realized by software executed by various types of processors, and can also be realized by the combination of the above hardware circuit and software, such as firmware.

[0187] Those skilled in the art can easily understand that the above only describes the preferred embodiments of the application and is not intended to limit the application. Any modification, equivalent replacement and improvement made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method for multi-objective optimization molecular generation combined with active learning, characterized in that, Comprise: Constructing and training the multi-objective optimization molecule generation model and property prediction model combined with active learning; After training the molecular generation model and the property prediction model, the weights of the generation model encoder and the property prediction model are frozen; then... From hidden space After random sampling, the data is input into the decoder. The prediction model uses the decoder's output as input to predict properties. The molecular generation model performs predictions relative to the latent space using the output of the prediction model. gradient descent optimization The optimized z is then input back into the decoder to obtain the molecular graph vector. Finally, the final optimized molecule is obtained through validity correction. The property prediction model is trained through the following steps: (1) ZINC250k dataset as an unlabeled sample pool , randomly select one thousand molecules as the initial training set , randomly select one thousand molecules as an independent test set , delete the extracted molecules from the sample pool ; calculate the molecular property label according to the given target type using different property calculation methods , and use the molecules in the training data set to build a molecular vector diagram; Among them, regarding the calculation of molecular properties, the QED value and the synthesis feasibility SA score are calculated by the built-in function of the RDkit package, and the binding affinity value with the target protein is calculated by AutoDock-GPU; (2) Input the molecular graph vector matrix of n*S obtained in step (1) into the feature extraction network to obtain an n*H feature vector matrix T; Input the feature vector matrix T into the fully connected network to obtain the final predicted property label vector y, which has a length of n, and n represents the number of data; (3) the predictive property label vector y and the real property label vector y obtained according to step (2) Calculate the loss, and train the generation model iteratively using the loss. Using Adam optimizer and ReduceLROnPlateau function to automatically update the learning rate, and finally obtaining the trained property prediction model; Among them, the formula of the mean square error loss function is as follows: ， (4) from the sample pool Randomly extract pool_num molecules, and use the trained generation model to generate five thousand legal molecules passing through the validity corrector, and merge the molecules to construct an unlabeled data set , pool_num size is set to 50000 in the case of QED or SA property prediction model training, and is set to 5000 in the case of binding affinity property prediction model training; data processing is performed on the unlabeled data set , and a molecular graph vector matrix is obtained ; using the trained property prediction model to predict the unlabeled data set, obtaining the prediction result y and the parameters output by the last second layer neural network Dropout layer of the prediction model , size n*H2; (5) input the molecular graph vector matrix obtained in step (4) and the parameter matrix into an active learning screening function; wherein the active learning screening function comprises: oversampling based on the boundary of the Fisher information optimized maximum likelihood estimation error to select 2*batch samples from the data set ; using a k-MEANS++ seed algorithm to cluster the 2*batch samples and screen out batch samples; the batch size is set to 10000 in the case of training a QED or SA property prediction model, and is set to 1000 in the case of training a binding affinity property prediction model.​​ (6) The batch of screened samples are calculated for property labels and added to the training set If the screened sample belongs to the sample in the unlabeled sample pool in the middle, the unlabeled sample is deleted from the sample pool, and the property prediction model is retrained. (7) Repeat steps (4) to (6) until the stopping condition is reached to obtain the final trained property prediction model, and the stopping condition is set to repeat 10 times.

2. The multi-objective optimization molecule generation method incorporating active learning of claim 1, wherein, Molecular generative models perform operations relative to the latent space by predicting the model's output. gradient descent optimization Includes: providing a set of property prediction models based on the set objectives. and corresponding weights Based on the obtained prediction results, the Adam optimizer is used to optimize the following loss: ， wherein denotes a decoder, denotes a property prediction model, denotes a given weight value.

3. The multi-objective optimization molecule generation method combined with active learning according to claim 1, characterized in that, The property prediction model comprises two parts of feature extraction network and fully connected network connected in turn; Among them, the structure of the feature extraction network is three linear layers and three RELU activation layers connected alternately, the input of the first linear layer is a molecular graph vector with a length of S, and the output is a feature vector with a length of H1; The input of the second linear layer is a feature vector with a length of H1, and the input is a feature vector with a length of H2; The input and output vectors of the third linear layer have a length of H2; S is the length of the molecular graph vector, which is 38*(10+4*38)=6156; H1 and H2 are the lengths of the hidden layers, H1 is 1024, and H2 is 512; The formula of the RELU activation function is as follows: ； The structure of the fully connected network is one layer of Dropout layer and one layer of linear layer, and the random inactivation proportion of Dropout is set to 0.1; The input of the linear layer is a feature vector with a length of H2, and the output is a property prediction result.

4. The multi-objective optimization molecule generation method incorporating active learning of claim 3, wherein, The implementation process of the screening function includes: 1) Using a parameter matrix , the Fisher information matrices I of the unlabeled dataset and the training set are calculated respectively, and the calculation formula is as follows: Fisher matrix calculation: ； ； ； where F denotes the selected Fisher matrix, at the initial time contains only the training set; 2) Greedy oversampling is performed, and 2*batch samples are added to F, and the calculation formula is as follows: ； ； In the formula, denotes the set of unlabelled samples selected by oversampling; 3) using k-MEANS++ seed algorithm, according to the molecular graph vector matrix corresponding to the sample perform clustering operation, and screen out batch cluster centers as final screened samples; Use the k-MEANS++ algorithm to randomly select a sample from the sample set as the first cluster center, and then sequentially sample k-1 new cluster centers from the sample set, where the probability of selecting each new cluster center is proportional to the square distance to the nearest cluster center.

5. A method of multi-objective optimization molecular generation combined with active learning, characterized in that, Comprise the following steps: Step one, obtain SMILES data and preprocess to generate an adjacency matrix and an annotation matrix, then merge into a molecular graph vector matrix; Input the molecular graph vector matrix into the GVAE model to train the generation model; The molecular generation model is realized by using the VAE model structure, and the molecular generation model is trained by the following steps: (1) Process SMILES data: after normalization operation on SMILES in ZINC250k dataset, convert to molecule; use RDkit package to construct graph representation matrix of molecule, annotation matrix X representing atomic features and adjacency matrix A representing bond features; use annotation matrix of size N*A to represent each row as one-hot encoding of atom, and use adjacency matrix of size N*N*B to represent whether there is a bond between atoms in each row and each column and the one-hot encoding of the bond; Perform dimension transformation and matrix splicing operation on the annotation matrix and the adjacency matrix to obtain the molecular graph vector M, which has a size of N*(A+(N*B)); wherein N represents the number of atoms, and the maximum number of atoms in the ZINC250k dataset is 38; A represents the number of atomic types, including 9 atomic types plus a virtual atomic type 0; B represents the number of bond types, which is 4; (2) Input the molecular graph vector M obtained in step (1) into the encoder to obtain a mean vector μ of length m and a standard deviation vector σ of length m; (3) Obtain the mean vector μ and the standard deviation vector σ from step (2), and sample e from the standard normal distribution N(0, 1) using the reparameterization trick according to the formula Obtain the latent variable sequence z of length m; (4) input the latent variable z obtained in step (3) into the decoder for decoding to obtain a generated molecular graph vector ; (5) Calculate the lower bound loss function under the evidence, and use the loss function to iteratively train the generation model; use the Adam optimizer to obtain the trained molecular generation model; Wherein, the loss function Loss is: ； In the formula, the first half of the Loss formula represents the reconstruction error, which is implemented using a cross-entropy loss function; the second half represents the KL divergence; E represents an expected value; represents a prior distribution of the latent variable z, representing an encoder; represents a posterior distribution of the latent variable z derived according to a real molecule graph vector, representing a decoder; Step two, after the training of the molecular generation model is completed, sample molecules from the latent space, correct the effectiveness of the generated molecules to obtain legal molecules; Step three, construct an initial labeled dataset according to the ZINC dataset, and train the property prediction model; Step four, construct an unlabeled dataset according to the ZINC dataset and the generated legal molecules, process the data into a molecular graph vector matrix, and input it into the trained prediction model to obtain the property prediction result; Step five, combine the Fisher information and the query function of the K-MEANS++ algorithm to filter using the last layer parameters of the prediction model and the molecular structure information to obtain high-quality unlabeled data; Step six, after labeling the data, add the data to the labeled dataset, retrain the prediction model, and repeat steps four to six until the iteration stopping condition is reached; Step seven, optimize the latent space using one or more trained predictors according to the optimization condition; Step eight, input the optimized latent vector into the decoder to generate molecules, and perform effectiveness check on the molecules to obtain the final optimized molecules.

6. A multi-objective optimization molecule generation system incorporating active learning, characterized by, Comprise: The first module is used for obtaining SMILES data and pre-processing to generate an adjacency matrix and an annotation matrix, and then merging into a molecular graph vector matrix; input the molecular graph vector matrix into the GVAE model to train the generation model; use the VAE model structure to realize the molecular generation model, which is trained by the following steps: (1) Processing SMILES data: After normalizing the SMILES in the ZINC250k dataset, the SMILES are converted into molecules; the RDkit package is used to construct the graph representation matrix of the molecules, the annotation matrix X representing the characteristics of the atoms and the adjacency matrix A representing the characteristics of the bonds; the size of the annotation matrix is N*A, and each row represents the one-hot encoding of the atoms; the size of the adjacency matrix is N*N*B, and each row and column represents whether there is a bond between the atoms and the one-hot encoding of the bond; The annotation matrix and the adjacency matrix are subjected to dimension transformation and matrix splicing operation to obtain the molecular graph vector M, the size of which is N*(A+(N*B)); wherein N represents the number of atoms, the maximum number of atoms in the ZINC250k dataset is 38; A represents the number of atom types, including 9 atom types plus a virtual atom type 0; B represents the number of bond types, the number of which is 4; (2) The molecular graph vector M obtained in step (1) is input into the encoder to obtain a mean vector μ of length m and a standard deviation vector σ of length m; (3) Obtain the mean vector μ and the standard deviation vector σ from step (2), and sample ϵ from the standard normal distribution N(0, 1) using the reparameterization trick according to the formula Obtain the latent variable sequence z of length m; (4) input the latent variable z obtained in step (3) into the decoder for decoding to obtain a generated molecular graph vector ; (5) Calculate the evidence lower bound loss function, and use the loss function to iteratively train the generation model; use the Adam optimizer, and finally obtain the trained molecular generation model; Wherein, the loss function Loss is: ； In the formula, the first half of the Loss formula represents the reconstruction error, which is implemented using a cross-entropy loss function; the second half represents the KL divergence; E represents an expected value; represents a prior distribution of the latent variable z, representing an encoder; represents a posterior distribution of the latent variable z derived according to a real molecular graph vector, representing a decoder; The second module is used to sample molecules from the latent space after the training of the molecular generation model is completed, and to correct the effectiveness of the generated molecules to obtain legal molecules; The third module is used to construct an initial labeled dataset according to the ZINC dataset, and train the property prediction model; The fourth module is used to construct an unlabeled dataset according to the ZINC dataset and the generated legal molecules, and to process the data into a molecular graph vector matrix and input it into the trained prediction model to obtain the property prediction result; The fifth module is used to combine the Fisher information and the K-MEANS++ algorithm query function to filter the last layer parameters of the prediction model and the molecular structure information to obtain high-quality unlabeled data; The sixth module is used to add the labeled data to the labeled dataset, retrain the prediction model, and repeat the fourth module to the sixth module until the iteration stopping condition is reached; The seventh module is used to optimize the latent space using one or more trained predictors according to the optimization condition; The eighth module is used to input the optimized latent vector into the decoder to generate molecules, and to perform effectiveness checking on the molecules to obtain the final optimized molecules.

Images