High-synthesizability molecule generation method and device based on multi-precision feedback driving

Abstract

The invention discloses a high-synthesizability molecule generation method and device based on multi-precision feedback driving, and the method comprises the following steps: constructing a hierarchical strategy network, selecting an available reaction in a current state through a reaction selection strategy network, and generating a high-synthesizability molecule; using the reactant selection strategy network to select available reactants in the current state and carrying out chemical reaction according to the reaction rule; constructing a multi-precision active learning framework, obtaining the generation molecules and the selected precision labels by using a hierarchical strategy network, inputting the generation molecules and the selected precision labels into a multi-precision substitution model, and scoring each generation molecule by using the multi-precision substitution model to obtain a multi-precision active learning framework; and giving a comprehensive score and feeding back the score to the hierarchical strategy network for gradient updating by using the acquisition function after considering the use cost required by the scoring tools corresponding to the precision labels of different molecules. The invention provides a more efficient, economic and reliable molecular design scheme, the discovery and optimization process of new molecular entities is accelerated, and technical progress and innovative development in related fields are promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of molecule generation, and in particular relates to a method and device for generating highly synthesizable molecules based on multi-precision feedback drive. Background Art

[0002] Inspired by combinatorial chemistry, some molecular design methods that combine combinatorial chemistry with molecular generation have emerged in recent years, which involves assembling molecules in the form of chemical reactions, combining selectable reactants through constrained chemical reaction rules. Due to the unique properties of reaction-based assembly methods, it gradually stands out from fragment-based methods and becomes a new basic method for molecular assembly (reaction-based methods). In recent years, several related algorithms have emerged, such as SynNet, PGFS, and SyntheMol. Among them, SyntheMol is currently the most cutting-edge and best-performing algorithm for molecular generation with high synthesizability.

[0003] SyntheMol uses the Monte Carlo Tree Search (MCTS) algorithm to identify high-scoring molecules by searching synthetic trees under the guidance of a property prediction model. During each MCTS expansion, SyntheMol constructs molecules by selecting building blocks and combining them through chemical reactions. The generated molecules are evaluated at their corresponding nodes to provide feedback to the MCTS algorithm. The node score includes four factors: molecular property prediction value P(N), development factor Q(N), exploration factor U(N), and building block diversity factor D(N). However, SyntheMol has three key problems:

[0004] 1) In the MCTS framework, each sample of generated molecules needs to perform a complete rollout operation, that is, starting from the initial node, gradually expanding and evaluating each branch of the search tree until the complete molecule is constructed. This means that if you want to sample 10,000 molecules, you need to perform 10,000 independent rollout operations. Although this method can effectively explore the chemical space, the rollout process involves a large number of tree searches and node evaluations, which has a high computational cost and significantly reduces the sampling speed, especially when the chemical space is large and the synthesis path is complex. Therefore, when generating large-scale molecular libraries, the efficiency of MCTS sampling is limited by the computational burden of each rollout, resulting in a very slow overall sampling process. It is not suitable for real-world application scenarios that require large-scale sampling of millions of molecules. This problem also exists in the SynNet model;

[0005] 2) Considering the complexity of molecular design (drug / material field) and the long cycle of R&D, it is not realistic to rely on the current generation model to find the so-called "optimal" molecule. Improving the diversity of sampled molecules is like not putting all your eggs in one basket, which is an effective way to reduce R&D risks. However, the existing generation algorithm has no sampling diversity advantage from the underlying algorithm logic, and SyntheMol also has this problem.

[0006] 3) SyntheMol and all iterative feedback generation models need to call the scoring tool every time they generate a sample of molecules during training. When a high-precision scoring tool is selected, its computational cost is also more expensive, which inevitably leads to a significant increase in training time costs. Part of the reason for this phenomenon is that the current quantitative structure-activity relationship (QSAR) model built based on deep learning methods has a high prediction accuracy, and its call cost is much lower than that of traditional docking scoring. When there are enough representative data as training sets, deep learning models can usually perform well. However, in real scenarios, molecular design tasks often face the dilemma of sparse small molecule label data. For the problem of data sparsity, active learning algorithms are a better solution. The key to its technology lies in how to effectively sample unlabeled data and maximize the prediction accuracy and generalization ability of the scoring model. Although there are many different sampling strategies, these methods are all based on the existing unlabeled data pool for sampling. Compared with the huge chemical space, the limited unlabeled data pool means that the structural diversity of the sampled molecules is limited, which will also affect the generalization ability of the scoring model to a certain extent.

[0007] In summary, the molecular generation algorithm based on chemical reaction splicing still has obvious problems in sampling efficiency, sampling diversity and training cost that need to be solved urgently. Summary of the invention

[0008] In view of the above, the purpose of the present invention is to provide a method and device for generating highly synthesizable molecules based on multi-precision feedback drive, so as to solve the problem of the difficulty of implementing the generation algorithm in the molecular (drug / material field) design scenario from the perspective of molecular synthesizability, improve the sampling diversity of the generation algorithm and improve the exploration coverage in the chemical space from the perspective of the underlying algorithm logic, and reduce the cost of calling expensive high-precision scoring tools by the generation model from the perspective of real application scenarios, so as to provide an efficient, economical and reliable molecular design solution for the fields of drug research and development and advanced material development, accelerate the discovery and optimization process of new molecular entities, and promote technological progress and innovative development in related fields.

[0009] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention is as follows:

[0010] In a first aspect, an embodiment of the present invention provides a method for generating highly synthesizable molecules based on multi-precision feedback drive, comprising the following steps:

[0011] Using a hierarchical strategy network including a reaction selection strategy network and a reactant selection strategy network to generate highly synthesizable molecules, including: using the reaction selection strategy network to select an available reaction in a current state according to an initial molecular building block or an intermediate product in a previous state, using the reactant selection strategy network to select an available reactant in a current state based on the selected available reaction and performing a chemical reaction according to a reaction rule of the selected available reaction to obtain an intermediate product or a final product;

[0012] A multi-precision active learning framework including a hierarchical strategy network, a Gaussian process as a multi-precision substitution model and a multi-precision maximum entropy search as an acquisition function is used to provide the required precision labels of the generated molecules while generating molecules, including: using a hierarchical strategy network to obtain the generated molecules and the selected precision labels and input them into the multi-precision substitution model, using the multi-precision substitution model to score each generated molecule, using the acquisition function to give a comprehensive score after considering the usage cost of the scoring tool corresponding to the precision labels of different molecules and feeding it back to the reaction selection strategy network and the reactant selection strategy network for gradient update.

[0013] Preferably, the method further comprises:

[0014] Construct labeled data including reactants and their valid reaction labels, and use the labeled data, reactant dataset and reaction dataset to pre-train the reactant selection strategy network for multi-label classification tasks. The output classification results are the valid reactants and invalid reactants corresponding to each reaction, so that the reactant selection strategy network can pre-learn to identify optional reactants that match the selected available reactions under different reaction labels and pre-screen invalid reactants through pre-training.

[0015] Preferably, the molecular generation process is defined as a Markov decision process in which reactants react step by step based on reaction rules to obtain the final product molecules, and the generation flow network model is used to implement this process. The generation flow network defines the molecular generation process as a directed acyclic graph G = (S, A), where S represents the state space and A represents the action space. The vertex s0∈S is defined as the initial state with no incoming edges and corresponds to no initial molecule or a specific reactant. The vertex s is defined. f ∈S is the final state without outgoing edges and corresponds to the final generated molecule. The state sequence τ from the initial state to a certain terminal state is defined as (s0→…→s n ) is a trajectory, and the molecular object x∈X realizes the forward state transfer s through sequential sampling actions t →s t+1, X represents the set of all molecules in the combinatorial chemical space, and the state transfer is to add new molecular fragments to the currently partially constructed molecules through chemical reactions;

[0016] In the hierarchical policy network, the state transfer s is based on the generation flow network t →s t+1 The action is divided into two steps. The first step is to select the available reactions in the current state based on the given reaction data set. The second step is to select the available reactants based on the given reactant library and use the reaction rules of the available reactions selected in the first step to perform chemical reactions. In the two steps, the reaction selection probability model is constructed respectively. The reaction selection strategy network is used to select reactions and build a reactant selection probability model The reactant selection strategy network is used to select reactants, and the reaction selection strategy network and the reactant selection strategy network are trained to sample molecular objects with a probability π(x) proportional to the reward value R(x), that is, π(x)∝R(x), to generate diverse high reward value molecular objects.

[0017] Preferably, the use of a reaction selection strategy network to select available reactions in the current state according to initial molecular building blocks or intermediate products in the previous state comprises:

[0018] When performing reaction selection, the initial molecular building blocks or the intermediate products of the previous state are input into the reaction selection strategy network to obtain the original probability values ​​of all reactions. The classification probability distribution is constructed by shielding invalid reactions and normalizing the remaining valid reactions. Then, random sampling is performed from the classification probability distribution to obtain the available reactions in the current state.

[0019] Preferably, the method of using a reactant selection strategy network to select available reactants in a current state based on available reactions and performing chemical reactions according to reaction rules of available reactions to obtain intermediate products or final products includes:

[0020] When selecting reactants, the available reactions in the current state obtained by the reaction selection strategy network are input into the reactant selection strategy network to obtain the original probability values ​​of all reactants. The classification probability distribution is constructed by shielding invalid reactants and normalizing all remaining valid reactants. Then, random sampling is performed from the classification probability distribution to obtain the available reactants in the current state. Chemical reactions are carried out according to the reaction rules of the available reactions to obtain intermediate products or final products. After the chemical reaction is completed, it is converted to the next state.

[0021] Preferably, the Gaussian process is modeled as a multi-precision surrogate model, wherein the kernel function adopts a multi-precision kernel.

[0022] In a second aspect, to achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a highly synthesizable molecule generation device based on multi-precision feedback drive, which is implemented by the above-mentioned highly synthesizable molecule generation method based on multi-precision feedback drive, and includes: a hierarchical strategy network module and a multi-precision feedback drive module;

[0023] The hierarchical strategy network module is used to generate highly synthesizable molecules using a hierarchical strategy network including a reaction selection strategy network and a reactant selection strategy network, including: using the reaction selection strategy network to select an available reaction in the current state according to an initial molecular building block or an intermediate product in a previous state, using the reactant selection strategy network to select an available reactant in the current state based on the selected available reaction and performing a chemical reaction according to the reaction rule of the selected available reaction to obtain an intermediate product or a final product;

[0024] The multi-precision feedback driving module is used to use a multi-precision active learning framework including a hierarchical strategy network, a Gaussian process as a multi-precision substitution model and a multi-precision maximum entropy search as an acquisition function to provide the required precision labels of the generated molecules while generating molecules, including: using the hierarchical strategy network to obtain the generated molecules and the selected precision labels and input them into the multi-precision substitution model, using the multi-precision substitution model to score each generated molecule, using the acquisition function to give a comprehensive score after considering the usage cost of the scoring tool corresponding to the precision labels of different molecules and feeding it back to the reaction selection strategy network and the reactant selection strategy network for gradient update.

[0025] In the third aspect, in order to achieve the above-mentioned purpose of the invention, an embodiment of the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to implement the above-mentioned method for generating highly synthesizable molecules based on multi-precision feedback drive when executing the computer program.

[0026] In a fourth aspect, in order to achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a computer, the above-mentioned method for generating highly synthesizable molecules based on multi-precision feedback drive is implemented.

[0027] In a fifth aspect, in order to achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a computer product, which includes a computer program. When the computer program is executed by a processor, it implements the above-mentioned method for generating highly synthesizable molecules based on multi-precision feedback drive.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] (1) The present invention uses a generative flow network as a basis and combines it with a Markov decision process to obtain an improved hierarchical strategy network for molecule generation. The generative flow network algorithm is used to assign rewards corresponding to the generated molecules to each trajectory in the probability flow network, ensuring that the generation probability of different molecules is proportional to their reward values, thereby creating a distribution related to feedback and achieving independent and identically distributed sampling from it. Starting from the underlying algorithm logic, the diversity of molecular sampling is improved, and the exploration coverage in the chemical space is improved.

[0030] (2) The present invention introduces a pre-training strategy to pre-train the reactant selection strategy network, which can effectively solve the problem of excessive dimension of discrete action space, so that the reactant selection strategy network can focus more on establishing a reasonable probability distribution for selecting effective reactants, thereby improving computational efficiency.

[0031] (3) The present invention introduces a multi-precision active learning framework, which is based on the active learning framework, uses Gaussian process regression as a multi-precision alternative model, and uses maximum entropy search as an acquisition function to train a hierarchical strategy network. Different scoring tools are selectively sampled by different molecules for scoring. While ensuring the overall scoring accuracy, the number of calls to high-precision scoring tools is reduced as much as possible, thereby reducing the overall training cost and ultimately achieving a balance between exploration, utilization, and cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 It is a schematic flow chart of a method for generating highly synthesizable molecules based on multi-precision feedback drive provided by an embodiment of the present invention;

[0034] Figure 2 It is a schematic diagram of the algorithm flow of the hierarchical strategy network provided by an embodiment of the present invention;

[0035] Figure 3 It is a schematic diagram of the algorithm flow of the multi-precision active learning framework provided by an embodiment of the present invention;

[0036] Figure 4 It is a schematic diagram of the algorithm framework of the hierarchical strategy network provided by an embodiment of the present invention;

[0037] Figure 5 It is a schematic diagram of the structure of a highly synthesizable molecule generation device based on multi-precision feedback drive provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0038] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0039] The inventive concept of the present invention is: in response to the problems of insufficient sampling efficiency and sampling diversity and high training cost of the molecule generation algorithm in the prior art, the embodiment of the present invention provides a method and device for generating highly synthesizable molecules based on multi-precision feedback drive, constructs a hierarchical strategy network based on a generation flow network and a Markov decision process, and generates molecules through two steps of reaction selection and reactant selection, thereby ensuring that the generated molecules are synthesizable and improving the diversity of molecular sampling, further improving the computing efficiency through a pre-training strategy, and reducing the overall training cost by introducing a multi-precision active learning framework, thereby providing an efficient, economical and reliable molecular design solution for the fields of drug research and development and advanced material development.

[0040] Figure 1 Schematic diagram of a method for generating highly synthesizable molecules based on multi-precision feedback drive provided by an embodiment of the present invention. Figure 1 As shown, the embodiment provides a method for generating highly synthesizable molecules based on multi-precision feedback drive, comprising the following steps:

[0041] S1, using a hierarchical strategy network including a reaction selection strategy network and a reactant selection strategy network to generate highly synthesizable molecules, including: using the reaction selection strategy network to select an available reaction in the current state based on an initial molecular building block or an intermediate product in a previous state, using the reactant selection strategy network to select an available reactant in the current state based on the selected available reaction and performing a chemical reaction according to the reaction rules of the selected available reaction to obtain an intermediate product or a final product.

[0042] In the embodiment, the constructed hierarchical strategy network is called SynGFN, which includes a reaction selection strategy network (strategy model 1) and a reactant selection strategy network (strategy model 2). SynGFN models molecular design as a series of cascade processes that simulate chemical reactions, gradually assembling novel molecules with synthesizable building blocks. SynGFN expresses goal-oriented molecular discovery as a Markov decision process that approximately satisfies a careful balance, thereby sampling a diverse set of high-performance molecules, each with a suggested synthetic route. During design, SynGFN faces two computational obstacles that cannot be ignored: (1) sampling from a probability distribution that is practically intractable; (2) evaluating reward functions (such as biological activity) is usually very expensive. SynGFN efficiently learns sampling by jointly adopting a custom pre-trained hierarchical strategy network, and uses a multi-precision active learning framework to strike a balance between accuracy and reward evaluation costs.

[0043] The underlying algorithm framework of SynGFN adopts the generative flow network (GFlowNet). Since the existing generative flow network does not add reasonable physical and chemical constraints, it is difficult to ensure that the generated molecules are synthesizable. SynGFN models molecular generation as a Markov decision process, in which reactants react in the form of a synthetic tree through sequential reactions (i.e., reaction templates) and finally reach the target molecule. In view of the complexity of assembling molecules through chemical reactions, SynGFN has designed a mechanism for executing hierarchical actions (decomposing each state transition in the Markov decision process into the selection of two sequential actions, first determining the reaction template and then selecting the corresponding appropriate reactants). The selection of hierarchical actions is guided by a hierarchical strategy network, that is, SynGFN contains strategy networks for reaction selection and reactant selection respectively. Chemical synthesis constraints are explicitly introduced in the design of SynGFN: during the generation process, only reactions and reactants that match the current state have a probability of being selected, and when invalid product molecules are generated, the reactions and reactants added in the current generation step will not be included in the generation process in the end. Figure 2As shown in a in , when selecting a reaction, the initial molecular building blocks or the intermediate products of the previous state are input into the reaction selection strategy network to obtain the original probability values ​​of all reactions, and the categorical probability distribution (specifically using softmax distribution) is constructed by shielding invalid reactions and normalizing the remaining valid reactions, and then random sampling is performed from the categorical probability distribution, and the temperature coefficient is introduced to adjust the probability distribution and the randomness of the sampling to obtain the available reactions in the current state. Once the appropriate reaction is selected, the reactant selection strategy network will be used to select the reactants compatible with the reactions of the subsequent reactions. When selecting reactants, the available reactions in the current state obtained by the reaction selection strategy network are input into the reactant selection strategy network to obtain the original probability values ​​of all reactants, and the categorical probability distribution is constructed by shielding invalid reactions and normalizing the remaining valid reactions, and then random sampling is performed from the categorical probability distribution (specifically using softmax distribution) to obtain the available reactants in the current state, and the chemical reaction is performed according to the reaction rules of the available reactions to obtain the intermediate products or final products, and the next state is converted after the chemical reaction is completed. If the layering is not performed, the action selected by the policy network needs to specify both the reactant and the reaction based on the current state. A natural approach is to design the action as reactant + reaction label, which will lead to the introduction of a large number of unnecessary dimensions. From the perspective of time complexity, the time complexity of a single-layer action network is O(n1*n2), while the time complexity of a layered action network becomes O(n1+n2). Therefore, the design of a layered action network can greatly reduce computational complexity and thus improve computational efficiency.

[0044] What is even more challenging is that in the current task, the number of selectable reactants may be very large. Once a reaction is selected, the space of new selectable reactants is actually updated. If a post-shielding operation is performed on the outputs of all reactants, the unnecessary information introduced by the extra dimensions will prevent the model from learning a reasonable probability distribution for selecting effective reactants, which will cause the high-dimensional discrete action space to be too large, making it difficult for the reactant selection strategy network to learn a reasonable probability distribution. To solve this problem, a pre-training strategy is designed for the reactant selection strategy network, and a multi-label classification task is designed to enable the reactant selection strategy network to pre-learn how to effectively identify selectable reactants that match the selected reaction under different reaction labels, thereby achieving a pre-shielding effect on invalid reactants. This allows the model to focus more on establishing a reasonable probability distribution for selecting effective reactants (such as Figure 2As shown in b), labeled data including reactants and their valid reaction labels are constructed, and the reactant selection strategy network is pre-trained for multi-label classification tasks using the labeled data, reactant datasets, and reaction datasets. The output classification results are valid reactants and invalid reactants corresponding to each reaction, and the reactant selection strategy network is optimized in combination with the multi-label classification loss, so that the reactant selection strategy network can pre-learn to identify optional reactants that match the selected available reactions under different reaction labels and pre-screen invalid reactants through pre-training.

[0045] Specifically, the molecular generation process is defined as a Markov decision process in which reactants react step by step based on reaction rules to obtain the final product molecules. The generation process of the generation flow network is defined as a directed acyclic graph G = (S, A), where S represents the state space and A represents the action space. The vertex s0∈S is defined as the initial state with no incoming edges. The corresponding initial molecule can be set to empty or to a specific reactant. In drug design, if there is a relatively good warhead structure for a certain target, the reactant containing the warhead substructure can be set as the common structure of all generated molecules, thereby flexibly transforming the problem from de novo molecule generation to a molecular optimization problem. In the embodiment, no initial molecule is uniformly used as the initial state. Define vertex s f ∈S is the final state without outgoing edges, corresponding to the final generated molecule. Define the state sequence τ from the initial state to a certain terminal state = (s0→…→s n ) is a trajectory, and the molecular object x∈X realizes the forward state transfer s through sequential sampling actions t →s t+1 , X represents the set of all molecules in the chemical space, and the state transition is to add new molecular fragments to the currently partially built molecules. Considering the complexity of reaction-based molecular assembly, the state transition s t →s t+1 The action is divided into two steps: the first step is to select the available reactions in the current state based on the given reaction data set; the second step is to select the appropriate reactants based on the given reactant library, and use the reaction rules selected in the first step to perform chemical reactions. In the embodiment, SynGFN uniformly adopts two-step reactions as the maximum number of reaction steps, but the maximum number of reaction steps can be adjusted at will in actual application. Considering the selection of dual actions, it is necessary to construct reaction selection probability models separately. Used to select reactions and build reactant selection probability models The two policy models of SynGFN are trained to sample molecular objects with a probability τ(x) proportional to the reward value R(x), i.e., τ(x)∝R(x), to generate diverse molecular objects with high reward value.

[0046] The random forward policy PF (s t+1 |s t ) is decomposed and parameterized into two independent neural networks with learnable parameters and To fit the state transition s t →s t+1 distribution, θ1 and θ2 represent the learnable parameters of the reaction selection strategy network and the reactant selection strategy network, respectively. Similarly, the backward state transfer strategy can be parameterized as and The probability τ(x) of generating object x is based on P F It can be defined as:

[0047]

[0048] That is, the sum of the probabilities of all trajectories τ mapped to the same object x. The goal of the generative flow network is to learn a strategy π to sample an object x∈X with a probability proportional to its reward value R(x). Currently, there are several optional learning objectives for the generative flow network to ensure that π(x)∝R(x) during sampling, among which Trajectory Balance (TB) is the current mainstream choice, and its formula is as follows:

[0049]

[0050] Among them, Z θ Represents the partition function ∑ x∈X An approximate estimate of R(x), where n is the length of the trajectory. TB loss Depending on the sampling of each trajectory, GFlowNet is trained based on off-policy trajectories, sampling trajectories τ~P F (τ; θ), and Perform a gradient descent step on it to make the state transition s in the trajectory τ t →s t+1 Based on the current two-step action design to achieve state transfer, the training objective is modified to:

[0051]

[0052] Furthermore, due to the characteristics of the current molecular generation order sampling, s t+1 After t can be uniquely determined, so the reverse probability and All are set to 1, and the final training goal is:

[0053]

[0054] Using the final training objective Train the hierarchical policy network.

[0055] S2, using a multi-precision active learning framework including a hierarchical strategy network, a Gaussian process as a multi-precision substitution model and a multi-precision maximum entropy search as an acquisition function to provide the required precision labels of the generated molecules while generating molecules, including: using a hierarchical strategy network to obtain the generated molecules and the selected precision labels and input them into the multi-precision substitution model, using the multi-precision substitution model to score each generated molecule, using the acquisition function to give a comprehensive score after considering the usage cost of the scoring tool corresponding to the precision labels of different molecules and feeding it back to the reaction selection strategy network and the reactant selection strategy network for gradient update.

[0056] In the embodiment, a multi-precision active learning framework SynGFN-MFA is further developed based on SynGFN, and a sampling scheme of molecules in the training process is performed using a multi-precision acquisition function (such as Figure 3 The multi-precision active learning framework is based on the active learning framework, uses Gaussian process regression as the multi-precision surrogate model, and uses maximum entropy search as the acquisition function to train SynGFN. The sampled high-scoring molecules are scored using computing tools of different precisions, and the labeled data are added to the dataset to update the multi-precision surrogate model, thereby performing iterative optimization of active learning.

[0057] Specifically, using Gaussian process as a multi-precision surrogate model, for a given series of data points z 1:n =(x1,m1),(x2,m2),…,(x n ,m n ) and the observed value y 1:n =y1,y2,…,y n , where m represents the precision label of the data and the observed value represents the score of the scoring tool. The objective function f is modeled as f|z using Gaussian process. 1:n ,y 1:n ~GP(μ n ,K n ),μ n represents the mean, and the kernel function is the multi-precision kernel K n , GP(μ n ,K n ) represents a Gaussian process. For a general Gaussian process, we have:

[0058] μ n (x) = μ(x) + K(x,x 1:n )(K(x 1:n ,x 1:n )+σ 2 I) -1 (y 1:n -μ(x1:n ))#(5)

[0059] K n (x1,x2)=K(x1,x2)-K(x,x 1:n )(K(x 1:n ,x 1:n )+σ 2 I) -1 K(x 1:n ,x2)#(6)

[0060] Among them, μ n (x) represents the posterior mean of the new evaluation point x after a given series of n data points, μ(x) represents the prior mean at position x, that is, the initial prediction mean, K(·) represents the kernel function used to measure the similarity between data points, σ ​​represents the noise in the observation, I represents the mutual information, and μ(x) represents the prior mean at position x. 1:n ) represents the mean vector based on the existing data points, K n (x1,x2) represents the posterior covariance function for new points x1 and x2 after a given series of n data points, K(x1,x2) represents the prior covariance function, x1 and x2 represent two independent new points to be evaluated, x 1:n Represents a collection of previously existing data points.

[0061] In the embodiment, for a new point to be evaluated z=(x, m), the kernel function uses a multi-precision kernel, which is expressed as:

[0062] K(z1,z2)=K1(x1,x2)×K2(m1,m2)#(7)

[0063] Among them, K(z1,z2) represents the multi-precision kernel function, which is used to measure the similarity between two data points z1 and z2. In multi-precision modeling, the multi-precision version of the kernel function considers both the input data point x and the corresponding precision information at the same time. K1(·) represents the square exponential kernel:

[0064]

[0065] Where l represents the length scale parameter, ‖x1-x2‖ 2 Represents the square of the Euclidean distance between x1-x2;

[0066] K2(·) represents the precision level kernel, which is used to measure the similarity between two precision levels m1 and m2:

[0067] K2(m1,m2)=c+(1-m1) 1+δ (1-m2) 1+δ #(9)

[0068] Among them, c,δ>0 is a hyperparameter.

[0069] The acquisition function uses the maximum entropy search (Max Value Entropy Search, MES), the formula is as follows:

[0070]

[0071] Among them, α(x,m) represents the acquisition function value of x when considering the accuracy level m. In a given data set D j (j represents the number of active learning iterations) The multi-precision surrogate model output f at the evaluation point x at the accuracy level m m The true objective function (here set to the highest precision M) the level of mutual information between is a regularization term used to balance the evaluation costs of different precision levels. Here, the precision is assumed to be positively correlated with the evaluation cost. Multi-precision MES aims to evaluate the amount of information that can be obtained from data tuples (x, m) outside the sampled dataset D. By selecting data points (x, m) that can maximize the mutual information I, it effectively guides the acquisition process. SynGFN is trained with the acquisition function as the reward function of the generative flow network, and the advantage of the generative flow network in diversity is used to provide high-scoring and diverse molecular objects. After the training, the top K (Top K) data points (molecules and their precision labels) of the generated N data points (Sample N) are scored using the scoring method corresponding to the precision label, and then added to the dataset D for the next round of active learning iterations.

[0072] In order to meet the requirements of the multi-precision active learning framework, SynGFN needs to be further modified. In the generative flow network algorithm, (x,m) is used as the sampling target object instead of just sampling the molecule itself, and the accuracy level M of the scoring method is ′ = {0,1,2,…,M} is introduced as a new dimension into the state, and the state space changes from S to S M′ =S∪M′, in state transition s t →s t+1 Added additional action options for selecting the accuracy level.

[0073] In summary, the embodiment of the present invention provides a method for generating highly synthesizable molecules based on multi-precision feedback drive. In the constructed SynGFN, synthesis planning is deeply integrated into the molecular generation process. Compared with the standard generation flow network, a hierarchical strategy network is introduced to guide the reaction-based molecular assembly (first select the reaction template, and then select the reaction-compatible reactants) in the process of synthetic constraint molecule generation under the generation flow network framework. By introducing the reaction-compatible reactant candidate mask and the unique pre-training technology of the strategy network, not only the computational complexity is reduced, but also the convergence speed of the strategy network in the training stage is accelerated. In short, SynGFN directly generates reasonable synthesis paths (such as Figure 4 It connects the design and manufacturing stages that were previously separated in molecular design, and accelerates the design-manufacturing-testing-analysis (DMTA) cycle (as shown in Figure 1). Figure 4 (b in Figure 1). SynGFN uses a generative flow network algorithm to assign rewards corresponding to the generated molecules to each trajectory in the probability flow network, which ensures that the generation probability of different molecules is proportional to their reward value (target characteristic), thereby creating a distribution related to feedback and achieving independent and identically distributed sampling from it. This method solves the problem of insufficient molecular diversity from a fundamental algorithmic perspective. At the same time, a multi-precision active learning framework (called SynGFN-MFA) is constructed based on SynGFN, a multi-precision substitution model, and an acquisition function, which reduces the cost of scoring and evaluation without affecting the quality of the molecule generation itself.

[0074] The following relevant verification experiments are designed in combination with the effects of SynGFN and SynGFN-MFA provided in the embodiments of the present invention to fully illustrate the effectiveness of the scheme.

[0075] (1) The pre-training strategy introduced by SynGFN can effectively solve the problem of excessive dimension of discrete action space and improve computational efficiency.

[0076] Taking soluble epoxide hydrolase (sEH) as the drug target (i.e., using the binding affinity of the generated molecule to the sEH target as the source of reward), the training effects of SynGFN with and without pre-training schemes were compared under four different sizes of reactant libraries (the number of molecules in the reactant library was 1,000, 10,000, 30,000, and 100,000, respectively). It was found that when the size of the reactant library reached 100,000 (i.e., the output dimension of the reactant selection strategy network was 100,000), the introduction of pre-training could effectively improve the exploration efficiency of SynGFN, enabling it to explore more reasonable high-dimensional chemical space regions more quickly and in a targeted manner. Specifically, the average reward score of the sampled molecules at each training iteration step during the training process and the average reward score of the top 10 molecules in the sampled molecules were used as monitoring indicators. When the training reached convergence, SynGFN with the pre-training scheme improved by 37.5% and 15.6% respectively in these two indicators.

[0077] (2) SynGFN has excellent spatial exploration efficiency.

[0078] In order to compare SynGFN with traditional virtual screening methods, the virtual chemical space constructed by the mini-reactant library scale (the number of molecules is 1879) was enumerated, and finally a chemical space with 300 million compounds was obtained. For SynGFN, its training mode is to traverse the molecules in the chemical space and learn a reasonable action probability distribution based on reward feedback. In order to reflect the spatial search advantage of SynGFN, 1 million compounds, 10 times the number of molecules traversed by SynGFN training, were randomly selected as virtual screening compound libraries for comparison. Here, taking the sEH target as an example, under the same QSAR activity prediction model, the chemical space coverage of the top 10,000 molecules searched during SynGFN training is higher than that of the top 10,000 molecules after virtual screening of 1 million. Only one-tenth of the number of molecules needs to be screened to effectively identify and discover the spatial region of highly active molecules. Naturally, it can be seen that the score distribution of molecules traversed by SynGFN will also be better than that of molecules obtained simply through virtual screening. It should be noted that for SynGFN, only about 100,000 molecules need to be traversed to complete the training (mean reward reaches a stable value), which means that there is no need to continue training in the future, and sufficiently diverse high-scoring molecules can be directly sampled. The search performance of SynGFN and virtual screening within different QSAR activity score thresholds was further compared (here, since the comparison is the number of high-scoring compounds screened, in order to ensure fairness, virtual screening uses a virtual library with the same number of molecules as SynGFN). Using the enrichment factor as an indicator (the ratio of the number of high-scoring compounds searched by SynGFN to the number of high-scoring compounds searched by virtual screening), it can be more intuitively seen that as the score threshold increases, the number of high-scoring compounds searched by SynGFN has increased by dozens of times. The results show that SynGFN can explore more high-scoring compounds.

[0079] (3) SynGFN is significantly better than SyntheMol in sampling molecular diversity (i.e., chemical space exploration).

[0080] The diversity of generated molecules essentially reflects the exploration coverage of the generative model in its chemical space. #Circles is used as a metric to evaluate the coverage of chemical space of different generative models. #Circles is inspired by the concept of packing number in topology. The packing number refers to the maximum number of non-overlapping spheres that can be used to fill a given volume. Given a distance threshold t, #Circles calculates the maximum number of non-overlapping circles that can be accommodated in the chemical space. By determining this maximum set of non-overlapping molecules, this metric can effectively evaluate the coverage of the chemical space. This evaluation method helps to understand the depth and breadth of the exploration of the generative model in the chemical space, and then evaluate its ability to generate diverse molecules. Using #Circles as a metric can provide a quantitative and comparable metric to help which generative models perform better in exploring and covering the chemical space. Aurora kinase A (AURKA), D2 dopamine receptor (DRD2) and soluble epoxide hydrolase (sEH) were selected as therapeutic targets for benchmarking the drug design quality of various methods. These three targets are currently very popular in the treatment of cancer, schizophrenia and Parkinson's syndrome, and Alzheimer's disease, respectively. First, the #Circles curves of molecules generated by different models (SynGFN and SyntheMol) at different distance thresholds t (10,000 sampled molecules and the top 1,000 molecules with the highest scores) were examined. SynGFN always showed the best #Circles values ​​at different thresholds and targets, especially in the medium and high threshold range (t>0.65), indicating its strong advantage in global space coverage. The performance of different generative models at three threshold settings, t=0.7, 0.75 and 0.8, was further summarized. The results showed that SynGFN showed high chemical space coverage in the process of searching for active molecules, targeting these three different targets. It is worth noting that compared with the baseline model SyntheMol, SynGFN achieved up to 70 times improvement in chemical space coverage on the top 1,000 generated molecules, while SynGFN-XL achieved an improvement of 2 to 13 times on the sampled 10,000 molecules set when targeting multiple targets and evaluation settings. In addition, the distribution of diversity-adjusted QSAR scores of generated molecules was examined, which were obtained after clustering the 10,000 sampled molecules and the top 1,000 subset molecules into #Circles, and then retaining the molecules with the highest scores from each circle. The top 1,000 molecules generated by SynGFN at all scales showed superior diversity-adjusted QSAR scores on all three targets. This highlights the excellent ability of SynGFN to explore and exploit the balance while searching for ideal molecules in chemical space.

[0081] (4) SynGFN-MFA can effectively reduce the annotation cost of the drug discovery process.

[0082] Using sEH as a target case, we investigated the effect of multi-fidelity active-learning (MF-AL) as a demonstration of the algorithm effect at the scale of a mini-reactor library (the number of molecules in the reactant library is 1879). Here, we used the QSAR activity prediction model of sEH (based on message passing neural network MPNN) as a low-precision scoring method and AutoDock Vina as a high-precision scoring method. In order to simulate the actual problems faced by drug design in real scenarios, the scarcity of activity data and the high cost of high-precision scoring, the initial data set size was set to 1000, and the amount of data annotated in each active learning round was investigated in four intervals from 50 to 200. The generation effect of SynGFN when using MF-AL and only using the high-precision scoring method (here called SF-AL) was compared. Specifically, the scores of the top 100 molecules of the docking scores of 1000 molecules sampled by SynGFN after each active learning round were investigated. Considering the time difference between QSAR activity scoring and docking scoring, the time cost of the two was set to 1:100. The final result is that under the multi-precision active learning framework, SynGFN can achieve the same effect (top-100 score on par with SF-AL) at a lower cost. Under different annotated data sizes (the amount of annotated data in each round is 50 / 100 / 150 / 200 respectively), MF-AL can reduce the cost by 30%-80%, and this advantage will become more obvious as the amount of annotated data decreases. When the amount of annotated data in each round is 50, MF-AL can reduce the cost by nearly 80%, which is very attractive for drug design in real scenarios.

[0083] (5) SynGFN has demonstrated its effectiveness and significant advantages in real drug molecule scenarios (the designed synthesis pathway is nearly 100% consistent with the actual synthesis pathway).

[0084] Based on the good performance of SynGFN in a series of previous tests, it was applied to the design of NMDAR inhibitors. In previous studies, traditional NMDAR is one of the hottest drug development targets for neurological diseases because it plays a key role in neuronal development and synaptic plasticity in the central nervous system. From the composition point of view, NMDAR is assembled by a group of different receptor subunits. The classic NMDA receptor is usually a heterotetramer composed of 2 GluN1 and 2 GluN2 subunits. GluN1 / GIuN3A is also one of the NMDA receptor subtypes, but as an excitatory glycine receptor (eGlyRs), it is significantly different from traditional NMDAR and is regarded as an NMDAR-related receptor subfamily that needs to be further studied. eGlyRs have been revealed to be true brain receptors, functionally expressed in multiple brain regions, and have become potential new therapeutic targets for neuropsychiatric diseases. However, eGlyR biology is still in its infancy and has not yet been widely developed as a target. One of the main obstacles is that the GluN1 / GluN3A crystal structure has not yet been resolved (lack of the structure of the full-length GluN3A receptor). At the same time, current protein structure prediction tools have large errors in predicting the active conformation of multi-subunit proteins. In this case, there is great uncertainty in starting from the perspective of structure-based drug design. Several compounds that can antagonize GluN1 / GIuN3A have been developed. Due to the high structural similarity between the orthosteric ligand binding pockets of GluN1 and GluN3A, it is difficult to design compounds that selectively target the orthosteric site of GluN3A. A more promising idea is to design negative allosteric modulators (NAMs), such as EU1180-438 and WZB117, which are considered to work by binding to the TMD site near the GluN3A pre-M1 region. Therefore, it is hoped that starting from the ligand-based drug design idea, SynGFN's ability to efficiently and accurately explore chemical space can be used to design new inhibitors of GluN1 / GIuN3A. The 3D conformations of the two molecules, WZB117 and EU1180-438, are used as shape queries for ROCS. OMEGA was used to generate conformations of the generated molecules, and then ROCS was used to calculate the similarity with the query based on the 3D shape superposition method, which was used as a reward score to guide the generation. Finally, 10 candidate compounds were selected for chemical synthesis and biological activity verification. In the end, all the compounds were successfully synthesized within 30 days, of which 6 compounds showed certain inhibitory activity against GluN1 / GluN3A, and the IC50 value of the most effective compound D1 was 2.68μM. 60% of the tested molecules had potential activity, and the synthesis route provided by SynGFN was nearly 100% consistent with the actual synthesis route.In the past, when drug molecules converted based on generative AI technology were actually screened, they often needed to be screened one by one by synthesis experts, and each molecule needed to go through tedious synthesis route confirmation (including reactants, catalyst prices, reaction conditions, post-reaction treatment methods, etc.). As far as the current test is concerned, one FTE (full-time equivalent) can only synthesize 2-3 compounds a month. With the help of SynGFN, the efficiency has been increased by 100%. With the help of SynGFN, this process can be greatly accelerated. SynGFN connects the two links of design and synthesis, breaking through the bottleneck of the slow cycle speed of DMTA.

[0085] Based on the same inventive concept, Figure 5 As shown, the embodiment of the present invention further provides a highly synthesizable molecule generation device 500 based on multi-precision feedback drive, including: a hierarchical strategy network module 510 and a multi-precision feedback drive module 520 .

[0086] The hierarchical strategy network module 510 is used to generate highly synthesizable molecules using a hierarchical strategy network including a reaction selection strategy network and a reactant selection strategy network, including: using the reaction selection strategy network to select an available reaction in the current state based on an initial molecular building block or an intermediate product in the previous state, using the reactant selection strategy network to select an available reactant in the current state based on the available reactions, and performing a chemical reaction according to the reaction rules of the available reactions to obtain an intermediate product or a final product.

[0087] The multi-precision feedback driving module 520 is used to optimize the reactant selection in the reactant selection strategy network using a multi-precision active learning framework including a hierarchical strategy network, a multi-precision substitution model and an acquisition function, including: using the reactant selection strategy network in the hierarchical strategy network to obtain optional reactants and the precision label of each optional reactant and input them into the multi-precision substitution model, using the multi-precision substitution model to select different scoring tools according to different precision labels to score each optional reactant, and using the acquisition function to sample the available reactants according to the scoring results and feed them back to the reactant selection strategy network.

[0088] Based on the same inventive concept, an embodiment of the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to implement the above-mentioned method for generating highly synthesizable molecules based on multi-precision feedback drive when executing the computer program.

[0089] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a computer, the above-mentioned method for generating highly synthesizable molecules based on multi-precision feedback drive is implemented.

[0090] Based on the same inventive concept, an embodiment of the present invention further provides a computer product, which includes a computer program. When the computer program is executed by a processor, the above-mentioned method for generating highly synthesizable molecules based on multi-precision feedback drive is implemented.

[0091] It should be noted that the highly synthesizable molecule generation device based on multi-precision feedback drive, electronic device, computer-readable storage medium and computer product provided in the above-mentioned embodiments all belong to the same inventive concept as the highly synthesizable molecule generation method based on multi-precision feedback drive. The specific implementation process is detailed in the embodiment of the highly synthesizable molecule generation method based on multi-precision feedback drive, which will not be repeated here.

[0092] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating highly synthesizable molecules based on multi-precision feedback drive, characterized in that: The following steps are involved: Using a hierarchical strategy network including a reaction selection strategy network and a reactant selection strategy network to generate highly synthesizable molecules, including: using the reaction selection strategy network to select an available reaction in a current state according to an initial molecular building block or an intermediate product in a previous state, using the reactant selection strategy network to select an available reactant in a current state based on the selected available reaction and performing a chemical reaction according to a reaction rule of the selected reaction to obtain an intermediate product or a final product; A multi-precision active learning framework including a hierarchical strategy network, a Gaussian process as a multi-precision substitution model and a multi-precision maximum entropy search as an acquisition function is used to provide the required precision labels of the generated molecules while generating molecules, including: using a hierarchical strategy network to obtain the generated molecules and the selected precision labels and input them into the multi-precision substitution model, using the multi-precision substitution model to score each generated molecule, using the acquisition function to give a comprehensive score after considering the usage cost of the scoring tool corresponding to the precision labels of different molecules and feeding it back to the reaction selection strategy network and the reactant selection strategy network for gradient update.

2. The method for generating highly synthesizable molecules based on multi-precision feedback drive according to claim 1, characterized in that: The method further comprises: Construct labeled data including reactants and their valid reaction labels, and use the labeled data, reactant dataset and reaction dataset to pre-train the reactant selection strategy network for multi-label classification tasks. The output classification results are the valid reactants and invalid reactants corresponding to each reaction, so that the reactant selection strategy network can pre-learn to identify optional reactants that match the selected available reactions under different reaction labels and pre-screen invalid reactants through pre-training.

3. The method for generating highly synthesizable molecules based on multi-precision feedback drive according to claim 1, characterized in that: The molecular generation process is defined as a Markov decision process in which reactants react step by step based on reaction rules to obtain the final product molecules, and this process is implemented using the generative flow network model. The generative flow network defines the molecular generation process as a directed acyclic graph G = (S, A), where S represents the state space and A represents the action space. The vertex s0∈S is defined as the initial state with no incoming edges and corresponds to no initial molecule or a specific reactant. The vertex s is defined f ∈S is the final state without outgoing edges and corresponds to the final generated molecule. The state sequence τ from the initial state to a certain terminal state is defined as (s0→…→s n ) is a trajectory, and the molecular object x∈X realizes the forward state transfer s through sequential sampling actions t →s t+1 , X represents the set of all molecules in the combinatorial chemical space, and the state transfer is to add new molecular fragments to the currently partially constructed molecules through chemical reactions; In the hierarchical policy network, the state transfer s is based on the generation flow network t →s t+1 The action is divided into two steps. The first step is to select the available reactions in the current state based on the given reaction data set. The second step is to select the available reactants based on the given reactant library and use the reaction rules of the available reactions selected in the first step to perform chemical reactions. In the two steps, the reaction selection probability model is constructed respectively. The reaction selection strategy network is used to select reactions and build a reactant selection probability model The reactant selection strategy network is used to select reactants, and the reaction selection strategy network and the reactant selection strategy network are trained to sample molecular objects with a probability π(x) proportional to the reward value R(x), that is, π(x)∝R(x), to generate diverse high reward value molecular objects.

4. The method for generating highly synthesizable molecules based on multi-precision feedback drive according to claim 1, characterized in that: The method of using the reaction selection strategy network to select available reactions in the current state according to the initial molecular building blocks or the intermediate products in the previous state includes: When performing reaction selection, the initial molecular building blocks or the intermediate products of the previous state are input into the reaction selection strategy network to obtain the original probability values ​​of all reactions. The classification probability distribution is constructed by shielding invalid reactions and normalizing the remaining valid reactions. Then, random sampling is performed from the classification probability distribution to obtain the available reactions in the current state.

5. The method for generating highly synthesizable molecules based on multi-precision feedback drive according to claim 1, characterized in that: The method of using the reactant selection strategy network to select available reactants in the current state based on the selected available reactions and performing chemical reactions according to the reaction rules of the selected available reactions to obtain intermediate products or final products includes: When selecting reactants, the available reactions in the current state obtained by the reaction selection strategy network are input into the reactant selection strategy network to obtain the original probability values ​​of all reactants. The classification probability distribution is constructed by shielding invalid reactants and normalizing all remaining valid reactants. Then, random sampling is performed from the classification probability distribution to obtain the available reactants in the current state. Chemical reactions are carried out according to the reaction rules of the available reactions to obtain intermediate products or final products. After the chemical reaction is completed, it is converted to the next state.

6. The method for generating highly synthesizable molecules based on multi-precision feedback drive according to claim 1, characterized in that: The Gaussian process is modeled as a multi-precision surrogate model, where the kernel function adopts a multi-precision kernel.

7. A highly synthesizable molecule generation device based on multi-precision feedback drive, implemented by the highly synthesizable molecule generation method based on multi-precision feedback drive according to any one of claims 1 to 6, characterized in that: include: Hierarchical strategy network module and multi-precision feedback drive module; The hierarchical strategy network module is used to generate highly synthesizable molecules using a hierarchical strategy network including a reaction selection strategy network and a reactant selection strategy network, including: using the reaction selection strategy network to select an available reaction in the current state according to an initial molecular building block or an intermediate product in a previous state, using the reactant selection strategy network to select an available reactant in the current state based on the selected available reaction and performing a chemical reaction according to the reaction rule of the selected available reaction to obtain an intermediate product or a final product; The multi-precision feedback driving module is used to use a multi-precision active learning framework including a hierarchical strategy network, a Gaussian process as a multi-precision substitution model and a multi-precision maximum entropy search as an acquisition function to provide the required precision labels of the generated molecules while generating molecules, including: using the hierarchical strategy network to obtain the generated molecules and the selected precision labels and input them into the multi-precision substitution model, using the multi-precision substitution model to score each generated molecule, using the acquisition function to give a comprehensive score after considering the usage cost of the scoring tool corresponding to the precision labels of different molecules and feeding it back to the reaction selection strategy network and the reactant selection strategy network for gradient update.

8. An electronic device comprising a memory and a processor, wherein the memory is used to store a computer program, wherein: The processor is used to implement the method for generating highly synthesizable molecules based on multi-precision feedback drive as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a computer, the method for generating highly synthesizable molecules based on multi-precision feedback drive as described in any one of claims 1 to 6 is implemented.

10. A computer product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for generating highly synthesizable molecules based on multi-precision feedback drive as described in any one of claims 1 to 6 is implemented.

Images