Deep learning method for designing molecular structures of green substitutes of chemicals

Abstract

The invention belongs to the technical field of green substitute design for chemical risk prevention and control, and discloses a deep learning method for designing a molecular structure of a chemical green substitute. On the basis of known chemical molecular structures and substitute target attributes, a green substitute molecular structure with synergistically optimized function and hazard attributes is obtained by applying the constructed method. The method is simple, convenient and efficient, resource investment of experimental testing can be saved, and regret replacement is prevented. The construction process of the method comprises the following steps: (1) generating a molecular structure of structural constraint based on a generative model; (2) realizing high-throughput prediction of multi-target attributes based on a decision-making model; and (3) realizing multi-target attribute collaborative optimization based on a Pareto optimization theory. The method established by the invention can effectively explore a wide chemical space, predict the multi-target attributes of the chemicals in a high-throughput manner, generate molecular structures with necessary functions and lower harmfulness, and provide a basic tool for green substitution of harmful chemicals.

Description

Technical Field

[0001] The present invention belongs to the technical field of green substitute design for chemical risk prevention and control, and relates to a deep learning method for designing the molecular structure of green substitutes for chemicals. Background Art

[0002] Chemicals are an important source of new pollutants. The rapid growth of chemical types and production will cause chemical pollution to exceed the earth's carrying capacity, becoming a major challenge to the sustainable development of human society. Therefore, it is necessary to prevent and control the risks of chemicals, reduce their environmental hazards from the source, and prevent them from becoming new pollutants. Designing and using green alternative chemicals can reduce their environmental hazards while ensuring the necessary functional properties of chemicals, and help achieve comprehensive and sustainable development of society, economy and environment.

[0003] Some studies have used empirical methods to design green alternative chemicals, but unfortunately chemical substitution is still happening, as described in the literature "Science, 2020, 367, 397-400; ACS Sustainable Chem. Eng. 2021, 9, 7749-7758". This can be attributed to the lack of tools to effectively explore the vast chemical space, so that when a harmful chemical is eliminated, it is likely to be replaced by a chemical with a similar structure. Since the goal of green substitution is to design chemicals with new structures to replace existing chemicals, the designed substitutes usually lack measured data on functional properties and hazard properties. Moreover, there may be correlations or contradictions between the functional and hazard properties of chemicals, making it difficult to achieve coordinated optimization of multi-objective properties.

[0004] Therefore, the present invention has created a green alternative molecular design method for chemicals based on generative and decision-making deep learning, which can effectively explore a vast chemical space, predict the multi-objective properties of chemicals with high throughput, and generate molecular structures with necessary functions and lower hazards. Compared with previous empirical substitution methods, the created method combines the Pareto multi-objective optimization theory with new deep learning technology to synergistically optimize the multi-objective properties of molecules, while ensuring the necessary functional properties of chemicals, reducing their environmental hazards. The green substitution of per / polyfluorinated substances (PFAS) surfactants is used as an example to demonstrate the application method of the model. Summary of the invention

[0005] The present invention creates a deep learning method for designing the molecular structure of green chemical alternatives. This method can efficiently predict the functional and hazardous properties of surfactants, including surface tension (T S ), the logarithm of the critical micelle concentration (logC MC ), the logarithm of the n-octanol-water partition coefficient (logKOW ), and 23 biological activities or toxicities, and efficiently explore the vast chemical space and synergistically optimize the multi-target properties of molecules according to the functional and hazard properties of chemicals. The method can design molecules with necessary functionality, such as T S Value and low logC MC values, and lower environmental hazards, such as low logK OW values, and chemicals that are not biologically active or toxic.

[0006] The technical solution of the present invention:

[0007] A deep learning method for designing the molecular structure of green alternatives to chemicals, the steps are as follows:

[0008] (1) Molecular structure generation based on structure constraints based on generative models

[0009] The present invention constructs a generative model based on a long short-term memory (LSTM) network to achieve high-throughput molecular design with structural constraints. The LSTM neural network is a variant of the traditional recurrent neural network, which is innovative in structure through gated recurrent units and cell states. Each gated recurrent unit contains three gates, which are composed of a single-layer neural network with a sigmoid activation function, which is used to decide whether to retain or forget input features, thereby achieving selective updating of features. Thanks to this architecture, the LSTM network can effectively capture semantic relevance in long sequences and alleviate problems such as gradient disappearance or explosion.

[0010] The constructed generative model consists of two LSTM layers, which can learn the conditional probability distribution of each character in the molecular SMILES code relative to all its previous characters. The generative model is trained using an autoregressive method, that is, it iteratively predicts the next character based on all the previous characters in the SMILES code. Therefore, the model only needs an unlabeled molecular structure dataset to achieve the learning goal.

[0011] During the training process, the generative model uses the Adam optimizer and the categorical cross entropy loss function; during the training process, all training data are completely traversed once, that is, 3486 PFAS molecular structures collected from the OECD PFAS structure list. During the generation process, starting from the initial character, the LSTM autoregressive method is repeatedly used to generate the probability distribution of the next character and sample it until the end character is generated, and the generation process stops.

[0012] Transfer learning is further introduced to realize the structural constraints of the generative model; first, the generative model is pre-trained on the collected PFAS molecular structures, and then the generative model is fine-tuned on the Pareto optimal frontier (POF) molecular structure dataset identified based on the Pareto optimization theory, so that the molecular structure generation is biased towards the target chemical space.

[0013] (2) Realize high-throughput prediction of multi-target attributes based on decision-making models

[0014] The present invention constructs a decision-making model to predict the target attribute values ​​of the generated molecules with high throughput. The target attributes of the molecular structure design of PFAS surfactants include two aspects, namely functional attributes and hazard attributes. Functional attributes include T S and logC MC , the hazard attribute includes logK OW and 23 biological activities or toxicities.

[0015] In order to enhance the robustness of the decision-making model on the target attribute dataset, an ensemble learning method was used for decision-making model training. The target attribute dataset was randomly divided 10 times and divided into training set, test set and validation set in a ratio of 8:1:1. The training set and test set were used to train the model and evaluate the goodness of fit of the model. After the decision-making model training was completed, the validation set was used to evaluate the prediction effect of the model. The decision-making model was constructed using these 10 randomly divided datasets, thereby creating an ensemble model containing 10 models for each endpoint. Then, the predicted values ​​of the 10 decision-making models in the ensemble model were averaged to improve the robustness of the prediction results.

[0016] Using the above target attribute data set, a decision-making prediction model was constructed based on the graph neural network algorithm. In the algorithm architecture of the graph neural network, the molecular structure of chemicals is represented in the form of a molecular graph. The first step in building the graph neural network model is to parse the structural information of chemicals according to the SMILES code and convert it into a molecular graph. In this process, a node feature vector is created for each atom in the molecular graph, which contains information such as the atom type, number of bonds, charge state, chiral identity, hybridization form, aromatic state, and the number of hydrogen atoms connected to the atom. At the same time, each chemical bond in the molecular graph is described by an edge feature vector, which covers the type of chemical bond and its conjugated state. Therefore, each molecular graph consists of an atomic feature matrix and a chemical bond feature matrix, which are used to gradually aggregate and update features in the subsequent training process. In addition, the molecular graph must also include the connection relationship between nodes to clarify the neighborhood range considered for each feature aggregation and update.

[0017] After converting the SMILES code into a molecular graph, the feature representation of a central node is updated by aggregating the features of adjacent nodes. This feature aggregation process is divided into two stages: the message passing stage and the readout stage. Specifically, the message passing stage implements the transmission and integration of the features of adjacent nodes to the central node according to specific rules, as shown in formulas (1) and (2).

[0018]

[0019] During the iteration process, and They represent the feature vectors of node v and node w after t iterations, respectively, which contain information such as the type of atom, the number of connected hydrogen atoms, etc. represents the characteristics of the edge connecting the node pair v and w after t iterations, including information such as the type of chemical bond; N v Represents a set of neighbor nodes of the central node v; in the message passing stage, M t The (·) function is used as a message passing mechanism to integrate the feature information of neighbor nodes and pass it to the central node. t (·) The function can be an operation such as taking the maximum, minimum or average value of the neighbor node features to generate the integrated feature representation. Finally, using the update function U t (·) to update the feature representation of the central node; U t (·) The function is implemented by the fully connected layer;

[0020] In the readout stage, the readout function R(·) is used to aggregate all node features according to the final feature Calculate the end point of the entire molecular graph G

[0021]

[0022] After completing the feature aggregation step of the atomic nodes, a virtual node connected to all atomic nodes is introduced as the global representative of the molecule. The same feature aggregation process as the atomic nodes is performed on the virtual nodes to integrate the feature information from all atomic nodes and update the feature representation of the virtual nodes accordingly, so that the virtual node features can comprehensively summarize the molecular characteristics related to the prediction target.

[0023] Next, using formula (3), combined with the virtual node features output by the molecular attention layer, the predicted endpoint value of the molecular graph is calculated In order to quantify the prediction accuracy of the model, a loss function is used to measure the predicted endpoint value. The difference between the predicted value and the actual value of the target attribute parameter of the chemical. The greater the deviation between the predicted value and the actual value, the higher the value of the loss function.

[0024] The training process of the neural network aims to continuously adjust all learnable parameters W within the model through optimization algorithms in order to minimize the value of the loss function. Based on the gradient information of the loss function and the gradient descent algorithm, the Adam optimizer is used to iteratively update the model parameters, gradually reducing the value of the loss function until the model performance is steadily improved, completing the entire "end-to-end" training process.

[0025] (3) Realizing multi-objective attribute collaborative optimization based on Pareto optimization theory

[0026] like Figure 1 As shown in Figure 1, based on the Pareto optimization theory, a method for identifying the optimal solution set, i.e., POF, for multi-objective optimization problems in molecular structure design is developed. For ease of explanation, the multi-objective optimization problem is first simplified to a two-dimensional optimization problem, and the calculation method is introduced as an example. When analyzing this example, a chemical space containing 10 molecules is considered.

[0027] Figure 1 A shows the scatter plot of these 10 molecules on a two-dimensional plane consisting of a hazard attribute and a functional attribute. One of the molecules (denoted as molecule 1) is randomly selected for the first iteration analysis. Figure 1 As shown in Figure 2, except for molecule 1, the molecules in the shaded area have higher hazards and lower functional properties. In the context of chemical structure design, molecule 1 is clearly superior to other molecules in the shaded area. Therefore, the molecules in the shaded area are removed from this chemical space, leaving Figure 1 C shows the five molecules.

[0028] Subsequently, this process is repeated, randomly selecting another molecule 2 that remains in the specified chemical space and removing molecules that molecule 2 completely outperforms, such as Figure 1 D. This iterative process continues until no molecule remaining in the specified chemical space is completely superior to any other molecule, as shown in Figure 1 E. When the iteration stops, all remaining molecules constitute the POF of the specified chemical space. This process can be generalized to three-dimensional or even higher-dimensional optimization problems by expanding the shaded area into a hyperrectangle, such as Figure 1 F. Therefore, this method is suitable for high-dimensional multi-objective optimization problems in molecular structure design. Based on the identified POF molecules, the generative model is iteratively trained to achieve collaborative optimization of the multi-objective properties of the generated molecules.

[0029] Hypervolume (V H) is a key indicator for evaluating the quality of POF. This indicator calculates the ratio of the volume of the space enclosed by POF to the maximum design space volume. The maximum design space volume is obtained by setting a reference point and calculating the volume enclosed by the reference point. The reference point is usually selected as the optimal value point in the target space and a point that is worse than all solutions. H It is an important benchmark for evaluating the POF quality of the generated molecular datasets in the iterative generation of molecules. H The larger the value, the larger the POF found by the developed method covers the larger target space, that is, the multi-objective collaborative optimization in molecular design is achieved. H The indicator is also used as a criterion for determining when to terminate the iterative training of the generative model and the generation of new molecular structures. When a new round of iterative training and generation of the generative model no longer produces V H When a set of molecular structures with higher values ​​is found, the developed method is considered to have fully explored the chemical space with Pareto optimality, and the iteration of the generative model is stopped.

[0030] Beneficial effects of the present invention: The constructed model innovatively combines multiple deep learning algorithms and Pareto multi-objective optimization theory to create a multi-objective optimization methodology suitable for the molecular design of green substitutes for chemicals. The created method can efficiently explore a vast chemical space, design molecular structures with specific properties in a high-throughput manner, and synergistically optimize the multi-objective properties of molecules, with strong functionality. The method is simple and efficient, and can be programmatically applied to the design of green substitute molecular structures for various chemicals, providing a basic tool for improving the source management of chemicals, and serving the major national needs of chemical risk management and control and new pollutant control. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a Pareto optimal boundary identification method for multi-objective optimization problems in molecular design. Figure 1 A is the initial molecular function and hazard attribute value, Figure 1 B is the dominant region of the first iteration molecule 1, Figure 1 C is the remaining numerator after the first iteration, Figure 1 D is the dominant region of molecule 2 in the second iteration, Figure 1 E is the POF when the iteration ends, Figure 1 F is a generalized example of a high-dimensional multi-objective optimization problem.

[0032] Figure 2 The construction process for the overall method. DETAILED DESCRIPTION

[0033] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0034] Example 1

[0035] As shown in Table 1, the target properties to be considered in the design of the molecular structure of green substitutes for PFAS surfactants include functional properties and hazard properties. S and logC MC , the hazard attribute is logK OW , and 23 biological activities or toxicities such as inhibition of cytochrome P450 (CYP) 2C9 activity. The measured values ​​of target properties of 59,819 chemicals were collected from the literature.

[0036] Table 1 Target attribute endpoints and significance of PFAS surfactants

[0037]

[0038]

[0039] Example 2

[0040] Based on the target attribute data set and the graph neural network algorithm, a decision-making prediction model for the target attribute is constructed. The performance of the model with a continuous endpoint is determined by the coefficient of determination (R 2 ) and root mean square error (E RMS The performance of the model predicting discrete endpoints was evaluated by the area under the receiver operating characteristic curve (A ROC ), true positive rate (R TP ) and accuracy (R A ) for evaluation. Table 2 shows the prediction performance of the constructed model on the validation set.

[0041] Table 2 Prediction performance of the graph neural network model on the validation set

[0042]

[0043]

[0044] Example 3

[0045] Given a chemical GenX (CAS number: 13252-13-6; SMILES code: C(=O)(C(C(F)(F)F)(OC(C(C(F)(F)F)(F)F)(F)F)F)O), a green alternative molecular structure with multiple target attributes better than GenX is to be designed. First, according to the SMILES code of GenX, it is converted into a molecular graph, and then the model constructed by the present invention is used to predict its target attributes. The model prediction results are shown in Table 3.

[0046] Table 3. Generated molecules with excellent performance as PFAS surfactant substitutes

[0047]

[0048] Then, based on the created multi-objective property collaborative optimization method, a new substitute molecular structure was designed. The results are shown in Table 4. The target properties of the generated molecules are better than the substitute target chemical, GenX.

[0049] Table 4. Generated molecules with excellent performance as PFAS surfactant substitutes

[0050]

[0051]

[0052] Example 4

[0053] Given a chemical F-53B (CAS number: 756426-58-1; SMILES code: C(C(C(C(F)(F)Cl)(F)F)(F)F)(C(C(OC(C(F)(F)S(=O)(=O)O)(F)F)(F)F)(F)F)(F)F)(F)F), a green alternative molecular structure with multiple target attributes better than F-53B is to be designed. First, according to the SMILES code of F-53B, it is converted into a molecular graph, and then the model constructed by the present invention is used to predict its target attributes. The model prediction results are shown in Table 5.

[0054] Table 5. Generated molecules with excellent performance as PFAS surfactant substitutes

[0055]

[0056] Then, based on the created multi-objective attribute collaborative optimization method, a new substitute molecular structure was designed. The results are shown in Table 6. The target attributes of the generated molecules are better than the substitute target chemical, F-53B.

[0057] Table 6. Generated molecules with excellent performance as PFAS surfactant replacements

[0058]

[0059]

[0060] Note: n toxicity Refers to the number of endpoints with biological activity / toxicity.

Claims

1. A deep learning method for designing the molecular structure of green alternatives to chemicals, characterized in that Follow these steps: (1) Molecular structure generation based on structure constraints based on generative models A generative model of chemical molecular structure is constructed based on LSTM network to achieve high-throughput molecular structure design with structural constraints. The constructed generative model consists of two LSTM layers, which is used to learn the conditional probability distribution of each character in the molecular SMILES code relative to all its previous characters. The generative model is trained using an autoregressive method, that is, it iteratively generates the next character based on the sequence of characters that have been generated in the molecular SMILES code. During the training process, the generative model uses the Adam optimizer and the categorical cross entropy loss function; during the training process, all training data are completely traversed once, that is, 3486 PFAS molecular structures collected from the PFAS structure list of the Organization for Economic Cooperation and Development; during the generation process, starting from the initial character, the autoregressive method is repeatedly used to generate the probability distribution of the next character and sample it until the end character is generated, at which point the generation process stops; Transfer learning is further introduced to realize the structural constraints of the generative model. First, the generative model is pre-trained on the collected PFAS molecular structures, and then the generative model is fine-tuned on the POF molecular structure identified based on the Pareto optimization theory, so that the molecular structure generation is biased towards the target chemical space. (2) Realize high-throughput prediction of multi-target attributes based on decision-making models Construct decision-making models to predict target property values ​​of generated molecular structures with high throughput; The target attributes of the molecular structure design of PFAS surfactants include two aspects: the functional attributes and hazard attributes of chemicals; the functional attributes include T S and logC MC , the hazard attribute includes logK OW and 23 biological activities or toxicities; In order to enhance the robustness of the decision-making model in predicting target attributes, the decision-making model training adopts an ensemble learning method; The dataset of the target attribute was randomly divided 10 times into training set, test set and validation set in the ratio of 8:1:1; Using the target attribute data set, a high-throughput prediction model is constructed based on the graph neural network algorithm. In the graph neural network algorithm architecture, the molecular structure of chemicals is represented as a molecular graph, as follows: First, the structural information of the chemical is parsed according to the molecular SMILES code and converted into a molecular graph. In this process, a node feature vector is created for each atom in the molecular graph, which contains the atom type, number of bonds, charge state, chiral identity, hybridization form, aromatic state, and the number of hydrogen atoms connected to the atom. At the same time, each chemical bond in the molecular graph is described by an edge feature vector, which covers the type of chemical bond and its conjugated state. Therefore, each molecular graph consists of an atomic feature matrix and a chemical bond feature matrix. In addition, the molecular graph must also include the connection relationship between nodes to clarify the neighborhood range considered for each feature aggregation and update. After converting the molecular SMILES code into a molecular graph, the feature representation of a central node is updated by aggregating the feature vectors of adjacent nodes. This feature aggregation process is divided into two stages: message passing stage and readout stage. The details are as follows: In the message transmission stage, the feature vectors of adjacent nodes are transmitted and integrated to the feature vector of the central node according to equations (1) and (2): in, and They represent the feature vectors of node v and node w after t iterations, respectively, which contain information such as the type of atoms, the number of connected hydrogen atoms, etc. N represents the feature vector of the edge connecting the node pair v and w after t iterations, including the type of chemical bond, etc. v Represents a set of neighbor nodes of the central node v; in the message passing stage, M t (·) function as a message passing mechanism, responsible for integrating the feature information of neighbor nodes and passing it to the central node; M t (·) The function takes the maximum, minimum or average value of the neighbor node features to generate the integrated feature representation. Finally, using the update function U t (·) to update the feature representation of the central node; U t (·) The function is implemented by the fully connected layer; In the readout phase, the readout function R(·) is used to aggregate all node features according to the final feature Calculate the end point of the entire molecular graph G After completing the feature aggregation step of the atomic nodes, a virtual node connected to all atomic nodes is introduced as the global representative of the molecule. The virtual node is subjected to the same feature aggregation process as the atomic node, and the feature information from all atomic nodes is integrated. The feature representation of the virtual node is updated accordingly, so that the virtual node features can comprehensively summarize the molecular characteristics related to the prediction target. Next, using formula (3), combined with the virtual node features output by the molecular attention layer, the predicted endpoint value of the molecular graph is calculated In order to quantify the prediction accuracy of the decision-making model, a loss function is used to measure the predicted endpoint value. The difference between the predicted value and the actual value of the chemical target attribute parameter; the greater the deviation between the predicted value and the actual value, the higher the value of the loss function; Based on the gradient information of the loss function and the gradient descent algorithm, the Adam optimizer is used to iteratively update the decision-making model parameters, gradually reducing the value of the loss function until the performance of the decision-making model is steadily improved, completing the entire "end-to-end" training process; (3) Realizing multi-objective attribute collaborative optimization based on Pareto optimization theory Based on Pareto optimization theory, a method for identifying the optimal solution set, namely POF, for multi-objective optimization problems in molecular structure design is proposed. First, the multi-objective optimization problem is simplified to a two-dimensional optimization problem, and the calculation method is introduced as an example. Consider a specific chemical space containing an arbitrary number of molecular structures. First, a two-dimensional scatter plot is drawn based on the performance of the molecular structure in the two dimensions of hazard and industriality. Then, one of the molecular structures is randomly selected as the starting point for the first iterative analysis. In the iterative process, all molecular structures that are more harmful than the starting molecule and have lower industrial performance than the starting molecule are first identified, and the identified molecular structures are defined as inferior molecules; then the identified inferior molecules are removed from the chemical space; then this process is repeated, another molecule remaining in the chemical space is randomly selected as a new starting molecule, and all inferior molecules defined under the new starting molecule are removed again; This iterative process continues until no molecular structure remaining in the chemical space is completely superior to any other molecular structure; when the iteration stops, all remaining molecular structures constitute the POF of the chemical space; The molecular structure on POF achieves the best balance between harmfulness and industrial performance, and is the optimal solution set in molecular structure design. By expanding the area surrounded by POF into a hyperrectangle, this method is extended to three-dimensional or even higher-dimensional chemical space design problems. The molecules on POF are used to iteratively train the generative model to achieve multi-objective collaborative optimization in molecular structure design. H The ratio of the space volume enclosed by POF to the maximum design space volume is calculated and used to evaluate the quality of POF; V H The indicator also serves as a criterion for determining when to terminate the iterative training of the generative model and the generation of new molecular structures; when a new round of iterative training and generation no longer produces V H When a set of molecular structures with higher values ​​is found, it is considered that the chemical space with Pareto optimality has been fully explored, and the iteration of the generative model is stopped immediately.

Images