A biological retrosynthesis method and system combining AND / OR tree and single-step reaction rule prediction

By combining the method with or tree and single-step reaction rule prediction model, the problems of low efficiency and low availability of bioreverse synthesis methods are solved, and the rapid finding of biosynthesis pathways is achieved, reducing experimental costs and time, and improving the availability of pathways.

CN114360659BActive Publication Date: 2024-11-08WUHAN UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202111570217.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-21
Publication Date
2024-11-08
Estimated Expiration
2041-12-21

AI Technical Summary

Technical Problem

The existing biological inverse synthesis methods are inefficient and have low actual availability, making it difficult to quickly find synthesis pathways, and require a lot of manpower, raw materials and time, and have high requirements for complex parameter settings and domain knowledge.

Method used

Combining the prediction model of EOR and single-step reaction rules, by selecting EOR and Tree nodes, predicting metabolic reaction rules, and expanding EOR and Tree to generate biosynthesis pathways, avoiding complex parameter settings such as thermodynamics and enzyme performance, and improving the Computational efficiency and pathway availability.

Benefits of technology

The rapid discovery of potential metabolic pathways is achieved, reducing experimental costs and time, improving the actual availability of biosynthetic pathways, and reducing dependence on complex parameters and domain knowledge.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114360659B_ABST
    Figure CN114360659B_ABST
Patent Text Reader

Abstract

The present invention provides a biological retrosynthesis method and system combining an AND / OR tree with single-step reaction rule prediction, wherein the method decomposes the retrosynthetic pathway generation of the target molecule into multiple single-step reaction rule predictions when performing retrosynthesis on the target molecule. When predicting the single-step reaction rule, the substrate molecule of the reaction rule predicted in the previous step is used as the product molecule of the reaction rule to be predicted next time. The SMILES sequence of the product molecule is used as input, and the molecular characteristics of the product molecule are calculated by custom calculation to be compatible with a variety of single-step reaction rule prediction models. The reaction rule is determined based on the prediction results of the model, and then expanded based on the AND / OR tree structure. Finally, the potential synthesis pathway of the target molecule is found. The present invention significantly improves the speed of finding a feasible synthesis pathway, does not require manual setting of complex parameters, can assist biologists in finding potential metabolic pathways more quickly, reduces experimental costs, and improves experimental efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of metabolic pathway analysis in biosynthesis, and in particular to a bioretrosynthesis method and system combining an AND / OR tree with single-step reaction rule prediction, and belongs to the application of an AND / OR tree and a machine learning method in the field of bioretrosynthesis. Background Art

[0002] In the 1960s, Corey proposed the concept of retrosynthesis, the core idea of ​​which is to decompose the target molecule into simpler molecules, which can be combined with chemistry to produce it, and iterate until all the desired compounds are commercially available. Chemical retrosynthesis has been very mature so far, and many commercially available software have been developed to help synthesize high-value compounds, but these software have hardly considered the principles of green chemistry, and the synthetic pathways they find may use or produce raw materials, catalysts, solvents and reagents, products, by-products, etc. that are harmful to human health, community safety, and the ecological environment. In the early 21st century, Hatzimanikatis V proposed biological retrosynthesis, which is different from chemical retrosynthesis mainly in that the reaction is limited to metabolic reactions and the final compound is required to be an available precursor of the chassis strain. Designing hypothetical metabolic pathways based on various chassis strains has also become an important topic in synthetic biology. Biological retrosynthesis involves designing, evaluating and optimizing metabolic pathways to produce high-value compounds from renewable resources and enzymes. Compared with traditional chemical retrosynthesis methods, it is more environmentally friendly and saves raw material costs.

[0003] According to statistics from the MetaNetX database, there are 1,045,319 chemical molecules and 75,699 reactions, of which 39,768 are enzyme-catalyzed reactions, involving more than 30,370 chemical molecules, accounting for only 2.9% of the total number of chemical molecules. There are still a large number of metabolic synthesis pathways for chemical molecules to be explored. Traditional pathway design methods based on biological retrosynthesis require biologists to conduct a large number of wet experiments, and finally find the synthesis pathway through a complete set of processes of design-build-test-learn. This work is very labor-intensive, raw material-intensive, and time-consuming. Therefore, it is of great practical significance to automate biological retrosynthesis with the help of computer technology.

[0004] Existing bioretrosynthesis methods mainly construct biosynthetic pathways by searching for reactions or reaction rules whose reaction products are similar to the molecules to be predicted. They have low computational efficiency and are not suitable for constructing biosynthetic pathways with long reaction steps. Moreover, most of them require complex parameter settings such as thermodynamics, auxiliary factors, and enzyme performance, which requires high domain knowledge. At the same time, existing methods do not evaluate the correctness of the metabolic reactions obtained at each step, making it difficult to guarantee the actual availability of the biosynthetic pathways finally obtained. Summary of the invention

[0005] The present invention proposes a bioretrosynthesis method and system combining an AND / OR tree with single-step reaction rule prediction. By searching and selecting molecular nodes through the AND / OR tree, predicting metabolic reaction rules that can generate product molecules based on a single-step reaction rule prediction model, and expanding the AND / OR tree, the generation of synthesis paths can be achieved, thereby solving the technical problems of low efficiency and low practical usability of the methods in the prior art.

[0006] In order to solve the above technical problems, the first aspect of the present invention provides a bioretrosynthetic method combining an AND / OR tree with single-step reaction rule prediction, comprising:

[0007] S1: Select an OR node from a pre-constructed AND-OR tree, and the molecule corresponding to the OR node is used as the product molecule to be predicted, wherein the pre-constructed AND-OR tree contains two types of nodes, namely, AND nodes and OR nodes, and the AND nodes represent reaction rules, or the OR nodes represent molecules;

[0008] S2: Use a pre-built single-step reaction rule prediction model to predict the k rules in a preset rule set that are most likely to synthesize product molecules, forming a rule set Top-k, and assign a weight value between (0, 1) to each rule, wherein the preset rule set is obtained based on the metabolic reaction structure in a known metabolic reaction data set;

[0009] S3: Expanding the constructed AND-OR tree, including: adding each rule in Top-k as a new AND node to the AND-OR tree, and the OR node selected in step S1 as the parent node of the newly added k AND nodes; for each new AND node, adding all the reaction substrate molecules corresponding to it as OR nodes to the AND-OR tree, and the new AND node as the parent node of these OR nodes;

[0010] S4: Determine whether there is a node in step S3, and the substrate molecules corresponding to its child nodes all belong to the known metabolite set: Sink-Compounds set. If so, find a biosynthetic pathway, stop the iterative retrosynthetic process, and generate the biosynthetic pathway; if not, determine whether the maximum number of iterations has been reached. If so, stop the iterative retrosynthetic process. If not, repeat steps S1 to S4 until a biosynthetic pathway is found or the maximum number of iterations has been reached.

[0011] In one embodiment, in step S1, the root node and leaf nodes of the AND-OR tree are both OR nodes, the intermediate nodes are AND nodes or OR nodes, the child nodes of each AND node are OR nodes, representing all substrate molecules in the reaction rule, and for each non-leaf OR node in the AND-OR tree, its child nodes are all AND nodes, each AND node represents a reaction rule that can produce the molecule corresponding to its parent node, and the root node of the AND-OR tree is the target molecule node whose biosynthetic pathway is to be predicted. The initially constructed AND-OR tree contains only one root node, which is the target molecule node for inversely predicting its biosynthetic pathway.

[0012] In one embodiment, in step S1, each node of the AND-OR tree has a weight value, wherein the weight value of the AND node is the weight value of its corresponding reaction rule, representing the predicted probability of the corresponding reaction rule, and the weight values ​​of other OR nodes except the root node are the weight values ​​of their parent nodes, and the weight value of the root node is set to 1.

[0013] In one embodiment, the OR node selected in step S1 is a leaf node in the OR tree that does not belong to a known Sink-Compounds set and has a maximum weight value. If there are multiple leaf nodes with the maximum weight value, one is randomly selected.

[0014] In one embodiment, the pre-built single-step reaction rule prediction model is a multi-classification model built based on machine learning, and step S2 includes:

[0015] S2.1: Using the constructed multi-classification model, predict the probability that all reaction rules in the preset rule set can produce the input product molecules;

[0016] S2.2: Select the first k reaction rules with the largest probability to form a set Top-k, and set the weight value of each reaction rule in the set Top-k to the corresponding probability value.

[0017] In one embodiment, in step S3, all newly added reaction substrate molecules with the node are obtained by calling a function in the open source library RDChiral, the input parameters of the function are the reaction rules and the SMILES sequence of the product molecules, and the output is a list of corresponding substrate molecules.

[0018] In one embodiment, the biosynthetic pathway process in step S4 comprises:

[0019] (1) For all leaf nodes, check whether they are in the Sink-Compounds. If so, mark the node as "true", otherwise, mark it as "false";

[0020] (2) For a non-leaf AND node in the tree, mark the node as "true" if and only if all its child nodes are marked as "true", otherwise mark it as "false"; for a non-leaf OR node in the tree, mark the node as "true" if and only if it contains at least one child node marked as "true", otherwise mark it as "false";

[0021] (3) If the root node is marked as “false”, it means that no synthetic pathway has been found and the output is “no solution”; otherwise, all nodes marked as “false” in the tree are deleted, and the remaining subtree represents the synthetic pathway of the target molecule.

[0022] Based on the same inventive concept, the second aspect of the present invention provides a bioretrosynthesis system combining an AND / OR tree with single-step reaction rule prediction, comprising:

[0023] A retrosynthesis planning module is used to select an OR node from a pre-constructed AND-OR tree, and the molecule corresponding to the OR node is used as the product molecule to be predicted, wherein the pre-constructed AND-OR tree includes two types of nodes, namely, an AND node and an OR node, and the AND node represents a reaction rule, or the OR node represents a molecule;

[0024] A reaction rule prediction module is used to use a pre-built single-step reaction rule prediction model to predict the k rules in a preset rule set that are most likely to synthesize product molecules, forming a rule set Top-k, and assigning a weight value between (0, 1) to each rule, wherein the preset rule set is obtained based on the metabolic reaction structure in a known metabolic reaction data set;

[0025] The AND-OR tree expansion module is used to expand the constructed AND-OR tree, including: adding each rule in Top-k as a new AND node to the AND-OR tree, and the OR node selected in the retrosynthesis planning module as the parent node of the newly added k AND nodes; for each new AND node, adding all the reaction substrate molecules corresponding to it as OR nodes to the AND-OR tree, and the new AND node as the parent node of these OR nodes;

[0026] The biosynthetic pathway generation module is used to determine whether there is an AND node in the AND-OR tree obtained by the AND-OR tree expansion module, and the substrate molecules corresponding to its child nodes all belong to the known metabolite set: the Sink-Compounds set. If so, a biosynthetic pathway is found, the iterative retrosynthetic process is stopped, and the biosynthetic pathway is generated; if not, it is determined whether the maximum number of iterations has been reached. If so, the iterative retrosynthetic process is stopped. If not, the corresponding steps from the retrosynthesis planning module to the biosynthetic pathway generation module are repeatedly executed until a biosynthetic pathway is found or the maximum number of iterations is reached.

[0027] In one embodiment, the system further comprises: a biosynthetic pathway visualization module for visually displaying the biosynthetic pathway of the target molecule finally obtained.

[0028] The above one or more technical solutions in the embodiments of the present application have at least one or more of the following technical effects:

[0029] The present invention proposes a bioretrosynthesis method combining an AND-OR tree and a single-step reaction rule prediction. First, an OR node is selected as the product molecule to be predicted, and then a pre-constructed single-step reaction rule prediction model is used to predict the product molecule to be predicted, and the k rules in the preset rule set that are most likely to synthesize the product molecule are predicted, and then the AND-OR tree is expanded, and finally the next step of reaction rule prediction is planned based on the AND-OR tree search and the final biosynthetic pathway is generated. The present invention adopts a single-step reaction rule prediction model, which does not require complex parameter settings such as thermodynamics, auxiliary factors, and enzyme performance, and does not require complex domain knowledge, and can achieve reaction rule prediction, thereby improving efficiency; at the same time, using the AND-OR tree to plan rule prediction is helpful to select the metabolic reaction that is most likely to participate in the biosynthetic pathway, thereby improving the actual availability of the biosynthetic pathway finally obtained.

[0030] In addition, the present invention also proposes a biological retrosynthesis system that combines the AND / OR tree with the prediction of single-step reaction rules to assist biologists in quickly finding potential metabolic pathways, reducing experimental costs and improving experimental efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] Figure 1 A flowchart of a bioretrosynthetic method combining an AND / OR tree with single-step reaction rule prediction in an embodiment of the present invention;

[0033] Figure 2 The figure is a schematic diagram of the construction and expansion operation of the AND / OR tree in the method of the embodiment of the present invention. In the AND / OR tree on the left, the node s is the product molecule, and the nodes with the largest weights are a, b, and c; the figure on the right is the AND / OR tree after randomly selecting node a and expanding it. Among them, the circle represents the OR node, the rectangle represents the AND node, the pentagon represents the node to be expanded, and the number on the right side of each node represents its node value;

[0034] Figure 3 SMILES sequence and molecular diagram of the product / substrate molecule in the embodiment of the present invention;

[0035] Figure 4 Schematic diagram of the reaction rule and its corresponding reaction predicted by the single-step reaction rule in the embodiment of the present invention. The upper half of the figure is the reaction rule, and the lower half of the figure is the reaction corresponding to the reaction rule;

[0036] Figure 5 A schematic diagram of the biosynthetic pathway of the target molecule obtained by the method of the embodiment of the present invention. The circles represent the target product molecules, the rectangles represent the intermediate molecules, and the diamonds represent the sink compounds;

[0037] Figure 6 This is a schematic diagram of the modules of the bioretrosynthesis system combining the AND / OR tree with the single-step reaction rule prediction in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In view of the deficiencies in the prior art, the present invention proposes a bioretrosynthesis method that combines an AND-OR tree with a single-step reaction rule prediction. The single-step reaction rule prediction model is used to predict the metabolic reaction rules that can produce product molecules and expand the AND-OR tree, and to predict the next step of reaction rules based on the AND-OR tree search and generate the final biosynthetic pathway. This method uses a single-step reaction rule prediction model, which does not require complex domain knowledge to predict metabolic reaction rules one by one, and can achieve the prediction of the Top-k reaction rules of product molecules. At the same time, using the AND-OR tree to plan rule predictions helps to select the metabolic reactions that are most likely to be involved in the biosynthetic pathway, thereby improving the actual availability of the biosynthetic pathway finally obtained.

[0039] The present invention also proposes a bioretrosynthesis system that combines the AND / OR tree with the prediction of single-step reaction rules, which is used to assist biologists in quickly finding potential metabolic pathways, reducing experimental costs and improving experimental efficiency.

[0040] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] Embodiment 1

[0042] The embodiment of the present invention provides a bioretrosynthesis method combining an AND / OR tree with single-step reaction rule prediction, comprising:

[0043] S1: Select an OR node from a pre-constructed AND-OR tree, and the molecule corresponding to the OR node is used as the product molecule to be predicted, wherein the pre-constructed AND-OR tree contains two types of nodes, namely, AND nodes and OR nodes, and the AND nodes represent reaction rules, or the OR nodes represent molecules;

[0044] S2: Use a pre-built single-step reaction rule prediction model to predict the k rules in a preset rule set that are most likely to synthesize product molecules, forming a rule set Top-k, and assign a weight value between (0, 1) to each rule, wherein the preset rule set is obtained based on the metabolic reaction structure in a known metabolic reaction data set;

[0045] S3: Expanding the constructed AND-OR tree, including: adding each rule in Top-k as a new AND node to the AND-OR tree, and the OR node selected in step S1 as the parent node of the newly added k AND nodes; for each new AND node, adding all the reaction substrate molecules corresponding to it as OR nodes to the AND-OR tree, and the new AND node as the parent node of these OR nodes;

[0046] S4: Determine whether there is a node in step S3, and the substrate molecules corresponding to its child nodes all belong to the known metabolite set: Sink-Compounds set. If so, find a biosynthetic pathway, stop the iterative retrosynthetic process, and generate the biosynthetic pathway; if not, determine whether the maximum number of iterations has been reached. If so, stop the iterative retrosynthetic process. If not, repeat steps S1 to S4 until a biosynthetic pathway is found or the maximum number of iterations has been reached.

[0047] Specifically, step S1 is retrosynthetic planning, S2 is reaction rule prediction, S3 is AND / OR tree expansion, and S4 is biosynthetic pathway generation.

[0048] The prediction preset rule set R is the set required for the prediction of biological retrosynthetic pathways. Each reaction rule in the set R is mainly composed of one or more substrate molecules and one product molecule, which means that one or more substrate molecules can produce corresponding product molecules after reaction.

[0049] The initial AND-OR tree in S1 contains only one root node, which is the OR node representing the target molecule to be predicted, and then it is expanded.

[0050] The present invention decomposes the generation of the retrosynthetic pathway of the target molecule into multi-step single-step reaction rule predictions. In the single-step reaction prediction, the product molecule input for the first time is the target molecule, and the product molecule input into the model afterwards is an intermediate molecule in the reaction chain, that is, the substrate molecule of the reaction rule predicted in the previous step is used as the product molecule of the reaction rule to be predicted at that time. For example, if it is necessary to predict the synthesis path of molecule C, molecule C is input as the target molecule into the single-step reaction rule prediction model, and k rules that are most likely to synthesize molecule C in the preset rule set are obtained. Taking the rule R1 with the highest probability as an example, the reaction corresponding to the reaction rule is selected, for example, its substrates are A and B, that is, A and B can form C through the reaction rule R1; then the next single-step reaction prediction is performed, and the substrate A is input as the product molecule into the single-step reaction rule prediction model, and the k rules that can synthesize A are predicted. For B, it is also a similar processing process to B, and so on, until the condition of step S3 is met: a biosynthetic pathway that can generate molecule C is found, or the maximum number of iterations is reached.

[0051] In one embodiment, in step S1, the root node and leaf nodes of the AND-OR tree are both OR nodes, the intermediate nodes are AND nodes or OR nodes, and the child nodes of each AND node are OR nodes, representing all substrate molecules in the reaction rule. For each non-leaf OR node in the AND-OR tree, its child nodes are all AND nodes, and each AND node represents a reaction rule that can produce the molecule corresponding to its parent node. The root node of the AND-OR tree is the product molecule node whose biosynthetic pathway is to be predicted.

[0052] Specifically, the AND-OR tree contains two types of nodes: AND and OR. The AND node represents a reaction rule, also called a rule node, and the OR node represents a molecule, also called a molecule node.

[0053] In one embodiment, in step S1, each node of the AND-OR tree has a weight value, wherein the weight value of the AND node is the weight value of its corresponding reaction rule, representing the predicted probability of the corresponding reaction rule, and the weight values ​​of other OR nodes except the root node are the weight values ​​of their parent nodes, and the weight value of the root node is set to 1.

[0054] In one embodiment, the OR node selected in step S1 is a leaf node in the OR tree that does not belong to a known Sink-Compounds set and has a maximum weight value. If there are multiple leaf nodes with the maximum weight value, one is randomly selected.

[0055] In one embodiment, the pre-built single-step reaction rule prediction model is a multi-classification model built based on machine learning, and step S2 includes:

[0056] S2.1: Using the constructed multi-classification model, predict the probability that all reaction rules in the preset rule set can produce the input product molecules;

[0057] S2.2: Select the first k reaction rules with the largest probability to form a set Top-k, and set the weight value of each reaction rule in the Top-k to the corresponding probability value.

[0058] In one embodiment, in step S3, all newly added reaction substrate molecules with the node are obtained by calling a function in the open source library RDChiral, the input parameters of the function are the reaction rules and the SMILES sequence of the product molecules, and the output is a list of corresponding substrate molecules.

[0059] The function in RDChiral is the rdchiralRunText function.

[0060] In one embodiment, the biosynthetic pathway process in step S4 comprises:

[0061] (1) For all leaf nodes, check whether they are in the Sink-Compounds. If so, mark the node as "true", otherwise, mark it as "false";

[0062] (2) For a non-leaf AND node in the tree, mark the node as "true" if and only if all its child nodes are marked as "true", otherwise mark it as "false"; for a non-leaf OR node in the tree, mark the node as "true" if and only if it contains at least one child node marked as "true", otherwise mark it as "false";

[0063] (3) If the root node is marked as “false”, it means that no synthetic pathway has been found and the output is “no solution”; otherwise, all nodes marked as “false” in the tree are deleted, and the remaining subtree represents the synthetic pathway of the target molecule.

[0064] In the specific implementation process, Sink-Compounds refers to metabolites available in precursors in the culture medium or in the host organism / strain. Its data comes from the BiGG database. By integrating the data of iML1515, iJO1366, Escherichia coli core metabolism and Bacillus subtilis iYO844, a Sink-Compounds data set is constructed to determine whether the substrate molecules of biological retrosynthesis are available and whether further iterative retrosynthesis is required.

[0065] See also Figure 1 , which is a flow chart of a bioretrosynthesis method combining an AND / OR tree with single-step rule prediction provided by an embodiment of the present invention.

[0066] In the specific application process, the AND-OR tree constructed by S1 contains two types of nodes, AND and OR. AND nodes represent metabolic reaction rule nodes, and OR nodes represent molecule nodes. The root node and leaf nodes of the AND-OR tree are both OR nodes, and the intermediate nodes have both AND nodes and OR nodes. The child nodes of each AND node in the AND-OR tree are OR nodes, representing all substrate molecules in the reaction rule. For each non-leaf OR node in the AND-OR tree, its child nodes are AND nodes, and each AND node represents a reaction rule that can produce the molecule corresponding to its parent node. The root node of the AND-OR tree is the product molecule node whose biosynthetic pathway is to be predicted. Each node of the AND-OR tree has a weight value. The weight value of the AND node is the weight value of its corresponding rule, and the weight value of other OR nodes except the root node is the weight value of its parent node. The weight value of the root node is set to 1.

[0067] The OR node selected in the AND-OR tree in step S1 is a leaf node that does not belong to the Sink-Compounds set and has the largest weight value. If there are multiple nodes, one is selected randomly. See Figure 2 , the left figure is the current AND-OR tree in the embodiment of the present invention, and the OR node a with the largest weight and not in the Sink-Compounds set is selected as the new product molecule node to be predicted, so as to realize the planning of single-step reaction rule prediction for the next step.

[0068] The single-step reaction rule prediction model B in step S2 includes two steps: (1) a multi-classification model constructed based on machine learning is used to simultaneously predict the probability that all reaction rules in the known metabolic reaction rule set R can produce the input product molecules; (2) the top k rules with the largest probability are selected to form a set Top-k, and the weight value of each rule is set to its probability value.

[0069] In the specific implementation process, the metabolic reaction rule set R in (1) is constructed based on the MetaNetX metabolic reaction database. First, the database is cleaned; then the template_extractor function in the open source library RDChiral is called to obtain the reaction rules corresponding to the reaction. The input of the function is the specific metabolic reaction, and the output is the corresponding reaction rule. The specific steps of data cleaning and rule set generation are as follows:

[0070] 1) If the reaction has multiple products, then represent the reaction as multiple reactions, each with only one product;

[0071] 2) Remove the reactions for which no corresponding SMILES sequences can be found for the products and substrates;

[0072] 3) Removal of reactions where the products are radicals or clusters of molecules;

[0073] 4) Removal of reactions whose products are cofactors (such as ions, water, carbon dioxide, ATP, NADP, etc.);

[0074] 5) Eliminate reactions where the number of substrates exceeds 3;

[0075] 6) Remove reactions where the product or substrate SMILES sequence length exceeds 250;

[0076] 7) Use the RXNMapper tool to add atomic mapping information to all reactions after cleaning and remove reactions that failed to map;

[0077] 8) Use the template_extractor function in RDChiral to extract the reaction rules for each reaction and remove the reactions for which the extraction rules failed;

[0078] 9) For each reaction rule, call the rdchiralRunText function in RDChiral to obtain all possible substrates, and remove the reaction rules that cannot obtain the real substrate in the reaction corresponding to the rule.

[0079] In specific implementations, the single-step reaction rule prediction model B can be constructed using a traditional machine learning algorithm or a deep learning algorithm.

[0080] As a specific implementation scheme, the single-step reaction rule prediction model B is constructed using a traditional machine learning algorithm and a deep learning algorithm. Specifically, the traditional machine learning algorithm uses a Gaussian naive Bayes classifier, and the deep learning algorithm uses a convolutional neural network ResNet50. The single-step reaction rule prediction model B is constructed based on the two algorithms as an example for explanation. The input of the model is the product molecule to be predicted represented by a SMILES sequence, and the output of the model is the probability that each reaction rule in the metabolic reaction rule set R can generate the product molecule.

[0081] Specific Example 1

[0082] The following is an introduction to the training process of the single-step reaction rule prediction model B using the traditional machine learning algorithm through a specific example, which specifically includes the following steps:

[0083] (1) Obtain training samples. Collect the product molecules corresponding to all reactions in the reaction rule set, remove duplicate reaction rules corresponding to the product molecules, and label them as their categories;

[0084] (2) Determine the characteristic attributes. Calculate the molecular characteristics of the product molecules, divide the molecular characteristics into multiple discrete characteristic attributes x, and input them into the Gaussian Naive Bayes classifier;

[0085] (3) The classifier calculates the frequency of each category in the training sample, and uses this frequency as the probability P(y i ), where y represents the category, y irepresents the i-th category, P(y i ) represents the probability that the training sample belongs to the i-th category;

[0086] (4) The classifier calculates the conditional probability P(x|y) of all partitions of each feature attribute x in the training sample. i ).

[0087] (5) The classifier calculates P(x|y) for each category i )·P(y i ), which is the probability that each reaction rule can synthesize a product molecule.

[0088] In the above step (2), the molecular features of the product molecules specifically refer to the MACCSKeys fingerprint, which has a total of 166 features, each of which has a corresponding specific meaning, such as the first feature indicates whether there are isomers, the 14th feature indicates whether there are disulfide bonds, etc. The MACCSKeys fingerprint with a length of 167 is calculated by the open source toolkit RDKit, where the 0th position is a placeholder and does not provide any information, so the 0th position is deleted and only 166 features are retained. Correspondingly, each of the 166 features is taken as a feature attribute x, x∈{0,1}.

[0089] In the above step (4), the conditional probability P(x|y i ) assumes that the feature attribute x follows a Gaussian distribution and is calculated based on the probability density function of the Gaussian distribution, which is:

[0090]

[0091] Among them, x represents the feature attribute value, σ represents the standard deviation of the Gaussian distribution, and μ represents the mean of the Gaussian distribution.

[0092] Specific Example 2

[0093] The following is an introduction to the training process of the single-step reaction rule prediction model B using a deep learning algorithm through a specific example, which specifically includes the following steps:

[0094] (1) Obtain training samples. Collect the product molecules corresponding to all reactions in the reaction rule set, remove duplicate reaction rules corresponding to the product molecules, and label them as their categories;

[0095] (2) Calculate the molecular features of the product molecules and input them into the convolutional neural network ResNet50 classifier;

[0096] (3) The classifier outputs the probability that each reaction rule can synthesize a product molecule;

[0097] In the above step (2), the molecular feature F2 of the product molecule is obtained by enhancing and expanding the molecular feature F1 in the specific example 1. The specific steps are as follows: the molecular feature F1 at position 166 is reversed to obtain a new molecular feature F1 ’ ; Longitudinal splicing molecular features F1 and F1 ’ Enhanced molecular features F aug , whose dimension is [2,166]; the longitudinal repeated molecular feature F aug , and the final molecular feature F is obtained, whose dimension is [166,166].

[0098] In the above step (2), in order to adapt to the dimension of the molecular feature F, the number of input channels of the ResNet50 classifier is changed to 1, and AvgPool2d is changed to AdaptiveAvgPool2d. In addition, the output of the last fully connected layer is modified to the number of reaction rule categories.

[0099] Specifically, the main idea of ​​ResNet50 is residual learning. It proposes two mapping methods, namely residual mapping and self mapping. The objective function combines these two parts. The specific formula is as follows:

[0100] h(x)=(h(x)-x))+x

[0101] Where h(x) represents the objective function, x represents the self-mapping, and h(x)-x represents the residual mapping.

[0102] ResNet50 uses the ReLU activation function, which is specifically expressed as:

[0103] f(x)=max(0,x)

[0104] Where x represents the input of the neuron, which can change all negative values ​​to 0 while keeping the positive values ​​unchanged. This unilateral inhibition function can make the neurons in the neural network sparsely activated.

[0105] During the specific implementation process, the single-step reaction rule prediction model data set in Specific Example 1 and Specific Example 2 consists of a given product molecule and a corresponding generated substrate molecule, and the reaction formed by the two. The data comes from the MetaNetX database, which stores metabolites and related biochemical reactions, provides cross-links between major public biochemical and genome-scale metabolic network databases, and contains approximately 80,000 single-step reactions (including reactions without chemical structures). Therefore, the data is cleaned at multiple levels, and the total data set after cleaning is 30,986, involving 25,415 reactions, of which 21,704 reactions have enzyme catalysis information, and a total of 15,930 reaction rules are extracted.

[0106] In the implementation of specific example 2, the training process of the ResNet50 classifier is to find the optimal value of the model based on the loss function and the gradient descent method. The loss function adopts the cross entropy loss function to quantify the difference between the model prediction and the true label. The gradient descent method adopts the adaptive momentum estimation algorithm Adam to find a set of parameters that can minimize structural risks.

[0107] The above cross entropy loss function is specifically expressed as:

[0108]

[0109] Among them, y represents the true label, which takes the value of 0 or 1. Indicates the probability that the sample prediction is positive; the greater the difference between the predicted output and y, the larger the J value.

[0110] The above-mentioned gradient descent method uses the adaptive momentum estimation algorithm Adam, which can iteratively update the neural network weights based on the training data. The Adam algorithm can be considered as a combination of RMSprop and Momentum. It not only uses the momentum method to update the parameters, but also can adaptively adjust the learning rate. In the back propagation process of the neural network, the momentum update method no longer relies solely on the descending gradient of the current parameters to update the parameters, but also relies on the parameters of the previous rounds and the descending gradient of the parameters to update the current parameters. Therefore, when the objective function is reached, the descent speed can be slowed down, making it easier to find the optimal value, and effectively alleviating the oscillation in the gradient descent process that seriously affects the optimization speed. At the tth iteration, the parameter update formula is:

[0111] θ t+1 =θ t -η·ΔJ(θ t )

[0112] Where η is the learning rate, θ t is the parameter of the tth round, J(θ t ) is the loss function, ΔJ(θ t ) is the updated gradient. The actual update difference of each parameter depends on the weighted average of the gradients in the recent period of time, θ t+1 Represents the parameters of the t+1th round.

[0113] In the implementation process of specific example 2, when training the single-step reaction rule prediction model, the training set, validation set, and test set are based on the single-step reaction rule prediction model data set, and are randomly divided in a ratio of 8:1:1. The number of training rounds L is set to 50, and the Adam optimizer is used for gradient descent calculation. The initial learning rate is 0.001, and the batch_size value is 1024. After L rounds of training, the parameters and results of the round with the smallest loss value in the validation set are obtained.

[0114] See also Figures 3-4 , Figure 3 The SMILES sequences of the product / substrate molecules used in Specific Example 1 and Specific Example 2 and their corresponding molecular graphs; Figure 4 The reaction rules and the reactions corresponding to the rules obtained by using the single-step reaction rule prediction model for specific example 1 and specific example 2.

[0115] In one implementation, the k value of Top-k in step S2 is set to 50, and the maximum number of iterations in step S4 is set to 50.

[0116] In step S3, i All reaction substrates are obtained by calling the rdchiralRunText function in the open source library RDChiral. The input parameters of this function are the reaction rules and the SMILES sequence of the product molecules, and the output is the corresponding substrate molecule list.

[0117] See also Figure 2 The right figure in FIG. 1 is an AND / OR tree obtained by selecting a leaf or node a in the left figure, using the prediction model of step S2 to obtain relevant rules, and then expanding it using step S3.

[0118] In this embodiment, the preset rule set in step S2 is a Sink-Compounds set, which refers to metabolites available in precursors in the culture medium or in the host organism / strain, and its data comes from the BiGG database. By integrating the data of iML1515, iJO1366, Escherichia coli core metabolism and Bacillus subtilis iYO844, a Sink-Compounds data set is constructed to determine whether the substrate molecules of biological retrosynthesis are available and whether further iterative retrosynthesis is required. The specific data distribution is shown in Table 1.

[0119] Table 1: Number of Sink Compounds of Various Bacteria

[0120] iML1515 iJO1366 iYO844 Escherichia coli core metabolism 738 727 466 53

[0121] The biosynthetic pathway of step S4 is obtained by searching the AND / OR tree obtained last. It includes the following steps:

[0122] (1) For all leaf nodes, check whether they are in the Sink-Compounds. If so, mark the node as "true", otherwise, mark it as "false";

[0123] (2) For a non-leaf AND node in the tree, mark the node as "true" if and only if all its child nodes are marked as "true", otherwise mark it as "false"; for a non-leaf OR node in the tree, mark the node as "true" if and only if it contains at least one child node marked as "true", otherwise mark it as "false";

[0124] (3) If the root node is marked as “false”, it means that no synthetic pathway has been found and the output is “no solution”; otherwise, all nodes marked as “false” in the tree are deleted, and the remaining subtree represents the synthetic pathway of the target molecule.

[0125] See also Figure 5 , is a schematic diagram of the biosynthetic pathway of the target molecule obtained in the examples of the present invention.

[0126] Compared with the prior art, the beneficial effects of the present invention are:

[0127] 1. The present invention is the first method in the field of bioretrosynthesis that combines the AND / OR tree and the single-step reaction rule prediction model to predict bioretrosynthesis. It does not require complex field parameter settings, solves the problems of high experimental cost and long time consumption of traditional pathway design methods, and can assist biologists in finding potential synthesis pathways more quickly.

[0128] 2. The present invention adopts an AND / OR tree structure to represent the retrosynthesis planning process. Reactions can appear explicitly as nodes in the tree, which can capture the relationship between candidate reactions and substrate molecules and solve the sparsity of variance estimation caused by representing each node in the Monte Carlo tree as a group of molecules.

[0129] Embodiment 2

[0130] Based on the same inventive concept, this embodiment provides a biological retrosynthesis system combining an AND / OR tree with single-step reaction rule prediction, comprising:

[0131] A retrosynthesis planning module is used to select an OR node from a pre-constructed AND-OR tree, and the molecule corresponding to the OR node is used as the product molecule to be predicted, wherein the pre-constructed AND-OR tree includes two types of nodes, namely, an AND node and an OR node, and the AND node represents a reaction rule, or the OR node represents a molecule;

[0132] A reaction rule prediction module is used to use a pre-built single-step reaction rule prediction model to predict the k rules in a preset rule set that are most likely to synthesize product molecules, forming a rule set Top-k, and assigning a weight value between (0, 1) to each rule, wherein the preset rule set is obtained based on the metabolic reaction structure in a known metabolic reaction data set;

[0133] The AND-OR tree expansion module is used to expand the constructed AND-OR tree, including: adding each rule in Top-k as a new AND node to the AND-OR tree, and the OR node selected in the retrosynthesis planning module as the parent node of the newly added k AND nodes; for each new AND node, adding all the reaction substrate molecules corresponding to it as OR nodes to the AND-OR tree, and the new AND node as the parent node of these OR nodes;

[0134] The biosynthetic pathway generation module is used to determine whether there is an AND node in the AND-OR tree obtained by the AND-OR tree expansion module, and the substrate molecules corresponding to its child nodes all belong to the known metabolite set: the Sink-Compounds set. If so, a biosynthetic pathway is found, the iterative retrosynthetic process is stopped, and the biosynthetic pathway is generated; if not, it is determined whether the maximum number of iterations has been reached. If so, the iterative retrosynthetic process is stopped. If not, the corresponding steps from the retrosynthesis planning module to the biosynthetic pathway generation module are repeatedly executed until a biosynthetic pathway is found or the maximum number of iterations is reached.

[0135] Specifically, before the synthesis pathway is generated, some pre-operations are included, including reaction rule set construction, Sink-Compounds set construction and data preprocessing, including: a reaction rule set construction module for constructing the rule set required for biosynthetic pathway prediction, a Sink-Compounds set construction for constructing a set of precursors available in the culture medium or metabolites available in the host organism / strain, and a data preprocessing module for generating molecular features (such as MACCSKeys fingerprints) based on the SMILES sequence of the molecule before the product molecule is input into the model.

[0136] In general, the first aspect of the present invention discloses a biological retrosynthesis method combining an AND-OR tree with a single-step reaction rule prediction. When the target molecule is retrosynthesized, the method decomposes the retrosynthetic pathway generation of the target molecule into multiple single-step reaction rule predictions. In the single-step reaction rule prediction, the substrate molecule of the reaction rule predicted in the previous step is used as the product molecule of the reaction rule to be predicted at that time. The SMILES sequence of the product molecule is used as input, and the molecular characteristics of the product molecule are calculated by custom calculation to be compatible with a variety of single-step reaction rule prediction models. According to the probability that each reaction rule predicted by the model can synthesize the product molecule, the top k reaction rules with the highest probability are selected, and then expanded based on the AND-OR tree structure. Finally, the potential synthesis pathway of the target molecule is found. The second aspect of the present invention also proposes a system of a biological retrosynthesis method based on the above-mentioned AND-OR tree search strategy and a convolutional network, which realizes the automation of the metabolic pathway retrosynthesis of the target molecule through modular processing processes such as reaction rule set construction, Sink-Compounds set construction, data preprocessing, reaction rule prediction, AND-OR tree expansion, retrosynthesis planning, biosynthetic pathway generation and visualization. The present invention significantly improves the speed of finding feasible synthetic pathways, does not require manual setting of complex parameters, and can assist biologists in finding potential metabolic pathways more quickly, thereby reducing experimental costs and improving experimental efficiency.

[0137] In one embodiment, the system further comprises: a biosynthetic pathway visualization module for visually displaying the biosynthetic pathway finally obtained.

[0138] Retrosynthesis planning module: Search the leaf nodes of the current AND / OR tree, select a leaf or node from it as the product molecule predicted by the next round of single-step reaction rules, and realize the planning search of the retrosynthetic pathway; Reaction rule prediction module: Use the built prediction model to predict the k rules that are most likely to produce product molecules; AND / OR tree expansion module: According to the type of node to be expanded, expand the AND / OR tree based on the OR node expansion rule or the AND node expansion rule. Biosynthetic pathway generation module: Search and finally obtain the AND / OR tree to obtain the biosynthetic pathway of the target molecule; Visualization of biosynthetic pathway: Visualize the final biosynthetic pathway for reference by experts in the field of biosynthesis.

[0139] See also Figure 6 , is a module schematic diagram of a bioretrosynthesis system combining an AND / OR tree with single-step reaction rule prediction in an embodiment of the present invention.

[0140] Since the system introduced in the second embodiment of the present invention is a system used to implement a biological retrosynthesis method combining an AND / OR tree and a single-step reaction rule prediction in the first embodiment of the present invention, those skilled in the art can understand the specific structure of the system based on the method introduced in the first embodiment of the present invention, and no further description is given here. All systems used in the method of the first embodiment of the present invention belong to the scope of protection of the present invention.

[0141] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. A bioretrosynthetic method combining an AND / OR tree with single-step reaction rule prediction, characterized in that: include: S1: Select an OR node from a pre-constructed AND-OR tree, and the molecule corresponding to the OR node is used as the product molecule to be predicted, wherein the pre-constructed AND-OR tree contains two types of nodes, namely, AND nodes and OR nodes, and the AND nodes represent reaction rules, or the OR nodes represent molecules; S2: Use a pre-built single-step reaction rule prediction model to predict the k rules in a preset rule set that are most likely to synthesize product molecules, forming a rule set Top-k, and assign a weight value between (0, 1) to each rule, wherein the preset rule set is obtained based on the metabolic reaction structure in a known metabolic reaction data set; S3: Expanding the constructed AND-OR tree, including: adding each rule in Top-k as a new AND node to the AND-OR tree, and the OR node selected in step S1 as the parent node of the newly added k AND nodes; for each new AND node, adding all the reaction substrate molecules corresponding to it as OR nodes to the AND-OR tree, and the new AND node as the parent node of these OR nodes; S4: Determine whether there is a node in step S3, and the substrate molecules corresponding to its child nodes all belong to the known metabolite set: Sink-Compounds set. If so, find a biosynthetic pathway, stop the iterative retrosynthetic process, and generate the biosynthetic pathway; if not, determine whether the maximum number of iterations has been reached. If so, stop the iterative retrosynthetic process. If not, repeat steps S1 to S4 until a biosynthetic pathway is found or the maximum number of iterations has been reached.

2. The bioretrosynthesis method according to claim 1, characterized in that: In step S1, the root node and leaf nodes of the AND-OR tree are both OR nodes, the intermediate nodes are AND nodes or OR nodes, the child nodes of each AND node are OR nodes, representing all substrate molecules in the reaction rule, for each non-leaf OR node in the AND-OR tree, its child nodes are all AND nodes, each AND node represents a reaction rule that can produce the molecule corresponding to its parent node, and the root node of the AND-OR tree is the target molecule node whose biosynthetic pathway is to be predicted; the initially constructed AND-OR tree contains only one root node, which represents the target molecule node for inverse prediction of its biosynthetic pathway.

3. The bioretrosynthesis method according to claim 1, characterized in that: In step S1, each node of the AND-OR tree has a weight value, wherein the weight value of the AND node is the weight value of its corresponding reaction rule, indicating the predicted probability of the corresponding reaction rule, and the weight values ​​of other OR nodes except the root node are the weight values ​​of their parent nodes, and the weight value of the root node is set to 1.

4. The bioretrosynthesis method according to claim 3, characterized in that: The OR node selected in step S1 is a leaf node in the AND tree that does not belong to the known Sink-Compounds set and has the maximum weight value. If there are multiple leaf nodes with the maximum weight value, one is randomly selected.

5. The bioretrosynthesis method according to claim 1, characterized in that: The pre-built single-step reaction rule prediction model is a multi-classification model built based on machine learning. Step S2 includes: S2.1: Using the constructed multi-classification model, predict the probability that all reaction rules in the preset rule set can produce the input product molecules; S2.2: Select the first k reaction rules with the largest probability to form a set Top-k, and set the weight value of each reaction rule in the set Top-k to the corresponding probability value.

6. The bioretrosynthesis method according to claim 1, characterized in that: In step S3, all newly added reaction substrate molecules with the node are obtained by calling a function in the open source library RDChiral, the input parameters of which are the reaction rules and the SMILES sequence of the product molecules, and the output is a list of corresponding substrate molecules.

7. The bioretrosynthesis method according to claim 1, characterized in that: The biosynthetic pathway process in step S4 includes: (1) For each leaf node, check whether it is in the Sink-Compounds. If so, mark the node as "true", otherwise, mark it as "false"; (2) For a non-leaf AND node in the tree, mark the node as "true" if and only if all its child nodes are marked as "true", otherwise mark it as "false"; for a non-leaf or node in the tree, mark the node as "true" if and only if it contains at least one child node marked as "true", otherwise mark it as "false"; (3) If the root node is marked as "false", it means that no synthetic pathway has been found and the output is "no solution". Otherwise, delete all nodes marked as "false" in the tree, and the remaining subtree represents the synthetic pathway of the target molecule.

8. A bioretrosynthesis system combining an AND / OR tree with single-step reaction rule prediction, characterized in that: include: A retrosynthesis planning module is used to select an OR node from a pre-constructed AND-OR tree, and the molecule corresponding to the OR node is used as the product molecule to be predicted, wherein the pre-constructed AND-OR tree includes two types of nodes, namely, an AND node and an OR node, and the AND node represents a reaction rule, or the OR node represents a molecule; A reaction rule prediction module is used to use a pre-built single-step reaction rule prediction model to predict the k rules in a preset rule set that are most likely to synthesize product molecules, forming a rule set Top-k, and assigning a weight value between (0, 1) to each rule, wherein the preset rule set is obtained based on the metabolic reaction structure in a known metabolic reaction data set; The AND-OR tree expansion module is used to expand the constructed AND-OR tree, including: adding each rule in Top-k as a new AND node to the AND-OR tree, and the OR node selected in the retrosynthesis planning module as the parent node of the newly added k AND nodes; for each new AND node, adding all the reaction substrate molecules corresponding to it as OR nodes to the AND-OR tree, and the new AND node as the parent node of these OR nodes; The biosynthetic pathway generation module is used to determine whether there is an AND node in the AND-OR tree obtained by the AND-OR tree expansion module, and the substrate molecules corresponding to its child nodes all belong to the known metabolite set: the Sink-Compounds set. If so, a biosynthetic pathway is found, the iterative retrosynthetic process is stopped, and the biosynthetic pathway is generated; if not, it is determined whether the maximum number of iterations has been reached. If so, the iterative retrosynthetic process is stopped. If not, the corresponding steps from the retrosynthesis planning module to the biosynthetic pathway generation module are repeatedly executed until a biosynthetic pathway is found or the maximum number of iterations is reached.

9. The bioretrosynthesis system according to claim 8, characterized in that: The system also includes: a biosynthetic pathway visualization module, which is used to visualize the biosynthetic pathway of the target molecule finally obtained.

Citation Information

Patent Citations

  • Biological reaction information processing system and biological reaction information processing method

    US20200303042A1

  • Bioretrosynthetic method and system based on and-or tree and single-step reaction template prediction

    US20230197205A1

  • Retrosynthesis prediction using deep highway networks and multiscale reaction classification

    WO2020023650A1

  • KR20210147862A