as more and more genome-scale metabolic networks are reconstructed,more and more data are mined in biology,so more theories,algorithms and tools are needed to support them.Enzymes,As elements with catalytic functions,have great advantages over chemical catalysts,and can form composite catalytic systems through rational design and multiple assembly to achieve cascade catalytic functions^[1,2][3,4]。 Enzyme cascade catalysis can produce various chemical products with high conversion rate in a short time^[5]。 Understanding how selected enzymes convert biological feedstocks into biochemicals in multienzyme cascade catalyzed reactions can serve as an important foundation for the application of biocatalysis,metabolic engineering,and synthetic biology to achieve environmentally friendly green bioprocess^[6~9],while the utilization of enzyme catalysis is also very valuable for the conversion of non-natural compounds^[10]。 the construction of multienzyme cascade reaction systems as well as in vitro organelles has further enabled The synthesis of complex structural compounds,thus solving a series of problems arising from biosynthetic applications^[11⇓~13]。 However,in the process of exploring multi-enzyme synthesis pathways,it is still a challenge to explore the best synthesis reaction pathway based on known products,and there are some limitations such as lack of experience and incomplete data in manual design based on personal prior knowledge.Therefore,based on the information related to multi-enzyme reactions(Figure 1),researchers have developed data-driven tool models for retrosynthetic path analysis of target compounds,and the results of these tool application tests are also excellent^[14]。 It can be seen that automated retrosynthetic pathway analysis and related tool models play a very important role in the study of multienzyme synthetic networks^[15]。 Based on this,in order to facilitate users to choose these tools reasonably and effectively,this paper briefly introduces and summarizes the calculation tools for constructing inverse synthesis paths。

multi-enzyme catalysis is a concurrent Multi-enzyme one-pot biocatalytic reaction and a reaction in which components are added in sequence or the process steps are stretched,because the reaction is environmentally friendly and the cost of the conversion of reaction intermediates is reduced.It has established a bridge between single enzyme catalysis and whole cell catalysis,which has been the focus of researchers in the field of biocatalysis^[16,17]。 the use of enzymatic cascades throughout The reaction is a particularly attractive aspect of biocatalysis because of its general reaction compatibility and ability to achieve multistep reactions in cell-free systems(in vitro)or in whole cells^[18]。 In addition to this,multienzyme catalysis not only reduces substrate transport and reaction time,but also reduces intermediate losses due to diffusion,producing fewer by-products and contaminants^[19]。 in recent years,the application of multi-enzyme catalytic reaction is of great significance to industrial development,and significant progress has been made In the synthesis of drugs,cosmetics and nutritional compounds^[20]。

At present,the multienzyme cascade catalytic system mainly includes in vivo multienzyme catalytic system and in vitro multienzyme catalytic system.multi-enzyme catalysis in vivo mainly refers to the process in which a substrate produces a specific product through Multi-step catalytic reactions under the action of a variety of enzymes in an organism.It can evaluate the applicability of the pathway of the endogenous metabolic system of the host organism in a specific environment,but the range of feasible pathways is limited due to its dependence on host cells.in vitro multienzyme catalysis is a rapidly developing research method in the field of synthetic biology,which is a process of converting a substrate into a target product by using purified enzymes and coenzymes through certain biochemical reactions^[21]。 in vitro multienzyme cascades are usually extracellular reconstitution of natural metabolic pathways,or assembly of different types of artificial enzymes to achieve conversion of substrates to target products.Cell extracts and pure enzymes are commonly used to construct multienzyme cascade catalytic systems In vitro^[22]。 Multienzyme cascades in vitro have many advantages:they have a greater ability to synthesize complex molecules,do not require intermediate separation,and can reduce the shift of unfavorable equilibria to products,such as toxic effects^[23,24]。 However,In a single system,it is challenging to achieve consistent reaction conditions that are optimal for all enzymes.in addition,the stability of enzymes is a big problem,and their performance may be impaired by unstable or inhibitory interactions between cascade components or uncoordinated reaction conditions^[25]。 Therefore,the optimization of such systems is usually inevitable,and many parameters,such as the design of synthetic routes,the selection of enzymes,reaction conditions,or process design,can change the performance of in vitro enzyme cascades(Figure 2),so in vitro multi-enzyme catalytic systems can be optimized by combining experimental methods with data-driven tool models^[26]。

based on the development of related applications of multi-enzyme catalytic reactions,multi-enzyme cascade reaction systems have been widely used in different fields,and there have been many successful cases of designing synthetic biological pathways,but the reaction steps are often relatively long.At the same time,some tool models for the design of biological retrosynthetic pathways have emerged in the field of retrosynthesis.Bioretrosynthesis is mainly used to produce high-value compounds from renewable resources or engineered enzymes,designing optimal synthetic pathways Based on microbial metabolic networks as a source of metabolic pathways in synthetic biology^[27]。 Biological retrosynthetic design first needs to define the target molecule to be synthesized,and then use the consistent biochemical reaction rules as a template to identify potential precursors and enzymes needed for the reaction.the final reaction path is relatively short and the reaction efficiency can be significantly improved^[28]。 In order to be able to explore some unknown areas of viable metabolite transformations,it is necessary to develop computational tools for new bioretrosynthetic pathways。

Through the analysis of the progress of the existing retrosynthetic path design methods,we find that there are three main methods of retrosynthetic path design(Figure 3):path search based on existing knowledge,template-based path construction method,and template-free path construction method^[29]。

Early researchers manually searched for pathways based on existing knowledge,extracting possible biosynthetic Pathways From existing reaction databases(MetaCyc(Metabolic Pathways From all Domains of Life),KEGG(Kyoto Encyclopedia of Genes and Genomes),etc.)and ranking pathways based on experience^[30][31]。 For complex compounds,these methods are usually not applicable when the synthetic path of the compound is not available in the database。

With the further development of data-driven,ML(Machine learning)has become the main tool for retrosynthetic path design,and has been applied to every step of the path design process^[32,33]。 Through the analysis of this kind of retrosynthetic pathway design tools,it is found that the main difference between machine learning based retrosynthetic tools in the process of designing metabolic pathways lies in whether they depend on templates or not.template-based construction of synthetic pathways is where Template-based or rule-based models(NovoStoic,RetroPathRL,etc.)match a query molecule to a collection of generalized reaction rules,i.e.,subgraph patterns of molecules highlight changes during a biochemical reaction^[34][35][36]。 Rules were manually summarized by experts or extracted from the reaction database^[37]。 Its principle is based on the template conversion of target molecules into reactants according to the reaction rules in the template library.While rule-based approaches have yielded promising results in reverse transcription biosynthesis,neglecting the effect of long-range substituents on the reaction center,in addition,reaction rules that are too specific or too general will result in predicted routes that are too conservative or unrealistic,respectively,requiring laborious and time-consuming optimization by experts,and they cannot predict reactions outside the database^[38][39]。 the Template-free synthetic pathway uses a model without a predefined reaction template to predict the precursor,and its principle is to convert the SMILES sequence of the product molecule into the SMILES sequence of the reactant molecule based on the molecular sequence(such as the SMILES sequence)or to divide the chemical bonds of the product based on the molecular map and add them appropriately to form the reactant.template-free retrosynthesis tools(BioNavi-NP et al.)use reaction databases to train ML models to predict precursors by inputting molecules of interest^[40]。

Whether template-dependent or template-free,the main workflow of the machine learning-based retrosynthetic pathway design tool is achieved through four steps(Figure 4)。

(1)Design metabolic pathways and generate metabolic networks.These tools first combine known reactions in the database,retrieve the desired molecule through homology alignment and construct new metabolic pathways based on reaction rules,or convert to possible precursors based on the sequence and structural random bond breakage of the target molecule,in which various prediction algorithms are a very important driver。

(2)Refine the data.the metabolic network generated in the first step will contain all possible reactants and enzymes to generate the target product,and random bond breaking based on the molecular structural formula alone may export candidate precursors that do not exist in nature.Retrosynthetic tools need to evaluate the most likely reactions and optimal reactants,during which unnecessary substances and metabolic pathways are removed。

(3)identify and prune paths.By comparing the path in the metabolic network with the information in the database containing biological metabolic pathways(MetaCyc,KEGG,etc.),we can Identify whether the path belongs to a new path,and in addition,we can obtain the reaction compounds and some biochemical characteristics of the reaction。

and(4)score and prioritizing that generate paths.the results obtained In the first three steps can be used to score and rank the pathways in order to pinpoint the optimal pathway that is most likely to produce the target molecule.in addition to this,the retrosynthetic tool will take into account the toxicity,yield,economics of the compound,and so on,thereby suggesting an optimal route suitable for putting into production^[41]。

Up to now,there are many computational tools for designing biosynthetic metabolic pathways,which are mainly divided into two categories based on practicality and user habits.One is to explore the retrosynthetic pathway of the target compound in the host(such as Escherichia coli,cyanobacteria,etc.),and finally obtain the reaction pathway with short reaction steps,easily available intermediates and low cost.Table 1 is a summary analysis of some host-based retrosynthetic tools。

in addition,a class of tools is the design of multiple pathways for bioretrosynthetic reactions based on known products in an in vitro environment,regardless of the host.Such tools(Table 2)are typically designed for host-free in vitro synthesis pathways。

in the data-driven environment,computer-aided design plays a vital role in enzyme design and pathway construction.At present,machine learning has been widely used in this field,but the traditional machine learning has shown some disadvantages,such as the limitation of autoregressive sampling strategy,the unclear classification of data set types,and the inaccurate information source of the algorithm.With the help of artificial intelligence,this problem has been well solved and applied to the artificial design of enzymes^[77⇓~79]。 the deep learning-based retrosynthetic path construction tool uses a series of algorithms to encode the input molecular objects based on molecular fingerprints(such as SMILES),IUPAC and molecular maps to identify and construct possible precursor structures and analyze the required enzymes^[80,81][82]。 Similar inverse synthesis path design tools have emerged,but the model is still a black box model,and it is difficult for laymen to explain the detailed construction process of the model without the necessary knowledge,and additional knowledge is still needed for interpretability methods^[83]。 In addition,the input of most models is mainly based on SMILES format,which still has problems such as semantic inconsistency and reading errors,which will lead to the reduction of model performance,and double verification can effectively solve this problem^[84,85]。 At present,some models use graph neural network to identify the input molecular structure,and 3D graph neural network has been applied in the medical field,so the double combination of SMILES and graph neural network is expected to solve this problem^[86,87][88]。 for the reverse synthesis tools that rely on the reaction database,there are still some problems in the process of constructing the precursor,such as the amount of reaction database data,the accuracy of the template,the inconsistency of database integration,and the incompleteness of the data,including ambiguity,errors,redundancy,or inconsistency with the literature.Therefore,a workaround is needed for a nonstandardized nomenclature for databases,which requires constant mapping and perhaps periodic updates as new knowledge is discovered.Therefore,artificial intelligence needs to be further optimized in helping multi-enzyme system。

the current most effective approach for designing synthetic routes to known target compounds is to combine chemical and biological retrosynthetic analysis into a single tool.This approach requires not only access to large databases containing catalytic reactions and understanding how biological and chemical catalysts can be combined In a productive way to shorten any synthetic process,Basic research is also needed to understand the advantages and limitations of biological and chemical catalysis in cascaded pathways,so as to effectively solve the problems of catalyst deactivation and harmful by-products in the reaction.in addition,the limitations of computer performance and the understanding of natural metabolism are common problems that slow the pace of rational design of synthetic pathways by retrosynthetic tools.Therefore,we still need to constantly mine new information to improve the database and update the database,and further optimize the development method of reverse synthesis tools to improve performance,achieve higher accuracy,and make greater contributions to the field of biological reverse synthesis。