RetroBioCat as a computer-aided synthesis planning tool for biocatalytic reactions and cascades.

As the enzyme toolbox for biocatalysis has expanded, so has the potential for the construction of powerful enzymatic cascades for efficient and selective synthesis of target molecules. Additionally, recent advances in computer-aided synthesis planning are revolutionising synthesis design in both synthetic biology and organic chemistry. However, the potential for biocatalysis is not well captured by tools currently available in either field. Here we present RetroBioCat, an intuitive and accessible tool for computer-aided design of biocatalytic cascades, freely available at retrobiocat.com. Our approach uses a set of expertly encoded reaction rules encompassing the enzyme toolbox for biocatalysis, and a system for identifying literature precedent for enzymes with the correct substrate specificity where this is available. Applying these rules for automated biocatalytic retrosynthesis, we show our tool to be capable of identifying promising biocatalytic pathways to target molecules, validated using a test-set of recent cascades described in the literature.

Introduction

Biocatalysis is at the nexus of rapidly expanding sequence data, cheaper DNA synthesis, advances in enzyme engineering and a strong need for more sustainable manufacturing processes [1]. Increasingly this means biocatalysis is an attractive option for organic synthesis, particularly where exquisite selectivity is required [2,3]. Mild operating conditions afford enzymes further advantages, in that they can be combined easily into multi-step cascades without costly purification steps, often in a single reactor [4]. Recent industrial examples include cascades for the production of the investigational HIV treatment drug islatravir, as well as a directed evolution campaign towards the synthesis of the Phase II clinical trial drug LSD1 inhibitor GSK2879552 [5,6].

In both organic chemistry and synthetic biology, computer-aided synthesis planning (CASP) tools are increasingly used to plan synthesis routes. Tools such as RetroPath have been deployed to develop new metabolic routes to molecules of interest in synthetic biology [7], whilst in chemistry, tools such as Chematica or ASKCOS have been shown to make useful suggestions for synthetic routes to a number of target molecules [8-10]. Despite successes in both of these fields, which biocatalysis spans, computer aided synthesis planning of biocatalytic cascades remains underdeveloped. Enzymatic steps are not well represented in chemical CASP tools, if they appear at all. In contrast, biological CASP tools predominantly feature biosynthetic enzymes, yet the objective of reaching a metabolic starting point, and the use of reaction rules describing transformations in metabolism, does not align well the use of enzymes for organic synthesis (Figure 1A). Indeed, enzymes in the toolbox for biocatalysis generally have a proven track-record for showing promiscuous substrate specificity, which may not be the case for all enzymes in metabolism.

Figure 1

An overview of the requirements for a biocatalysis CASP tool.

A. A CASP tool for biocatalysis requires elements of both Synthetic Biology and Chemistry. B. In the process of generating a pathway, manual and automated processes can be used synergistically for maximum benefit. CAR: Carboxylic acid reductase, ATP: Adenosine triphosphate, NADPH: Nicotinamide adenine dinucleotide phosphate.

Here we present RetroBioCat (available at retrobiocat.com), a tool for computer-aided synthesis planning of biocatalytic cascades, which builds upon CASP elements developed in both organic chemistry and synthetic biology (Figure 1A). We began the development of RetroBioCat by considering the process a scientist undergoes when planning a new biocatalytic pathway, which typically follows three stages (Figure 1B). In the first step, pathways are generated by biocatalytic retrosynthesis. Importantly, this should incorporate some scope for what enzymes could do with the application of enzyme engineering. Secondly, specific enzymes are identified for each step, before finally, potential pathways are evaluated based on factors such as the availability of starting materials or the number of steps required. In this work we describe tools which attempt to automate parts of this workflow, seeking to augment the abilities of chemists wishing to exploit the power of biocatalysis.

Results

Reaction rules for biocatalysis

For the automated generation of pathways to a target molecule, CASP tools typically rely on the use of reaction rules or templates to describe potential retrosynthetic steps, which are iteratively applied until a suitable stopping point is reached (Figure 1B, Step 1) [11,12]. Rules can be manually entered [9,13,14], or automatically extracted from a database of known reactions [15,16]. Critical to the success of this approach is the level of generalisation applied to the reaction rules. Rules which are too specific limit the potential to predict new routes, whilst rules which are too general can lead to unrealistic suggestions [17,18].

Rules to describe biosynthetic reactions have previously been automatically extracted from databases of metabolic reactions, with the option to specify the level of generalisation through the selection of a diameter from the reaction centre [16]. Integrating these rules into RetroBioCat, we found that whilst this method has clearly been successful in generating new biosynthetic routes [19], literature examples of biocatalytic cascades were either not represented, or required the most extreme promiscuity setting in order to be captured. Crucially, treating all enzymes in metabolism as very promiscuous is unrealistic, yielding a large number of unhelpful results. Additionally, such algorithmically extracted rules typically lack the nomenclature associated with enzymes for biocatalysis. Algorithmic extraction of reaction rules from a thorough database of synthetic biotransformations relevant to biocatalysis may be an option in the future, pending the development of such a database.

As an alternative, we developed a smaller set of expertly encoded reaction rules to describe the enzyme toolbox for biocatalysis (Figure 2A) [20-22]. These rules were made relatively general in most cases to reflect established substrate promiscuity, and to highlight the scope for enzyme engineering. Importantly, the enzymes that these rules represent have been shown to be amenable to enzyme engineering in many cases, for the acceptance of highly synthetic substrates [2,23-26]. In other instances, stricter limits on substrate specificity are known, and here greater context was incorporated into the reaction SMARTS. In addition, the rules include some necessary spontaneous chemical steps. An indication of whether a suggested transformation is immediately applicable, or may require enzyme engineering, is provided by the automatic identification of literature precedents by RetroBioCat (Figure 2B).

Figure 2

Critical components of RetroBioCat.

A. A selection of exemplar reaction rules for industrially relevant enzymes, written as reaction SMARTS. B. An example query for similar reactions present in a database of literature precedents. For each molecule, visualisation of the atomic contributions to the Morgan fingerprint similarity is shown. Below each molecule, the Tanimoto similarity [27] is calculated for (a) RDKit fingerprints and (b) Morgan fingerprints, each using the default settings in RDKit. The best CAR enzyme identified for each reaction is shown. TA: Transaminase, AmDH: Amine dehydrogenase, IRED: Imine reductase, CAR: Carboxylic acid reductase, ADH: Alcohol dehydrogenase, KRED: Keto reductase.

The current rule set consists of 99 reactions, described using 135 reaction SMARTS. RetroBioCat also includes the facility to accept submissions for new rules by members of the biocatalysis community, offering the potential for the community-driven development of reaction rules as the enzyme toolbox expands. Furthermore, reaction suggestions which users feel are unrealistic can be flagged for review.

Molecular similarity for identification of reaction precedents

A system for identifying specific enzyme sequences to carry out each step is also necessary for automated cascade design. Manually, scientists typically rely on extensive literature searches or enzyme screening panels to compete this step (Figure 1B, Step 2). To automate this process, a database of literature precedents for synthetic biotransformations is required. However, whilst there are many well-established enzyme databases, these tend to focus on biosynthetic reactions rather than examples of synthetic biotransformations utilised in biocatalysis. Therefore, to demonstrate this step we have begun the manual curation of a database of synthetic biotransformations, which we aim to expand upon in future work. We created a module within RetroBioCat to score reactions based on their similarity to recorded reactions [27] through the use of fingerprint similarity (Figure 2B), as has been demonstrated in both biology [7,28], and chemistry [8]. Where many enzymes have been shown to catalyse a specific reaction, our approach selects the best as ranked by activity. Whilst the determinants of substrate specificity may be more complex than can be captured by fingerprint similarity alone, the selection of similar substrates allows a chemist to quickly access the relevant information to make the final decision.

Prioritising reactions by change in molecular complexity

Finally, many chemistry CASP tools feature a metric for molecular complexity to help guide the retrosynthetic search towards a simpler starting material. RetroBioCat uses the recently described SC-Score, which utilises a neural network trained on a large number of synthetic chemistry reactions to score the complexity of each molecule between 1 and 5 [29]. Applied to biocatalysis, this score appears to function well in guiding pathway suggestions towards synthetically useful routes (Extended Data Figure 1).

Extended Data Fig. 1

An example generated using Network Explorer to illustrate changes in molecular complexity.

Arrows and reactions are coloured by the change in molecular complexity, determined using the SC-Score. Green indicates a negative change in molecule complexity, which in most cases corresponds to a synthetically useful transformation. Red indicates a positive change in molecular complexity. Colours are determined relative to the other transformations leading to a specific molecule. Some reactions have been removed for clarity. Pathway published in reference 68.

Network and pathway explorer

Having established a set of rules describing important reactions in biocatalysis (Figure 2A), a method for searching for similar reaction literature precedents (Figure 2B), and a complexity metric by which to guide retrosynthetic searches (Extended Data Figure 1), we developed two complementary approaches for exploring potential biocatalytic pathways. Firstly, a network exploration mode for human-led CASP, in which the user can explore different routes to a target molecule by expanding a network of biocatalytic disconnections (Figure 3). Alternatively, a pathway exploration mode, in which pathways are automatically generated before being ranked according to a user-defined weighted score (Figure 4). Importantly, both approaches are primarily available through an interactive web-app, but also as an open-source python package for expert users.

Figure 3

Human-led exploration of a network of potential biotransformations using network explorer.

Each substrate node can be iteratively expanded to reveal further possible biotransformations. Reaction nodes in green indicate high similarity to a literature reported reaction (currently a proof-of-principle dataset). The target molecule is outlined in orange, and buyable compounds outlined in purple. Interactions possible in network explorer are shown. IRED: Imine reductase, AlOx: Alcohol oxidase, ATP: Adenosine triphosphate, AMP: Adenosine monophosphate, PPi: Pyrophosphate, NAD(P): Nicotinamide adenine dinucleotide (phosphate).

Figure 4

- An example selection of some of the biocatalytic cascades identified in the literature and used as a test-set for Pathway explorer.

The remainder of the 52 cascades are available in Extended Data Figures 2–5 [5,31–72]. In some cases, a number of cascades were demonstrated with different R groups, for which we have chosen a single example, highlighted in blue. Pathway rankings by RetroBioCat using a maximum of 4 steps and either the default scoring weights, or default weights but with the weight for number of steps with literature precedent set to zero, are shown. Pathways are marked as identified even where RetroBioCat suggests additional steps. * indicates pathways where the data from the relevant paper has not been added to the database of literature precedent reactions in RetroBioCat. TDL: thiamine-dependent lyase, TA: transaminase, TPL: tyrosine phenol lyase, TAL: tyrosine ammonia lyase, DC: decarboxylase, P450: cytochrome P450, ADH: alcohol dehydrogenase, AmDH: amine dehydrogenase, IRED: imine reductase, PPM: phosphopentomutase, PNP: purine nucleoside phosphorylase, AlOx: alcohol oxidase, BVMO: Baeyer-Villiger monooxygenase, CAR: carboxylic acid reductase, ERED: ene reductase, CMT: C-methyltransferase, AAD: amino acid deaminase, AADH: amino acid dehydrogenase, PAL: phenylalanine ammonia lyase, AmOx: amine oxidase, TrpS: tryptophan synthase, ThDP: thiamine diphosphate, ATP: Adenosine triphosphate, NADP: Nicotinamide adenine dinucleotide phosphate.

In particular, the network exploration mode can be useful for scientists who may not be familiar with biocatalysis, allowing them to visualise potential biocatalytic disconnections to their target molecule. Integrated is the enzyme identification module, which colours reaction nodes green where a similar literature precedent is identified, or red where only negative data has been reported. Further data on substrate specificity, buy-ability or molecular complexity is also available through the interactive graph. For example, hovering over a green node displays further data on the activity and literature source for that enzyme. For more suggestions and detail on a particular reaction, clicking and holding a reaction node launches a pop-up window with further information (Figure 3). In addition, custom reactions may be added to the graph, allowing custom chemical steps to be included by the user. Alternatively, the reaction rules which are applied can be switched over to suggestions which make use of the recently described chemistry CASP tool AIZynthfinder [30], for the creation of powerful chemo-enzymatic cascades.

Alternatively, the pathway exploration mode seeks to automatically generate useful suggestions for possible pathways to a target molecule. To do this, a network is first generated by applying reaction rules iteratively up to a user-defined maximum length. Networks which reach a user-defined limit to the number of nodes are reduced in size by removing the outer-most worst reactions, as scored by the change in molecular complexity. Pathways are then generated by a best first search approach, which prioritises steps with higher changes in molecular complexity until all possible pathways have been generated, or a limit to number of pathways is reached. A weighted score is used to evaluate and rank each pathway, taking into account the change in molecular complexity, the number of steps, whether the starting material appears in a catalogue of buyable building blocks, and the number of steps with a similar literature precedent.

Most chemistry CASP tools utilise a stopping criterion (other than maximum pathway length), such as the commercial availability of starting materials [8-10]. Whilst starting material availability is clearly of relevance to the design of biocatalytic cascades, often experimenters are also interested in demonstrating enzymatic cascades with commercially available intermediates, inhibiting the use of starting material availability as a stopping criterion for RetroBioCat. Instead, RetroBioCat generates pathways of all lengths from a network, relying on the weighted score to determine which pathways are the most promising. Changing the weighted score might result in longer or shorter pathways being suggested more highly. For example, shorter pathways to buyable starting materials can be favoured by increasing both the weight for the number of steps, and the weight for whether a buyable starting material is available. However, in general, the default weights are a good starting point.

Evaluation using a test-set of 52 literature cascades

To test both network and pathway explorer, we carried out a thorough review of the biocatalytic cascades reported in the literature, generating a test-set of 52 pathways (Figure 4, Extended Data Figures 2–5) [5,31-72]. Except for C-H oxidation by P450 enzymes, all of the reactions in the test-set were correctly predicted by RetroBioCat. Importantly, the majority of pathways were suggested within the top few suggestions using pathway explorer with only the default settings for the weighted score, validating this as a useful approach for the automated design of biocatalytic cascades.

Extended Data Fig. 2

Rankings for pathways 12 to 20 of the test-set for Pathway explorer.

A continuation of Figure 4, showing rankings for pathways 12 to 20 by RetroBioCat using a maximum of 4 steps and either the default scoring weights, or default weights but with the weight for number of steps with literature precedent set to zero, are shown. Pathways are marked as identified even where RetroBioCat suggests additional steps. * indicates pathways where the data from the relevant paper has not been added to the database of literature precedent reactions in RetroBioCat. TPL: tyrosine phenol lyase, AAD: amino acid deaminase, AADH: amino acid dehydrogenase, TDL: thiamine-dependent lyase, TA: transaminase, PSase: Pictet-Spenglerase, PAL: phenylalanine ammonia lyase, P450: cytochrome P450, ADH: alcohol dehydrogenase, CumDO: cumene dioxygenase, ERED: ene reductase, BVMO: Baeyer-Villiger monooxygenase, SMO: styrene monooxygenase, AlDH: aldehyde dehydrogenase, AlOx: alcohol oxidase.

Extended Data Fig. 5

Rankings for pathways 41 to 52 of the test-set for Pathway explorer.

A continuation of Figure 4, showing rankings for pathways 41 to 52 by RetroBioCat using a maximum of 4 steps and either the default scoring weights, or default weights but with the weight for number of steps with literature precedent set to zero, are shown. Pathways are marked as identified even where RetroBioCat suggests additional steps. * indicates pathways where the data from the relevant paper has not been added to the database of literature precedent reactions in RetroBioCat. TA: transaminase, IRED: imine reductase, PAL: phenylalanine ammonia lyase, DC: decarboxylase, AmDH: amine dehydrogenase, TPL: tyrosine phenol lyase, AAD: amino acid deaminase, TAM: tyrosine aminomutase, TDL: thiamine-dependent lyase, ADH: alcohol dehydrogenase, AlOx: alcohol oxidase, EH: epoxide hydrolase, CAR: carboxylic acid reductase, ATP: Adenosine triphosphate, NADP: Nicotinamide adenine dinucleotide phosphate.

Discussion

CASP tools should strive to augment the abilities of scientists seeking to design new routes to a target molecule. An intuitive and easy to use user-interface, as we have developed for RetroBioCat, is therefore crucial. Furthermore, the manually curated reaction rules utilised by RetroBioCat strike a balance between being general enough, so that the potential for enzyme engineering or discovery is captured, whilst providing context where necessary so as to be realistic.

Suggestions for potential biocatalytic transformations even without literature precedent are themselves a valuable resource, as in many cases enzyme screening panels can be employed to find the right enzyme for a specific reaction. However, where there is literature precedent for a reaction, suggestions are more robust and easier to implement if these are automatically identified. Here, we have demonstrated the use of molecular similarity to automate this process and have begun the construction of a database of synthetic biotransformations described in the literature, with further contributions to be reported in future work.

Pathway explorer offers automated ranking of suggested pathways using a selection of metrics. We have shown that this functions well in suggesting previously reported pathways early in the ranking system. Future improvements could seek to provide further information on the suitability of each suggested pathway. For example, thermodynamics, cofactor usage, substrate and product solubility or stability [74], toxicity, reaction conditions, starting material price and predicted pathway kinetics [75], could all offer more insight into which pathway is the most promising for experimental characterisation. Substantial advances are being made in the pathway searching algorithms utilised in organic chemistry. As we seek to incorporate organic chemistry or biosynthetic steps into RetroBioCat, or simply as we expand the reactions rules for biocatalysis, it may become necessary to exploit more advanced algorithms for pathway generation, such as the Monte Carlo tree search (MCTS) [7,10,76,77].

Several challenges still remain for the refinement of RetroBioCat. For example, at present enzymatic C-H activations such as hydroxylations and halogenations are currently not fully included, as the context in these reactions rules requires more careful consideration. Additionally, larger, more complex target molecules are sometimes handled inadequately by RetroBioCat, possibly highlighting the need for increased research into bond forming enzymes in the biocatalysis field as a whole. Indeed, with exceptions, most biocatalytic transformations are performed on small molecules of typically less than 500 Da. To help mitigate this issue, RetroBioCat features an option to fragment a molecule along synthetically accessible bonds prior to the generation of pathways or networks. Additionally, chemical steps can be suggested in network explorer [30]. Future work to include chemistry steps in pathway explorer will allow better automated suggestions to be made where some chemical steps are necessary, although care must be taken that enzymatic steps are well represented amongst the more numerous chemical options. Additionally, incorporating the reaction rules developed for metabolic engineering could further blur the lines between biocatalysis and biosynthesis, and open up access to a broad pool of renewable resources for use as substrates. Crucially, many recent CASP tools are written in python using open-source libraries such as RDKit [7,8,11,30]. Furthermore, the use of reaction SMARTS to describe reaction templates is relatively common across many tools, which should facilitate the combination of approaches from different fields into a single solution in the future.

In summary, RetroBioCat offers an accessible set of tools for computer-aided design of biocatalytic cascades. These tools should be useful in highlighting the potential of enzymes for organic synthesis, and for the design of de novo biocatalytic pathways.

Methods

Overview

Both network explorer and pathway explorer utilise the creation of a bipartite directional graph using the NetworkX Python package, to hold all the possible transformations to the target molecule. The first node in the network is the target molecule as a SMILES string. On applying the reaction rules, reactions are added as nodes with edges between the new reaction nodes and the molecule the rules were applied to. The products of the reactions are then also added as nodes, in the form of SMILES strings, with edges between these new molecules and the reaction node that produced them. Applying the reaction rules iteratively creates a network of potential routes leading back to the target molecule. A maximum number of nodes can be set to limit combinatorial explosion, above which the outer-most reactions are deleted according to which has the worst change in molecule complexity. Nodes in the network are scored as described below, with the results held in a dictionary for each node. The web interface offers a network exploration mode, in which double clicking on a molecule applies reaction rules to that molecule to expand the network in this location. Alternatively, pathway explorer automatically generates pathways by automatically expanding a network out to a specified number of steps, before generating the possible pathways present in the network and ranking them, as described below.

Chemistry

The RDKit chemoinformatics library is used to implement all chemistry-related methods, such as reaction transformations or calculating molecular similarity. Molecules are stored in pathways or networks as SMILES strings, as generated by RDKit using the default settings. For processing, SMILES are first translated into a mol object as defined in RDKit.

Reaction rules are defined using reaction SMARTS. Reaction SMARTS are developed manually, often making use of the capability of Marvin JS (https://chemaxon.com/products/marvin-js) to draw the proposed reaction and extracting it as a reaction SMARTS. Positive and negative tests, in the form of SMILES strings which should or should not be transformed by the rules, are defined to ensure reaction SMARTS act as planned. Reaction rules are applied using a modified version of rdChiral [78] (https://github.com/connorcoley/rdchiral).

Node scoring

In both network and pathway explorer, molecule and reaction nodes are scored to allow scoring of individual steps or entire pathways, as detailed below.

Whether a particular molecule is available as a buyable building block is identified by querying a database of buyable SMILES strings. We used a combination of the ‘in-stock building blocks’ list in the ZINC database, the building blocks listed by emolecules, and the building blocks available from molport to construct this database. All SMILES strings were pre-processed to be in the form generated by RDKit using the default settings. If a SMILES string is in the database, the is_buyable attribute is marked as 1.

To calculate molecular complexity, we use the SC-Score [29]. Code for this module is taken from https://github.com/connorcoley/scscore/. We use the standalone numpy version of the SC-Scorer, utilising Boolean fingerprints with a length of 1024. Every molecule is scored using the SC-Scorer, with the result saved in an attribute on the node as ‘complexity’. For every molecule, a ‘relative complexity’ is also calculated, by taking the difference between the complexity of the current molecule and the complexity of the target molecule. The difference in complexity between a reaction substrate and product is used to calculate ‘change in complexity’ for every reaction node. Where a reaction has multiple substrates, the substrate with the highest complexity is used.

A module for comparing molecule similarity is available within RetroBioCat for comparing suggested reactions against a database of literature precedent. To do this, fingerprints are constructed for every molecule in the database, and for the molecules in the query reaction. RDKit fingerprints are used with the default settings as implemented in RDKit. Fingerprints are compared by calculating Tanimoto similarity, again implemented using RDKit. Molecules with similarity below a cut-off value are discarded. Similarity is scored using either only the similarity of the products, or the average of similarities for both the products and the substrates. The highest scoring reaction is used as a suggestion in network or pathway explorer, with the enzyme with the highest activity chosen. Optionally, negative data can be included in this search. A number of alternative fingerprints are available within RDKit, such as Morgan fingerprints, Avalon Fingerprints, or Atom-Pair and Topological-Torsion Fingerprints. Each of these fingerprints extracts features of a molecule in a different way, and could be used to calculate similarity. In our hands, the RDKit fingerprint functions well in identifying similar molecules in our dataset.

Pathway generation

To automatically generate pathways to a target molecule, a network is first generated by iteratively applying the reaction rules up a specified number of steps. Pathways are generated by applying a best first search on the network, using molecular complexity as the selection criteria, adjusted to only include positive values. At each step, the option to stop the search is also available. Once all the possible pathways in the network have been generated, or the maximum number of pathways reached, pathways are scored and ranked. Pathways are scored on their total change in complexity, the number of enzymatic steps in the pathway, the percentage of starting material which is marked as ‘buyable’, and the number of steps which have been identified with similar literature precedent. Each score is normalised to between 0 and 1, to which a user-specified weight can be applied. The pathways are ranked in order of the total of the weighted scores, presenting the user with the highest scoring pathways first. In addition, a diversity score is applied which penalises reactions which have appeared in the prior suggestions.

Pathway explorer test set evaluation

The test set of 52 pathways consists of the target molecules, the starting materials used and the enzymes in the published pathway. For each test, pathways are automatically generated and ranked. Each generated pathway is then compared with the published version. Pathways which contain the starting molecules and include the recorded enzymes for the published pathway are marked as a match, with the ranking of the pathway recorded.

Rankings for pathways 21 to 30 of the test-set for Pathway explorer.

A continuation of Figure 4, showing rankings for pathways 21 to 30 by RetroBioCat using a maximum of 4 steps and either the default scoring weights, or default weights but with the weight for number of steps with literature precedent set to zero, are shown. Pathways are marked as identified even where RetroBioCat suggests additional steps. * indicates pathways where the data from the relevant paper has not been added to the database of literature precedent reactions in RetroBioCat. TDL: thiamine-dependent lyase, ADH: alcohol dehydrogenase, PAL: phenylalanine ammonia lyase, CAR: carboxylic acid reductase, ERED: ene reductase, IRED: imine reductase, AmOx: amine oxidase, P450: cytochrome P450, ATP: Adenosine triphosphate, NADP: Nicotinamide adenine dinucleotide phosphate.

Rankings for pathways 31 to 40 of the test-set for Pathway explorer.

A continuation of Figure 4, showing rankings for pathways 31 to 40 by RetroBioCat using a maximum of 4 steps and either the default scoring weights, or default weights but with the weight for number of steps with literature precedent set to zero, are shown. Pathways are marked as identified even where RetroBioCat suggests additional steps. * indicates pathways where the data from the relevant paper has not been added to the database of literature precedent reactions in RetroBioCat. ADH: alcohol dehydrogenase, BVMO: Baeyer-Villiger monooxygenase, SMO: styrene monooxygenase, EH: epoxide hydrolase, AmDH: amine dehydrogenase, CAR: carboxylic acid reductase, TA: transaminase, AlOx: alcohol oxidase, TA: transaminase, ERED: ene reductase, TrpS: tryptophan synthase, XOR: xanthine oxidoreductase, AAD: amino acid deaminase, ATP: Adenosine triphosphate, NADP: Nicotinamide adenine dinucleotide phosphate.