Summary

Our company offers a service called Spotlight that uses a machine learning model to suggest mutants that improve properties such as enzyme activity and thermostability. We input target enzyme sequences into a pre-trained model using various enzymes to predict mutants with improved activity and thermostability for that enzyme. In this tech blog, we have verified the predictive accuracy of Spotlight™ compared to previous research.

The previous research used for comparison.

In Li et al., 2022, a machine learning model (DLKcat) was created to predict kcat using enzyme amino acid sequences and compounds as input information. To ensure equality in the comparison, we utilized the DLKcat machine learning model algorithm and reconstructed the model using the same training data as Spotlight™, namely the kcat entries from BRENDA. We compared the predicted kcat values of the mutants by the reconstructed DLKcat and Spotlight and evaluated which values were closer to the actual measured values. For this study, we extracted entries from BRENDA, ensuring that only wild type (WT) and single mutant variants were included. Our focus was to compare the sensitivity for a single mutation between the two models.

Results

1. Construction of the machine learning model using BRENDA’s kcat (Turnover Number) data.

Entries for variants with reported kcat values were extracted from BRENDA, including entries for the corresponding wild-type (WT) sequences and information about the compounds used to measure kcat. While ensuring no bias toward specific enzyme families, the entries were divided into a 3:1 ratio of training to test data. Training data consisted of 3,969 entries with an increased kcat, 2,985 entries with an unchanged kcat, and 8,296 entries with a decreased kcat (Figure 1). The test data consisted of 792 entries with an increased kcat, 748 entries with an unchanged kcat, and 1,926 entries with a decreased kcat (Figure 2).

2. Evaluation of mutant/WT ratio of predicted kcat by DLKcat and Spotlight™.

The information from the training data was converted into the format of features required by DLKcat, and a machine learning model was constructed. We converted the training data into the required format of features for Spotlight™ and built a machine learning model (Figure 3).

In the case of DLKcat, the Pearson correlation coefficient between the measured and predicted values of the ratio of the kcat of the mutant to the wild type (WT) kcat was 0.18 (Figure 3). We believe that the reason why the predicted values in DLKcat did not correlate well with the measured values is that DLKcat converts the entire length of the sequence into a vector as a feature, making it difficult for the difference of one amino acid to be reflected in the feature.

In the case of Spotlight™, the Pearson correlation coefficient between the measured and predicted values of the ratio of the kcat of the mutant to the WT kcat was 0.66 (Figure 3). We believe that our Spotlight™ is able to accurately predict the changes caused by single mutations from the WT because it has been devised to reflect the properties of the mutant as a feature.

Conclusion

Our Spotlight™ model was found to more accurately predict changes in activity compared to previous research in cases where only one amino acid mutated.

Acknowledgments

We are grateful for the use of data from the following paper to compare the accuracy of enzyme activity prediction in this study.

Li et al., (2022) Deep learning-based kcat prediction enables improved enzyme-constrained model reconstruction. Nature Catalysis.

Introduction

I am Isozaki from the Business Development Department. Our company conducts explorations of artificial synthetic routes from "raw materials" to "target products" using enzymatic reactions. By simply inputting the compound structure data of the "target products" and "raw materials", we can output potential synthetic route candidates for producing the target product from the starting compound. In this blog, I will introduce a specific example where we predict a route to synthesize 4-amino-cinnamic acid, a which is used in the production of high-strength polymers  for high-strength polymers, from glucose and the enzymes involved in the reactions.

Materials Used for Synthetic Pathway Exploration

In Tateyama et al. (2016), 4-amino-cinnamic acid is used as a which is used in the production of high-strength polymers for producing high-strength polymers. The pathway used to synthesize this 4-amino-cinnamic acid is shown in Figure 1. Glucose serves as the raw material, and 4-amino-phenylalanine is produced using Escherichia coli engineered with Aminodeoxychorismate synthase (PapA) derived from Streptomyces venezuelae and Aminodeoxychorismate synthase (PapBC) derived from S. pristinaespiralis. Furthermore, this 4-amino-phenylalanine is used as a raw material, along with E. coli engineered with Phenylalanine ammonia-lyase (RgPAL) derived from Rhodotorula glutinis, to produce 4-amino-cinnamic acid.

Results

1. Biosynthetic Pathway Exploration

By inputting glucose as the Starting compound and 4-amino cinnamic acid as the product, an artificial synthesis pathway, as shown in Figure 1, was output. The output pathway was identical to the known synthesis pathway of chorismate from glucose, leading to the synthesis of 4-amino cinnamic acid via 4-amino phenyl alanine.

2. Similar Reaction Exploration

Among the artificial synthesis pathways identified in Result 1, the similar reaction from 4-amino phenyl alanine to 4-amino cinnamic acid was explored.

Through the exploration of similar reactions, a reaction that removes an amino group and generates a double bond was identified. Some of the similar reactions with a high degree of similarity to the target reaction and their rankings are shown in Figure 2. Similar reactions were extracted, including those that match the target reaction exactly.

3. Exploration of Corresponding Enzymes for Similar Reactions

In Result 2, similar reactions for the target reaction were extracted. The enzyme sequences responsible for these similar reactions were extracted by taxon. The filtered sequences were then compared with the enzymes used in the paper. Sequences were extracted at three levels: Rhodotorula genus, Eukaryota domain, and all taxa (Table 1). The extracted sequences included those that exhibited over 90% sequence homology with the sequences used in the paper.

Conclusion

In this blog, we demonstrated the exploration of artificial synthetic pathways. We explored an artificial route to synthesize the compound 4-amino cinnamic acid, which serves as a raw material for high-strength polymers, from glucose. We aimed to determine whether we could find enzymes that synthesize 4-amino cinnamic acid from 4-amino phenyl alanine using similar reaction enzyme exploration techniques. For the above reactions, we extracted sequences by taxon and presented the number of sequences for each. We successfully extracted multiple sequences that included several with high similarity to the enzymes used in the paper.

Acknowledgments

We utilized data from the following paper for this synthetic pathway exploration:

Tateyama et al. (2016). Ultrastrong, Transparent Polytruxillamides Derived from Microbial Photodimers. Macromolecules.

Introduction

This is Isozaki from the Business Development Department. At our company, we specialize in exploring enzymes that catalyze unknown reactions. By analyzing the reaction similarity to known enzymatic reactions and the sequence homology to enzymes responsible for these reactions, we can predict candidate enzyme sequences for target unknown reactions. In this blog, we present a specific example where we predicted the enzyme sequences responsible for the synthetic reaction of Islatravir, a candidate compound for HIV treatment.

Materials Used for Enzyme Exploration

We utilized data from Huffman et al., 2019. This study designed a novel synthetic pathway for Islatravir, identified enzymes catalyzing each reaction in the pathway, and validated them experimentally. The synthetic pathway of Islatravir is illustrated in Figure 1. The synthesis follows the sequence: Compound 6 → Compound 7 or 8 → Compound 5 → Compound 4 → Compound 3 + Compound 2 → Islatravir. Using our enzyme exploration technology, we predicted the enzymes responsible for each reaction in this pathway and compared them with the enzymes used in the study.

Results

First, we explored reactions similar to each of the five reactions in the synthesis pathway.

1. Similar Reaction Search

Similar Reaction for Phosphorylation Reaction of Starting Material 6→8 (or 7→5)

Several reactions that phosphorylate a hydroxyl group were extracted as similar reactions. Figure 3 shows a portion of the similar reactions with high similarity to the target reaction and their rankings. Similar to the reaction mentioned above, this reaction is not found in known metabolic pathways, so multiple similar reaction candidates were extracted.

Synthesis Reaction of Intermediate 5→4 (Ribose)

A reaction that forms deoxyribose by cyclization upon the addition of acetaldehyde was extracted (Figure 4). In the paper, this reaction mimics known metabolic reactions, and therefore, a reaction that is identical except for the alkyne group was obtained as a similar reaction.

Phosphorylation Reaction of Intermediate 4→3

A reaction that transfers a phosphate group from a hydroxyalkyl group to a hydroxyl group was extracted (Figure 5). Similar to the ribose synthesis reaction mentioned above, this reaction also mimics known metabolic reactions, and therefore, a reaction that is identical except for the alkyne group was obtained as a similar reaction.

Intermediate 3 → Nucleoside Synthesis Reaction of Islatravir

A reaction that adds purine to deoxyribose was extracted (Figure 6). Similar to the above phosphate group transfer reaction, this reaction also mimics known metabolic reactions, and therefore, a reaction that is identical except for the alkyne group and fluorine was obtained as a similar reaction.

2. Search for enzymes corresponding to similar reactions

In Result 1, similar reactions for each of the five reactions were extracted. Enzyme sequences responsible for these similar reactions were extracted by taxon. For each of the five reactions, we checked whether the enzymes used in the study were included among the narrowed-down sequences. Additionally, for each of the five reactions, the enzyme sequences were further filtered using phylogenetic position screening from all taxa-derived enzyme sequences.

Extraction of enzyme sequences for similar reactions by taxon

We searched for enzymes responsible for the similar reactions in Result 1 and extracted them in three stages: from the genus Escherichia, bacteria, and all taxa. The number of sequences extracted is shown in Table 1 below. We checked whether the enzymes used in the current study were included. In four of the five reactions, the enzymes used in the study were included among the enzyme sequences extracted by our enzyme discovery technology.

Screening based on phylogenetic position

From the similar reaction enzyme sequences extracted from all taxa, we further narrow down the sequences based on their phylogenetic positions. All sequences were clustered and a phylogenetic tree was generated. From phylogenetically grouped clusters, one sequence was selected from each group. In this selection process, priority was given to sequences with high conservation across species (Table 2, Figure 7).

Considerations on the oxidation reaction from 6→7 (8→5)

As shown in the results above, the enzyme used for this reaction in the referenced paper was not identified in this study. One possible reason is that the target reaction, 6→7 (8→5), and the enzymatic reaction used in the paper are not sufficiently similar (Figure 8). However, the similar reactions identified through the current search also utilize O2-dependent oxidoreductases, which may catalyze the target reaction. The enzyme used in the paper for catalyzing the reaction shown in Figure 9 is estimated to be UniParc ID: UPI0001E112C2. This sequence is a member of the UniRef50 cluster, which includes sequences confirmed to catalyze RHEA_24161. However, UPI0001E112C2 itself has not been curated to confirm its catalytic activity for this reaction.

Conclusion

In this blog, we demonstrated the search for enzymes responsible for unknown reactions. We used a novel synthetic pathway for Islatravir, where we identified the enzymes that catalyze each reaction in this pathway. We attempted to find similar reaction enzymes for five unknown reactions, successfully extracting multiple similar reactions for each. In this process, we extracted sequences by arbitrary taxa and provided the number of sequences for each. For four of the reactions, we were able to extract multiple sequences that included the enzymes used in the paper. We then narrowed down the similar reaction enzymes extracted from all taxa by considering their phylogenetic positions. In standard screenings, we can further refine candidate sequences using other indicators, such as the properties of the enzymes (cellular localization, etc.) and their three-dimensional structures.

Acknowledgments

We utilized data from the following paper for the similar reaction enzyme search:

Huffman et al., (2019) Design of an in vitro biocatalytic cascade for the manufacture of islatravir. Science.

Introduction

I am Isosaki from the Business Development Department. At our company, we are conducting enzyme activity prediction as part of our useful enzyme exploration efforts using molecular dynamics (MD) simulations. From unknown enzyme sequences, we predict their structures and subject the enzyme-ligand complexes to MD simulations. Based on the results, we calculate a proprietary digzyme score to predict enzyme activity. In this blog, we will present an example of predicting the activity of a thiolase-like enzyme, OleA, from its homologous sequences.

Materials Used for Enzyme Activity Prediction

The natural substrate for OleA is acyl-CoA. The enzyme's Cys143 residue cleaves this acyl group. In investigating this activity, a p-nitrophenolate-based experimental system is used (Figure 1).

Results

We predicted whether 59 homologous OleA sequences would hydrolyze one type of p-nitrophenolate, specifically 4-nitrophenyl-hexanoate.

1. Prediction of the 3D Structure of 59 Homologous Sequences

First, as all 59 homologous sequences had unknown structures, we predicted their 3D structures. From the predicted structures, we also identified the location of the active residues and the substrate-binding pocket. Figure 2 shows the predicted 3D structure of the homologous sequences and the location of the active residue, Cys. Figure 3 shows the predicted location of the substrate-binding pocket.

2. Molecular Dynamics Simulation

Next, we placed the enzyme-substrate complex, consisting of the enzyme and 4-nitrophenyl-hexanoate, in a system containing water molecules and ions, and ran molecular dynamics simulations (Figure 4).

3. Calculation of digzyme's Proprietary Enzyme Activity Prediction Score

Finally, based on the results of the molecular dynamics simulations, we calculate digzyme's proprietary score. Figure 5 shows the predicted scores for all 59 sequences, sorted in descending order. Sequences with confirmed activity from experimental validation are highlighted in pink, while those without activity are shown in gray. Sequences with a score of 70 or higher were classified as active (above the red line in Figure 5). In this case, we predicted 9 sequences to be active, 3 of which were experimentally confirmed to have activity, resulting in a positive predictive value (PPV) of 0.30. The true positive rate (TPR) was 0.6, and the false positive rate (FPR) was 0.13. This indicates that inactive sequences were successfully ranked lower, while active sequences were included in the higher ranks.

Conclusion

In this blog, we demonstrated the prediction of enzyme activity using our enzyme activity prediction technology, and validated the predictions with experimental results. Typically, 5 to 10 sequences are synthesized for experimental validation. In this case, two of the top five ranked sequences were confirmed to have activity through experiments, demonstrating the practical accuracy of our enzyme activity prediction method. The dataset was specifically selected to simulate a scenario where only a small fraction of the enzyme sequences in the population exhibit activity (in this case, 5 out of 59 sequences). Since the false negative rate was kept low, we successfully predicted enzyme activity with high accuracy.

Acknowledgments

We would like to thank the following paper for providing the experimental data used in this enzyme activity prediction:

Robinson et al., (2020) Machine learning-based prediction of activity and substrate specificity for OleA enzymes in the thiolase superfamily. Synthetic Biology.