Method and device for discovering enzyme

Abstract

A method and a device for discovering an enzyme are disclosed. The enzyme discovery method is an enzyme discovery method performed by a computing device including a processor and a memory. The method may comprise the steps in which: the processor loads a first encoder trained on the basis of functional characteristics of an enzyme into the memory; the processor loads a second encoder trained on the basis of structural characteristics of the compound into the memory; the processor constructs a combination network for combining the first encoder and the second encoder through transfer learning; the processor inputs an amino acid sequence of a candidate enzyme and graph data of a target compound to the first encoder and the second encoder, calculates a reaction probability of a reaction between the candidate enzyme and the target compound through the combination network, and stores same in the memory; and the processor selects a final candidate enzyme that performs a reaction on a biosynthetic pathway for producing the target compound from among the candidate enzymes on the basis of the calculated reaction probability, and stores biosynthetic pathway data in which the biosynthetic pathway is optimized using the final candidate enzyme in the memory.

Description

Enzyme screening method and device

[0001] Cross-citation with related applications

[0002] This application claims the benefit of priority to Korean Patent Application No. 10-2024-0095089, filed July 18, 2024, and Korean Patent Application No. 10-2024-0100992, filed July 30, 2024, the entire contents of which are incorporated herein by reference.

[0003] The disclosure relates to an enzyme screening method and device, and more particularly, to an enzyme screening method and device for screening an enzyme through screening a candidate enzyme and selecting a final candidate enzyme.

[0004] As global interest in sustainability grows, research into ways to mitigate climate change by reducing greenhouse gases is actively underway. Bio-based materials can play a crucial role in reducing greenhouse gas emissions by replacing conventional fossil fuel-based materials. As an example of a bio-based material, polylactic acid (PLA), a type of bioplastic, is produced through microbial fermentation from renewable plant resources such as corn, sugarcane, and potatoes. This process converts carbohydrates such as glucose and starch into lactic acid, synthesizing and polymerizing lactide. Another example is bio-based 1,4-butanediol (Bio-BDO), which is produced from renewable biomass and can be produced through microbial fermentation without relying on fossil fuels. Bio-1,4-butanediol (Bio-BDO) can be applied in various applications, such as being converted to tetrahydrofuran (THF) through a dehydration reaction.

[0005] Industrial strains are microbial strains selected and optimized for the mass production of industrially useful products or materials. They are essential for the production of bio-based materials and can play a key role in converting biomass into various high-value-added products. Traditional methods for developing industrial strains include mutation induction, genetic recombination, metabolic engineering, and adaptive evolution. However, recent efforts are leveraging rapidly advancing artificial intelligence technology to maximize research efficiency through data-driven prediction and optimization.

[0006] The task to be solved is to provide an enzyme search method and device that can search for enzymes capable of reacting with a designated target compound using artificial intelligence technology based on the functional characteristics of enzymes and the structural characteristics of compounds.

[0007] The task to be solved is to provide an enzyme search method and device that can provide enzyme function prediction that can improve the prediction accuracy for functional annotations having a hierarchical structure and solve the data imbalance problem.

[0008] According to one embodiment, an enzyme search method is performed by a computing device including a processor and a memory, the method comprising: a step in which the processor loads a first encoder, for which learning has been performed based on functional characteristics of an enzyme, into the memory; a step in which the processor loads a second encoder, for which learning has been performed based on structural characteristics of a compound, into the memory; a step in which the processor constructs a coupling network that couples the first encoder and the second encoder through transfer learning; a step in which the processor inputs an amino acid sequence of a candidate enzyme and graph data of a target compound into the first encoder and the second encoder, and calculates a reaction probability for the candidate enzyme to react with the target compound through the coupling network and stores the calculated reaction probability in the memory; and a step in which the processor selects, based on the calculated reaction probability, a final candidate enzyme that performs a reaction on a biosynthetic pathway for producing the target compound from among the candidate enzymes, and stores biosynthetic pathway data in which the biosynthetic pathway is optimized using the final candidate enzyme in the memory.

[0009] The step of constructing the combined network may include a step in which the processor combines the first feature extracted from the first encoder and the second feature extracted from the second encoder into one vector through a fully connected layer, and the step of calculating the response probability and storing it in the memory may include a step in which the processor calculates the response probability by nonlinearly transforming the vector by applying a nonlinear activation function to the fully connected layer.

[0010] The above first encoder learns the reaction characteristics of the enzyme by predicting the EC number (Enzyme Commission Number) from the enzyme amino acid sequence, and can perform multi-class prediction for multi-levels of the EC number using contrastive learning.

[0011] The first encoder may be trained to generate two positive samples through data augmentation, maximize the similarity between the positive samples through a contrastive loss function (Contrastive Learning), and minimize the similarity between the negative samples.

[0012] The second encoder can map the graph of the target compound into a continuous number-based latent space based on a generative model, and extract features for junctions in units of junction trees larger than atomic units.

[0013] The second encoder may further include a network that predicts the energy of the reactant from the latent space mapping values.

[0014] The first encoder includes an enzyme function prediction model that predicts an enzyme function through prediction of a functional annotation of a protein, and the enzyme search method may further include: a step in which the processor obtains learning data associated with a functional characteristic of the enzyme; a step in which the processor introduces a plurality of base losses for each of a plurality of activation layers corresponding to a plurality of levels; a step in which the processor applies data augmentation to the enzyme function prediction model and introduces an additional loss; a step in which the processor defines a total loss function based on the plurality of base losses and the additional loss; and a step in which the processor trains the enzyme function prediction model based on the learning data and the total loss function.

[0015] The above multiple basic losses include the first basic loss to the Mth basic loss (M is an integer greater than or equal to 2), and the step of defining the total loss function is to define the total loss function according to the following mathematical expression 1. may include steps to define:

[0016] (Equation 1)

[0017]

[0018] Here, is the kth basic loss among the above basic losses (k is an integer greater than or equal to 1 and less than or equal to M), The above additional loss, is a predetermined weight for the above k basic loss, is a predetermined weight for the above additional loss.

[0019] The above additional loss can be defined according to the following mathematical equation 2:

[0020] (Equation 2)

[0021]

[0022] Here, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 is indicated.

[0023] The above additional loss can be defined according to the following mathematical formula 3:

[0024] (Equation 3)

[0025]

[0026] Here, is a multilevel term, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 is indicated.

[0027] An enzyme search device according to one embodiment comprises one or more non-transitory computer-readable media comprising instructions; And one or more processors that execute the instructions to perform an operation, wherein the operation may include loading a first encoder, in which learning is performed based on the functional characteristics of an enzyme, onto the medium, loading a second encoder, in which learning is performed based on the structural characteristics of a compound, onto the medium, constructing a coupling network that couples the first encoder and the second encoder through transfer learning, inputting an amino acid sequence of a candidate enzyme and graph data of a target compound into the first encoder and the second encoder, calculating a reaction probability that the candidate enzyme and the target compound will react through the coupling network, and storing the calculated reaction probability in the medium, and selecting a final candidate enzyme that performs a reaction on a biosynthetic pathway for producing the target compound from among the candidate enzymes based on the calculated reaction probability, and storing biosynthetic pathway data that optimizes the biosynthetic pathway using the final candidate enzyme in the medium.

[0028] Building the above combined network may include combining the first feature extracted from the first encoder and the second feature extracted from the second encoder into one vector through a fully connected layer, and the calculating may include nonlinearly transforming the vector by applying a nonlinear activation function to the fully connected layer to predict the response probability.

[0029] The above first encoder learns the reaction characteristics of the enzyme by predicting the EC number (Enzyme Commission Number) from the enzyme amino acid sequence, and can perform multi-class prediction for multi-levels of the EC number using contrastive learning.

[0030] The first encoder may be trained to generate two positive samples through data augmentation, maximize the similarity between the positive samples through a contrastive loss function (Contrastive Learning), and minimize the similarity between the negative samples.

[0031] The second encoder can map the graph of the target compound into a continuous number-based latent space based on a generative model, and extract features for junctions in units of junction trees larger than atomic units.

[0032] The second encoder may further include a network that predicts the energy of the reactant from the latent space mapping values.

[0033] The first encoder may include an enzyme function prediction model that predicts an enzyme function through prediction of a functional annotation of a protein, and the operation may further include obtaining learning data associated with a functional characteristic of the enzyme, introducing a plurality of base losses for each of a plurality of activation layers corresponding to a plurality of levels, applying data augmentation to the enzyme function prediction model and introducing an additional loss, defining a total loss function based on the plurality of base losses and the additional loss, and training the enzyme function prediction model based on the learning data and the total loss function.

[0034] The above multiple basic losses include the first basic loss to the Mth basic loss (M is an integer greater than or equal to 2),

[0035] Defining the above total loss function is:

[0036] The total loss function according to the following mathematical formula 1 This may include defining:

[0037] (Equation 1)

[0038]

[0039] Here, is the kth basic loss among the above basic losses (k is an integer greater than or equal to 1 and less than or equal to M), The above additional loss, is a predetermined weight for the above k basic loss, is a predetermined weight for the above additional loss.

[0040] The above additional loss can be defined according to the following mathematical equation 2:

[0041] (Equation 2)

[0042]

[0043] Here, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 is indicated.

[0044] The above additional loss can be defined according to the following mathematical formula 3:

[0045] (Equation 3)

[0046]

[0047] Here, is a multilevel term, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 is indicated.

[0048] According to the embodiments, AI technology based on the functional characteristics of enzymes and the structural characteristics of compounds can be used to search for enzymes capable of reacting with a specified target compound. Using AI models that have learned enzyme function and transfer learning, the potential for reactions with target compounds can be predicted, beyond simply predicting enzyme function. This can identify promising enzymes that excel at specific reactions and accelerate the development of microbial strains that produce eco-friendly bio-based materials.

[0049] Furthermore, according to embodiments, enzyme function prediction is provided, and to train an enzyme function prediction model, a total loss function can be defined by introducing multiple base losses corresponding to the hierarchical structure and additional losses related to data augmentation. Accordingly, the hierarchical structure of functional annotations can be reflected, ensuring that features of enzymes with the same class at each level are sufficiently close together, thereby improving prediction accuracy for functional annotations with a hierarchical structure and resolving data imbalance issues.

[0050] Figure 1 is a drawing for explaining an enzyme search device according to one embodiment.

[0051] Figure 2 is a drawing for explaining an enzyme search device according to one embodiment.

[0052] FIG. 3 is a diagram for explaining a generation model related to a second encoder according to one embodiment.

[0053] FIG. 4 is a drawing for explaining a second encoder according to one embodiment.

[0054] FIG. 5 is a drawing for explaining an enzyme search method according to one embodiment.

[0055] FIG. 6 is a drawing for explaining an enzyme function prediction device according to one embodiment.

[0056] FIG. 7 is a drawing for explaining an enzyme function prediction device according to one embodiment.

[0057] FIG. 8 is a drawing for explaining an enzyme function prediction device according to one embodiment.

[0058] FIG. 9 is a diagram illustrating an enzyme search method based on enzyme function prediction according to one embodiment.

[0059] FIG. 10 is a drawing for explaining a computing device according to one embodiment.

[0060] Below, with reference to the attached drawings, embodiments of the present invention are described in detail so that those skilled in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In addition, in the drawings, parts irrelevant to the description are omitted for clarity of description, and similar parts are designated with similar reference numerals throughout the specification.

[0061] Throughout the specification and claims, whenever a part is referred to as "comprising" a component, this does not exclude other components, but rather includes other components, unless otherwise stated. Terms including ordinal numbers, such as "first," "second," etc., may be used to describe various components, but these components are not limited by these terms. These terms are used solely to distinguish one component from another.

[0062] Terms such as "... part", "... device", and "module" described in the specification may refer to a unit capable of processing at least one function or operation described in the specification, which may be implemented by hardware or a circuit, software, or a combination of hardware or a circuit and software. In addition, at least some of the components or functions of the enzyme search method and device according to the embodiments described below may be implemented by a program or software, and the program or software may be stored in a computer-readable recording medium or storage medium.

[0063] Figure 1 is a drawing for explaining an enzyme search device according to one embodiment.

[0064] Referring to FIG. 1, an enzyme search device (10) according to an embodiment may be implemented as a computing device including a processor and a memory. For example, the enzyme search device (10) may be implemented as a computing device (50) as described below with reference to FIG. 10. In this case, the processor may correspond to the processor (510) of the computing device (50), and the memory may correspond to the memory (520) of the computing device (50). Alternatively, in some embodiments, the enzyme search device (10) may include one or more non-transitory computer-readable media including instructions and one or more processors that execute the instructions to perform operations. Here, the operations may include configurations, functions, steps, etc. described herein with respect to enzyme search methods and devices according to embodiments. In this specification, the term "module" is used to logically distinguish these operations performed by the control logic derivation method and device according to the embodiments.

[0065] The development of a strain producing a target compound can be accomplished through several processes, including strain selection, metabolic pathway design, screening for highly resistant strains, metabolic flux optimization, fermentation, and isolation / purification / scale-up. Strain selection refers to selecting a microbial strain capable of producing the target compound. This can be done by exploring strains that produce the target compound in various natural environments, selecting strains that are advantageous for target compound production from already studied strains, or using strains held in a microbial strain bank. Metabolic pathway design refers to designing and optimizing the strain's metabolic pathway to efficiently produce the target compound. This can include steps such as target compound biosynthetic pathway exploration, candidate enzyme screening, and high-performance enzyme development through mutation. High-resistant strain selection refers to selecting strains that can survive high concentrations of the target compound, and metabolic flux optimization refers to optimizing the target compound production flow within the metabolic pathway to maximize productivity. Fermentation is the production of target compounds by culturing optimized strains in large quantities, and separation / purification / scale-up can be the efficient separation and purification of target compounds after fermentation and production on an industrial scale.

[0066] As an example of a metabolic pathway, the process of converting glucose into ethanol can occur through glycolysis and ethanol fermentation. Specifically, glucose can be broken down into pyruvate through glycolysis, which can be a series of enzyme-catalyzed reactions that break down one glucose molecule into two pyruvate molecules. Pyruvate can be decarboxylated by pyruvate decarboxylase to form acetaldehyde. Acetaldehyde can then be reduced to ethanol by alcohol dehydrogenase. Thus, the final product, ethanol, can be produced as a result of alcoholic fermentation.

[0067] Advances in genomics and proteomics technologies are leading to the entry of numerous novel protein sequences into databases. However, only a small fraction of these proteins have direct experimental functional confirmation. Therefore, functional annotation of most proteins relies on computer-based predictions. The more accurate the functional annotation of proteins, the more likely it is to accelerate the development of new drugs and novel bio-based materials.

[0068] Protein functional annotation can be the biological meaning assigned to a protein sequence or structure to explain and understand its function. Specifically, functional annotation can encompass molecular functional aspects, including the specific activity or function performed by the protein; biological process aspects, including the life processes or pathways involved in the protein; and cellular structural aspects, including the cellular location where the protein functions. A representative example of functional annotation is the Enzyme Commission (EC) number.

[0069] The EC number is an internationally used numbering system for systematically classifying enzymes. The EC numbering system can be used to classify and name enzymes based on their chemical reaction mechanisms. An EC number consists of four digits, each separated by a period, that can indicate specific properties and reaction types of the enzyme. Specifically, the first digit represents the main hierarchy entry, indicating the general type of reaction catalyzed by the enzyme. The second digit represents a sub-hierarchy entry, indicating a more specific reaction type within the main hierarchy entry. The third digit represents a sub-sub-hierarchy entry, indicating a more specific reaction type within the sub-hierarchy entry. The fourth digit represents a naming hierarchy entry, which can be a unique number that identifies a specific enzyme.

[0070] In the following, for the sake of clarity and convenience of explanation, examples will be described mainly using EC numbers as examples. However, the scope of functional annotations to which the idea of ​​the present invention is applied is not limited to EC numbers, and may extend to any functional annotation having a hierarchical structure.

[0071] In the metabolic pathway where glucose is converted to ethanol, the EC number of pyruvate decarboxylase may be EC 4.1.1.1, and the gene ontology representing pyruvate decarboxylase activity may be GO:0004736. EC 4.1.1.1 may provide a specific classification of pyruvate decarboxylase as follows.

[0072] EC 4: Lyases - Enzymes that break down compounds by creating or removing double bonds in the substrate.

[0073] EC 4.1: Carbon-Carbon Lyases - Enzymes that cleave carbon-carbon bonds

[0074] EC 4.1.1: Carboxy-Lyases - Enzymes that catalyze the decarboxylation of carboxylic acids

[0075] EC 4.1.1.1: Pyruvate decarboxylase

[0076] Meanwhile, GO:0004736 is a GO that represents pyruvate decarboxylase activity to explain the molecular function of this enzyme, and can define the activity of an enzyme that catalyzes the reaction that converts pyruvate into acetaldehyde and carbon dioxide.

[0077] Meanwhile, the EC number of alcohol dehydrogenase may be EC 1.1.1.1, and the GO representing alcohol dehydrogenase activity may be GO:0004022. EC 1.1.1.1 may provide a specific classification of alcohol dehydrogenase as follows.

[0078] EC 1: Oxidoreductases - enzymes that catalyze redox reactions

[0079] EC 1.1: Oxidoreductases whose functional group is alcohol or polyol

[0080] EC 1.1.1: NAD+ or NADP + Oxidoreductases that use as receptors

[0081] EC 1.1.1.1: Alcohol Dehydrogenase

[0082] Meanwhile, GO:0004022 is a GO that represents alcohol dehydrogenase activity to explain the molecular function of this enzyme, NAD + can be defined as the enzyme activity that oxidizes alcohols to aldehydes or ketones using .

[0083] Once the target compound biosynthetic pathway has been explored, candidate enzymes capable of performing the corresponding reactions at each step must be identified. Biochemical reactions mediated by enzymes are classified by EC number or GO, as exemplified above, and these are used to select enzyme candidates. In other words, for previously unknown biochemical reactions, the EC number or GO that performs the most similar reaction is searched for based on the chemical similarity between the reactants and products. However, the method of narrowing down candidate enzymes by EC number or GO has limitations in that it requires numerous trials and errors to find highly productive enzymes.

[0084] In order to overcome such limitations and reduce trial and error in high-productivity enzyme exploration, the enzyme exploration device (10) provides an artificial intelligence model that inputs a graph of a compound together with the amino acid sequence of the enzyme and outputs the probability that the enzyme and the compound will react. This is different from most conventional enzyme function prediction models that input only the amino acid sequence of the enzyme and output the probability distribution of the classified enzyme functions. Specifically, the enzyme exploration device (10) may include a first encoder provision module (110), a second encoder provision module (120), a combined network construction module (130), a reaction probability prediction module (140), a learning module (150), and a final candidate enzyme selection module (160) to explore enzymes that can react with compounds.

[0085] The first encoder providing module (110) may provide a first encoder in which learning is performed based on the functional characteristics of an enzyme, and the second encoder providing module (120) may provide a second encoder in which learning is performed based on the structural characteristics of a compound. Specifically, the first encoder providing module (110) may load the first encoder into a memory, and the second encoder providing module (120) may load the second encoder into a memory. Since the complexity of the amino acid sequence of an enzyme and the molecular graph of a compound is very high, the first encoder and the second encoder may first learn the function of the enzyme (e.g., EC number) and the structural characteristics of the compound, respectively. Here, the first encoder may be also called a protein encoder as an enzyme function learning model, and the second encoder may be also called a compound encoder as a compound structural characteristic learning model.

[0086] The combined network construction module (130) can construct a combined network that combines the first encoder provided from the first encoder providing module (110) and the second encoder provided from the second encoder providing module (120). In some embodiments, the combined network construction module (130) can use transfer learning to combine the first encoder and the second encoder. Transfer learning is a technique that utilizes the weights of an already learned model in the initialization stage of a new model, and can construct a new model by combining the independently learned first encoder and second encoder. For example, it can be implemented to combine the outputs of the first encoder and the second encoder and perform the final prediction through an additional prediction layer. Accordingly, whether an enzyme and a compound react can be learned by combining an enzyme function learning model and a compound structure characteristic learning model through transfer learning. By combining enzyme function learning models with compound structure feature learning models through transfer learning, the reactivity of enzymes and compounds can be learned. Each encoder learns the characteristics of a specific data type, and then combines these to learn an integrated representation of various data sources. This enables the processing of complex multimodal data and achieves better predictive performance.

[0087] In some embodiments, the combined network construction module (130) may combine the first feature extracted from the first encoder and the second feature extracted from the second encoder into a single vector through a fully connected layer. That is, an additional prediction layer that receives as input a vector that merges the outputs of the first encoder and the second encoder may be defined as a fully connected layer. The combined feature vector from the outputs of the first encoder and the outputs of the second encoder may be input to the fully connected layer and processed as integrated information required for the final prediction. Here, the fully connected layer is a layer in which all nodes in one layer are connected to all nodes in the next layer, and may be used to process input data and generate the final output. Each connection may be assigned a weight and a bias may be added, which may be applied when the input value is converted to the output value.

[0088] In some embodiments, the combined network (33) may receive a vector that combines the first feature extracted from the first encoder and the second feature extracted from the second encoder as input, and may apply a nonlinear activation function to add nonlinearity to the fully connected layer, perform a linear transformation by applying weights and biases, and may predict the response probability by nonlinearly transforming it. In some embodiments, the combined network (33) may apply a sigmoid activation function as a nonlinear activation function. The sigmoid activation function may transform all input values ​​into values ​​between 0 and 1, and may have an S-shaped curve shape.

[0089] The reaction probability prediction module (140) inputs graph data regarding the amino acid sequence of a candidate enzyme and the graph of a target compound into the first encoder and the second encoder, and can perform calculations to predict the reaction probability of the candidate enzyme and the target compound reacting through a coupling network. The reaction probability prediction module (140) can store the calculated reaction probability in memory.

[0090] The learning module (150) can train the first encoder, the second encoder, and the combined network using a pre-prepared data set.

[0091] In some embodiments, the first encoder can predict enzyme properties from enzyme amino acid sequences. Here, the enzyme properties can be classified according to various methods. For example, the enzyme properties can be classified through GO (Gene Ontology), PRD (Protein Ontology), KEGG (Kyto Encyclopedia of Genes and Genomes) Brite / Ontology, InterPRO, MetaCyc, COGs (Cluster of Orthologous Groups of proteins), SEED, SCOPe, CATH, etc. In other words, the first encoder can predict enzyme properties from enzyme amino acid sequences in a form according to the classification methods listed above. For clarity and convenience of explanation, the following embodiments will be described mainly using EC numbers as an example. However, the classification method for predicting enzyme properties to which the invention is applied is not limited to the form of EC numbers, and may include various types of classification methods listed above.

[0092] In some embodiments, the first encoder may learn the reaction characteristics of the enzyme by predicting the EC number from the enzyme amino acid sequence. The EC number may include four levels (main level, sub-level, sub-sub-level, and naming level). For example, EC 1.2.1.3 is for acetaldehyde dehydrogenase (NAD). + -dependent) and can provide specific classifications as follows.

[0093] EC 1: Oxidoreductases - Enzymes that catalyze redox reactions.

[0094] EC 1.2: Using aldehyde or oxo group as donor, NAD + or NADP + An enzyme that uses as a receptor.

[0095] EC 1.2.1: NAD + or NADP + Aldehyde dehydrogenase that uses . as a receptor.

[0096] EC 1.2.1.3: Acetaldehyde dehydrogenase

[0097] Here, EC 1 can be the main level (or main hierarchy item), EC 1.2 can be the sub-level (or sub-hierarchy item), EC 1.2.1 can be the sub-sub-level (or sub-sub-hierarchy item), and EC 1.2.1.3 can be the naming level (or naming hierarchy item). Since each level consists of multiple classes, the problem of predicting EC numbers can correspond to a multi-level multi-class problem.

[0098] To solve multi-class problems, learning can be performed using Softmax Activation as the output layer. However, when using Softmax Activation Learning, predicting the EC number of a candidate enzyme, say acetaldehyde dehydrogenase, as 1.2.1.4 or 6.3.5.10 can both be treated as incorrect answers. However, from the perspective of multi-level, multi-class EC numbers, the incorrect answer of 1.2.1.4 is similar to the correct answer for the candidate enzyme, while the incorrect answer of 6.3.5.10 can be completely different from the correct answer. Therefore, a method based on Softmax Activation Learning may not be suitable for learning the reaction characteristics of enzymes.

[0099] In some embodiments, the first encoder may perform multi-class prediction for multiple levels of EC numbers using contrastive learning. The first encoder may be trained to generate two positive samples through data augmentation, maximize the similarity between the positive samples, and minimize the similarity between the negative samples through a contrastive loss function. Due to the nature of EC number data, the number of data for each number can vary significantly. For example, in the Swiss-Prot data of the UniProt database, the number of enzyme data corresponding to EC 2.1.3.15 is 980, while the number of enzyme data for EC 1.14.14.138 is 1, resulting in data imbalance. This data imbalance may hinder the neural network from learning the reaction characteristics of the enzyme. To address this data imbalance issue, contrastive learning, which can learn similarities and dissimilarities between data, may be introduced. In some embodiments, SimCLR (Simple Framework for Contrastive Learning of Visual Representations) may be used as a contrastive learning framework, and a loss function such as the following mathematical expression 1 may be used as a contrastive loss function.

[0100] (Equation 1)

[0101]

[0102] Here, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 can be represented.

[0103] In some other embodiments, a loss function such as Equation 2 below may be used.

[0104] (Equation 2)

[0105]

[0106] Here, is a multi-level term, and the enzyme The EC number can have a value of 0 if it is the same, 0.25 if the sub-subclasses are the same, 0.5 if it is the same up to the subclasses, 0.75 if the main classes are the same, and 1 if they are different.

[0107] In some embodiments, the second encoder may map the graph of the target compound into a continuous, number-based latent space based on a generative model. The generative model may include an encoder that maps material structures, which are typically discrete information, into a continuous, number-based latent space, and a decoder that converts values ​​in the latent space back into material structures. For example, a Variational AutoEncoder (VAE)-based generative model may encode material structures in training data into a latent space, add small noise, and then decode them back into structures, thereby reflecting the error from the original structure in the loss function to train the model. Through this training process, materials with similar structures are encoded in similar locations in the latent space, and similar structures can be generated when similar latent spaces are decoded. By utilizing the encoder part of such a generative model as a second encoder, latent space values ​​that reflect the structural similarity between compounds can be obtained.

[0108] In some embodiments, the second encoder can extract features for junctions at the junction tree level, which is larger than the atomic level. That is, a substance is divided into meaningful substructures, junctions, rather than atomic units. Input data such as junction labels, inter-junction coupling relationships, and junction positions are generated. A latent space is then obtained through the encoder, and the decoder can be made to predict the junction labels, order, and junction positions to be used for substance decoding. In this case, a loss function, such as the following mathematical expression 3, can be introduced.

[0109] (Equation 3)

[0110]

[0111] Here, is the Kullback-Leibler Divergence Loss (KL loss) between the encoder transform value distribution of x and the normal distribution, is the prediction accuracy of the junction, is the inter-junction adjacent prediction accuracy, is the accuracy of prediction of the junction-to-junction joint location, can be the L1 loss of the energy prediction value. Here, the L1 loss can represent a loss function that computes the difference between the predicted value and the actual value as the sum of the absolute errors.

[0112] In some embodiments, the second encoder may further include a network that predicts the energy of a reactant from the latent space mapping values. That is, by adding a neural network that predicts energy from the latent space transformation values ​​of a substance to the generative model and incorporating the property prediction results into the loss function, the latent space learned by the model can be aligned according to the material properties, i.e., energy. Energy is utilized to align the latent space because, in the reaction between a substance and an enzyme, not only structure but also energy changes are important factors. Through this process, reactants with similar structures and energies have similar values ​​in the latent space, and the enzyme-reactant reaction prediction model can reflect reactants and energies in its learning by utilizing the latent space values ​​of the reactants for learning.

[0113] The final candidate enzyme selection module (160) can select a final candidate enzyme based on the reaction probability calculated by the reaction probability prediction module (140). Specifically, the final candidate enzyme selection module (160) can select a final candidate enzyme that performs a reaction in a biosynthetic pathway for producing a target compound from among the candidate enzymes, and can store biosynthetic pathway data that optimizes the biosynthetic pathway using the final candidate enzyme in memory. In some embodiments, the final candidate enzyme selection module (160) can be implemented to calculate a score for the candidate enzyme and select the final candidate enzyme from among the candidate enzymes based on the score.

[0114] According to this embodiment, based on AI technology, enzymes capable of reacting with a designated target compound are searched for. Furthermore, AI models that have learned enzyme functions and transfer learning can be used to predict the likelihood of a reaction with the target compound, beyond simply predicting enzyme function. This allows for the discovery of promising enzymes that excel at specific reactions and accelerates the development of microbial strains that produce eco-friendly bio-based materials.

[0115] Figure 2 is a drawing for explaining an enzyme search device according to one embodiment.

[0116] Referring to FIG. 2, an enzyme search device according to an embodiment may include a candidate enzyme search model (20). The candidate enzyme search model (20) may include a first encoder (21), a second encoder (22), and a combination network (23). The first encoder (21) may be a protein encoder that receives an amino acid sequence of an enzyme as input and outputs a first embedding, and the second encoder (22) may be a compound encoder that receives a molecular graph of a compound as input and outputs a second embedding. The combination network (23) may combine the first embedding and the second embedding into one vector and perform a final prediction to output a reaction probability between a candidate enzyme and a target compound.

[0117] For details on the first encoder (21), the second encoder (22) and the coupling network (23), reference may be made to the description of the embodiments described herein.

[0118] FIG. 3 is a drawing for explaining a generation model related to a second encoder according to one embodiment, and FIG. 4 is a drawing for explaining a second encoder according to one embodiment.

[0119] Referring to FIG. 3, the generation model (30) related to the second encoder includes an encoder (31) and a decoder (32), and the latent space value, which is the output of the encoder, can be utilized for learning an enzyme-reactant prediction model.

[0120] Referring to Figure 4, the latent space distribution and energy calculation values ​​are shown. The graph depicts the first and second components of the PCA (Principal Component Analysis) of the latent space values ​​of the learning material after the latent space values ​​are calculated using the generative model after training has been completed. The color of each point represents the calculated energy value, and it can be confirmed that it is well-aligned from the left (dark, high energy) to the right (light, low energy).

[0121] FIG. 5 is a drawing for explaining an enzyme search method according to one embodiment.

[0122] Referring to FIG. 5, an enzyme search method according to an embodiment may include a step (S501) of providing a first encoder for which learning is performed based on functional characteristics of an enzyme, a step (S502) of providing a second encoder for which learning is performed based on structural characteristics of a compound, a step (S503) of constructing a combined network that combines the first encoder and the second encoder, and a step (S504) of inputting an amino acid sequence of a candidate enzyme and a graph of a target compound into the first encoder and the second encoder and predicting a reaction probability that the candidate enzyme and the target compound will react through the combined network.

[0123] The above method may be implemented to be performed by a computing device including a processor and a memory. In this case, in step (S501), the processor may load a first encoder, for which learning has been performed based on the functional characteristics of the enzyme, into the memory, and in step (S502), the processor may load a second encoder, for which learning has been performed based on the structural characteristics of the compound, into the memory. In step (S503), the processor may build a combining network that combines the first encoder and the second encoder through transfer learning, and in step (S504), the processor may input the amino acid sequence of the candidate enzyme and the graph data of the target compound into the first encoder and the second encoder, and may calculate a reaction probability for the candidate enzyme and the target compound to react through the combining network and store the result in the memory. In step (S505), the processor may select a final candidate enzyme that performs a reaction on a biosynthetic pathway for producing a target compound from among the candidate enzymes based on the calculated reaction probability, and store biosynthetic pathway data obtained by optimizing the biosynthetic pathway using the final candidate enzyme in memory.

[0124] For more detailed information on the above method, reference may be made to the description of the embodiments described in this specification, so redundant description is omitted here.

[0125] FIG. 6 is a drawing for explaining an enzyme function prediction device according to one embodiment, and FIG. 7 is a drawing for explaining an enzyme function prediction device according to one embodiment.

[0126] As explained above, when learning is performed using softmax activation as the output layer to solve multi-class problems, incorrect answers that are similar to the correct answer and incorrect answers that are completely different from the correct answer are not distinguished and treated in terms of the EC number of multi-level multi-classes. Therefore, the softmax activation learning-based method is not suitable for learning the reaction characteristics of enzymes. To overcome these limitations, improve the prediction accuracy for functional annotations with a hierarchical structure, and address the data imbalance problem, a total loss function is defined by introducing multiple base losses corresponding to the hierarchical structure of functional annotations and additional losses related to data augmentation. Accordingly, an enzyme function prediction model can be trained so that the features of enzymes with the same class at each level are sufficiently close to each other by reflecting the hierarchical structure of functional annotations.

[0127] Referring to FIGS. 6 and 7 together, the enzyme function prediction device (60) may include a learning data acquisition module (610), an enzyme function prediction model provision module (620), a basic loss introduction module (630), an additional loss introduction module (640), a learning module (650), and an enzyme function prediction module (660). The enzyme function prediction device (60) of FIG. 6 may correspond to the enzyme function prediction device (70) of FIGS. 7 and 8. The first encoder described above with reference to FIGS. 1 to 5 may include an enzyme function prediction model that predicts enzyme function through prediction of functional annotations of proteins. In this way, the enzyme function prediction model included in the first encoder may be one for which learning has been performed by the enzyme function prediction device (60, 70). The first encoder may perform feature extraction based on the functional characteristics of the enzyme using the enzyme function prediction model.

[0128] The learning data acquisition module (610) can acquire learning data associated with the functional characteristics of an enzyme. In some embodiments, the learning data acquisition module (610) can acquire data containing information about the amino acid sequence of an enzyme as learning data.

[0129] The enzyme function prediction model providing module (620) can provide an enzyme function prediction model (71). In addition, the enzyme function prediction model providing module (620) can provide a plurality of activation layers (721, 722, 723, 724) corresponding to a plurality of levels following the enzyme function prediction model (71).

[0130] The base loss introduction module (130) can introduce multiple base losses for each of the multiple activation layers (221, 222, 223, 224). The multiple base losses can include the first base loss to the Mth base loss (M is an integer greater than or equal to 2). In FIG. 2, the multiple base losses include the first base loss ( ), second basic loss ( ), third basic loss ( ) and the fourth basic loss ( ) is shown to include.

[0131] As described above, the functional annotation may include first to M-th layer items (M is an integer greater than or equal to 2) corresponding to multiple levels. In this case, the first to M-th basic losses may correspond to the first to M-th layer items, respectively. For example, when the functional annotation includes an EC number, the functional annotation may include a main layer item, a sub-layer item, a sub-sub-layer item, and a naming layer item. In this case, the main layer item may include the first basic loss ( ) corresponds to the second basic loss among the multiple basic losses, and the sub-layer items correspond to the second basic loss ( ) corresponds to the third basic loss among the multiple basic losses, and the sub-sublayer items correspond to the third basic loss ( ) corresponds to the naming hierarchy entry, and the 4th basic loss among the multiple basic losses ( ) can respond.

[0132] The additional loss introduction module (640) can apply data augmentation to the enzyme function prediction model (71) and introduce additional loss. In this case, the enzyme function prediction model (71) may be for performing multi-class prediction for multiple levels of functional annotation using contrastive learning. Due to the characteristics of EC number data, it can be seen that the number of data for each number greatly differs. For example, in the Swiss-Prot data of the UniProt database, the number of enzyme data corresponding to EC 2.1.3.15 is 980, whereas the number of enzyme data for EC 1.14.14.138 is 1, which means that there is a data imbalance. Such data imbalance can hinder the neural network from learning the reaction characteristics of the enzyme. To solve this data imbalance problem, contrastive learning can be introduced that can learn similarities and dissimilarities between data. Specifically, the additional loss introduction module (640) can introduce additional loss to generate positive samples through data augmentation, maximize the similarity between positive samples, and minimize the similarity between negative samples.

[0133] The learning module (150) can define a total loss function based on a plurality of base losses introduced by the base loss introduction module (130) and an additional loss introduced by the additional loss introduction module (140). In some embodiments, the learning module (150) can define a total loss function according to the following mathematical expression 4. can be defined.

[0134] (Equation 4)

[0135]

[0136] Here, is the kth basic loss among the multiple basic losses (k is an integer greater than or equal to 1 and less than or equal to M), may be an additional loss.

[0137] In some embodiments, the additional loss introduction module (140) may define the additional loss according to the following mathematical expression (5).

[0138] (Equation 5)

[0139]

[0140] Here, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 can be represented.

[0141] In some embodiments, the additional loss introduction module (140) may define the additional loss according to the following mathematical expression (6).

[0142] (Equation 6)

[0143]

[0144] Here, is a multilevel term, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 can be represented.

[0145] In some embodiments, the learning module (650) calculates a total loss function according to the following mathematical expression 7: can also be defined. The difference from Equation 4 is that a predetermined weight is applied to each of the basic loss and additional loss.

[0146]

[0147] (Equation 7)

[0148]

[0149] Here, is the kth basic loss among the multiple basic losses (k is an integer greater than or equal to 1 and less than or equal to M), Silver additional loss, is a predefined weight for the k-th basic loss, may be a predefined weight for additional losses.

[0150] If the functional annotation includes an EC number, is an enzyme For EC numbers, it can be set to have a value of 0 if they are the same, 0.25 if the sub-subclasses are the same, 0.5 if they are the same up to the subclasses, 0.75 if the main classes are the same, and 1 if they are different. According to this setting, the prediction accuracy for functional annotations with a hierarchical structure can be improved and the data imbalance problem can be solved by ensuring that the features of enzymes with the same class at each level are located sufficiently close to each other by reflecting the hierarchical structure of the functional annotation.

[0151] Additionally, the learning module (150) can train the enzyme function prediction model (31) based on the learning data acquired by the learning data acquisition module (110) and the total loss function. Specifically, the learning module (150) can change the parameter values ​​of the enzyme function prediction model (31) based on the learning data and the total loss function.

[0152] The enzyme function prediction module (660) can predict functional annotations using a trained enzyme function prediction model (71). In some embodiments, as previously described with reference to FIGS. 1 to 5 , the results predicted by the enzyme function prediction module (660) based on the functional characteristics of an enzyme may be combined with other results predicted based on the structural characteristics of a compound to predict the likelihood of a reaction between a candidate enzyme and a target compound.

[0153] According to the present embodiment, by defining a total loss function by introducing multiple base losses corresponding to the hierarchical structure and additional losses related to data augmentation to perform learning of an enzyme function prediction model, the hierarchical structure of functional annotations is reflected, so that features of enzymes with the same class at each level are positioned sufficiently close to each other, thereby improving the prediction accuracy for functional annotations with a hierarchical structure and solving the data imbalance problem. In addition, candidate enzymes that can react with a designated target compound are searched for based on artificial intelligence technology, and through an artificial intelligence model that has learned enzyme function and transfer learning, it is applied not only to enzyme function prediction but also to predict the possibility of reaction with a target compound, thereby discovering promising enzymes that can perform a specific reaction well, and accelerating the development of microbial strains that produce eco-friendly bio-based materials.

[0154] FIG. 8 is a drawing for explaining an enzyme function prediction device according to one embodiment.

[0155] Referring to FIG. 8, in an enzyme function prediction device (70) according to one embodiment, an enzyme function prediction model (71) may include an embedding layer (711), an attention layer (712), and a feedforward network layer (713). As described in FIG. 7, a plurality of activation layers (721, 722, 723, 724) may be formed following the enzyme function prediction model (71).

[0156] The embedding layer (711) can convert input data, for example, data containing information about the amino acid sequence of an enzyme, into a high-dimensional vector. In other words, the embedding layer can digitize the input data so that the model can process it.

[0157] The attention layer (712) may receive a vector embedded by the embedding layer (711) and assign weights to each element of the input sequence to emphasize important parts. In some embodiments, the attention layer (712) may include a FAVOR+ (Fast Attention via Orthogonal Random Feature) attention layer. By adopting the FAVOR+ attention layer, computational efficiency can be improved by accelerating attention calculation using orthogonal random features.

[0158] The feedforward network layer (713) processes the output of the attention layer (712) to apply a nonlinear transformation and can independently transform features at each location. For example, the feedforward network layer (713) can generate an output vector by applying two linear transformations (e.g., a fully connected layer) and a nonlinear activation function (e.g., ReLU) between them.

[0159] According to this embodiment, the embedding layer (711), the attention layer (712), and the feedforward network layer (713) complement each other to learn complex patterns and interactions of amino acid sequence data of an enzyme, and high performance, efficient computation, and powerful expression capabilities can be expected in functional annotation prediction.

[0160] FIG. 9 is a diagram illustrating an enzyme search method based on enzyme function prediction according to one embodiment.

[0161] Referring to FIG. 9, an enzyme search method according to an embodiment may include a step of obtaining learning data associated with functional characteristics of an enzyme (S901), a step of obtaining an enzyme function prediction model (S902), a step of obtaining a plurality of activation layers corresponding to a plurality of levels following the enzyme function prediction model (S903), a step of introducing a plurality of base losses for each of the plurality of activation layers (S904), a step of applying data augmentation to the enzyme function prediction model and introducing additional losses (S905), a step of defining a total loss function based on the plurality of base losses and the additional losses (S906), and a step of training the enzyme function prediction model based on the learning data and the total loss function (S907). The method may be implemented to be performed by a computing device including a processor and a memory. In this case, each of the steps (S901-S907) may be performed by the processor.

[0162] For more detailed information on the above method, reference may be made to the description of the embodiments described in this specification, so redundant description is omitted here.

[0163] FIG. 10 is a drawing for explaining a computing device according to one embodiment.

[0164] Referring to FIG. 10, the enzyme search method and device according to the embodiments may be implemented using a computing device (50). This computing device (50) may be implemented as various types of electronic devices, servers, or similar devices, and its functions may be implemented through a combination of software and hardware.

[0165] The computing device (50) may include at least one of a processor (510), a memory (530), a user interface input device (540), a user interface output device (550), and a storage device (560) that communicate via a bus (520). The computing device (50) may also include a network interface (570) electrically connected to a network (40). The network interface (570) may transmit or receive signals to or from other entities via the network (40).

[0166] The processor (510) may be implemented as various types of computing devices, such as an MCU (Micro Controller Unit), an AP (Application Processor), a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an NPU (Neural Processing Unit), a QPU (Quantum Processing Unit), etc. The processor (510) is a semiconductor device that executes instructions stored in a memory (530) or a storage device (560), and may play a core role in the system. The program code and data stored in the memory (530) or the storage device (560) instruct the processor (510) to perform a specific task, thereby enabling the overall operation of the system. The processor (510) may be configured to implement various functions and methods described above with reference to FIGS. 1 to 9.

[0167] The memory (530) and storage device (560) may include various types of volatile or non-volatile storage media for storing and accessing data of the system. For example, the memory (530) may include read-only memory (ROM) (531) and random access memory (RAM) (532). In some embodiments, the memory (530) may be built into the processor (510), in which case the data transfer speed between the memory (530) and the processor (510) may be very fast. In other embodiments, the memory (530) may be located external to the processor (510), in which case the memory (530) may be connected to the processor (510) through various data buses or interfaces. Such connection may be made through various means already known in the art, for example, a Peripheral Component Interconnect Express (PCIe) interface for high-speed data transfer or a memory controller.

[0168] In some embodiments, at least some components or functions of the enzyme search method and device according to the embodiments may be implemented as a program or software running on a computing device (50), and the program or software may be stored on a computer-readable recording medium or storage medium. Specifically, the computer-readable recording medium or storage medium according to one embodiment may be a program recorded on a computer including a processor (510) that executes a program or instructions stored in a memory (530) or a storage device (560) to execute steps included in the implementation of the enzyme search method and device according to the embodiments.

[0169] In some embodiments, at least some components or functions of the enzyme search methods and devices according to the embodiments may be implemented using hardware or circuitry of the computing device (50), or may be implemented as separate hardware or circuitry that can be electrically connected to the computing device (50).

[0170] In some embodiments, a computing device (50) is provided with one or more non-transitory computer-readable media containing executable instructions, which, when executed by one or more processors of the computing device (50), cause the computing device (50) to perform operations. Here, the operations may include the components, functions, steps, etc. described herein with respect to the enzyme search methods and devices according to the embodiments.

[0171] According to the embodiments, AI technology based on the functional characteristics of enzymes and the structural characteristics of compounds can be used to search for enzymes capable of reacting with a specified target compound. Using AI models that have learned enzyme function and transfer learning, the potential for reactions with target compounds can be predicted, beyond simply predicting enzyme function. This can identify promising enzymes that excel at specific reactions and accelerate the development of microbial strains that produce eco-friendly bio-based materials.

[0172] Furthermore, according to embodiments, enzyme function prediction is provided, and to train an enzyme function prediction model, a total loss function can be defined by introducing multiple base losses corresponding to the hierarchical structure and additional losses related to data augmentation. Accordingly, the hierarchical structure of functional annotations can be reflected, ensuring that features of enzymes with the same class at each level are sufficiently close together, thereby improving prediction accuracy for functional annotations with a hierarchical structure and addressing data imbalance issues.

[0173] Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements made by a person of ordinary skill in the art to which the present invention pertains using the basic concept of the present invention defined in the following claims also fall within the scope of the present invention.

Claims

1. An enzyme search method performed by a computing device including a processor and a memory, The step of the above processor loading the first encoder, in which learning is performed based on the functional characteristics of the enzyme, into the memory; The step of the processor loading a second encoder, in which learning is performed based on the structural characteristics of the compound, into the memory; A step in which the processor constructs a combined network that combines the first encoder and the second encoder through transfer learning; The step of the processor inputting the amino acid sequence of the candidate enzyme and the graph data of the target compound into the first encoder and the second encoder, calculating the reaction probability that the candidate enzyme and the target compound will react through the coupling network, and storing the result in the memory; and The processor comprises a step of selecting a final candidate enzyme that performs a reaction on a biosynthetic pathway for producing the target compound from among the candidate enzymes based on the calculated reaction probability, and storing biosynthetic pathway data that optimizes the biosynthetic pathway using the final candidate enzyme in the memory. Enzyme search method.

2. In paragraph 1, The steps of constructing the above combined network are: The above processor includes a step of combining the first feature extracted from the first encoder and the second feature extracted from the second encoder into one vector through a fully connected layer, The step of calculating the above reaction probability and storing it in the memory is: An enzyme search method, comprising a step of the processor applying a nonlinear activation function to the fully connected layer to nonlinearly transform the vector and calculate the reaction probability.

3. In paragraph 1, The above first encoder, An enzyme search method that learns the reaction characteristics of an enzyme by predicting the EC number (Enzyme Commission Number) from the enzyme amino acid sequence, and performs multi-class prediction for multi-levels of the EC number using contrastive learning.

4. In paragraph 3, The above first encoder, Generate two positive samples through data augmentation, An enzyme search method that is learned to maximize the similarity between positive samples and minimize the similarity between negative samples through a contrastive learning function.

5. In paragraph 1, The above second encoder, An enzyme search method that maps the graph of the target compound to a continuous number-based latent space based on a generative model and extracts features for junctions in units of junction trees larger than atomic units.

6. In paragraph 5, The above second encoder, An enzyme search method further comprising a network for predicting the energies of reactants from latent space mapping values.

7. In paragraph 1, The first encoder includes an enzyme function prediction model that predicts enzyme function through prediction of functional annotation of a protein, A step in which the above processor acquires learning data associated with functional characteristics of an enzyme; A step in which the processor introduces a plurality of base losses for each of a plurality of activation layers corresponding to a plurality of levels; The step of the processor applying data augmentation to the enzyme function prediction model and introducing additional loss; The step of the processor defining a total loss function based on the plurality of basic losses and the additional losses; and An enzyme search method, further comprising a step of allowing the processor to train the enzyme function prediction model based on the learning data and the total loss function.

8. In paragraph 7, The above multiple basic losses include the first basic loss to the Mth basic loss (M is an integer greater than or equal to 2), The steps for defining the above total loss function are: The total loss function according to the following mathematical formula 1 A method for searching an enzyme, comprising the steps of: (Equation 1) Here, is the kth basic loss among the above basic losses (k is an integer greater than or equal to 1 and less than or equal to M), The above additional loss, is a predetermined weight for the above k basic loss, is a predetermined weight for the above additional loss.

9. In paragraph 8, The above additional loss is defined according to the following mathematical formula 2, enzyme search method: (Equation 2) Here, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 is indicated.

10. In paragraph 8, The above additional loss is defined according to the following mathematical formula 3, enzyme search method: (Equation 3) Here, is a multilevel term, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 is indicated.

11. One or more non-transitory computer-readable media containing instructions; and One or more processors that perform an operation by executing the above instructions, wherein the operation comprises: A first encoder, in which learning is performed based on the functional characteristics of the enzyme, is loaded onto the medium, A second encoder, in which learning is performed based on the structural characteristics of the compound, is loaded onto the medium, A combined network is constructed by combining the first encoder and the second encoder through transfer learning, The amino acid sequence of the candidate enzyme and the graph data of the target compound are input to the first encoder and the second encoder, and the reaction probability that the candidate enzyme and the target compound will react is calculated through the coupling network and stored in the medium, Based on the calculated reaction probability, a final candidate enzyme that performs a reaction on a biosynthetic pathway for producing the target compound is selected from among the candidate enzymes, and biosynthetic pathway data that optimizes the biosynthetic pathway using the final candidate enzyme is stored in the medium. Enzyme search device.

12. In paragraph 11, Building the above combined network is: Combining the first feature extracted from the first encoder and the second feature extracted from the second encoder into one vector through a fully connected layer, The above calculation is, An enzyme search device comprising a nonlinear transformation of the vector by applying a nonlinear activation function to the fully connected layer to predict the reaction probability.

13. In paragraph 11, The above first encoder, An enzyme search device that learns the reaction characteristics of an enzyme by predicting the EC number (Enzyme Commission Number) from the enzyme amino acid sequence, and performs multi-class prediction for multi-levels of the EC number using contrastive learning.

14. In paragraph 13, The above first encoder, Generate two positive samples through data augmentation, An enzyme search device that is trained to maximize the similarity between positive samples and minimize the similarity between negative samples through a contrastive learning function.

15. In paragraph 11, The above second encoder, An enzyme search device that maps the graph of the target compound to a continuous number-based latent space based on a generative model and extracts features for junctions in units of junction trees larger than atomic units.

16. In paragraph 15, The above second encoder, An enzyme search device further comprising a network for predicting the energy of a reactant from latent space mapping values.

17. In paragraph 11, The first encoder includes an enzyme function prediction model that predicts enzyme function through prediction of functional annotation of a protein, The above action is, Obtain learning data associated with the functional characteristics of enzymes, Introducing multiple base losses for each of multiple activation layers corresponding to multiple levels, Applying data augmentation to the above enzyme function prediction model and introducing additional loss, Define a total loss function based on the above multiple basic losses and the above additional losses, An enzyme search device further comprising training the enzyme function prediction model based on the learning data and the total loss function.

18. In paragraph 17, The above multiple basic losses include the first basic loss to the Mth basic loss (M is an integer greater than or equal to 2), Defining the above total loss function is: The total loss function according to the following mathematical formula 1 An enzyme search device comprising: (Equation 1) Here, is the kth basic loss among the above basic losses (k is an integer greater than or equal to 1 and less than or equal to M), The above additional loss, is a predetermined weight for the above k basic loss, is a predetermined weight for the above additional loss.

19. In paragraph 18, The above additional loss is defined by the following mathematical equation 2, enzyme search device: (Equation 2) Here, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 is indicated.

20. In paragraph 18, The above additional loss is defined by the following mathematical formula 3, enzyme search device: (Equation 3) Here, is a multilevel term, and is the embedding vector of the positive pair, Is and is the cosine similarity of , is the temperature parameter, is the batch size, Is When 1 is indicated.

Images