Enzyme mining method, system and storage medium

Abstract

The application provides an enzyme mining method, system and medium. The enzyme mining method comprises the following steps: inputting to-be-queried data; inputting the to-be-queried data into a pre-trained VenusRXN model; when the input to-be-queried data is a reaction, determining a reaction CLS embedding vector by using a reaction encoder, and then mapping the reaction CLS embedding vector into a reaction query embedding vector through a linear layer corresponding to the reaction encoder, calculating the similarity of the reaction query embedding vector and each protein embedding vector and sorting, and outputting protein sequences corresponding to multiple protein embedding vectors with the highest similarity; or when the input to-be-queried data is a template enzyme, determining a template enzyme CLS embedding vector by using a protein encoder, and then mapping the template enzyme CLS embedding vector into a template enzyme query embedding vector through a linear layer corresponding to the protein encoder, calculating the similarity of the template enzyme query embedding vector and each protein embedding vector and sorting, and outputting protein sequences corresponding to multiple protein embedding vectors with the highest similarity.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and more particularly to an enzyme mining method, an enzyme mining system, and a computer-readable storage medium. Background Technology

[0002] Enzymes, a subclass of proteins, act as catalysts in biological systems and are the core executors of various biological functions. They can accelerate and regulate almost all chemical processes and metabolic pathways in organisms, and have wide applications in drug synthesis, gene editing, bioremediation, and plastic degradation. Although billions of proteins have been sequenced, less than 1% of them have been fully studied and have functional annotations. Therefore, function-based large-scale enzyme mining has become particularly important and a key step in many scientific and applied researches. Recent enzyme mining research has mainly focused on two aspects: inferring the function of unlabeled proteins by comparing the sequence or structural similarity between unlabeled proteins and enzymes with known functions; or training supervised learning models on functionally annotated protein databases to classify and screen unlabeled proteins. However, the first type of method has limitations because protein sequence or structural similarity is not necessarily consistent with functional similarity, and there is a lack of template enzymes as a starting point when studying novel catalytic reactions. On the other hand, the second type of method also has limitations because current protein functional annotation systems cannot accurately reflect the reaction or substrate specificity of enzymes, and related annotation databases are still incomplete. In summary, there is an urgent need to develop an enzyme discovery method that can directly link enzymes to their core function—catalyzing chemical reactions—and that can be generalized to entirely new reactions.

[0003] To overcome the aforementioned deficiencies in the existing technology, the present invention provides an enzyme mining method, an enzyme mining system, and a computer-readable storage medium, which can mine enzymes with specific catalytic functions from protein databases based on a given chemical reaction or template enzyme, effectively reducing the dependence on protein similarity and enzyme function annotation, and balancing the accuracy, speed, and practicality of enzyme mining. Summary of the Invention

[0004] The following provides a brief overview of one or more aspects to offer a basic understanding of them. This overview is not an exhaustive summary of all conceived aspects, nor is it intended to identify key or decisive elements of all aspects, nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed descriptions that follow.

[0005] To overcome the aforementioned deficiencies in the existing technology, the present invention provides an enzyme mining method, an enzyme mining system, and a computer-readable storage medium, which can mine enzymes with specific catalytic functions from protein databases based on a given chemical reaction or template enzyme, effectively reducing the dependence on protein similarity and enzyme function annotation, and balancing the accuracy, speed, and practicality of enzyme mining.

[0006] Specifically, the enzyme mining method provided by the first aspect of the present invention is characterized by comprising the following steps: inputting query data, the query data including a reaction or a template enzyme, the reaction being represented by the string SMILES and the template enzyme being represented by an amino acid sequence; inputting the query data into a pre-trained VenusRXN model, the VenusRXN model including a reaction encoder and a protein encoder; when the input query data is a reaction, using the reaction encoder to determine the reaction CLS embedding vector, and then mapping the reaction CLS embedding vector to a reaction query embedding vector via the linear layer corresponding to the reaction encoder, and calculating the reaction... The algorithm queries the cosine similarity between the embedded vector and each protein embedded vector in the protein embedded vector database, sorts them, and outputs the protein sequences corresponding to the multiple protein embedded vectors with the highest cosine similarity. Alternatively, when the input query data is a template enzyme, the protein encoder is used to determine the template enzyme CLS embedded vector, and then the template enzyme CLS embedded vector is mapped to a template enzyme query embedded vector through the linear layer corresponding to the protein encoder. The cosine similarity between the template enzyme query embedded vector and each protein embedded vector in the protein embedded vector database is calculated and sorted, and the protein sequences corresponding to the multiple protein embedded vectors with the highest cosine similarity are output.

[0007] Furthermore, in some embodiments of the present invention, the step of constructing the protein embedding vector database includes: obtaining a protein sequence database, wherein the protein sequence database preferably contains all enzyme sequences in a reaction-enzyme pairing dataset; using the protein encoder to determine the CLS embedding vector of each protein sequence in the protein sequence database; and mapping the protein sequence CLS embedding vector to the corresponding protein embedding vector via a linear layer corresponding to the protein encoder, thereby constructing the protein embedding vector database.

[0008] Furthermore, in some embodiments of the present invention, the training steps of the VenusRXN model include: constructing a reaction-enzyme pairing dataset, the reaction-enzyme pairing dataset including multiple reactions and enzyme sequences corresponding to the reactions; performing multimodal training of the VenusRXN model on the reaction-enzyme pairing dataset, the multimodal training including calculating cross-modal feature alignment loss and calculating cross-modal feature fusion loss; and calculating multimodal training loss, optimizing the learnable parameters of the VenusRXN model based on the multimodal training loss until the maximum number of training epochs is reached or the multimodal training loss does not decrease within a specified number of epochs.

[0009] Furthermore, in some embodiments of the present invention, the step of calculating the cross-modal feature alignment loss value for the VenusRXN model includes: inputting each of the responses in the training batch into the response encoder to obtain the response CLS embedding vector of each response, and inputting the response CLS embedding vector of each response into the linear layer corresponding to the response encoder to obtain the response embedding vector of each response. ,in, This represents the number of reaction-enzyme pairs in the training batch. Each enzyme sequence in the training batch is input into the protein encoder to obtain the protein CLS embedding vector for each enzyme sequence. The protein CLS embedding vectors for each enzyme sequence are then input into the corresponding linear layer of the protein encoder to obtain the enzyme embedding vector for each enzyme sequence. ,in, The number of reaction-enzyme pairs in the training batch is represented; the contrastive learning loss value and the soft label alignment loss value are calculated based on the reaction embedding vector r and the enzyme embedding vector e; and the cross-modal feature alignment loss value is calculated based on the contrastive learning loss value and the soft label alignment loss value.

[0010] Furthermore, in some embodiments of the present invention, the step of calculating the contrastive learning loss value includes: calculating the normalized reaction-enzyme similarity in the same training batch. Similarity to enzyme-reaction :

[0011]

[0012]

[0013] in, This represents the number of reaction-enzyme pairs in the training batch. Let be the cosine similarity function of vectors. For trainable temperature parameters; and based on the reaction-enzyme similarity. and the enzyme-reaction similarity Calculate the contrastive learning loss value, where the loss for the i-th reaction-enzyme pair is:

[0014] .

[0015] Furthermore, in some embodiments of the present invention, the step of calculating the soft label alignment loss value includes: calculating the normalized response-response similarity in the same training batch. Enzyme-enzyme similarity :

[0016]

[0017]

[0018] in, This represents the number of reaction-enzyme pairs in the training batch. Let be the cosine similarity function of vectors. For trainable temperature parameters; and based on the reaction-enzyme similarity. The enzyme-reaction similarity The reaction-reaction similarity and the enzyme-enzyme similarity Calculate the soft label alignment loss value: .

[0019] Furthermore, in some embodiments of the present invention, the cross-modal feature alignment loss value is:

[0020]

[0021] in, The cross-modal feature alignment loss value is... To compare the first loss value of learning, This is the second loss value for soft label alignment.

[0022] Further, in some embodiments of the present invention, the step of calculating the cross-modal feature fusion loss value includes: inputting the reactions in the reaction-enzyme pairing dataset into the reaction encoder to obtain an atomic embedding matrix Hr composed of stacked atomic embedding vectors in the reaction compression map of the reaction; inputting the atomic embedding matrix Hr into the reaction-protein joint encoder, wherein the reaction-protein joint encoder inserts a cross-attention layer in the Transformer block of the protein language model; inputting each enzyme sequence corresponding to each reaction in the reaction-enzyme pairing dataset into the reaction-protein joint encoder; obtaining an initial amino acid embedding matrix He composed of stacked initial amino acid embedding vectors in the enzyme sequence in each Transformer block of the reaction-protein joint encoder; and calculating an updated amino acid embedding matrix using a cross-attention mechanism in each Transformer block of the reaction-protein joint encoder based on the atomic embedding matrix Hr and the initial amino acid embedding matrix He. :

[0023]

[0024] in, To query the parameter matrix, For the key parameter matrix, For the value parameter matrix, For the hidden layer dimension of the reaction-protein co-encoder, the amino acid-updated embedding matrix is ​​used to update the embedding matrix. The amino acid initial embedding matrix is ​​used as the next Transformer block in the reaction-protein joint encoder; and the reaction-protein joint CLS embedding vector output by the reaction-protein joint encoder is input into the matching head to calculate the matching score of the reaction-enzyme pairing. Based on the matching score, the cross-modal feature fusion loss value is calculated.

[0025]

[0026] in, This is the true label for the reaction-enzyme pairing. A value of 0 indicates a mismatch. A value of 1 indicates a match. It is the sigmoid activation function. The matching score is the score for the reaction-enzyme pairing.

[0027] Furthermore, in some embodiments of the present invention, the step of calculating the multimodal training loss value includes: calculating the multimodal training loss value based on the cross-modal feature alignment loss value and the cross-modal feature fusion loss value; calculating the multimodal training loss value.

[0028]

[0029] in, The multimodal training loss value is... The cross-modal feature alignment loss value is... This is the cross-modal feature fusion loss value.

[0030] Furthermore, the enzyme mining system provided according to a second aspect of the present invention includes a memory and a processor. The memory stores computer instructions. The processor is connected to the memory and configured to execute the computer instructions stored thereon to implement the enzyme mining method as provided in the first aspect of the present invention.

[0031] Furthermore, the computer-readable storage medium provided according to the third aspect of the present invention stores computer instructions thereon. When the computer instructions are executed by a processor, the enzyme mining method as provided in the first aspect of the present invention is implemented. Attached Figure Description

[0032] The above-described features and advantages of the present invention will be better understood after reading the following detailed description of embodiments of the present disclosure in conjunction with the accompanying drawings. In the drawings, components are not necessarily drawn to scale, and components having similar related characteristics or features may have the same or similar reference numerals.

[0033] Figure 1 A schematic diagram of a reaction data preprocessing flow provided according to some embodiments of the present invention is shown.

[0034] Figure 2 A schematic diagram of the architecture of a molecular graph feature extractor provided according to some embodiments of the present invention is shown.

[0035] Figure 3 A schematic diagram of the pre-training process of a molecular graph feature extractor provided according to some embodiments of the present invention is shown.

[0036] Figure 4 A schematic diagram of the architecture of a reaction encoder provided according to some embodiments of the present invention is shown.

[0037] Figure 5 A schematic diagram of the training process of a reaction encoder provided according to some embodiments of the present invention is shown.

[0038] Figure 6 A flowchart illustrating a method for predicting enzyme classification and numbering of enzyme-catalyzed reactions according to some embodiments of the present invention is shown.

[0039] Figure 7A schematic diagram showing a performance comparison of enzyme classification and numbering prediction methods for enzyme-catalyzed reactions provided according to some embodiments of the present invention is illustrated.

[0040] Figure 8 A schematic diagram of the training process of the VenusRXN model provided according to some embodiments of the present invention is shown.

[0041] Figure 9 A schematic flowchart of an enzyme mining method according to some embodiments of the present invention is shown.

[0042] Figure 10 The diagrams show pretreated reactant molecular diagrams and product molecular diagrams provided according to some embodiments of the present invention.

[0043] Figure 11 A schematic diagram showing a performance comparison of reaction-based enzyme mining methods provided according to some embodiments of the present invention is illustrated.

[0044] Figure 12 A schematic diagram showing a performance comparison of template enzyme-based enzyme mining methods provided according to some embodiments of the present invention is illustrated.

[0045] Figure 13 A schematic diagram of the fine-tuning process of a reaction encoder and a reaction-protein co-encoder provided according to some embodiments of the present invention is shown.

[0046] Figure 14 A performance comparison diagram of enzyme catalytic performance prediction methods based on chemical reactions provided according to some embodiments of the present invention is shown. Detailed Implementation

[0047] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Although the description of the present invention is presented in conjunction with preferred embodiments, this does not mean that the features of the invention are limited to these embodiments. On the contrary, the purpose of describing the invention in conjunction with embodiments is to cover other options or modifications that may be derived based on the claims of the present invention. To provide a thorough understanding of the invention, many specific details will be included in the following description. The invention may also be implemented without using these details. Furthermore, to avoid confusion or obscuring the focus of the invention, some specific details will be omitted in the description.

[0048] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0049] Furthermore, the terms "upper," "lower," "left," "right," "top," "bottom," "horizontal," and "vertical" used in the following description should be understood as the orientations shown in the relevant paragraphs and accompanying drawings. These relative terms are for illustrative purposes only and do not imply that the described apparatus must be manufactured or operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0050] It is understood that although terms such as "first," "second," and "third" may be used herein to describe various components, regions, layers, and / or parts, these components, regions, layers, and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers, and / or parts. Therefore, the first components, regions, layers, and / or parts discussed below may be referred to as second components, regions, layers, and / or parts without departing from some embodiments of the present invention.

[0051] As mentioned above, existing enzyme numbering prediction techniques for enzyme-catalyzed reactions have significant shortcomings, limiting their accuracy and widespread application. Existing machine learning-based enzyme classification and numbering prediction methods for enzyme-catalyzed reactions are often limited to predicting a few major categories or only performing binary classification tasks, resulting in insufficient generalization ability. Furthermore, recently developed deep learning models have not adequately addressed the problem of imbalanced enzyme classification and numbering, and most rely solely on simple information about substrates and products, failing to fully learn the deep features of the reaction, leading to unsatisfactory prediction accuracy for rare or complex reactions.

[0052] To overcome the aforementioned deficiencies in the prior art, this invention provides a method for predicting enzyme classification numbers for enzyme-catalyzed reactions, a system for predicting enzyme classification numbers for enzyme-catalyzed reactions, and a computer-readable storage medium, which can effectively extract molecular structural features and chemical reaction rules in enzyme-catalyzed reactions to accurately predict their corresponding enzyme classification numbers.

[0053] In some non-limiting embodiments, the enzyme classification and numbering prediction system for the above-described enzyme-catalyzed reaction provided in the second aspect of the present invention can be implemented based on the enzyme classification and numbering prediction method for the above-described enzyme-catalyzed reaction provided in the first aspect of the present invention.

[0054] Specifically, in some non-limiting embodiments, the enzyme classification and numbering prediction system for the above-mentioned enzymatic reactions provided in the second aspect of the present invention includes a memory and a processor. Here, the memory includes, but is not limited to, the computer-readable storage medium provided in the third aspect above, on which computer instructions are stored. The processor is connected to the memory and configured to execute the computer instructions stored in the memory to implement the enzyme classification and numbering prediction method for enzymatic reactions as provided in the first aspect of the present invention.

[0055] The working principle of the enzyme classification and numbering prediction system for enzyme-catalyzed reactions will be described below with reference to some embodiments of enzyme classification and numbering prediction methods for enzyme-catalyzed reactions. Those skilled in the art will understand that these embodiments of prediction methods are merely non-limiting implementations provided by the present invention, intended to clearly demonstrate the main concepts of the invention and provide specific solutions convenient for public implementation, rather than limiting all functions or all working methods of the enzyme classification and numbering prediction system for enzyme-catalyzed reactions. Similarly, the enzyme classification and numbering prediction system for enzyme-catalyzed reactions is also only one non-limiting implementation provided by the present invention, and does not constitute a limitation on the executing entity and execution order of each step in these enzyme classification and numbering prediction methods for enzyme-catalyzed reactions.

[0056] In some non-limiting embodiments, the enzyme classification number prediction method for the enzyme-catalyzed reaction provided in the first aspect of the present invention can be implemented in two stages: training and prediction. Specifically, a person skilled in the art can first train a reaction encoder for extracting chemical reaction features in the training stage, and then, in the prediction stage, use the reaction encoder to extract the CLS embedding vector of the test reaction, use a multilayer perceptron to map the CLS embedding vector of the test reaction to a query embedding vector of the test reaction, calculate the cosine similarity between the query embedding vector of the test reaction and each training reaction embedding vector in the training reaction embedding vector database, and finally output the enzyme classification number corresponding to the training reaction embedding vector with the highest cosine similarity to predict the enzyme classification number corresponding to the test reaction.

[0057] Furthermore, the aforementioned reaction encoder includes a molecular graph feature extractor and a reaction rule extractor, which are connected in series. Here, the molecular feature extractor is used to determine the reactant CLS embedding vector and the embedding vectors of all atoms in the reactant in the reactant molecular graph, and the product CLS embedding vector and the embedding vectors of all atoms in the product in the product molecular graph, respectively, based on the reactant molecular graph and the product molecular graph. The reaction rule extractor is used to determine the reaction CLS embedding vector based on the Condensed Graph of Reaction (CGR).

[0058] First, please refer to Figures 1-3 . Figure 1A schematic diagram of a reaction data preprocessing flow provided according to some embodiments of the present invention is shown. Figure 2 A schematic diagram of the architecture of a molecular graph feature extractor provided according to some embodiments of the present invention is shown. Figure 3 A schematic diagram of the pre-training process of a molecular graph feature extractor provided according to some embodiments of the present invention is shown.

[0059] like Figure 1 As shown, a reaction-enzyme pairing dataset is first constructed for the subsequent training process of the VenusRXN model.

[0060] Specifically, technicians can first extract enzyme sequences and their catalyzed chemical reactions from the BRENDA database (version 2024.1) and the Rhea database (version 2025.2.5). In the BRENDA database, the reaction expressions for the raw data consist of the molecule name, and the enzyme is given by its UniProt ID. The Rhea database directly provides the SMILES expressions and enzyme sequences for the reactions.

[0061] For the reaction data in the BRENDA database, the original reaction expressions are parsed, standardized, and repaired according to the EnzymeMap reaction preprocessing workflow, converted into valid SMILES representations, and the corresponding enzyme sequences are downloaded from UniProt.

[0062] The processor can then use RXNMapper to label the atom-to-atom mapping (AAM) between reactants and products in each reaction data sample to identify the reaction centers and describe the atomic changes before and after the reaction. If the reaction is not in equilibrium, the rules provided by EnzymeMap are used to attempt to repair the reaction and relabel the atom maps. The atom maps will be used to identify the reaction centers and describe the molecular changes before and after the reaction. After filtering out reactions that failed to be labeled with atom maps and merging and deduplicating the paired data from the two databases, a total of 265,920 reaction-enzyme pairs were obtained, including 23,595 different reactions and 178,035 different enzymes.

[0063] Next, the technicians built an unlabeled reaction dataset for the pre-training process of the molecular graph feature extractor.

[0064] Specifically, technicians can merge the 23,595 different reactions in the aforementioned reaction-enzyme pairing dataset with the reaction data from the USPTO and CARE datasets. For reactions lacking atom mapping in USPTO and CARE, the same atom mapping labeling process as in the reaction-enzyme pairing dataset is used, and the merged data is deduplicated, resulting in a total of 1,145,180 reaction data.

[0065] Next, the technicians built a training reaction dataset for training the reaction encoder. The training reaction dataset used the reaction-enzyme classification number pairing data from the CARE dataset.

[0066] After that, such as Figure 1 As shown, the technicians obtained the SMILES expressions for all reactions in the unlabeled reaction dataset and the training reaction dataset, and used the RDKit tool to preprocess the expressions to construct the reactant molecular diagrams and product molecular diagrams for each reaction. Here, hydrogen atoms are not treated as independent nodes.

[0067] Please refer to Tables 1 and 2. Table 1 shows the node feature table of the reactant molecular diagram and product molecular diagram provided according to some embodiments of the present invention. Table 2 shows the edge feature table of the reactant molecular diagram and product molecular diagram provided according to some embodiments of the present invention.

[0068] Table 1. Characteristics of nodes in reactant and product molecular diagrams.

[0069]

[0070] Table 2. Edge characteristics of reactant and product molecular diagrams

[0071]

[0072] In the embodiments shown in Tables 1 and 2, the reactant molecular diagrams include the node features and edge features of the corresponding reactants, and the product molecular diagrams include the node features and edge features of the corresponding products. Here, node features include atom type, chirality type, number of connected heavy atoms, formal charge, number of connected hydrogen atoms, number of free radical electrons, hybridization, whether it is an aromatic atom, and whether it is within a ring. Edge features include chemical bond type, stereochemistry, and whether it is conjugated.

[0073] Specifically, if a reactant or product contains multiple molecules, the molecular maps of these molecules will be merged into a single, disconnected molecular map. Based on this, the molecular map feature extractor can model the relationships between different reactant or product molecules in the same reaction. The constructed reactant and product molecular maps are then input into the molecular map feature extractor, respectively.

[0074] After that, such as Figure 2 As shown, the processor can first construct the molecular graph feature extractor described above. Here, the processor can change the sequence self-attention mechanism of each layer of the Transformer to a graph self-attention mechanism based on the original Graphormer encoder model architecture, introducing the encoding of the distance between nodes in the molecular graph and the encoding of edge features.

[0075] Specifically, the processor can first determine each node in the molecular graph. Shortest path length between Then, based on the shortest path length, determine each node. The first scalar between and the second scalar :

[0076]

[0077]

[0078] in, For the node To the node The feature vectors of all edges on the shortest path, where w is a learnable weight vector.

[0079] Then, the processor can base its decisions on the first scalar. and the second scalar Determine the node To the node Attention scores between:

[0080]

[0081] in, For node embedding vectors, The hidden layer dimension of the molecular graph feature extractor. To query the parameter matrix, The key parameter matrix, superscript This represents the transpose of a matrix.

[0082] Furthermore, building upon the original Graphormer, the processor adds a virtual CLS node to the molecular graph to connect all nodes and extract the global embedding vector. Compared to traditional Graph Neural Networks (GNNs), this molecular graph feature extractor can better model the interactions between distant nodes in the molecular graph. The molecular graph feature extractor has 6 layers and 512 hidden units.

[0083] After that, such as Figure 3 As shown, the processor can perform self-supervised pre-training on the constructed molecular graph feature extractor. Specifically, the self-supervised pre-training of the molecular graph feature extractor is performed simultaneously on reactant and product molecular graphs, employing three self-supervised training objectives: Masked Atom Property Prediction (MAP), Reaction Center Atom Recognition and Property Prediction (RCP), and Graph Contrastive Learning (GCL) between reactants and products. The pre-training of the molecular graph feature extractor is performed using the aforementioned constructed unlabeled reaction dataset.

[0084] Furthermore, such as Figure 3 As shown, the processor performs a masked atom property prediction task on the molecular graph feature extractor to calculate the first loss value of the molecular graph feature extractor.

[0085] In the MAP task, the processor randomly masks a certain proportion of atoms in the molecular graph of reactant or product molecules, forcing the molecular graph feature extractor to reconstruct the various properties of these atoms based on contextual information. This task borrows from the Masked Language Modeling (MLM) task in the BERT model to force the molecular graph feature extractor to learn the contextual representations of graph nodes, thereby improving its ability to model local graph structures.

[0086] Specifically, the masking strategy references the BERT model, which randomly selects 15% of the atoms in each molecular graph for masking to generate a masked molecular graph. For each selected masked atom, there is an 80% probability of replacing it with a MASK token, a 10% probability of replacing it with a random atom feature vector, and a 10% probability of leaving it unchanged.

[0087] Furthermore, in some preferred embodiments, since the values ​​of various features of atoms in the reaction dataset are highly unbalanced, in order to increase the difficulty of prediction, the processor can combine the features numbered 1 (atom type), 3 (number of connected heavy atoms), 4 (formal charge), and 5 (number of connected hydrogen atoms) in the molecular graph node feature table into a single attribute C1 for prediction. The C1 attribute focuses on the electronic and bonding properties of the atom; and combine the features numbered 1 (atom type), 2 (chirality type), 3 (number of connected heavy atoms), 8 (whether it is an aromatic atom), and 9 (whether it is in a ring) in the molecular graph node feature table into a single attribute C2 for prediction. The C2 attribute focuses on the topological and stereochemical properties of the atom.

[0088] The processor can then input the masked reactant molecular map and the masked product molecular map into the molecular map feature extractor to obtain the masked atom embedding vector. This masked atom embedding vector is then input into a corresponding multilayer perceptron (MLP) to predict the probability of each of its attribute values. The first loss value is calculated using the cross-entropy loss function:

[0089]

[0090] Among them, among them, The set of subscripts of the masked atoms. This refers to the molecular diagram of the masked reactant or the molecular diagram of the masked product. For the first The true C1 property value of the masked atom For the first The true C2 property value of the masked atom.

[0091] Simultaneously, the processor also performs tasks for identifying reaction center atoms and predicting their properties on the molecular graph feature extractor, and calculates the second loss value of the molecular graph feature extractor.

[0092] In the RCP task, the molecular graph feature extractor needs to identify the reaction center atoms in both the reactant and product molecular graphs—atoms whose properties or linked chemical bonds change during the reaction—and then predict the properties of these atoms on the other side of the reaction equation. This task forces the molecular graph feature extractor to locate the positions of key atoms and groups in chemical reactions, helping it to abstract general reaction rules.

[0093] Specifically, the reactant molecular maps and product molecular maps are input into a molecular map feature extractor to obtain all atom embedding vectors. All atom embedding vectors are then input into a corresponding MLP, and the second loss value of the molecular map feature extractor is calculated.

[0094]

[0095] Among them, The actual reaction center marker representing atom i, It is the set of subscripts of the atoms at the reaction center. and This represents the actual property value of the atom corresponding to atom i on the other side of the reaction. This is the prediction target of the task. and Obtained by atomic mapping.

[0096] Simultaneously, the processor also performs a graph comparison learning task between reactants and products on the molecular graph feature extractor, and calculates the third loss value of the molecular graph feature extractor.

[0097] In the GCL task, reactant-product pairs from the same reaction are considered positive pairs, while those from different reactions are considered negative pairs. The molecular graph feature extractor needs to distinguish molecular graphs from different reactions as much as possible, i.e., identify reactions that do not qualify. This task forces the molecular graph feature extractor to extract discriminative global features of the molecular graph, helping it learn the relationships between reactants and products.

[0098] Specifically, assuming the number of reactions in a training batch is N, the reactant molecular maps and product molecular maps of all reactions in the training batch are input into a molecular map feature extractor to obtain the CLS embedding vectors of the reactant molecular maps and the product molecular maps for each reaction. These CLS embedding vectors are then input into different corresponding MLPs to obtain the reactant embedding vectors for each reaction. and product embedding vector ; Calculate the third loss value of the molecular graph feature extractor:

[0099]

[0100] in, The cosine similarity function represents the similarity between vectors. This is a trainable temperature parameter. Afterwards, the processor can use the first loss value on the reactant and product molecular maps. , The second loss value on the reactant molecular diagram and product molecular diagram , And the third loss value between the reactant molecular diagram and the product molecular diagram. Calculate the fourth loss value of the molecular map feature extractor:

[0101]

[0102] Then, the processor can optimize the learnable parameters of the molecular graph feature extractor based on the fourth loss value until the maximum number of pre-training rounds is reached or the fourth loss value does not decrease within a specified number of rounds on the validation set. 5000 samples from the unlabeled reaction dataset are randomly selected as the validation set.

[0103] Furthermore, in some preferred embodiments, considering computational efficiency, the masked reaction molecular maps and masked product molecular maps constructed for the MAP task are also used as model inputs for the RCP and GCL tasks. In this way, the reactant and product molecular maps only need to undergo a single forward propagation in each training iteration, and this masked input can be considered a data augmentation for the RCP and GCL tasks. During pre-training, the molecular map feature extractors encoding the reactant and product molecular maps share parameters, but the MLPs used for different tasks differ.

[0104] The molecular graph feature extractor was pre-trained on four H100 GPUs using Distributed Data Parallel (DDP). The optimizer was AdamW with an initial learning rate of 0.0001 and a total training batch size of 256. 5000 samples from the reaction dataset were randomly selected as the validation set. After 100 rounds of pre-training, the checkpoints with the lowest validation loss were used in subsequent processes.

[0105] Please refer to further information. Figure 4 . Figure 4 A schematic diagram of the architecture of a reaction encoder provided according to some embodiments of the present invention is shown.

[0106] In cheminformatics, the condensed graph of reaction (CGR) is a good way to represent reactions and is used in various tasks, such as reaction condition prediction and reaction similarity search. Figure 4 As shown, the CGR is an overlap of the reactant molecular map and the product molecular map based on atomic mapping. Specifically, in the reaction encoder, the feature of each node in the CGR is formed by concatenating the embedding vectors of the corresponding nodes in the reactant molecular map and the product molecular map, and the feature of each edge is formed by concatenating the embedding vectors of the corresponding edges in the reactant and product maps. Furthermore, in some embodiments, if an edge only appears on one side of the reaction, i.e., the chemical bond at that position is broken or formed after the reaction, the zero vector is used as the edge embedding vector of the other side before concatenation. Figure 4 In the graph, edges represented by dashed lines represent bonds that change after the reaction, and edge embedding vectors represented by dashed lines are the zero-filled vectors. An additional shallow grapher layer, called the reaction rule extractor, is used in the reaction encoder to encode the CGR, to determine the reaction CLS embedding vector. The model architecture of the reaction rule extractor is the same as that of the molecular graph feature extractor, but the number of layers is reduced to two and it is not pre-trained; its parameters will be optimized in subsequent reaction-enzyme multimodal training.

[0107] Typically, to construct a CGR for a reaction, the reaction needs to have correct atom mappings and be in equilibrium. However, despite the use of heuristic rules to repair reactions when constructing the unlabeled reaction dataset, non-equilibrium reactions still exist in the unlabeled reaction dataset. In some embodiments, for non-equilibrium reactions, the CGR is constructed by selecting the set of atoms in the reactants and products that can be successfully mapped. In this case, the input to the reaction rule extractor consists of a subgraph of the reactant molecular map or the product molecular map. Although the reaction rule extractor cannot read the entire reaction from the CGR, the attention mechanism of the molecular map feature extractor, which still takes the complete reactant and product molecular maps as input, allows the node features of the CGR to implicitly contain global information about the reaction. This makes the reaction encoder robust to non-equilibrium reactions.

[0108] Please refer to further information. Figure 5 . Figure 5 A schematic diagram of the training process of a reaction encoder provided according to some embodiments of the present invention is shown.

[0109] like Figure 5 As shown, the processor can train the reaction encoder described above. Here, the processor can use the reaction-enzyme classification number pairing data in the CARE dataset as the training reaction dataset, and use supervised contrastive learning techniques to train the reaction encoder, bringing reactions under the same enzyme classification number closer together and pushing reactions under different enzyme classification numbers apart in its embedding space.

[0110] Specifically, the processor can first construct a reaction encoder and input each group of training reactions in the training reaction training set into the reaction encoder to determine the CLS embedding vector of each training reaction.

[0111] The processor can then input the embedding vectors of each training response (CLS) into the MLP to determine the embedding vectors of each training response. To construct a training response embedding vector library, where The number of training responses in the training set.

[0112] Then, the processor can embed vectors based on each training response. Calculate the fifth loss value of the reactive encoder:

[0113]

[0114] in, In each training response, except for the first The set of indices for all training responses other than the training response itself. For each training response and the first The set of indices of other training reactions that have the same enzyme classification number as a training reaction. The cosine similarity function represents the similarity between vectors. This is a learnable temperature parameter.

[0115] Then, the processor can optimize the learnable parameters of the reaction encoder based on the fifth loss value until the maximum number of training epochs is reached or the fifth loss value does not decrease within the specified number of epochs in the validation set of the above-mentioned training reaction dataset. This is to increase the cosine similarity between training reaction embedding vectors of enzymes with the same classification number and decrease the cosine similarity between training reaction embedding vectors of enzymes with different classification numbers.

[0116] By performing the above steps, those skilled in the art can obtain a trained reaction encoder. This encoder can then be stored on an optical disc, hard disk, cloud drive, or various other storage media for access and retrieval during the online prediction phase of an enzyme classification and numbering prediction system for enzyme-catalyzed reactions.

[0117] Next, please refer to Figure 6 . Figure 6 A flowchart illustrating a method for predicting enzyme classification and numbering of enzyme-catalyzed reactions according to some embodiments of the present invention is shown.

[0118] like Figure 6 As shown, the processor can acquire the reaction data to be tested and use the reaction encoder and MLP to sequentially determine the reaction CLS embedding vector and the reaction query embedding vector to be tested.

[0119] The processor can then calculate the cosine similarity between the query embedding vector of the response to be tested and each training response embedding vector in the training response embedding vector database.

[0120] Then, the processor can output the enzyme classification number corresponding to the training reaction embedding vector with the highest cosine similarity, as the enzyme classification number prediction result for the reaction to be tested.

[0121] Please refer to Figure 7 . Figure 7 A schematic diagram showing a performance comparison of enzyme classification and numbering prediction methods for enzyme-catalyzed reactions provided according to some embodiments of the present invention is illustrated.

[0122] like Figure 7 As shown, in order to verify the predictive performance of the enzyme classification number prediction method for the above-mentioned enzyme-catalyzed reaction provided in the first aspect of the present invention, a person skilled in the art can directly verify the accuracy of the reaction encoder in extracting reaction features by predicting the enzyme classification number corresponding to the reaction, that is, predicting the reaction EC number. Specifically, if the reaction encoder can accurately predict the reaction EC number, it indicates that it can capture the core chemical rules of the reaction with high quality, providing a reliable foundation for subsequent enzyme mining.

[0123] To verify the predictive performance of the enzyme classification number prediction method for enzyme-catalyzed reactions provided in the first aspect of the present invention, specifically, those skilled in the art can use the reaction-EC number prediction dataset of the CARE dataset as the evaluation dataset. The CARE dataset contains three levels of difficulty (e.g., easy, medium, hard) of training-test set splits. The higher the difficulty, the lower the similarity between the reaction in the test set and the reaction in the training set, and the lower the overlap of the EC numbers of the reaction. Therefore, it is also better able to test the generalization ability of the model.

[0124] The reaction encoder (without subsequent multimodal learning) was used as the test model. The training task was supervised contrastive learning, which brought reaction features of the same EC number closer together and pushed reaction features of different EC numbers apart. Nearest neighbor prediction was used to test the model. For each test reaction, the reaction in the training set with the highest cosine similarity to the test reaction was found in the embedding space, and the EC number of that reaction was used as the model prediction result.

[0125] The baseline models used were DRFP, CLIPZyme, and CREEP. DRFP is an efficient reaction fingerprint model, used for nearest neighbor prediction based on reaction similarity calculated during testing. CLIPZyme and CREEP are both multimodal models, trained using contrastive learning with reaction, enzyme, and EC numbers. Ablation experiments were also conducted on the reaction encoder, considering two variants. The reaction encoder (without CGR) represents a reaction compression graph CGR representation without using the reaction, directly concatenating the reactant and product CLS embedding vectors from the molecular graph feature extractor as the reaction embedding. The reaction encoder (without CGR and untrained) represents a reaction compression graph CGR without using the reaction, and the molecular graph feature extractor is untrained and randomly initialized.

[0126] like Figure 7As shown, under different difficulty levels (easy, medium, and hard) of the reaction classification number prediction dataset, each model has different prediction accuracies for different levels of classification numbers. The experimental results for the DRFP, CLIPZyme, and CREEP models are provided in the original CARE paper. It is evident that in most cases, the reaction encoder's prediction performance is significantly better than these three baseline models, especially for predicting complete 4-level EC numbers. Compared to the CLIPZyme and CREEP models, the reaction encoder's training data is unimodal, i.e., it only uses reactions and not enzymes, and the amount of training data is smaller. Due to the absence of a protein encoder, the reaction encoder's model parameter count is also significantly lower. Therefore, its higher prediction performance indicates that it can efficiently extract richer and more discriminative reaction features for classification. When CGR input is not used, the reaction encoder's prediction performance decreases at medium and hard difficulty levels, indicating that CGR helps the model learn reaction rules better, thus generalizing better to reactions that differ significantly from the training set. Overall, the untrained reaction encoder shows a significant performance decrease compared to the original version. The results are in line with expectations, demonstrating that the proposed pre-training strategy enables the molecular graph feature extractor to learn high-quality, reaction-contextualized molecular graph features. In summary, the reaction encoder is a reliable reaction encoding model that can support enzyme reaction classification and enzyme mining.

[0127] In summary, the method, system, and computer-readable storage medium for predicting enzyme classification numbers of enzyme-catalyzed reactions provided by this invention can all be used to capture chemical reaction rules in order to accurately predict the enzyme classification numbers corresponding to enzyme-catalyzed reactions.

[0128] Recent enzyme discovery research has focused on two main areas: inferring the function of unlabeled proteins by comparing their sequence or structural similarity with enzymes of known function; and training supervised learning models on functionally annotated protein databases to classify and screen unlabeled proteins. However, the first approach has limitations because protein sequence or structural similarity does not necessarily align with functional similarity, and because there is a lack of template enzymes as a starting point when studying novel catalytic reactions. On the other hand, the second approach also has limitations because current protein functional annotation systems cannot accurately reflect enzyme reaction or substrate specificity, and related annotation databases are still incomplete.

[0129] To overcome the aforementioned deficiencies in the existing technology, the present invention provides an enzyme mining method, an enzyme mining system, and a computer-readable storage medium, which can mine enzymes with specific catalytic functions from protein databases based on a given chemical reaction or template enzyme, effectively reducing the dependence on protein similarity and enzyme function annotation, and balancing the accuracy, speed, and practicality of enzyme mining.

[0130] In some non-limiting embodiments, the enzyme mining system provided in the fifth aspect of the present invention can be implemented based on the enzyme mining method provided in the fourth aspect of the present invention.

[0131] Specifically, in some non-limiting embodiments, the enzyme mining system provided in the fifth aspect of the present invention includes a memory and a processor. Here, the memory includes, but is not limited to, the computer-readable storage medium provided in the sixth aspect above, on which computer instructions are stored. The processor is connected to the memory and configured to execute the computer instructions stored in the memory to implement the enzyme mining method as provided in the fourth aspect of the present invention.

[0132] The working principle of the enzyme mining system described above will be described below with reference to some embodiments of enzyme mining prediction methods. Those skilled in the art will understand that these embodiments of enzyme mining methods are merely non-limiting implementations provided by the present invention, intended to clearly demonstrate the main concepts of the invention and provide specific solutions convenient for public implementation, rather than limiting all functions or all working methods of the enzyme mining system. Similarly, the enzyme mining system is also only one non-limiting implementation provided by the present invention, and does not constitute a limitation on the executing entity and execution order of the steps in these enzyme mining methods.

[0133] In some non-limiting embodiments, the enzyme mining method provided in the fourth aspect of the present invention can be implemented in two phases: training and prediction. Specifically, a person skilled in the art can first train the VenusRXN model in the training phase, and then, in the prediction phase, use the reaction encoder and protein encoder in the VenusRXN model to retrieve the target enzyme based on a template enzyme with a known function, respectively.

[0134] The training steps of the VenusRXN model include constructing a reaction-enzyme pairing dataset, which includes multiple reactions and their corresponding enzyme sequences; performing multimodal training on the reaction-enzyme pairing dataset, which includes calculating cross-modal feature alignment loss and cross-modal feature fusion loss; and calculating the multimodal training loss value and optimizing the learnable parameters of the VenusRXN model accordingly until the maximum number of training epochs is reached or the multimodal training loss value does not decrease within a specified number of epochs on the validation set.

[0135] Specifically, the VenusRXN model, based on the previously constructed reaction-enzyme pairing dataset, jointly trains the reaction encoder and the Protein Language Model (PLM) to achieve cross-modal modeling of reaction-enzyme associations. The VenusRXN model modifies the PLM to include both a protein encoder and a reaction-protein co-enzyme encoder. The reaction-protein co-enzyme encoder inserts a cross-attention layer into the PLM's Transformer block, performing fine-grained matching and fusion of the reaction encoder's output with the protein embeddings. The protein encoder shares parameters with the reaction-protein co-enzyme encoder, but its cross-attention layer is not activated and only receives protein sequences as input, used to extract the enzyme's single-modal embedding. During multimodal training, the embedding spaces of the reaction encoder and protein encoder are aligned across modalities, while the reaction-enzyme positive-negative pairing classification task is used to train the reaction-protein co-enzyme encoder. Through this multi-task training approach, the VenusRXN model can simultaneously learn global and local matching patterns between reactions and enzymes. The multimodal learning of the VenusRXN model is performed on the previously constructed reaction-enzyme pairing dataset, mainly including two parts: calculating the cross-modal feature alignment loss value and calculating the cross-modal feature fusion loss value.

[0136] Please refer to the details. Figure 8 . Figure 8 A schematic diagram of the training process of the VenusRXN model provided according to some embodiments of the present invention is shown.

[0137] When calculating the cross-modal feature alignment loss, contrastive learning and soft label alignment techniques are used simultaneously to align the embedding spaces of the reactive encoder and the protein encoder, enabling direct comparison of their generated embedding vectors and creating conditions for high-speed vector retrieval and fine-grained feature matching. For example... Figure 8 As shown on the left, assuming the number of reaction-enzyme pairs in a training batch is N, each reaction in the training batch is input into the reaction encoder to obtain the reaction CLS embedding vector of each reaction. Then, the reaction CLS embedding vector of each reaction is input into the corresponding linear layer of the reaction encoder to obtain the reaction embedding vector of each reaction. .

[0138] Simultaneously, the enzyme sequences from the training batch are input into the protein encoder to obtain the protein CLS embedding vectors of each enzyme sequence. These protein CLS embedding vectors are then input into the corresponding linear layer of the protein encoder to obtain the enzyme embedding vectors of each enzyme sequence. .

[0139] In this way, the reaction CLS embedding vectors and enzyme CLS embedding vectors generated by the reaction encoder and protein encoder are respectively input into different linear layers and mapped to the same low-dimensional embedding space, so that they can be directly compared, creating conditions for high-speed vector retrieval and fine-grained feature matching.

[0140] Then, the processor can calculate the contrastive learning loss and the soft label alignment loss based on each reaction embedding vector r and each enzyme embedding vector e, and calculate the cross-modal feature alignment loss based on the contrastive learning loss and the soft label alignment loss.

[0141] Specifically, the loss function for contrastive learning is InfoNCE, which aims to bring paired reactions and enzymes as close as possible and push away unpaired reactions and enzymes in the embedding space. Assuming there are N reaction-enzyme pairs in a training batch, the vector of the i-th pair of reactions and enzymes in the mapped embedding space is... Calculate the normalized reaction-enzyme similarity and enzyme-reaction similarity within the same training batch:

[0142]

[0143]

[0144] in, Let be the cosine similarity function of vectors. These are trainable temperature parameters.

[0145] Then, the processor determines the reaction-enzyme similarity. and enzyme-reaction similarity Calculate the contrastive learning loss value The loss in the i-th reaction-enzyme pair is:

[0146]

[0147] Furthermore, the InfoNCE loss function requires pairing in the normalized denominator:

[0148] Here, All are incorrect pairings. However, an enzyme-catalyzed reaction can often be catalyzed by multiple enzymes, and an enzyme may catalyze multiple reactions, making it difficult to guarantee this condition when constructing batches with complete randomization. To satisfy this condition, when constructing training batches, for each reaction-enzyme pairing sample (where the reaction-enzyme pairing is in the form of <reaction SMILES, enzyme sequence>) sampled from the reaction-enzyme pairing dataset, all enzymes corresponding to that reaction and all reactions corresponding to those enzymes are excluded from the reaction-enzyme pairing dataset before sampling the next sample, until the number of samples in the current batch reaches N. If the reaction-enzyme pairing dataset is exhausted before the number of samples in the current batch reaches N, the state of the reaction-enzyme pairing dataset is reset and sampling begins again.

[0149] Contrastive learning is a type of hard-label alignment method that directly treats mismatched reactions and enzymes as negative examples, ignoring potential weak correlations between mismatched sample pairs. To mitigate this issue, soft-label alignment is used for cross-modal feature alignment. This method defines soft labels for cross-modal matching based on the correlation between reactions and enzymes within their respective modalities. This avoids the reaction encoder and protein encoder only considering perfect match between reactions and enzymes while ignoring weak correlations, allowing the reaction encoder and protein encoder to learn more nuanced alignment relationships between the two modalities.

[0150] Specifically, the normalized response-response similarity is first calculated within the same training batch. Enzyme-enzyme similarity :

[0151]

[0152]

[0153] Where N represents the number of reaction-enzyme pairs in the training batch. Let be the cosine similarity function of vectors. These are trainable temperature parameters.

[0154] When the embedding spaces of reactions and enzymes are fully aligned, the distribution of sample similarity within and between modalities should be consistent. Therefore, it can be determined based on reaction-enzyme similarity. Enzyme-reaction similarity Reaction-reaction similarity and enzyme-enzyme similarity Calculate the soft label alignment loss value based on KL divergence. :

[0155]

[0156] in, It is a vector. .

[0157] Then, the processor can learn the loss value based on the contrast. and soft label alignment loss value Determine the cross-modal feature alignment loss value :

[0158]

[0159] In the aforementioned feature alignment process, reactions and enzymes are encoded as single global embedding vectors r and e, respectively. While this supports fast vector retrieval, it may result in the loss of significant local information and reduced matching accuracy. In addition to CLS embedding, the reaction encoder can provide atomic-scale embeddings, while the PLM can provide amino acid-scale embeddings. The reaction-protein joint encoder, by inserting a cross-attention layer between the self-attention layer and the feedforward layer of the Transformer block in the PLM, enables detailed matching and fusion of atomic embeddings of reactions and amino acid embeddings of proteins.

[0160] Specifically, in calculating the cross-modal feature fusion loss value, such as Figure 8 As shown on the right, the reactions in the reaction-enzyme pairing dataset are input into the reaction encoder to obtain the atomic embedding matrix Hr, which is composed of the stacked atomic embedding vectors of all atoms in the reaction compression map of the reaction. The atomic embedding matrix Hr is then input into the reaction-protein joint encoder.

[0161] Input the enzyme sequences corresponding to each reaction in the reaction-enzyme pairing dataset into the reaction-protein joint encoder. In each Transformer block of the reaction-protein joint encoder, obtain the amino acid initial embedding matrix He, which is composed of the stacked initial embedding vectors of all amino acids in the enzyme sequence.

[0162] In each Transformer block of the reaction-protein co-encoder, the amino acid-updated embedding matrix is ​​calculated using a cross-attention mechanism based on the atom embedding matrix Hr and the initial amino acid embedding matrix He. :

[0163]

[0164] in, To query the parameter matrix, For the key parameter matrix, For the value parameter matrix, The hidden layer dimension of the reaction-protein co-encoder is indicated by the superscript. Represents matrix transpose, amino acid update embedding matrix The amino acid initial embedding matrix will be used as the next Transformer block in the reaction-protein co-encoder. The reaction-protein co-encoder provides a multimodal fusion representation of reaction-enzyme pairings. The output reaction-protein co-CLS embedding vector is input into a matching head, and the matching score of the reaction-enzyme pairings is calculated. Based on the matching score, the cross-modal feature fusion loss value is calculated. :

[0165]

[0166] in, This is a true label for reaction-enzyme pairing. A value of 0 indicates a mismatch. A value of 1 indicates a match. It is the sigmoid activation function. The matching score is the result of the reaction-enzyme pairing.

[0167] The VenusRXN model dynamically samples hard negative samples in each training batch to perform a binary classification task. Specifically, for each reaction in a batch, it calculates reaction-enzyme similarity. Comparison similarity distribution calculated by formula Sample an unpaired enzyme to form a negative pair. Similarly, for each enzyme, calculate the enzyme-reaction similarity. Distribution calculated by formula Sample an unpaired reaction to form a negative pair. In this way, unpaired reactions and enzymes that are close in distance in the current model embedding space are more likely to be used as negative examples, thus encouraging the model to better identify positive pairings.

[0168] The final multimodal training loss value for VenusRXN is:

[0169]

[0170] in, This represents the multimodal training loss value. The cross-modal feature alignment loss value, This represents the cross-modal feature fusion loss value.

[0171] Then, the learnable parameters of the VenusRXN model are optimized based on the multimodal training loss value until the maximum number of training epochs is reached or the multimodal training loss value does not decrease within a specified number of epochs on the validation set.

[0172] The reaction-protein joint encoder and protein encoder were initialized using ESM C (600M version), with a cross-attention layer inserted after 18 Transformer blocks. Multimodal training was performed on eight H100 GPUs using DDP. The optimizer was AdamW with an initial learning rate of 0.0001, a total training batch size of 80, and a training duration of 320,000 steps. Thus, the multimodal training of the VenusRXN model was completed.

[0173] By performing the above steps, those skilled in the art can obtain the trained VenusRXN model. This model can then be stored on an optical disc, hard drive, cloud storage, or other storage media for access and retrieval by the enzyme mining system during the online prediction phase.

[0174] Because it employs a protein sequence-based protein sequence model (PLM), the VenusRXN model is suitable for retrieving large protein sequence databases. Specifically, in practice, the protein sequence database can be arbitrary; for example, it can be the UniProt database obtained online, or a self-constructed protein sequence database. In some embodiments of this invention, the protein sequence database used is all the enzyme sequences in the previously constructed reaction-enzyme pairing dataset.

[0175] Next, please refer to Figure 9 . Figure 9 A schematic flowchart of an enzyme mining method according to some embodiments of the present invention is shown.

[0176] like Figure 9 As shown in the middle section, the processor can acquire a protein sequence database, use a pre-trained protein encoder to determine the CLS embedding vectors of each protein sequence in the database, and then map these protein sequence CLS embedding vectors to their corresponding protein embedding vectors via a linear layer corresponding to the protein encoder. Construct a protein embedding vector database, where M is the total number of protein sequences in the protein embedding vector database.

[0177] like Figure 9 As shown, the processor can retrieve the data to be queried. Specifically, in Figure 9 In the illustrated embodiment, the data to be queried includes reaction data, which is a chemical reaction.

[0178] When the input query data is a reaction, the reaction encoder in the VenusRXN model is used to determine the reaction CLS embedding vector, and then the reaction CLS embedding vector is mapped to the reaction query embedding vector through the linear layer corresponding to the reaction encoder.

[0179] Please refer to further information. Figure 10 . Figure 10 The diagrams show pretreated reactant molecular diagrams and product molecular diagrams provided according to some embodiments of the present invention.

[0180] exist Figure 10 In the illustrated embodiment, the input response to be queried is represented by the string SMILES, and the initial input is:

[0181] COc1ccc(C[C@@H]2[NH+](C)CCc3cc(OC)c(O)cc23)cc1O.O=O>>[H][C@@]12Cc3ccc(OC)c(O)c3CN1CCc1cc(OC)c(O)cc21.[H]OO[H].[H+]

[0182] Since the embedding spaces of the reaction encoder and protein encoder are highly aligned after training, enzymes that may catalyze the reaction can be efficiently retrieved directly by calculating the cosine similarity between the reaction embedding vector and protein embedding vectors in the protein sequence database. Specifically, the processor can calculate and sort the cosine similarity between the reaction query embedding vector and each protein embedding vector in the protein embedding vector database, and output the protein sequences corresponding to the k protein embedding vectors with the highest cosine similarity to determine the enzymes that may catalyze the reaction. Here, k can be customized according to actual needs.

[0183] For example, based on the response query embedding vector, the top k candidate protein sequences can be retrieved from a protein embedding vector database. Only one of these sequences is shown below:

[0184] MENKTPIFFSLSIFLSLLNCALGGNDLLSCLTFNGVRNHTVFSADSDSDFNRFLHLSIQNPLFQNSLISKPSAIILPGSKEELSNTIRCIRKGSWTIRLRSGGHSYEGLSYTSDTPFILIDLMNLNRVSIDLES ETAWVESGSTLGELYYAITESSSKLGFTAGWCPTVGTGGHISGGGFGMMSRKYGLAADNVVDAILIDANGAILDRQAMGEDVFWAIRGGGGGVWGAIYAWKIKLLPVPEKVTVFRVTKNVAIDEATSLLHKWQFV AEELEEDFTLSVLGGADEKQVWLTMLGFHFGLKTVAKSTFDLLFPELGLVEEDYLEMSWGESFAYLAGLETVSQLNNRFLKFDERAFKTKVDLTKEPLPSKAFYGLLERLSKEPNGFIALNGFGGQMSKISSDF TPFPHRSGTRLMVEYIVAWNQSEQKKKTEFLDWLEKVYEFMKPFVSKNPRLGYVNHIDLDLGGIDWGNKTVVNNAIEISRSWGESYFLSNYERLIRAKTLIDPNNFNHPQSIPPMANFDYLEKTLGSDGGEVVI

[0185] Please refer to Figure 11 . Figure 11 A schematic diagram showing a performance comparison of reaction-based enzyme mining methods provided according to some embodiments of the present invention is illustrated.

[0186] like Figure 11As shown, to verify the predictive performance of the enzyme mining method provided in the fourth aspect of the present invention, the technician can divide the previously constructed reaction-enzyme pairing dataset into a training set and a test set according to the reaction. Specifically, firstly, 500 different reactions are randomly selected from the reaction-enzyme pairing dataset, ensuring that each selected reaction corresponds to no less than 4 enzymes in the dataset. Then, all pairings containing these reactions are used as the test set, and the remaining pairings are used as the training set. The final constructed test set contains 15,526 reaction-enzyme pairs, which contain 14,893 different enzymes. During testing, for each test reaction, all 178,035 enzymes in the reaction-enzyme pairing dataset are used as a protein sequence database for retrieval, and the model's ability to screen out the positive enzymes corresponding to the reaction is evaluated. The evaluation metrics used are BEDROC, Enrichment Factor (EF), Top-k Recall, and Top-k Hit Rate. Here, BEDROC measures early retrieval efficiency, EF measures the enrichment of retrieval results, Top-k recall is defined as the proportion of positive enzymes among the top k enzymes retrieved by the model out of the total number of positive enzymes, and Top-k hit rate is defined as the probability that at least one positive enzyme is found among the top k enzymes retrieved by the model. In addition to scoring all test reactions, the test set is further divided into several subsets based on the maximum similarity between the test reactions and the reactions in the training set. Scores are then calculated on these subsets to observe the model's generalization ability. The similarity between reactions during the split is measured using the cosine value of DRFP.

[0187] The baseline model used in the tests was the CLIPZyme model, which aligns the molecular graph representation of the reaction with the tertiary structure representation of the protein based on contrastive learning. For fair comparison, the CLIPZyme model was retrained on the training set constructed in this approach and tested using the same methodology as the VenusRXN model. Since the CLIPZyme model requires the tertiary structure of the protein as input, we first attempted to download the tertiary structures of all enzymes in the dataset from AlphaFoldDB. For enzymes not present in AlphaFoldDB, AlphaFold2 was used to predict their structures. Furthermore, the VenusRXN model (without a pre-trained molecular graph feature extractor) was also used as the baseline model and referred to as the VenusRXN model.

[0188] like Figure 11As shown, although the VenusRXN model did not utilize protein structure during training, it significantly outperformed the CLIPZyme model across all metrics, regardless of whether a pre-trained molecular graph feature extractor was used. Furthermore, the VenusRXN model was much less affected by training-test reaction similarity than the CLIPZyme model. The difference was most pronounced when the training-test reaction similarity was as low as 0.2, with the VenusRXN model achieving more than double the Top-20 recall and hit rate compared to the CLIPZyme model. This indicates that the multimodal learning strategy employed by the VenusRXN model can more effectively associate enzymes with their catalyzed reactions and exhibits good generalization ability for reactions that differ significantly from the training set. Compared to using only contrastive learning, the VenusRXN model simultaneously aligns and finely fuses the embedding vectors of the molecular graph feature extractor and the PLM during multimodal learning, thereby better capturing the matching patterns between reactions and enzymes and achieving more accurate cross-modal retrieval. Furthermore, the VenusRXN model (with an untrained molecular graph feature extractor) performed significantly worse than the VenusRXN model, indicating the effectiveness of the molecular graph feature extractor pre-training strategy for enzyme mining.

[0189] Or, specifically, in Figure 9 In the illustrated embodiment, the data to be queried includes template enzymes.

[0190] When the input query data is a template enzyme, the protein encoder in the VenusRXN model is used to determine the template enzyme CLS embedding vector. Then, through the linear layer corresponding to the protein encoder, the template enzyme CLS embedding vector is mapped to the template enzyme query embedding vector.

[0191] The algorithm calculates and sorts the cosine similarity between the template enzyme query embedding vector and each protein embedding vector in the protein embedding vector database. It then outputs the protein sequences corresponding to the k protein embedding vectors with the highest cosine similarity to identify enzymes with similar functions to the template enzyme. Here, k can be customized according to actual needs.

[0192] For example, the amino acid sequence of the input template enzyme is:

[0193] MDINDKLVADLVANDVSFVTTVPCKQLAGVIDKIDAHPDILHIPSNKEDEGMGLCAGAFMGGKRPAIIMQNTALGVTINTLATLIQYYRMPLPMLISYRGELGEPVACQVEMAVHTKALLDQMHIPTFHFHRQSDAEEFDKILKYTFMCNKPVAILTDASFWGGY

[0194] For example, based on the template enzyme query embedding vector, the top k candidate protein sequences can be retrieved from a protein vector database. Only one of these sequences is shown below:

[0195] MIRSDVLKTLIPMISDHFVVCNIGLPSQELHMLDDQPTNFYMLGTMGLSSSIGLGLALAQKQKVISIDGDGSVLTNFGTLPTIANNVADNFILLIIDNGSYGSTGDQPTYAGKKTSLTAVAQACGCENVIECAAEDCADVMEQALTGDKMTIIVCKCESGNIPVPVITKDPVTIRDRFMTALREKNAQ

[0196] Please refer to Figure 12 . Figure 12 A schematic diagram showing a performance comparison of template enzyme-based enzyme mining methods provided according to some embodiments of the present invention is illustrated.

[0197] like Figure 12 As shown, to verify the predictive performance of the enzyme mining method provided in the fourth aspect of the present invention, similar to the performance comparison of reaction-based enzyme mining methods, the technician can divide the previously constructed reaction-enzyme pairing dataset into a training set and a test set according to the reaction. For each test reaction constructed above, a corresponding positive enzyme is randomly selected as a template enzyme and input into the protein encoder in the VenusRXN model. The selected template enzyme does not appear in the protein vector database used for retrieval. The evaluation metrics used are BEDROC, EF, Top-k recall, and Top-k hit rate. Here, BEDROC measures the early retrieval efficiency, EF measures the enrichment of retrieval results, Top-k recall is defined as the proportion of positive enzymes in the top k enzymes retrieved by the model to the total number of positive enzymes, and Top-k hit rate is defined as the probability that at least one positive enzyme is among the top k enzymes retrieved by the model. To verify whether the VenusRXN model can identify the functional similarity of enzymes in addition to sequence similarity, the test set is divided into several subsets according to the maximum sequence identity between each template enzyme and its corresponding other positive enzymes, and the model's test score is calculated on each subset.

[0198] The baseline models considered included ESM-2 and ESM-C, which retrieve protein embeddings based on their original vectors; the sequence alignment method BLASTp; and the structure alignment method Foldseek. For ESM-2 and ESM-C, their outputs were averaged and pooled to obtain protein embedding vectors, and retrieval was performed based on the cosine similarity between these vectors. Foldseek requires the protein's tertiary structure as input. First, the tertiary structures of all enzymes in the AlphaFoldDB database were downloaded. For enzymes not present in the AlphaFoldDB database, their structures were predicted using the AlphaFold2 protein prediction model.

[0199] like Figure 12 As shown, although only protein sequence input was used, the VenusRXN model achieved the best performance in all cases, surpassing traditional sequence / structure alignment methods and retrieval methods based on the original PLM. In particular, its advantage over the baseline model was most pronounced when the sequence identity between the positive enzyme and the template enzyme was low (≤0.4), at which point its BEDROC reached 0.86, while Foldseek's BEDROC was only 0.61. Compared to the baseline model, which only had access to protein information, the VenusRXN model's protein encoder was guided and constrained by reaction information during multimodal learning. Therefore, the enzyme embedding vector it provided better reflected enzyme function, thus enabling the screening of enzymes with low sequence or structural similarity but similar functions. On the other hand, because the VenusRXN model explicitly considered the semantic similarity of reactions and enzymes within their respective modalities through soft label alignment during training, its cross-modal and intramodal retrieval capabilities were guaranteed. In summary, when retrieving based on template enzymes, the VenusRXN model has advantages over traditional protein search methods and exhibits good generalization performance.

[0200] In summary, the enzyme mining method, enzyme mining system, and computer-readable storage medium provided by this invention can all mine enzymes with specific catalytic functions from protein databases based on a given chemical reaction or template enzyme, effectively reducing the dependence on protein similarity and enzyme function annotation, and taking into account the accuracy, speed, and practicality of enzyme mining.

[0201] Existing intelligent prediction methods for enzyme catalytic performance parameters have significant limitations. First, existing methods rely solely on enzyme sequences and single substrate information for prediction, neglecting substrate molecular structure and complete reaction information, including products and cofactors. This makes it difficult for them to capture the enzyme catalytic process. Second, due to the small size of known enzyme catalytic performance annotation datasets, models trained directly on these datasets generally exhibit weak generalization ability, with a significant drop in prediction performance for novel enzymes or reactions with low similarity to the training set.

[0202] To overcome the aforementioned deficiencies in the prior art, this invention provides a method for predicting enzyme catalytic performance based on chemical reactions, a system for predicting enzyme catalytic performance based on chemical reactions, and a computer-readable storage medium. Through multimodal learning on reaction-enzyme pairing data and efficient fine-tuning on enzyme catalytic performance labeled data, the catalytic performance parameters of an enzyme under a given chemical reaction can be accurately predicted.

[0203] In some non-limiting embodiments, the enzyme catalytic performance prediction system based on chemical reactions provided in the eighth aspect of the present invention can be implemented based on the enzyme catalytic performance prediction method based on chemical reactions provided in the seventh aspect of the present invention.

[0204] Specifically, in some non-limiting embodiments, the enzyme catalytic performance prediction system based on chemical reactions provided in the eighth aspect of the present invention includes a memory and a processor. Here, the memory includes, but is not limited to, the computer-readable storage medium provided in the ninth aspect, on which computer instructions are stored. The processor is connected to the memory and configured to execute the computer instructions stored in the memory to implement the enzyme catalytic performance prediction method based on chemical reactions as provided in the seventh aspect of the present invention.

[0205] The working principle of the above-mentioned enzyme catalytic performance prediction system based on chemical reactions will be described below with reference to some embodiments of enzyme catalytic performance prediction methods based on chemical reactions. Those skilled in the art will understand that these embodiments of enzyme catalytic performance prediction methods based on chemical reactions are merely non-limiting implementations provided by the present invention, intended to clearly demonstrate the main concepts of the invention and provide specific solutions convenient for public implementation, rather than limiting all functions or all working methods of the enzyme catalytic performance prediction system based on chemical reactions. Similarly, the enzyme catalytic performance prediction system based on chemical reactions is also merely a non-limiting implementation provided by the present invention, and does not constitute a limitation on the executing entity and execution order of the steps in these enzyme catalytic performance prediction methods based on chemical reactions.

[0206] In some non-limiting embodiments, the enzyme catalytic performance prediction method based on chemical reactions provided in the seventh aspect of the present invention can be implemented in two stages: offline fine-tuning and online prediction. Specifically, in the offline fine-tuning stage, the reaction encoder and reaction-protein joint encoder in the previously pre-trained VenusRXN model can be fine-tuned. Then, in the online prediction stage, the reaction-protein joint encoder is used to calculate the matching score between the reaction and enzyme in each test sample to predict the enzyme catalytic performance based on chemical reactions.

[0207] The training steps of the VenusRXN model include constructing a reaction-enzyme pairing dataset, which includes multiple reactions and their corresponding enzyme sequences; performing multimodal training on the reaction-enzyme pairing dataset, which includes calculating cross-modal feature alignment loss and cross-modal feature fusion loss; and calculating the multimodal training loss value and optimizing the learnable parameters of the VenusRXN model accordingly until the maximum number of training epochs is reached or the multimodal training loss value does not decrease within a specified number of epochs on the validation set.

[0208] Specifically, the VenusRXN model, based on the previously constructed reaction-enzyme pairing dataset, jointly trains the reaction encoder and PLM to achieve cross-modal modeling of reaction-enzyme associations. The VenusRXN model modifies the PLM to include both a protein encoder and a reaction-protein co-enzyme encoder. The reaction-protein co-enzyme encoder inserts a cross-attention layer into the PLM's Transformer block, performing fine-grained matching and fusion of the reaction encoder's output with the protein embeddings. The protein encoder shares parameters with the reaction-protein co-enzyme encoder, but its cross-attention layer is not activated and only receives protein sequences as input, used to extract the enzyme's single-modal embedding. During multimodal training, the embedding spaces of the reaction encoder and protein encoder are aligned across modalities, while the reaction-enzyme positive-negative pairing classification task is used to train the reaction-protein co-enzyme encoder. Through this multi-task training approach, the VenusRXN model can simultaneously learn global and local matching patterns between reactions and enzymes. The multimodal learning of the VenusRXN model is performed on the previously constructed reaction-enzyme pairing dataset, mainly including two parts: calculating the cross-modal feature alignment loss value and calculating the cross-modal feature fusion loss value.

[0209] Please refer to the details. Figure 8 . Figure 8 A schematic diagram of the training process of the VenusRXN model provided according to some embodiments of the present invention is shown.

[0210] When calculating the cross-modal feature alignment loss, contrastive learning and soft label alignment techniques are used simultaneously to align the embedding spaces of the reactive encoder and the protein encoder, enabling direct comparison of their generated embedding vectors and creating conditions for high-speed vector retrieval and fine-grained feature matching. For example... Figure 8 As shown on the left, assuming the number of reaction-enzyme pairs in a training batch is N, each reaction in the training batch is input into the reaction encoder to obtain the reaction CLS embedding vector of each reaction. Then, the reaction CLS embedding vector of each reaction is input into the corresponding linear layer of the reaction encoder to obtain the reaction embedding vector of each reaction. .

[0211] Simultaneously, the enzyme sequences from the training batch are input into the protein encoder to obtain the protein CLS embedding vectors of each enzyme sequence. These protein CLS embedding vectors are then input into the corresponding linear layer of the protein encoder to obtain the enzyme embedding vectors of each enzyme sequence. .

[0212] In this way, the reaction CLS embedding vectors and enzyme CLS embedding vectors generated by the reaction encoder and protein encoder are respectively input into different linear layers and mapped to the same low-dimensional embedding space, so that they can be directly compared, creating conditions for high-speed vector retrieval and fine-grained feature matching.

[0213] Then, the processor can calculate the contrastive learning loss and the soft label alignment loss based on each reaction embedding vector r and each enzyme embedding vector e, and calculate the cross-modal feature alignment loss based on the contrastive learning loss and the soft label alignment loss.

[0214] In the aforementioned feature alignment process, reactions and enzymes are encoded as single global embedding vectors r and e, respectively. While this supports fast vector retrieval, it may result in the loss of significant local information and reduced matching accuracy. In addition to CLS embedding, the reaction encoder can provide atomic-scale embeddings, while the PLM can provide amino acid-scale embeddings. The reaction-protein joint encoder, by inserting a cross-attention layer between the self-attention layer and the feedforward layer of the Transformer block in the PLM, enables detailed matching and fusion of atomic embeddings of reactions and amino acid embeddings of proteins.

[0215] Specifically, in calculating the cross-modal feature fusion loss value, such as Figure 8 As shown on the right, the reactions in the reaction-enzyme pairing dataset are input into the reaction encoder to obtain the atomic embedding matrix Hr, which is composed of the stacked atomic embedding vectors of all atoms in the reaction compression map. The atomic embedding matrix Hr is then input into the reaction-protein joint encoder.

[0216] The enzyme sequences corresponding to each reaction in the reaction-enzyme pairing dataset are input into the reaction-protein joint encoder. In each Transformer block of the reaction-protein joint encoder, the initial amino acid embedding matrix He, which is composed of the stacked initial embedding vectors of all amino acids in the enzyme sequence, is obtained.

[0217] In each Transformer block of the reaction-protein co-encoder, the amino acid-updated embedding matrix is ​​calculated using a cross-attention mechanism based on the atom embedding matrix Hr and the initial amino acid embedding matrix He. :

[0218]

[0219] in, To query the parameter matrix, For the key parameter matrix, For the value parameter matrix, The hidden layer dimension of the reaction-protein co-encoder is indicated by the superscript. Represents matrix transpose, amino acid update embedding matrix The amino acid initial embedding matrix will be used as the next Transformer block in the reaction-protein co-encoder. The reaction-protein co-encoder provides a multimodal fusion representation of reaction-enzyme pairings. The output reaction-protein co-CLS embedding vector is input into a matching head, and the matching score of the reaction-enzyme pairings is calculated. Based on the matching score, the cross-modal feature fusion loss value is calculated. :

[0220]

[0221] in, This is a true label for reaction-enzyme pairing. A value of 0 indicates a mismatch. A value of 1 indicates a match. It is the sigmoid activation function. The matching score is the result of the reaction-enzyme pairing.

[0222] The final multimodal training loss value for VenusRXN is:

[0223]

[0224] in, This represents the multimodal training loss value. The cross-modal feature alignment loss value, This represents the cross-modal feature fusion loss value.

[0225] Then, the learnable parameters of the VenusRXN model are optimized based on the multimodal training loss value until the maximum number of training epochs is reached or the multimodal training loss value does not decrease within a specified number of epochs on the validation set.

[0226] The reaction-protein joint encoder and protein encoder were initialized using ESM C (600M version), with a cross-attention layer inserted after 18 Transformer blocks. Multimodal training was performed on eight H100 GPUs using DDP. The optimizer was AdamW with an initial learning rate of 0.0001, a total training batch size of 80, and a training duration of 320,000 steps. Thus, the multimodal training of the VenusRXN model was completed.

[0227] Furthermore, since the reaction-protein co-encoder has already demonstrated good transfer learning capabilities through the reaction-enzyme matching task, it can be efficiently applied to supervised learning tasks that use reaction-enzyme pairing as input. When enzyme catalytic performance labeled data is available, the reaction encoder and the reaction-protein co-encoder can be fine-tuned to achieve enzyme catalytic performance prediction based on chemical reactions.

[0228] Please refer to Figure 13 . Figure 13 A schematic diagram of the fine-tuning process of a reaction encoder and a reaction-protein co-encoder provided according to some embodiments of the present invention is shown.

[0229] Specifically, the processor can fine-tune the response encoder and response-protein co-encoder in the trained VenusRXN model, such as... Figure 13 As shown, a fine-tuning sample set is first constructed, consisting of triplet samples of <reaction SMILES, enzyme sequence, and catalytic performance label>, where the catalytic performance label is a real scalar. Based on the fine-tuning sample set, the reaction encoder and the reaction-protein joint encoder are fine-tuned using the mean squared error (MSE) loss function. The maximum number of training rounds, learning rate, and training batch size are determined based on the cross-validation performance on the fine-tuning sample set.

[0230] Specifically, in some preferred embodiments, the processor can perform z-score standardization on the catalytic performance labels in the fine-tuning sample set, that is, for all labels, subtract the mean of the training set labels and divide by the standard deviation of the training set labels to make them conform to a standard normal distribution.

[0231] Next, the processor can input the sampled reactions SMILES from the fine-tuning sample set into the reaction encoder in the VenusRXN model to obtain an atomic embedding matrix composed of stacked atomic embedding vectors of all the sampled reaction's reaction compression graph. The processor then inputs the atomic embedding matrix and the enzyme sequences corresponding to the sampled reactions from the fine-tuning sample set into the reaction-protein joint encoder in the VenusRXN model to obtain a reaction-protein joint CLS embedding vector. Finally, the joint CLS embedding vector is input into the matching head to calculate the matching score between the reaction and the corresponding enzyme sequence. Based on matching score Calculate the loss value of the reaction-protein co-enzyme encoder on reaction-enzyme paired samples. :

[0232]

[0233] in, It is a standardized label for catalytic performance.

[0234] The training batch size for fine-tuning is 32, and the learning rate and number of epochs are determined by the 4-fold cross-validation performance on the training set. The search range for the learning rate is {0.0001, 0.0005, 0.001, 0.005, 0.01}, and the search range for the number of epochs is 1 to 80.

[0235] By performing the above steps, those skilled in the art can obtain the fine-tuned reaction encoder and reaction-protein joint encoder from the trained VenusRXN model. They can then store this data on an optical disc, hard disk, cloud drive, or other storage medium for online access and retrieval during the prediction phase of a chemical reaction-based enzyme catalysis performance prediction system.

[0236] When performing enzyme catalytic performance prediction based on chemical reactions:

[0237] The processor can acquire input test samples, which are reaction-enzyme paired samples, input in the form of <reaction SMILES, enzyme sequence> pairs.

[0238] The processor then inputs the reaction SMILES from the test sample into the pre-trained reaction encoder in the VenusRXN model to obtain an atomic embedding matrix composed of stacked atomic embedding vectors in the reaction compression map of the reaction. The processor then inputs the atomic embedding matrix and the corresponding enzyme sequence from the paired test sample into the pre-tuned reaction-protein joint encoder in the VenusRXN model to obtain the reaction-protein joint CLS embedding vector.

[0239] Next, the processor inputs the aforementioned joint CLS embedding vector into the matching head and calculates the matching score between the reaction and the corresponding enzyme sequence.

[0240] The final output is a matching score, which serves as the enzyme catalytic performance predicted based on chemical reactions.

[0241] Specifically, before inputting the reaction SMILES from the sample to be tested into the above reaction encoder, the RDKit tool is needed to construct the reactant molecular map and product molecular map of the reaction respectively. The molecular map includes node features and edge features. The node features include atom type, chirality type, number of connected heavy atoms, formal charge, number of connected hydrogen atoms, number of free radical electrons, hybridization, whether it is an aromatic atom and whether it is in a ring. The edge features include chemical bond type, stereoconfiguration and whether it is conjugated.

[0242] Please refer to Figure 14 . Figure 14 A performance comparison diagram of enzyme catalytic performance prediction methods based on chemical reactions provided according to some embodiments of the present invention is shown.

[0243] like Figure 14 As shown, to verify the predictive performance of the enzyme catalytic performance prediction method based on chemical reactions provided in the seventh aspect of this invention, those skilled in the art can use the kcat prediction dataset provided in the TurNup model paper for performance evaluation. This dataset contains 4271 <reaction, enzyme sequence, kcat> triplet data, including 2977 different reactions and 2827 different enzymes. The dataset is split into training and testing sets in an 8:2 ratio, ensuring that enzymes from the testing set do not appear in the training set. According to the TurNup model settings, label preprocessing is performed before training and testing. Specifically, all kcat labels are transformed by log10 to make them approximately follow a normal distribution, improving the model's training stability. On the training set, the reaction encoder and reaction-protein co-encoder in the VenusRXN model are fine-tuned using the mean squared error loss function. The maximum number of training rounds, learning rate, and training batch size are determined based on the cross-validation performance on the fine-tuned sample set.

[0244] The baseline model used was the TurNup model, which uses the DRFP fingerprint of the reaction and the ESM-1b embedding vector of the enzyme as input to train an XGBoost regression model to predict kcat. Following the TurNup model, the kcat prediction performance of the VenusRXN model was also tested when using only reaction or enzyme input. When using reaction as input, a regression head was added to the reaction encoder of the VenusRXN model and fine-tuned and tested; when using enzyme as input, the protein encoder of the VenusRXN model was similarly fine-tuned and tested. Evaluation metrics included the mean squared error (MSE) measuring the deviation of predicted values ​​from true values, the coefficient of determination (R²) measuring the model's ability to explain data variation, the Pearson coefficient measuring linear correlation, and the Spearman coefficient measuring monotonic correlation and being more robust to outliers.

[0245] like Figure 14As shown, regardless of the input format (reaction-only, enzyme-only, or reaction-enzyme combined input), the VenusRXN model significantly outperforms TurNup across all metrics, with Pearson and Spearman coefficients exceeding 0.6. Specifically, the performance of the VenusRXN model with only enzyme input is close to that of the TurNup model with the reaction-enzyme combined input in terms of Spearman coefficient. This indicates that through multimodal learning, the reaction encoder, protein encoder, and reaction-protein combined encoder in the VenusRXN model can be efficiently transferred to the kcat prediction task, enabling enzyme catalytic performance prediction based on chemical reactions. Furthermore, for both the VenusRXN and TurNup models, the performance using the reaction-enzyme combined input is significantly better than that using only reaction or only enzyme input. This result is expected because enzyme kinetic parameters are highly correlated not only with the enzyme itself but also with the catalyzed reaction, especially the substrate. This also verifies that the VenusRXN model's reaction-protein combined encoder can effectively fuse reaction and enzyme characterization through cross-attention, making it suitable for enzyme catalytic performance prediction tasks related to chemical reactions.

[0246] In summary, the above-mentioned enzyme catalytic performance prediction method, enzyme catalytic performance prediction system, and computer-readable storage medium based on chemical reactions provided by this invention can all use the reaction-protein joint encoder in the VenusRXN model to calculate the matching score between the reaction and enzyme in the test sample, and are used to predict the enzyme catalytic performance based on chemical reactions.

[0247] Although the methods described above are illustrated and depicted as a series of actions for the sake of simplicity, it should be understood and appreciated that these methods are not limited by the order of the actions, as some actions may occur in a different order and / or concurrently with other actions from the illustrations and descriptions herein or not illustrated and described herein but which may be understood by those skilled in the art, according to one or more embodiments.

[0248] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0249] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or via a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0250] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for enzyme mining, characterized in that, Includes the following steps: Input the data to be queried, which includes a reaction or a template enzyme, wherein the reaction is represented by the string SMILES and the template enzyme is represented by an amino acid sequence; The query data is input into a pre-trained VenusRXN model, which includes a reaction encoder and a protein encoder. The training steps of the VenusRXN model include: constructing a reaction-enzyme pairing dataset, which includes multiple reactions and corresponding enzyme sequences; and performing multimodal training on the reaction-enzyme pairing dataset. The multimodal training includes calculating cross-modal feature alignment loss and cross-modal feature fusion loss. Calculating the cross-modal feature alignment loss includes calculating contrastive learning loss and soft label alignment loss. The step of calculating the contrastive learning loss includes calculating the normalized reaction-enzyme similarity in the same training batch. Similarity to enzyme-reaction : in, This represents the number of reaction-enzyme pairs in the training batch. Let be the cosine similarity function of vectors. For trainable temperature parameters; and based on the reaction-enzyme similarity. and the enzyme-reaction similarity Calculate the contrastive learning loss value, where the loss for the i-th reaction-enzyme pair is: ; When the input query data is a reaction, the reaction encoder is used to determine the reaction CLS embedding vector. Then, through the linear layer corresponding to the reaction encoder, the reaction CLS embedding vector is mapped to a reaction query embedding vector. The cosine similarity between the reaction query embedding vector and each protein embedding vector in the protein embedding vector database is calculated and sorted. The protein sequences corresponding to the multiple protein embedding vectors with the highest cosine similarity are output; or When the input query data is a template enzyme, the protein encoder is used to determine the template enzyme CLS embedding vector. Then, through the linear layer corresponding to the protein encoder, the template enzyme CLS embedding vector is mapped to the template enzyme query embedding vector. The cosine similarity between the template enzyme query embedding vector and each protein embedding vector in the protein embedding vector database is calculated and sorted. The protein sequences corresponding to the multiple protein embedding vectors with the highest cosine similarity are output.

2. The method as described in claim 1, characterized in that, The steps for constructing the protein embedding vector database include: Obtain a protein sequence database, wherein the protein sequence database consists of all enzyme sequences in the reaction-enzyme pairing dataset; Using the protein encoder, the CLS embedding vectors of each protein sequence in the protein sequence database are determined; and The protein sequence CLS embedding vector is mapped to the corresponding protein embedding vector through the linear layer corresponding to the protein encoder, thereby constructing the protein embedding vector database.

3. The method as described in claim 1, characterized in that, The training steps for the VenusRXN model also include: Calculate the multimodal training loss value, and optimize the learnable parameters of the VenusRXN model based on the multimodal training loss value until the maximum number of training epochs is reached or the multimodal training loss value does not decrease within a specified number of epochs.

4. The method as described in claim 3, characterized in that, The steps for calculating the cross-modal feature alignment loss value include: Each reaction in the training batch is input into the reaction encoder to obtain the reaction CLS embedding vector for each reaction. The reaction CLS embedding vectors for each reaction are then input into the linear layer corresponding to the reaction encoder to obtain the reaction embedding vector for each reaction. ,in, This represents the number of reaction-enzyme pairs in the training batch; The enzyme sequences from the training batch are input into the protein encoder to obtain the protein CLS embedding vectors of each enzyme sequence. These CLS embedding vectors are then input into the corresponding linear layer of the protein encoder to obtain the enzyme embedding vectors of each enzyme sequence. ,in, This represents the number of reaction-enzyme pairs in the training batch; Based on the reaction embedding vector r and the enzyme embedding vector e, calculate the contrastive learning loss and the soft label alignment loss; and Calculate the cross-modal feature alignment loss value based on the contrastive learning loss value and the soft label alignment loss value.

5. The method as described in claim 4, characterized in that, The steps for calculating the soft label alignment loss include: Calculate normalized response-response similarity in the same training batch Enzyme-enzyme similarity : in, This represents the number of reaction-enzyme pairs in the training batch. Let be the cosine similarity function of vectors. For trainable temperature parameters; and Based on the reaction-enzyme similarity The enzyme-reaction similarity The reaction-reaction similarity and the enzyme-enzyme similarity Calculate the soft label alignment loss value: 。 6. The method as described in claim 4, characterized in that, The cross-modal feature alignment loss value is: in, The cross-modal feature alignment loss value is... To compare the learning loss values, This represents the soft label alignment loss value.

7. The method as described in claim 1, characterized in that, The steps for calculating the cross-modal feature fusion loss value include: The reactions in the reaction-enzyme pairing dataset are input into the reaction encoder to obtain the atomic embedding matrix Hr, which is composed of all the atomic embedding vectors stacked in the reaction compression map of the reaction. The atomic embedding matrix Hr is then input into the reaction-protein joint encoder, wherein the reaction-protein joint encoder inserts a cross attention layer in the Transformer block of the protein language model. The enzyme sequences corresponding to each reaction in the reaction-enzyme pairing dataset are input into the reaction-protein joint encoder. In each Transformer block of the reaction-protein joint encoder, the initial amino acid embedding matrix He, which is formed by stacking the initial embedding vectors of all amino acids in the enzyme sequence, is obtained. In each Transformer block of the reaction-protein co-encoder, the amino acid-updated embedding matrix is ​​calculated using a cross-attention mechanism based on the atom embedding matrix Hr and the initial amino acid embedding matrix He. : in, To query the parameter matrix, For the key parameter matrix, For the value parameter matrix, For the hidden layer dimension of the reaction-protein co-encoder, the amino acid-updated embedding matrix is ​​used to update the embedding matrix. The amino acid initial embedding matrix will serve as the next Transformer block in the reaction-protein co-encoder; and The reaction-protein joint CLS embedding vector output by the reaction-protein joint encoder is input into the matching head to calculate the matching score of the reaction-enzyme pairing. Based on the matching score, the cross-modal feature fusion loss value is calculated. in, The true label for the reaction-enzyme pairing, A value of 0 indicates a mismatch. A value of 1 indicates a match. It is the sigmoid activation function. The matching score is the score for the reaction-enzyme pairing.

8. The method as described in claim 3, characterized in that, The steps for calculating the multimodal training loss value include: Based on the cross-modal feature alignment loss value and the cross-modal feature fusion loss value, the multimodal training loss value is calculated: in, The multimodal training loss value is... The cross-modal feature alignment loss value is... This is the cross-modal feature fusion loss value.

9. An enzyme mining system, characterized in that, include: Memory, on which computer instructions are stored; and A processor, connected to the memory, and configured to execute computer instructions stored thereon to implement the enzyme mining method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed by the processor, the enzyme mining method as described in any one of claims 1 to 8 is implemented.

Images